fs/btrfs/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/bio.h>
   8 #include <linux/buffer_head.h>
   9 #include <linux/file.h>
  10 #include <linux/fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/highmem.h>
  13 #include <linux/time.h>
  14 #include <linux/init.h>
  15 #include <linux/string.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/writeback.h>
  18 #include <linux/compat.h>
  19 #include <linux/xattr.h>
  20 #include <linux/posix_acl.h>
  21 #include <linux/falloc.h>
  22 #include <linux/slab.h>
  23 #include <linux/ratelimit.h>
  24 #include <linux/btrfs.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/posix_acl_xattr.h>
  27 #include <linux/uio.h>
  28 #include <linux/magic.h>
  29 #include <linux/iversion.h>
  30 #include <linux/swap.h>
  31 #include <linux/sched/mm.h>
  32 #include <asm/unaligned.h>
  33 #include "misc.h"
  34 #include "ctree.h"
  35 #include "disk-io.h"
  36 #include "transaction.h"
  37 #include "btrfs_inode.h"
  38 #include "print-tree.h"
  39 #include "ordered-data.h"
  40 #include "xattr.h"
  41 #include "tree-log.h"
  42 #include "volumes.h"
  43 #include "compression.h"
  44 #include "locking.h"
  45 #include "free-space-cache.h"
  46 #include "inode-map.h"
  47 #include "backref.h"
  48 #include "props.h"
  49 #include "qgroup.h"
  50 #include "delalloc-space.h"
  51 #include "block-group.h"
  52
  53 struct btrfs_iget_args {
  54         struct btrfs_key *location;
  55         struct btrfs_root *root;
  56 };
  57
  58 struct btrfs_dio_data {
  59         u64 reserve;
  60         u64 unsubmitted_oe_range_start;
  61         u64 unsubmitted_oe_range_end;
  62         int overwrite;
  63 };
  64
  65 static const struct inode_operations btrfs_dir_inode_operations;
  66 static const struct inode_operations btrfs_symlink_inode_operations;
  67 static const struct inode_operations btrfs_dir_ro_inode_operations;
  68 static const struct inode_operations btrfs_special_inode_operations;
  69 static const struct inode_operations btrfs_file_inode_operations;
  70 static const struct address_space_operations btrfs_aops;
  71 static const struct file_operations btrfs_dir_file_operations;
  72 static const struct extent_io_ops btrfs_extent_io_ops;
  73
  74 static struct kmem_cache *btrfs_inode_cachep;
  75 struct kmem_cache *btrfs_trans_handle_cachep;
  76 struct kmem_cache *btrfs_path_cachep;
  77 struct kmem_cache *btrfs_free_space_cachep;
  78 struct kmem_cache *btrfs_free_space_bitmap_cachep;
  79
  80 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  81 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
  82 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  83 static noinline int cow_file_range(struct inode *inode,
  84                                    struct page *locked_page,
  85                                    u64 start, u64 end, int *page_started,
  86                                    unsigned long *nr_written, int unlock);
  87 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
  88                                        u64 orig_start, u64 block_start,
  89                                        u64 block_len, u64 orig_block_len,
  90                                        u64 ram_bytes, int compress_type,
  91                                        int type);
  92
  93 static void __endio_write_update_ordered(struct inode *inode,
  94                                          const u64 offset, const u64 bytes,
  95                                          const bool uptodate);
  96
  97 /*
  98  * Cleanup all submitted ordered extents in specified range to handle errors
  99  * from the btrfs_run_delalloc_range() callback.
 100  *
 101  * NOTE: caller must ensure that when an error happens, it can not call
 102  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 103  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 104  * to be released, which we want to happen only when finishing the ordered
 105  * extent (btrfs_finish_ordered_io()).
 106  */
 107 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 108                                                  struct page *locked_page,
 109                                                  u64 offset, u64 bytes)
 110 {
 111         unsigned long index = offset >> PAGE_SHIFT;
 112         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 113         u64 page_start = page_offset(locked_page);
 114         u64 page_end = page_start + PAGE_SIZE - 1;
 115
 116         struct page *page;
 117
 118         while (index <= end_index) {
 119                 page = find_get_page(inode->i_mapping, index);
 120                 index++;
 121                 if (!page)
 122                         continue;
 123                 ClearPagePrivate2(page);
 124                 put_page(page);
 125         }
 126
 127         /*
 128          * In case this page belongs to the delalloc range being instantiated
 129          * then skip it, since the first page of a range is going to be
 130          * properly cleaned up by the caller of run_delalloc_range
 131          */
 132         if (page_start >= offset && page_end <= (offset + bytes - 1)) {
 133                 offset += PAGE_SIZE;
 134                 bytes -= PAGE_SIZE;
 135         }
 136
 137         return __endio_write_update_ordered(inode, offset, bytes, false);
 138 }
 139
 140 static int btrfs_dirty_inode(struct inode *inode);
 141
 142 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 143 void btrfs_test_inode_set_ops(struct inode *inode)
 144 {
 145         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 146 }
 147 #endif
 148
 149 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 150                                      struct inode *inode,  struct inode *dir,
 151                                      const struct qstr *qstr)
 152 {
 153         int err;
 154
 155         err = btrfs_init_acl(trans, inode, dir);
 156         if (!err)
 157                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 158         return err;
 159 }
 160
 161 /*
 162  * this does all the hard work for inserting an inline extent into
 163  * the btree.  The caller should have done a btrfs_drop_extents so that
 164  * no overlapping inline items exist in the btree
 165  */
 166 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 167                                 struct btrfs_path *path, int extent_inserted,
 168                                 struct btrfs_root *root, struct inode *inode,
 169                                 u64 start, size_t size, size_t compressed_size,
 170                                 int compress_type,
 171                                 struct page **compressed_pages)
 172 {
 173         struct extent_buffer *leaf;
 174         struct page *page = NULL;
 175         char *kaddr;
 176         unsigned long ptr;
 177         struct btrfs_file_extent_item *ei;
 178         int ret;
 179         size_t cur_size = size;
 180         unsigned long offset;
 181
 182         ASSERT((compressed_size > 0 && compressed_pages) ||
 183                (compressed_size == 0 && !compressed_pages));
 184
 185         if (compressed_size && compressed_pages)
 186                 cur_size = compressed_size;
 187
 188         inode_add_bytes(inode, size);
 189
 190         if (!extent_inserted) {
 191                 struct btrfs_key key;
 192                 size_t datasize;
 193
 194                 key.objectid = btrfs_ino(BTRFS_I(inode));
 195                 key.offset = start;
 196                 key.type = BTRFS_EXTENT_DATA_KEY;
 197
 198                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 199                 path->leave_spinning = 1;
 200                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 201                                               datasize);
 202                 if (ret)
 203                         goto fail;
 204         }
 205         leaf = path->nodes[0];
 206         ei = btrfs_item_ptr(leaf, path->slots[0],
 207                             struct btrfs_file_extent_item);
 208         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 209         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 210         btrfs_set_file_extent_encryption(leaf, ei, 0);
 211         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 212         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 213         ptr = btrfs_file_extent_inline_start(ei);
 214
 215         if (compress_type != BTRFS_COMPRESS_NONE) {
 216                 struct page *cpage;
 217                 int i = 0;
 218                 while (compressed_size > 0) {
 219                         cpage = compressed_pages[i];
 220                         cur_size = min_t(unsigned long, compressed_size,
 221                                        PAGE_SIZE);
 222
 223                         kaddr = kmap_atomic(cpage);
 224                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 225                         kunmap_atomic(kaddr);
 226
 227                         i++;
 228                         ptr += cur_size;
 229                         compressed_size -= cur_size;
 230                 }
 231                 btrfs_set_file_extent_compression(leaf, ei,
 232                                                   compress_type);
 233         } else {
 234                 page = find_get_page(inode->i_mapping,
 235                                      start >> PAGE_SHIFT);
 236                 btrfs_set_file_extent_compression(leaf, ei, 0);
 237                 kaddr = kmap_atomic(page);
 238                 offset = offset_in_page(start);
 239                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 240                 kunmap_atomic(kaddr);
 241                 put_page(page);
 242         }
 243         btrfs_mark_buffer_dirty(leaf);
 244         btrfs_release_path(path);
 245
 246         /*
 247          * we're an inline extent, so nobody can
 248          * extend the file past i_size without locking
 249          * a page we already have locked.
 250          *
 251          * We must do any isize and inode updates
 252          * before we unlock the pages.  Otherwise we
 253          * could end up racing with unlink.
 254          */
 255         BTRFS_I(inode)->disk_i_size = inode->i_size;
 256         ret = btrfs_update_inode(trans, root, inode);
 257
 258 fail:
 259         return ret;
 260 }
 261
 262
 263 /*
 264  * conditionally insert an inline extent into the file.  This
 265  * does the checks required to make sure the data is small enough
 266  * to fit as an inline extent.
 267  */
 268 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 269                                           u64 end, size_t compressed_size,
 270                                           int compress_type,
 271                                           struct page **compressed_pages)
 272 {
 273         struct btrfs_root *root = BTRFS_I(inode)->root;
 274         struct btrfs_fs_info *fs_info = root->fs_info;
 275         struct btrfs_trans_handle *trans;
 276         u64 isize = i_size_read(inode);
 277         u64 actual_end = min(end + 1, isize);
 278         u64 inline_len = actual_end - start;
 279         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 280         u64 data_len = inline_len;
 281         int ret;
 282         struct btrfs_path *path;
 283         int extent_inserted = 0;
 284         u32 extent_item_size;
 285
 286         if (compressed_size)
 287                 data_len = compressed_size;
 288
 289         if (start > 0 ||
 290             actual_end > fs_info->sectorsize ||
 291             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 292             (!compressed_size &&
 293             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 294             end + 1 < isize ||
 295             data_len > fs_info->max_inline) {
 296                 return 1;
 297         }
 298
 299         path = btrfs_alloc_path();
 300         if (!path)
 301                 return -ENOMEM;
 302
 303         trans = btrfs_join_transaction(root);
 304         if (IS_ERR(trans)) {
 305                 btrfs_free_path(path);
 306                 return PTR_ERR(trans);
 307         }
 308         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 309
 310         if (compressed_size && compressed_pages)
 311                 extent_item_size = btrfs_file_extent_calc_inline_size(
 312                    compressed_size);
 313         else
 314                 extent_item_size = btrfs_file_extent_calc_inline_size(
 315                     inline_len);
 316
 317         ret = __btrfs_drop_extents(trans, root, inode, path,
 318                                    start, aligned_end, NULL,
 319                                    1, 1, extent_item_size, &extent_inserted);
 320         if (ret) {
 321                 btrfs_abort_transaction(trans, ret);
 322                 goto out;
 323         }
 324
 325         if (isize > actual_end)
 326                 inline_len = min_t(u64, isize, actual_end);
 327         ret = insert_inline_extent(trans, path, extent_inserted,
 328                                    root, inode, start,
 329                                    inline_len, compressed_size,
 330                                    compress_type, compressed_pages);
 331         if (ret && ret != -ENOSPC) {
 332                 btrfs_abort_transaction(trans, ret);
 333                 goto out;
 334         } else if (ret == -ENOSPC) {
 335                 ret = 1;
 336                 goto out;
 337         }
 338
 339         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 340         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 341 out:
 342         /*
 343          * Don't forget to free the reserved space, as for inlined extent
 344          * it won't count as data extent, free them directly here.
 345          * And at reserve time, it's always aligned to page size, so
 346          * just free one page here.
 347          */
 348         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 349         btrfs_free_path(path);
 350         btrfs_end_transaction(trans);
 351         return ret;
 352 }
 353
 354 struct async_extent {
 355         u64 start;
 356         u64 ram_size;
 357         u64 compressed_size;
 358         struct page **pages;
 359         unsigned long nr_pages;
 360         int compress_type;
 361         struct list_head list;
 362 };
 363
 364 struct async_chunk {
 365         struct inode *inode;
 366         struct page *locked_page;
 367         u64 start;
 368         u64 end;
 369         unsigned int write_flags;
 370         struct list_head extents;
 371         struct btrfs_work work;
 372         atomic_t *pending;
 373 };
 374
 375 struct async_cow {
 376         /* Number of chunks in flight; must be first in the structure */
 377         atomic_t num_chunks;
 378         struct async_chunk chunks[];
 379 };
 380
 381 static noinline int add_async_extent(struct async_chunk *cow,
 382                                      u64 start, u64 ram_size,
 383                                      u64 compressed_size,
 384                                      struct page **pages,
 385                                      unsigned long nr_pages,
 386                                      int compress_type)
 387 {
 388         struct async_extent *async_extent;
 389
 390         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 391         BUG_ON(!async_extent); /* -ENOMEM */
 392         async_extent->start = start;
 393         async_extent->ram_size = ram_size;
 394         async_extent->compressed_size = compressed_size;
 395         async_extent->pages = pages;
 396         async_extent->nr_pages = nr_pages;
 397         async_extent->compress_type = compress_type;
 398         list_add_tail(&async_extent->list, &cow->extents);
 399         return 0;
 400 }
 401
 402 /*
 403  * Check if the inode has flags compatible with compression
 404  */
 405 static inline bool inode_can_compress(struct inode *inode)
 406 {
 407         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW ||
 408             BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
 409                 return false;
 410         return true;
 411 }
 412
 413 /*
 414  * Check if the inode needs to be submitted to compression, based on mount
 415  * options, defragmentation, properties or heuristics.
 416  */
 417 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 418 {
 419         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 420
 421         if (!inode_can_compress(inode)) {
 422                 WARN(IS_ENABLED(CONFIG_BTRFS_DEBUG),
 423                         KERN_ERR "BTRFS: unexpected compression for ino %llu\n",
 424                         btrfs_ino(BTRFS_I(inode)));
 425                 return 0;
 426         }
 427         /* force compress */
 428         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 429                 return 1;
 430         /* defrag ioctl */
 431         if (BTRFS_I(inode)->defrag_compress)
 432                 return 1;
 433         /* bad compression ratios */
 434         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 435                 return 0;
 436         if (btrfs_test_opt(fs_info, COMPRESS) ||
 437             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 438             BTRFS_I(inode)->prop_compress)
 439                 return btrfs_compress_heuristic(inode, start, end);
 440         return 0;
 441 }
 442
 443 static inline void inode_should_defrag(struct btrfs_inode *inode,
 444                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 445 {
 446         /* If this is a small write inside eof, kick off a defrag */
 447         if (num_bytes < small_write &&
 448             (start > 0 || end + 1 < inode->disk_i_size))
 449                 btrfs_add_inode_defrag(NULL, inode);
 450 }
 451
 452 /*
 453  * we create compressed extents in two phases.  The first
 454  * phase compresses a range of pages that have already been
 455  * locked (both pages and state bits are locked).
 456  *
 457  * This is done inside an ordered work queue, and the compression
 458  * is spread across many cpus.  The actual IO submission is step
 459  * two, and the ordered work queue takes care of making sure that
 460  * happens in the same order things were put onto the queue by
 461  * writepages and friends.
 462  *
 463  * If this code finds it can't get good compression, it puts an
 464  * entry onto the work queue to write the uncompressed bytes.  This
 465  * makes sure that both compressed inodes and uncompressed inodes
 466  * are written in the same order that the flusher thread sent them
 467  * down.
 468  */
 469 static noinline int compress_file_range(struct async_chunk *async_chunk)
 470 {
 471         struct inode *inode = async_chunk->inode;
 472         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 473         u64 blocksize = fs_info->sectorsize;
 474         u64 start = async_chunk->start;
 475         u64 end = async_chunk->end;
 476         u64 actual_end;
 477         u64 i_size;
 478         int ret = 0;
 479         struct page **pages = NULL;
 480         unsigned long nr_pages;
 481         unsigned long total_compressed = 0;
 482         unsigned long total_in = 0;
 483         int i;
 484         int will_compress;
 485         int compress_type = fs_info->compress_type;
 486         int compressed_extents = 0;
 487         int redirty = 0;
 488
 489         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 490                         SZ_16K);
 491
 492         /*
 493          * We need to save i_size before now because it could change in between
 494          * us evaluating the size and assigning it.  This is because we lock and
 495          * unlock the page in truncate and fallocate, and then modify the i_size
 496          * later on.
 497          *
 498          * The barriers are to emulate READ_ONCE, remove that once i_size_read
 499          * does that for us.
 500          */
 501         barrier();
 502         i_size = i_size_read(inode);
 503         barrier();
 504         actual_end = min_t(u64, i_size, end + 1);
 505 again:
 506         will_compress = 0;
 507         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 508         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 509         nr_pages = min_t(unsigned long, nr_pages,
 510                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 511
 512         /*
 513          * we don't want to send crud past the end of i_size through
 514          * compression, that's just a waste of CPU time.  So, if the
 515          * end of the file is before the start of our current
 516          * requested range of bytes, we bail out to the uncompressed
 517          * cleanup code that can deal with all of this.
 518          *
 519          * It isn't really the fastest way to fix things, but this is a
 520          * very uncommon corner.
 521          */
 522         if (actual_end <= start)
 523                 goto cleanup_and_bail_uncompressed;
 524
 525         total_compressed = actual_end - start;
 526
 527         /*
 528          * skip compression for a small file range(<=blocksize) that
 529          * isn't an inline extent, since it doesn't save disk space at all.
 530          */
 531         if (total_compressed <= blocksize &&
 532            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 533                 goto cleanup_and_bail_uncompressed;
 534
 535         total_compressed = min_t(unsigned long, total_compressed,
 536                         BTRFS_MAX_UNCOMPRESSED);
 537         total_in = 0;
 538         ret = 0;
 539
 540         /*
 541          * we do compression for mount -o compress and when the
 542          * inode has not been flagged as nocompress.  This flag can
 543          * change at any time if we discover bad compression ratios.
 544          */
 545         if (inode_need_compress(inode, start, end)) {
 546                 WARN_ON(pages);
 547                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 548                 if (!pages) {
 549                         /* just bail out to the uncompressed code */
 550                         nr_pages = 0;
 551                         goto cont;
 552                 }
 553
 554                 if (BTRFS_I(inode)->defrag_compress)
 555                         compress_type = BTRFS_I(inode)->defrag_compress;
 556                 else if (BTRFS_I(inode)->prop_compress)
 557                         compress_type = BTRFS_I(inode)->prop_compress;
 558
 559                 /*
 560                  * we need to call clear_page_dirty_for_io on each
 561                  * page in the range.  Otherwise applications with the file
 562                  * mmap'd can wander in and change the page contents while
 563                  * we are compressing them.
 564                  *
 565                  * If the compression fails for any reason, we set the pages
 566                  * dirty again later on.
 567                  *
 568                  * Note that the remaining part is redirtied, the start pointer
 569                  * has moved, the end is the original one.
 570                  */
 571                 if (!redirty) {
 572                         extent_range_clear_dirty_for_io(inode, start, end);
 573                         redirty = 1;
 574                 }
 575
 576                 /* Compression level is applied here and only here */
 577                 ret = btrfs_compress_pages(
 578                         compress_type | (fs_info->compress_level << 4),
 579                                            inode->i_mapping, start,
 580                                            pages,
 581                                            &nr_pages,
 582                                            &total_in,
 583                                            &total_compressed);
 584
 585                 if (!ret) {
 586                         unsigned long offset = offset_in_page(total_compressed);
 587                         struct page *page = pages[nr_pages - 1];
 588                         char *kaddr;
 589
 590                         /* zero the tail end of the last page, we might be
 591                          * sending it down to disk
 592                          */
 593                         if (offset) {
 594                                 kaddr = kmap_atomic(page);
 595                                 memset(kaddr + offset, 0,
 596                                        PAGE_SIZE - offset);
 597                                 kunmap_atomic(kaddr);
 598                         }
 599                         will_compress = 1;
 600                 }
 601         }
 602 cont:
 603         if (start == 0) {
 604                 /* lets try to make an inline extent */
 605                 if (ret || total_in < actual_end) {
 606                         /* we didn't compress the entire range, try
 607                          * to make an uncompressed inline extent.
 608                          */
 609                         ret = cow_file_range_inline(inode, start, end, 0,
 610                                                     BTRFS_COMPRESS_NONE, NULL);
 611                 } else {
 612                         /* try making a compressed inline extent */
 613                         ret = cow_file_range_inline(inode, start, end,
 614                                                     total_compressed,
 615                                                     compress_type, pages);
 616                 }
 617                 if (ret <= 0) {
 618                         unsigned long clear_flags = EXTENT_DELALLOC |
 619                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 620                                 EXTENT_DO_ACCOUNTING;
 621                         unsigned long page_error_op;
 622
 623                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 624
 625                         /*
 626                          * inline extent creation worked or returned error,
 627                          * we don't need to create any more async work items.
 628                          * Unlock and free up our temp pages.
 629                          *
 630                          * We use DO_ACCOUNTING here because we need the
 631                          * delalloc_release_metadata to be done _after_ we drop
 632                          * our outstanding extent for clearing delalloc for this
 633                          * range.
 634                          */
 635                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 636                                                      clear_flags,
 637                                                      PAGE_UNLOCK |
 638                                                      PAGE_CLEAR_DIRTY |
 639                                                      PAGE_SET_WRITEBACK |
 640                                                      page_error_op |
 641                                                      PAGE_END_WRITEBACK);
 642
 643                         for (i = 0; i < nr_pages; i++) {
 644                                 WARN_ON(pages[i]->mapping);
 645                                 put_page(pages[i]);
 646                         }
 647                         kfree(pages);
 648
 649                         return 0;
 650                 }
 651         }
 652
 653         if (will_compress) {
 654                 /*
 655                  * we aren't doing an inline extent round the compressed size
 656                  * up to a block size boundary so the allocator does sane
 657                  * things
 658                  */
 659                 total_compressed = ALIGN(total_compressed, blocksize);
 660
 661                 /*
 662                  * one last check to make sure the compression is really a
 663                  * win, compare the page count read with the blocks on disk,
 664                  * compression must free at least one sector size
 665                  */
 666                 total_in = ALIGN(total_in, PAGE_SIZE);
 667                 if (total_compressed + blocksize <= total_in) {
 668                         compressed_extents++;
 669
 670                         /*
 671                          * The async work queues will take care of doing actual
 672                          * allocation on disk for these compressed pages, and
 673                          * will submit them to the elevator.
 674                          */
 675                         add_async_extent(async_chunk, start, total_in,
 676                                         total_compressed, pages, nr_pages,
 677                                         compress_type);
 678
 679                         if (start + total_in < end) {
 680                                 start += total_in;
 681                                 pages = NULL;
 682                                 cond_resched();
 683                                 goto again;
 684                         }
 685                         return compressed_extents;
 686                 }
 687         }
 688         if (pages) {
 689                 /*
 690                  * the compression code ran but failed to make things smaller,
 691                  * free any pages it allocated and our page pointer array
 692                  */
 693                 for (i = 0; i < nr_pages; i++) {
 694                         WARN_ON(pages[i]->mapping);
 695                         put_page(pages[i]);
 696                 }
 697                 kfree(pages);
 698                 pages = NULL;
 699                 total_compressed = 0;
 700                 nr_pages = 0;
 701
 702                 /* flag the file so we don't compress in the future */
 703                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 704                     !(BTRFS_I(inode)->prop_compress)) {
 705                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 706                 }
 707         }
 708 cleanup_and_bail_uncompressed:
 709         /*
 710          * No compression, but we still need to write the pages in the file
 711          * we've been given so far.  redirty the locked page if it corresponds
 712          * to our extent and set things up for the async work queue to run
 713          * cow_file_range to do the normal delalloc dance.
 714          */
 715         if (page_offset(async_chunk->locked_page) >= start &&
 716             page_offset(async_chunk->locked_page) <= end)
 717                 __set_page_dirty_nobuffers(async_chunk->locked_page);
 718                 /* unlocked later on in the async handlers */
 719
 720         if (redirty)
 721                 extent_range_redirty_for_io(inode, start, end);
 722         add_async_extent(async_chunk, start, end - start + 1, 0, NULL, 0,
 723                          BTRFS_COMPRESS_NONE);
 724         compressed_extents++;
 725
 726         return compressed_extents;
 727 }
 728
 729 static void free_async_extent_pages(struct async_extent *async_extent)
 730 {
 731         int i;
 732
 733         if (!async_extent->pages)
 734                 return;
 735
 736         for (i = 0; i < async_extent->nr_pages; i++) {
 737                 WARN_ON(async_extent->pages[i]->mapping);
 738                 put_page(async_extent->pages[i]);
 739         }
 740         kfree(async_extent->pages);
 741         async_extent->nr_pages = 0;
 742         async_extent->pages = NULL;
 743 }
 744
 745 /*
 746  * phase two of compressed writeback.  This is the ordered portion
 747  * of the code, which only gets called in the order the work was
 748  * queued.  We walk all the async extents created by compress_file_range
 749  * and send them down to the disk.
 750  */
 751 static noinline void submit_compressed_extents(struct async_chunk *async_chunk)
 752 {
 753         struct inode *inode = async_chunk->inode;
 754         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 755         struct async_extent *async_extent;
 756         u64 alloc_hint = 0;
 757         struct btrfs_key ins;
 758         struct extent_map *em;
 759         struct btrfs_root *root = BTRFS_I(inode)->root;
 760         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
 761         int ret = 0;
 762
 763 again:
 764         while (!list_empty(&async_chunk->extents)) {
 765                 async_extent = list_entry(async_chunk->extents.next,
 766                                           struct async_extent, list);
 767                 list_del(&async_extent->list);
 768
 769 retry:
 770                 lock_extent(io_tree, async_extent->start,
 771                             async_extent->start + async_extent->ram_size - 1);
 772                 /* did the compression code fall back to uncompressed IO? */
 773                 if (!async_extent->pages) {
 774                         int page_started = 0;
 775                         unsigned long nr_written = 0;
 776
 777                         /* allocate blocks */
 778                         ret = cow_file_range(inode, async_chunk->locked_page,
 779                                              async_extent->start,
 780                                              async_extent->start +
 781                                              async_extent->ram_size - 1,
 782                                              &page_started, &nr_written, 0);
 783
 784                         /* JDM XXX */
 785
 786                         /*
 787                          * if page_started, cow_file_range inserted an
 788                          * inline extent and took care of all the unlocking
 789                          * and IO for us.  Otherwise, we need to submit
 790                          * all those pages down to the drive.
 791                          */
 792                         if (!page_started && !ret)
 793                                 extent_write_locked_range(inode,
 794                                                   async_extent->start,
 795                                                   async_extent->start +
 796                                                   async_extent->ram_size - 1,
 797                                                   WB_SYNC_ALL);
 798                         else if (ret)
 799                                 unlock_page(async_chunk->locked_page);
 800                         kfree(async_extent);
 801                         cond_resched();
 802                         continue;
 803                 }
 804
 805                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 806                                            async_extent->compressed_size,
 807                                            async_extent->compressed_size,
 808                                            0, alloc_hint, &ins, 1, 1);
 809                 if (ret) {
 810                         free_async_extent_pages(async_extent);
 811
 812                         if (ret == -ENOSPC) {
 813                                 unlock_extent(io_tree, async_extent->start,
 814                                               async_extent->start +
 815                                               async_extent->ram_size - 1);
 816
 817                                 /*
 818                                  * we need to redirty the pages if we decide to
 819                                  * fallback to uncompressed IO, otherwise we
 820                                  * will not submit these pages down to lower
 821                                  * layers.
 822                                  */
 823                                 extent_range_redirty_for_io(inode,
 824                                                 async_extent->start,
 825                                                 async_extent->start +
 826                                                 async_extent->ram_size - 1);
 827
 828                                 goto retry;
 829                         }
 830                         goto out_free;
 831                 }
 832                 /*
 833                  * here we're doing allocation and writeback of the
 834                  * compressed pages
 835                  */
 836                 em = create_io_em(inode, async_extent->start,
 837                                   async_extent->ram_size, /* len */
 838                                   async_extent->start, /* orig_start */
 839                                   ins.objectid, /* block_start */
 840                                   ins.offset, /* block_len */
 841                                   ins.offset, /* orig_block_len */
 842                                   async_extent->ram_size, /* ram_bytes */
 843                                   async_extent->compress_type,
 844                                   BTRFS_ORDERED_COMPRESSED);
 845                 if (IS_ERR(em))
 846                         /* ret value is not necessary due to void function */
 847                         goto out_free_reserve;
 848                 free_extent_map(em);
 849
 850                 ret = btrfs_add_ordered_extent_compress(inode,
 851                                                 async_extent->start,
 852                                                 ins.objectid,
 853                                                 async_extent->ram_size,
 854                                                 ins.offset,
 855                                                 BTRFS_ORDERED_COMPRESSED,
 856                                                 async_extent->compress_type);
 857                 if (ret) {
 858                         btrfs_drop_extent_cache(BTRFS_I(inode),
 859                                                 async_extent->start,
 860                                                 async_extent->start +
 861                                                 async_extent->ram_size - 1, 0);
 862                         goto out_free_reserve;
 863                 }
 864                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 865
 866                 /*
 867                  * clear dirty, set writeback and unlock the pages.
 868                  */
 869                 extent_clear_unlock_delalloc(inode, async_extent->start,
 870                                 async_extent->start +
 871                                 async_extent->ram_size - 1,
 872                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 873                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 874                                 PAGE_SET_WRITEBACK);
 875                 if (btrfs_submit_compressed_write(inode,
 876                                     async_extent->start,
 877                                     async_extent->ram_size,
 878                                     ins.objectid,
 879                                     ins.offset, async_extent->pages,
 880                                     async_extent->nr_pages,
 881                                     async_chunk->write_flags)) {
 882                         struct page *p = async_extent->pages[0];
 883                         const u64 start = async_extent->start;
 884                         const u64 end = start + async_extent->ram_size - 1;
 885
 886                         p->mapping = inode->i_mapping;
 887                         btrfs_writepage_endio_finish_ordered(p, start, end, 0);
 888
 889                         p->mapping = NULL;
 890                         extent_clear_unlock_delalloc(inode, start, end,
 891                                                      NULL, 0,
 892                                                      PAGE_END_WRITEBACK |
 893                                                      PAGE_SET_ERROR);
 894                         free_async_extent_pages(async_extent);
 895                 }
 896                 alloc_hint = ins.objectid + ins.offset;
 897                 kfree(async_extent);
 898                 cond_resched();
 899         }
 900         return;
 901 out_free_reserve:
 902         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 903         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 904 out_free:
 905         extent_clear_unlock_delalloc(inode, async_extent->start,
 906                                      async_extent->start +
 907                                      async_extent->ram_size - 1,
 908                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 909                                      EXTENT_DELALLOC_NEW |
 910                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 911                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 912                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 913                                      PAGE_SET_ERROR);
 914         free_async_extent_pages(async_extent);
 915         kfree(async_extent);
 916         goto again;
 917 }
 918
 919 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 920                                       u64 num_bytes)
 921 {
 922         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 923         struct extent_map *em;
 924         u64 alloc_hint = 0;
 925
 926         read_lock(&em_tree->lock);
 927         em = search_extent_mapping(em_tree, start, num_bytes);
 928         if (em) {
 929                 /*
 930                  * if block start isn't an actual block number then find the
 931                  * first block in this inode and use that as a hint.  If that
 932                  * block is also bogus then just don't worry about it.
 933                  */
 934                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 935                         free_extent_map(em);
 936                         em = search_extent_mapping(em_tree, 0, 0);
 937                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 938                                 alloc_hint = em->block_start;
 939                         if (em)
 940                                 free_extent_map(em);
 941                 } else {
 942                         alloc_hint = em->block_start;
 943                         free_extent_map(em);
 944                 }
 945         }
 946         read_unlock(&em_tree->lock);
 947
 948         return alloc_hint;
 949 }
 950
 951 /*
 952  * when extent_io.c finds a delayed allocation range in the file,
 953  * the call backs end up in this code.  The basic idea is to
 954  * allocate extents on disk for the range, and create ordered data structs
 955  * in ram to track those extents.
 956  *
 957  * locked_page is the page that writepage had locked already.  We use
 958  * it to make sure we don't do extra locks or unlocks.
 959  *
 960  * *page_started is set to one if we unlock locked_page and do everything
 961  * required to start IO on it.  It may be clean and already done with
 962  * IO when we return.
 963  */
 964 static noinline int cow_file_range(struct inode *inode,
 965                                    struct page *locked_page,
 966                                    u64 start, u64 end, int *page_started,
 967                                    unsigned long *nr_written, int unlock)
 968 {
 969         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 970         struct btrfs_root *root = BTRFS_I(inode)->root;
 971         u64 alloc_hint = 0;
 972         u64 num_bytes;
 973         unsigned long ram_size;
 974         u64 cur_alloc_size = 0;
 975         u64 blocksize = fs_info->sectorsize;
 976         struct btrfs_key ins;
 977         struct extent_map *em;
 978         unsigned clear_bits;
 979         unsigned long page_ops;
 980         bool extent_reserved = false;
 981         int ret = 0;
 982
 983         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 984                 WARN_ON_ONCE(1);
 985                 ret = -EINVAL;
 986                 goto out_unlock;
 987         }
 988
 989         num_bytes = ALIGN(end - start + 1, blocksize);
 990         num_bytes = max(blocksize,  num_bytes);
 991         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
 992
 993         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 994
 995         if (start == 0) {
 996                 /* lets try to make an inline extent */
 997                 ret = cow_file_range_inline(inode, start, end, 0,
 998                                             BTRFS_COMPRESS_NONE, NULL);
 999                 if (ret == 0) {
1000                         /*
1001                          * We use DO_ACCOUNTING here because we need the
1002                          * delalloc_release_metadata to be run _after_ we drop
1003                          * our outstanding extent for clearing delalloc for this
1004                          * range.
1005                          */
1006                         extent_clear_unlock_delalloc(inode, start, end, NULL,
1007                                      EXTENT_LOCKED | EXTENT_DELALLOC |
1008                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1009                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1010                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1011                                      PAGE_END_WRITEBACK);
1012                         *nr_written = *nr_written +
1013                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1014                         *page_started = 1;
1015                         goto out;
1016                 } else if (ret < 0) {
1017                         goto out_unlock;
1018                 }
1019         }
1020
1021         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1022         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1023                         start + num_bytes - 1, 0);
1024
1025         while (num_bytes > 0) {
1026                 cur_alloc_size = num_bytes;
1027                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1028                                            fs_info->sectorsize, 0, alloc_hint,
1029                                            &ins, 1, 1);
1030                 if (ret < 0)
1031                         goto out_unlock;
1032                 cur_alloc_size = ins.offset;
1033                 extent_reserved = true;
1034
1035                 ram_size = ins.offset;
1036                 em = create_io_em(inode, start, ins.offset, /* len */
1037                                   start, /* orig_start */
1038                                   ins.objectid, /* block_start */
1039                                   ins.offset, /* block_len */
1040                                   ins.offset, /* orig_block_len */
1041                                   ram_size, /* ram_bytes */
1042                                   BTRFS_COMPRESS_NONE, /* compress_type */
1043                                   BTRFS_ORDERED_REGULAR /* type */);
1044                 if (IS_ERR(em)) {
1045                         ret = PTR_ERR(em);
1046                         goto out_reserve;
1047                 }
1048                 free_extent_map(em);
1049
1050                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1051                                                ram_size, cur_alloc_size, 0);
1052                 if (ret)
1053                         goto out_drop_extent_cache;
1054
1055                 if (root->root_key.objectid ==
1056                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1057                         ret = btrfs_reloc_clone_csums(inode, start,
1058                                                       cur_alloc_size);
1059                         /*
1060                          * Only drop cache here, and process as normal.
1061                          *
1062                          * We must not allow extent_clear_unlock_delalloc()
1063                          * at out_unlock label to free meta of this ordered
1064                          * extent, as its meta should be freed by
1065                          * btrfs_finish_ordered_io().
1066                          *
1067                          * So we must continue until @start is increased to
1068                          * skip current ordered extent.
1069                          */
1070                         if (ret)
1071                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1072                                                 start + ram_size - 1, 0);
1073                 }
1074
1075                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1076
1077                 /* we're not doing compressed IO, don't unlock the first
1078                  * page (which the caller expects to stay locked), don't
1079                  * clear any dirty bits and don't set any writeback bits
1080                  *
1081                  * Do set the Private2 bit so we know this page was properly
1082                  * setup for writepage
1083                  */
1084                 page_ops = unlock ? PAGE_UNLOCK : 0;
1085                 page_ops |= PAGE_SET_PRIVATE2;
1086
1087                 extent_clear_unlock_delalloc(inode, start,
1088                                              start + ram_size - 1,
1089                                              locked_page,
1090                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1091                                              page_ops);
1092                 if (num_bytes < cur_alloc_size)
1093                         num_bytes = 0;
1094                 else
1095                         num_bytes -= cur_alloc_size;
1096                 alloc_hint = ins.objectid + ins.offset;
1097                 start += cur_alloc_size;
1098                 extent_reserved = false;
1099
1100                 /*
1101                  * btrfs_reloc_clone_csums() error, since start is increased
1102                  * extent_clear_unlock_delalloc() at out_unlock label won't
1103                  * free metadata of current ordered extent, we're OK to exit.
1104                  */
1105                 if (ret)
1106                         goto out_unlock;
1107         }
1108 out:
1109         return ret;
1110
1111 out_drop_extent_cache:
1112         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1113 out_reserve:
1114         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1115         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1116 out_unlock:
1117         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1118                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1119         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1120                 PAGE_END_WRITEBACK;
1121         /*
1122          * If we reserved an extent for our delalloc range (or a subrange) and
1123          * failed to create the respective ordered extent, then it means that
1124          * when we reserved the extent we decremented the extent's size from
1125          * the data space_info's bytes_may_use counter and incremented the
1126          * space_info's bytes_reserved counter by the same amount. We must make
1127          * sure extent_clear_unlock_delalloc() does not try to decrement again
1128          * the data space_info's bytes_may_use counter, therefore we do not pass
1129          * it the flag EXTENT_CLEAR_DATA_RESV.
1130          */
1131         if (extent_reserved) {
1132                 extent_clear_unlock_delalloc(inode, start,
1133                                              start + cur_alloc_size,
1134                                              locked_page,
1135                                              clear_bits,
1136                                              page_ops);
1137                 start += cur_alloc_size;
1138                 if (start >= end)
1139                         goto out;
1140         }
1141         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1142                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1143                                      page_ops);
1144         goto out;
1145 }
1146
1147 /*
1148  * work queue call back to started compression on a file and pages
1149  */
1150 static noinline void async_cow_start(struct btrfs_work *work)
1151 {
1152         struct async_chunk *async_chunk;
1153         int compressed_extents;
1154
1155         async_chunk = container_of(work, struct async_chunk, work);
1156
1157         compressed_extents = compress_file_range(async_chunk);
1158         if (compressed_extents == 0) {
1159                 btrfs_add_delayed_iput(async_chunk->inode);
1160                 async_chunk->inode = NULL;
1161         }
1162 }
1163
1164 /*
1165  * work queue call back to submit previously compressed pages
1166  */
1167 static noinline void async_cow_submit(struct btrfs_work *work)
1168 {
1169         struct async_chunk *async_chunk = container_of(work, struct async_chunk,
1170                                                      work);
1171         struct btrfs_fs_info *fs_info = btrfs_work_owner(work);
1172         unsigned long nr_pages;
1173
1174         nr_pages = (async_chunk->end - async_chunk->start + PAGE_SIZE) >>
1175                 PAGE_SHIFT;
1176
1177         /* atomic_sub_return implies a barrier */
1178         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1179             5 * SZ_1M)
1180                 cond_wake_up_nomb(&fs_info->async_submit_wait);
1181
1182         /*
1183          * ->inode could be NULL if async_chunk_start has failed to compress,
1184          * in which case we don't have anything to submit, yet we need to
1185          * always adjust ->async_delalloc_pages as its paired with the init
1186          * happening in cow_file_range_async
1187          */
1188         if (async_chunk->inode)
1189                 submit_compressed_extents(async_chunk);
1190 }
1191
1192 static noinline void async_cow_free(struct btrfs_work *work)
1193 {
1194         struct async_chunk *async_chunk;
1195
1196         async_chunk = container_of(work, struct async_chunk, work);
1197         if (async_chunk->inode)
1198                 btrfs_add_delayed_iput(async_chunk->inode);
1199         /*
1200          * Since the pointer to 'pending' is at the beginning of the array of
1201          * async_chunk's, freeing it ensures the whole array has been freed.
1202          */
1203         if (atomic_dec_and_test(async_chunk->pending))
1204                 kvfree(async_chunk->pending);
1205 }
1206
1207 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1208                                 u64 start, u64 end, int *page_started,
1209                                 unsigned long *nr_written,
1210                                 unsigned int write_flags)
1211 {
1212         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1213         struct async_cow *ctx;
1214         struct async_chunk *async_chunk;
1215         unsigned long nr_pages;
1216         u64 cur_end;
1217         u64 num_chunks = DIV_ROUND_UP(end - start, SZ_512K);
1218         int i;
1219         bool should_compress;
1220         unsigned nofs_flag;
1221
1222         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1223
1224         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1225             !btrfs_test_opt(fs_info, FORCE_COMPRESS)) {
1226                 num_chunks = 1;
1227                 should_compress = false;
1228         } else {
1229                 should_compress = true;
1230         }
1231
1232         nofs_flag = memalloc_nofs_save();
1233         ctx = kvmalloc(struct_size(ctx, chunks, num_chunks), GFP_KERNEL);
1234         memalloc_nofs_restore(nofs_flag);
1235
1236         if (!ctx) {
1237                 unsigned clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC |
1238                         EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1239                         EXTENT_DO_ACCOUNTING;
1240                 unsigned long page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1241                         PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
1242                         PAGE_SET_ERROR;
1243
1244                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1245                                              clear_bits, page_ops);
1246                 return -ENOMEM;
1247         }
1248
1249         async_chunk = ctx->chunks;
1250         atomic_set(&ctx->num_chunks, num_chunks);
1251
1252         for (i = 0; i < num_chunks; i++) {
1253                 if (should_compress)
1254                         cur_end = min(end, start + SZ_512K - 1);
1255                 else
1256                         cur_end = end;
1257
1258                 /*
1259                  * igrab is called higher up in the call chain, take only the
1260                  * lightweight reference for the callback lifetime
1261                  */
1262                 ihold(inode);
1263                 async_chunk[i].pending = &ctx->num_chunks;
1264                 async_chunk[i].inode = inode;
1265                 async_chunk[i].start = start;
1266                 async_chunk[i].end = cur_end;
1267                 async_chunk[i].locked_page = locked_page;
1268                 async_chunk[i].write_flags = write_flags;
1269                 INIT_LIST_HEAD(&async_chunk[i].extents);
1270
1271                 btrfs_init_work(&async_chunk[i].work, async_cow_start,
1272                                 async_cow_submit, async_cow_free);
1273
1274                 nr_pages = DIV_ROUND_UP(cur_end - start, PAGE_SIZE);
1275                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1276
1277                 btrfs_queue_work(fs_info->delalloc_workers, &async_chunk[i].work);
1278
1279                 *nr_written += nr_pages;
1280                 start = cur_end + 1;
1281         }
1282         *page_started = 1;
1283         return 0;
1284 }
1285
1286 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1287                                         u64 bytenr, u64 num_bytes)
1288 {
1289         int ret;
1290         struct btrfs_ordered_sum *sums;
1291         LIST_HEAD(list);
1292
1293         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1294                                        bytenr + num_bytes - 1, &list, 0);
1295         if (ret == 0 && list_empty(&list))
1296                 return 0;
1297
1298         while (!list_empty(&list)) {
1299                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1300                 list_del(&sums->list);
1301                 kfree(sums);
1302         }
1303         if (ret < 0)
1304                 return ret;
1305         return 1;
1306 }
1307
1308 /*
1309  * when nowcow writeback call back.  This checks for snapshots or COW copies
1310  * of the extents that exist in the file, and COWs the file as required.
1311  *
1312  * If no cow copies or snapshots exist, we write directly to the existing
1313  * blocks on disk
1314  */
1315 static noinline int run_delalloc_nocow(struct inode *inode,
1316                                        struct page *locked_page,
1317                                        const u64 start, const u64 end,
1318                                        int *page_started, int force,
1319                                        unsigned long *nr_written)
1320 {
1321         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1322         struct btrfs_root *root = BTRFS_I(inode)->root;
1323         struct btrfs_path *path;
1324         u64 cow_start = (u64)-1;
1325         u64 cur_offset = start;
1326         int ret;
1327         bool check_prev = true;
1328         const bool freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
1329         u64 ino = btrfs_ino(BTRFS_I(inode));
1330         bool nocow = false;
1331         u64 disk_bytenr = 0;
1332
1333         path = btrfs_alloc_path();
1334         if (!path) {
1335                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1336                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1337                                              EXTENT_DO_ACCOUNTING |
1338                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1339                                              PAGE_CLEAR_DIRTY |
1340                                              PAGE_SET_WRITEBACK |
1341                                              PAGE_END_WRITEBACK);
1342                 return -ENOMEM;
1343         }
1344
1345         while (1) {
1346                 struct btrfs_key found_key;
1347                 struct btrfs_file_extent_item *fi;
1348                 struct extent_buffer *leaf;
1349                 u64 extent_end;
1350                 u64 extent_offset;
1351                 u64 num_bytes = 0;
1352                 u64 disk_num_bytes;
1353                 u64 ram_bytes;
1354                 int extent_type;
1355
1356                 nocow = false;
1357
1358                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1359                                                cur_offset, 0);
1360                 if (ret < 0)
1361                         goto error;
1362
1363                 /*
1364                  * If there is no extent for our range when doing the initial
1365                  * search, then go back to the previous slot as it will be the
1366                  * one containing the search offset
1367                  */
1368                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1369                         leaf = path->nodes[0];
1370                         btrfs_item_key_to_cpu(leaf, &found_key,
1371                                               path->slots[0] - 1);
1372                         if (found_key.objectid == ino &&
1373                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1374                                 path->slots[0]--;
1375                 }
1376                 check_prev = false;
1377 next_slot:
1378                 /* Go to next leaf if we have exhausted the current one */
1379                 leaf = path->nodes[0];
1380                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1381                         ret = btrfs_next_leaf(root, path);
1382                         if (ret < 0) {
1383                                 if (cow_start != (u64)-1)
1384                                         cur_offset = cow_start;
1385                                 goto error;
1386                         }
1387                         if (ret > 0)
1388                                 break;
1389                         leaf = path->nodes[0];
1390                 }
1391
1392                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1393
1394                 /* Didn't find anything for our INO */
1395                 if (found_key.objectid > ino)
1396                         break;
1397                 /*
1398                  * Keep searching until we find an EXTENT_ITEM or there are no
1399                  * more extents for this inode
1400                  */
1401                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1402                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1403                         path->slots[0]++;
1404                         goto next_slot;
1405                 }
1406
1407                 /* Found key is not EXTENT_DATA_KEY or starts after req range */
1408                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1409                     found_key.offset > end)
1410                         break;
1411
1412                 /*
1413                  * If the found extent starts after requested offset, then
1414                  * adjust extent_end to be right before this extent begins
1415                  */
1416                 if (found_key.offset > cur_offset) {
1417                         extent_end = found_key.offset;
1418                         extent_type = 0;
1419                         goto out_check;
1420                 }
1421
1422                 /*
1423                  * Found extent which begins before our range and potentially
1424                  * intersect it
1425                  */
1426                 fi = btrfs_item_ptr(leaf, path->slots[0],
1427                                     struct btrfs_file_extent_item);
1428                 extent_type = btrfs_file_extent_type(leaf, fi);
1429
1430                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1431                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1432                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1433                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1434                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1435                         extent_end = found_key.offset +
1436                                 btrfs_file_extent_num_bytes(leaf, fi);
1437                         disk_num_bytes =
1438                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1439                         /*
1440                          * If extent we got ends before our range starts, skip
1441                          * to next extent
1442                          */
1443                         if (extent_end <= start) {
1444                                 path->slots[0]++;
1445                                 goto next_slot;
1446                         }
1447                         /* Skip holes */
1448                         if (disk_bytenr == 0)
1449                                 goto out_check;
1450                         /* Skip compressed/encrypted/encoded extents */
1451                         if (btrfs_file_extent_compression(leaf, fi) ||
1452                             btrfs_file_extent_encryption(leaf, fi) ||
1453                             btrfs_file_extent_other_encoding(leaf, fi))
1454                                 goto out_check;
1455                         /*
1456                          * If extent is created before the last volume's snapshot
1457                          * this implies the extent is shared, hence we can't do
1458                          * nocow. This is the same check as in
1459                          * btrfs_cross_ref_exist but without calling
1460                          * btrfs_search_slot.
1461                          */
1462                         if (!freespace_inode &&
1463                             btrfs_file_extent_generation(leaf, fi) <=
1464                             btrfs_root_last_snapshot(&root->root_item))
1465                                 goto out_check;
1466                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1467                                 goto out_check;
1468                         /* If extent is RO, we must COW it */
1469                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1470                                 goto out_check;
1471                         ret = btrfs_cross_ref_exist(root, ino,
1472                                                     found_key.offset -
1473                                                     extent_offset, disk_bytenr);
1474                         if (ret) {
1475                                 /*
1476                                  * ret could be -EIO if the above fails to read
1477                                  * metadata.
1478                                  */
1479                                 if (ret < 0) {
1480                                         if (cow_start != (u64)-1)
1481                                                 cur_offset = cow_start;
1482                                         goto error;
1483                                 }
1484
1485                                 WARN_ON_ONCE(freespace_inode);
1486                                 goto out_check;
1487                         }
1488                         disk_bytenr += extent_offset;
1489                         disk_bytenr += cur_offset - found_key.offset;
1490                         num_bytes = min(end + 1, extent_end) - cur_offset;
1491                         /*
1492                          * If there are pending snapshots for this root, we
1493                          * fall into common COW way
1494                          */
1495                         if (!freespace_inode && atomic_read(&root->snapshot_force_cow))
1496                                 goto out_check;
1497                         /*
1498                          * force cow if csum exists in the range.
1499                          * this ensure that csum for a given extent are
1500                          * either valid or do not exist.
1501                          */
1502                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1503                                                   num_bytes);
1504                         if (ret) {
1505                                 /*
1506                                  * ret could be -EIO if the above fails to read
1507                                  * metadata.
1508                                  */
1509                                 if (ret < 0) {
1510                                         if (cow_start != (u64)-1)
1511                                                 cur_offset = cow_start;
1512                                         goto error;
1513                                 }
1514                                 WARN_ON_ONCE(freespace_inode);
1515                                 goto out_check;
1516                         }
1517                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr))
1518                                 goto out_check;
1519                         nocow = true;
1520                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1521                         extent_end = found_key.offset + ram_bytes;
1522                         extent_end = ALIGN(extent_end, fs_info->sectorsize);
1523                         /* Skip extents outside of our requested range */
1524                         if (extent_end <= start) {
1525                                 path->slots[0]++;
1526                                 goto next_slot;
1527                         }
1528                 } else {
1529                         /* If this triggers then we have a memory corruption */
1530                         BUG();
1531                 }
1532 out_check:
1533                 /*
1534                  * If nocow is false then record the beginning of the range
1535                  * that needs to be COWed
1536                  */
1537                 if (!nocow) {
1538                         if (cow_start == (u64)-1)
1539                                 cow_start = cur_offset;
1540                         cur_offset = extent_end;
1541                         if (cur_offset > end)
1542                                 break;
1543                         path->slots[0]++;
1544                         goto next_slot;
1545                 }
1546
1547                 btrfs_release_path(path);
1548
1549                 /*
1550                  * COW range from cow_start to found_key.offset - 1. As the key
1551                  * will contain the beginning of the first extent that can be
1552                  * NOCOW, following one which needs to be COW'ed
1553                  */
1554                 if (cow_start != (u64)-1) {
1555                         ret = cow_file_range(inode, locked_page,
1556                                              cow_start, found_key.offset - 1,
1557                                              page_started, nr_written, 1);
1558                         if (ret) {
1559                                 if (nocow)
1560                                         btrfs_dec_nocow_writers(fs_info,
1561                                                                 disk_bytenr);
1562                                 goto error;
1563                         }
1564                         cow_start = (u64)-1;
1565                 }
1566
1567                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1568                         u64 orig_start = found_key.offset - extent_offset;
1569                         struct extent_map *em;
1570
1571                         em = create_io_em(inode, cur_offset, num_bytes,
1572                                           orig_start,
1573                                           disk_bytenr, /* block_start */
1574                                           num_bytes, /* block_len */
1575                                           disk_num_bytes, /* orig_block_len */
1576                                           ram_bytes, BTRFS_COMPRESS_NONE,
1577                                           BTRFS_ORDERED_PREALLOC);
1578                         if (IS_ERR(em)) {
1579                                 if (nocow)
1580                                         btrfs_dec_nocow_writers(fs_info,
1581                                                                 disk_bytenr);
1582                                 ret = PTR_ERR(em);
1583                                 goto error;
1584                         }
1585                         free_extent_map(em);
1586                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1587                                                        disk_bytenr, num_bytes,
1588                                                        num_bytes,
1589                                                        BTRFS_ORDERED_PREALLOC);
1590                         if (ret) {
1591                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1592                                                         cur_offset,
1593                                                         cur_offset + num_bytes - 1,
1594                                                         0);
1595                                 goto error;
1596                         }
1597                 } else {
1598                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1599                                                        disk_bytenr, num_bytes,
1600                                                        num_bytes,
1601                                                        BTRFS_ORDERED_NOCOW);
1602                         if (ret)
1603                                 goto error;
1604                 }
1605
1606                 if (nocow)
1607                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1608                 nocow = false;
1609
1610                 if (root->root_key.objectid ==
1611                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1612                         /*
1613                          * Error handled later, as we must prevent
1614                          * extent_clear_unlock_delalloc() in error handler
1615                          * from freeing metadata of created ordered extent.
1616                          */
1617                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1618                                                       num_bytes);
1619
1620                 extent_clear_unlock_delalloc(inode, cur_offset,
1621                                              cur_offset + num_bytes - 1,
1622                                              locked_page, EXTENT_LOCKED |
1623                                              EXTENT_DELALLOC |
1624                                              EXTENT_CLEAR_DATA_RESV,
1625                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1626
1627                 cur_offset = extent_end;
1628
1629                 /*
1630                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1631                  * handler, as metadata for created ordered extent will only
1632                  * be freed by btrfs_finish_ordered_io().
1633                  */
1634                 if (ret)
1635                         goto error;
1636                 if (cur_offset > end)
1637                         break;
1638         }
1639         btrfs_release_path(path);
1640
1641         if (cur_offset <= end && cow_start == (u64)-1)
1642                 cow_start = cur_offset;
1643
1644         if (cow_start != (u64)-1) {
1645                 cur_offset = end;
1646                 ret = cow_file_range(inode, locked_page, cow_start, end,
1647                                      page_started, nr_written, 1);
1648                 if (ret)
1649                         goto error;
1650         }
1651
1652 error:
1653         if (nocow)
1654                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1655
1656         if (ret && cur_offset < end)
1657                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1658                                              locked_page, EXTENT_LOCKED |
1659                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1660                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1661                                              PAGE_CLEAR_DIRTY |
1662                                              PAGE_SET_WRITEBACK |
1663                                              PAGE_END_WRITEBACK);
1664         btrfs_free_path(path);
1665         return ret;
1666 }
1667
1668 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1669 {
1670
1671         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1672             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1673                 return 0;
1674
1675         /*
1676          * @defrag_bytes is a hint value, no spinlock held here,
1677          * if is not zero, it means the file is defragging.
1678          * Force cow if given extent needs to be defragged.
1679          */
1680         if (BTRFS_I(inode)->defrag_bytes &&
1681             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1682                            EXTENT_DEFRAG, 0, NULL))
1683                 return 1;
1684
1685         return 0;
1686 }
1687
1688 /*
1689  * Function to process delayed allocation (create CoW) for ranges which are
1690  * being touched for the first time.
1691  */
1692 int btrfs_run_delalloc_range(struct inode *inode, struct page *locked_page,
1693                 u64 start, u64 end, int *page_started, unsigned long *nr_written,
1694                 struct writeback_control *wbc)
1695 {
1696         int ret;
1697         int force_cow = need_force_cow(inode, start, end);
1698         unsigned int write_flags = wbc_to_write_flags(wbc);
1699
1700         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1701                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1702                                          page_started, 1, nr_written);
1703         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1704                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1705                                          page_started, 0, nr_written);
1706         } else if (!inode_can_compress(inode) ||
1707                    !inode_need_compress(inode, start, end)) {
1708                 ret = cow_file_range(inode, locked_page, start, end,
1709                                       page_started, nr_written, 1);
1710         } else {
1711                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1712                         &BTRFS_I(inode)->runtime_flags);
1713                 ret = cow_file_range_async(inode, locked_page, start, end,
1714                                            page_started, nr_written,
1715                                            write_flags);
1716         }
1717         if (ret)
1718                 btrfs_cleanup_ordered_extents(inode, locked_page, start,
1719                                               end - start + 1);
1720         return ret;
1721 }
1722
1723 void btrfs_split_delalloc_extent(struct inode *inode,
1724                                  struct extent_state *orig, u64 split)
1725 {
1726         u64 size;
1727
1728         /* not delalloc, ignore it */
1729         if (!(orig->state & EXTENT_DELALLOC))
1730                 return;
1731
1732         size = orig->end - orig->start + 1;
1733         if (size > BTRFS_MAX_EXTENT_SIZE) {
1734                 u32 num_extents;
1735                 u64 new_size;
1736
1737                 /*
1738                  * See the explanation in btrfs_merge_delalloc_extent, the same
1739                  * applies here, just in reverse.
1740                  */
1741                 new_size = orig->end - split + 1;
1742                 num_extents = count_max_extents(new_size);
1743                 new_size = split - orig->start;
1744                 num_extents += count_max_extents(new_size);
1745                 if (count_max_extents(size) >= num_extents)
1746                         return;
1747         }
1748
1749         spin_lock(&BTRFS_I(inode)->lock);
1750         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1751         spin_unlock(&BTRFS_I(inode)->lock);
1752 }
1753
1754 /*
1755  * Handle merged delayed allocation extents so we can keep track of new extents
1756  * that are just merged onto old extents, such as when we are doing sequential
1757  * writes, so we can properly account for the metadata space we'll need.
1758  */
1759 void btrfs_merge_delalloc_extent(struct inode *inode, struct extent_state *new,
1760                                  struct extent_state *other)
1761 {
1762         u64 new_size, old_size;
1763         u32 num_extents;
1764
1765         /* not delalloc, ignore it */
1766         if (!(other->state & EXTENT_DELALLOC))
1767                 return;
1768
1769         if (new->start > other->start)
1770                 new_size = new->end - other->start + 1;
1771         else
1772                 new_size = other->end - new->start + 1;
1773
1774         /* we're not bigger than the max, unreserve the space and go */
1775         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1776                 spin_lock(&BTRFS_I(inode)->lock);
1777                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1778                 spin_unlock(&BTRFS_I(inode)->lock);
1779                 return;
1780         }
1781
1782         /*
1783          * We have to add up either side to figure out how many extents were
1784          * accounted for before we merged into one big extent.  If the number of
1785          * extents we accounted for is <= the amount we need for the new range
1786          * then we can return, otherwise drop.  Think of it like this
1787          *
1788          * [ 4k][MAX_SIZE]
1789          *
1790          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1791          * need 2 outstanding extents, on one side we have 1 and the other side
1792          * we have 1 so they are == and we can return.  But in this case
1793          *
1794          * [MAX_SIZE+4k][MAX_SIZE+4k]
1795          *
1796          * Each range on their own accounts for 2 extents, but merged together
1797          * they are only 3 extents worth of accounting, so we need to drop in
1798          * this case.
1799          */
1800         old_size = other->end - other->start + 1;
1801         num_extents = count_max_extents(old_size);
1802         old_size = new->end - new->start + 1;
1803         num_extents += count_max_extents(old_size);
1804         if (count_max_extents(new_size) >= num_extents)
1805                 return;
1806
1807         spin_lock(&BTRFS_I(inode)->lock);
1808         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1809         spin_unlock(&BTRFS_I(inode)->lock);
1810 }
1811
1812 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1813                                       struct inode *inode)
1814 {
1815         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1816
1817         spin_lock(&root->delalloc_lock);
1818         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1819                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1820                               &root->delalloc_inodes);
1821                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1822                         &BTRFS_I(inode)->runtime_flags);
1823                 root->nr_delalloc_inodes++;
1824                 if (root->nr_delalloc_inodes == 1) {
1825                         spin_lock(&fs_info->delalloc_root_lock);
1826                         BUG_ON(!list_empty(&root->delalloc_root));
1827                         list_add_tail(&root->delalloc_root,
1828                                       &fs_info->delalloc_roots);
1829                         spin_unlock(&fs_info->delalloc_root_lock);
1830                 }
1831         }
1832         spin_unlock(&root->delalloc_lock);
1833 }
1834
1835
1836 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1837                                 struct btrfs_inode *inode)
1838 {
1839         struct btrfs_fs_info *fs_info = root->fs_info;
1840
1841         if (!list_empty(&inode->delalloc_inodes)) {
1842                 list_del_init(&inode->delalloc_inodes);
1843                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1844                           &inode->runtime_flags);
1845                 root->nr_delalloc_inodes--;
1846                 if (!root->nr_delalloc_inodes) {
1847                         ASSERT(list_empty(&root->delalloc_inodes));
1848                         spin_lock(&fs_info->delalloc_root_lock);
1849                         BUG_ON(list_empty(&root->delalloc_root));
1850                         list_del_init(&root->delalloc_root);
1851                         spin_unlock(&fs_info->delalloc_root_lock);
1852                 }
1853         }
1854 }
1855
1856 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1857                                      struct btrfs_inode *inode)
1858 {
1859         spin_lock(&root->delalloc_lock);
1860         __btrfs_del_delalloc_inode(root, inode);
1861         spin_unlock(&root->delalloc_lock);
1862 }
1863
1864 /*
1865  * Properly track delayed allocation bytes in the inode and to maintain the
1866  * list of inodes that have pending delalloc work to be done.
1867  */
1868 void btrfs_set_delalloc_extent(struct inode *inode, struct extent_state *state,
1869                                unsigned *bits)
1870 {
1871         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1872
1873         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1874                 WARN_ON(1);
1875         /*
1876          * set_bit and clear bit hooks normally require _irqsave/restore
1877          * but in this case, we are only testing for the DELALLOC
1878          * bit, which is only set or cleared with irqs on
1879          */
1880         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1881                 struct btrfs_root *root = BTRFS_I(inode)->root;
1882                 u64 len = state->end + 1 - state->start;
1883                 u32 num_extents = count_max_extents(len);
1884                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1885
1886                 spin_lock(&BTRFS_I(inode)->lock);
1887                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1888                 spin_unlock(&BTRFS_I(inode)->lock);
1889
1890                 /* For sanity tests */
1891                 if (btrfs_is_testing(fs_info))
1892                         return;
1893
1894                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1895                                          fs_info->delalloc_batch);
1896                 spin_lock(&BTRFS_I(inode)->lock);
1897                 BTRFS_I(inode)->delalloc_bytes += len;
1898                 if (*bits & EXTENT_DEFRAG)
1899                         BTRFS_I(inode)->defrag_bytes += len;
1900                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1901                                          &BTRFS_I(inode)->runtime_flags))
1902                         btrfs_add_delalloc_inodes(root, inode);
1903                 spin_unlock(&BTRFS_I(inode)->lock);
1904         }
1905
1906         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1907             (*bits & EXTENT_DELALLOC_NEW)) {
1908                 spin_lock(&BTRFS_I(inode)->lock);
1909                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1910                         state->start;
1911                 spin_unlock(&BTRFS_I(inode)->lock);
1912         }
1913 }
1914
1915 /*
1916  * Once a range is no longer delalloc this function ensures that proper
1917  * accounting happens.
1918  */
1919 void btrfs_clear_delalloc_extent(struct inode *vfs_inode,
1920                                  struct extent_state *state, unsigned *bits)
1921 {
1922         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
1923         struct btrfs_fs_info *fs_info = btrfs_sb(vfs_inode->i_sb);
1924         u64 len = state->end + 1 - state->start;
1925         u32 num_extents = count_max_extents(len);
1926
1927         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1928                 spin_lock(&inode->lock);
1929                 inode->defrag_bytes -= len;
1930                 spin_unlock(&inode->lock);
1931         }
1932
1933         /*
1934          * set_bit and clear bit hooks normally require _irqsave/restore
1935          * but in this case, we are only testing for the DELALLOC
1936          * bit, which is only set or cleared with irqs on
1937          */
1938         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1939                 struct btrfs_root *root = inode->root;
1940                 bool do_list = !btrfs_is_free_space_inode(inode);
1941
1942                 spin_lock(&inode->lock);
1943                 btrfs_mod_outstanding_extents(inode, -num_extents);
1944                 spin_unlock(&inode->lock);
1945
1946                 /*
1947                  * We don't reserve metadata space for space cache inodes so we
1948                  * don't need to call delalloc_release_metadata if there is an
1949                  * error.
1950                  */
1951                 if (*bits & EXTENT_CLEAR_META_RESV &&
1952                     root != fs_info->tree_root)
1953                         btrfs_delalloc_release_metadata(inode, len, false);
1954
1955                 /* For sanity tests. */
1956                 if (btrfs_is_testing(fs_info))
1957                         return;
1958
1959                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1960                     do_list && !(state->state & EXTENT_NORESERVE) &&
1961                     (*bits & EXTENT_CLEAR_DATA_RESV))
1962                         btrfs_free_reserved_data_space_noquota(
1963                                         &inode->vfs_inode,
1964                                         state->start, len);
1965
1966                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1967                                          fs_info->delalloc_batch);
1968                 spin_lock(&inode->lock);
1969                 inode->delalloc_bytes -= len;
1970                 if (do_list && inode->delalloc_bytes == 0 &&
1971                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1972                                         &inode->runtime_flags))
1973                         btrfs_del_delalloc_inode(root, inode);
1974                 spin_unlock(&inode->lock);
1975         }
1976
1977         if ((state->state & EXTENT_DELALLOC_NEW) &&
1978             (*bits & EXTENT_DELALLOC_NEW)) {
1979                 spin_lock(&inode->lock);
1980                 ASSERT(inode->new_delalloc_bytes >= len);
1981                 inode->new_delalloc_bytes -= len;
1982                 spin_unlock(&inode->lock);
1983         }
1984 }
1985
1986 /*
1987  * btrfs_bio_fits_in_stripe - Checks whether the size of the given bio will fit
1988  * in a chunk's stripe. This function ensures that bios do not span a
1989  * stripe/chunk
1990  *
1991  * @page - The page we are about to add to the bio
1992  * @size - size we want to add to the bio
1993  * @bio - bio we want to ensure is smaller than a stripe
1994  * @bio_flags - flags of the bio
1995  *
1996  * return 1 if page cannot be added to the bio
1997  * return 0 if page can be added to the bio
1998  * return error otherwise
1999  */
2000 int btrfs_bio_fits_in_stripe(struct page *page, size_t size, struct bio *bio,
2001                              unsigned long bio_flags)
2002 {
2003         struct inode *inode = page->mapping->host;
2004         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2005         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
2006         u64 length = 0;
2007         u64 map_length;
2008         int ret;
2009         struct btrfs_io_geometry geom;
2010
2011         if (bio_flags & EXTENT_BIO_COMPRESSED)
2012                 return 0;
2013
2014         length = bio->bi_iter.bi_size;
2015         map_length = length;
2016         ret = btrfs_get_io_geometry(fs_info, btrfs_op(bio), logical, map_length,
2017                                     &geom);
2018         if (ret < 0)
2019                 return ret;
2020
2021         if (geom.len < length + size)
2022                 return 1;
2023         return 0;
2024 }
2025
2026 /*
2027  * in order to insert checksums into the metadata in large chunks,
2028  * we wait until bio submission time.   All the pages in the bio are
2029  * checksummed and sums are attached onto the ordered extent record.
2030  *
2031  * At IO completion time the cums attached on the ordered extent record
2032  * are inserted into the btree
2033  */
2034 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
2035                                     u64 bio_offset)
2036 {
2037         struct inode *inode = private_data;
2038         blk_status_t ret = 0;
2039
2040         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2041         BUG_ON(ret); /* -ENOMEM */
2042         return 0;
2043 }
2044
2045 /*
2046  * extent_io.c submission hook. This does the right thing for csum calculation
2047  * on write, or reading the csums from the tree before a read.
2048  *
2049  * Rules about async/sync submit,
2050  * a) read:                             sync submit
2051  *
2052  * b) write without checksum:           sync submit
2053  *
2054  * c) write with checksum:
2055  *    c-1) if bio is issued by fsync:   sync submit
2056  *         (sync_writers != 0)
2057  *
2058  *    c-2) if root is reloc root:       sync submit
2059  *         (only in case of buffered IO)
2060  *
2061  *    c-3) otherwise:                   async submit
2062  */
2063 static blk_status_t btrfs_submit_bio_hook(struct inode *inode, struct bio *bio,
2064                                           int mirror_num,
2065                                           unsigned long bio_flags)
2066
2067 {
2068         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2069         struct btrfs_root *root = BTRFS_I(inode)->root;
2070         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2071         blk_status_t ret = 0;
2072         int skip_sum;
2073         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2074
2075         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2076
2077         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2078                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2079
2080         if (bio_op(bio) != REQ_OP_WRITE) {
2081                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2082                 if (ret)
2083                         goto out;
2084
2085                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2086                         ret = btrfs_submit_compressed_read(inode, bio,
2087                                                            mirror_num,
2088                                                            bio_flags);
2089                         goto out;
2090                 } else if (!skip_sum) {
2091                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2092                         if (ret)
2093                                 goto out;
2094                 }
2095                 goto mapit;
2096         } else if (async && !skip_sum) {
2097                 /* csum items have already been cloned */
2098                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2099                         goto mapit;
2100                 /* we're doing a write, do the async checksumming */
2101                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2102                                           0, inode, btrfs_submit_bio_start);
2103                 goto out;
2104         } else if (!skip_sum) {
2105                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2106                 if (ret)
2107                         goto out;
2108         }
2109
2110 mapit:
2111         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2112
2113 out:
2114         if (ret) {
2115                 bio->bi_status = ret;
2116                 bio_endio(bio);
2117         }
2118         return ret;
2119 }
2120
2121 /*
2122  * given a list of ordered sums record them in the inode.  This happens
2123  * at IO completion time based on sums calculated at bio submission time.
2124  */
2125 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2126                              struct inode *inode, struct list_head *list)
2127 {
2128         struct btrfs_ordered_sum *sum;
2129         int ret;
2130
2131         list_for_each_entry(sum, list, list) {
2132                 trans->adding_csums = true;
2133                 ret = btrfs_csum_file_blocks(trans,
2134                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2135                 trans->adding_csums = false;
2136                 if (ret)
2137                         return ret;
2138         }
2139         return 0;
2140 }
2141
2142 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2143                               unsigned int extra_bits,
2144                               struct extent_state **cached_state)
2145 {
2146         WARN_ON(PAGE_ALIGNED(end));
2147         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2148                                    extra_bits, cached_state);
2149 }
2150
2151 /* see btrfs_writepage_start_hook for details on why this is required */
2152 struct btrfs_writepage_fixup {
2153         struct page *page;
2154         struct btrfs_work work;
2155 };
2156
2157 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2158 {
2159         struct btrfs_writepage_fixup *fixup;
2160         struct btrfs_ordered_extent *ordered;
2161         struct extent_state *cached_state = NULL;
2162         struct extent_changeset *data_reserved = NULL;
2163         struct page *page;
2164         struct inode *inode;
2165         u64 page_start;
2166         u64 page_end;
2167         int ret;
2168
2169         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2170         page = fixup->page;
2171 again:
2172         lock_page(page);
2173         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2174                 ClearPageChecked(page);
2175                 goto out_page;
2176         }
2177
2178         inode = page->mapping->host;
2179         page_start = page_offset(page);
2180         page_end = page_offset(page) + PAGE_SIZE - 1;
2181
2182         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2183                          &cached_state);
2184
2185         /* already ordered? We're done */
2186         if (PagePrivate2(page))
2187                 goto out;
2188
2189         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2190                                         PAGE_SIZE);
2191         if (ordered) {
2192                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2193                                      page_end, &cached_state);
2194                 unlock_page(page);
2195                 btrfs_start_ordered_extent(inode, ordered, 1);
2196                 btrfs_put_ordered_extent(ordered);
2197                 goto again;
2198         }
2199
2200         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2201                                            PAGE_SIZE);
2202         if (ret) {
2203                 mapping_set_error(page->mapping, ret);
2204                 end_extent_writepage(page, ret, page_start, page_end);
2205                 ClearPageChecked(page);
2206                 goto out;
2207          }
2208
2209         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2210                                         &cached_state);
2211         if (ret) {
2212                 mapping_set_error(page->mapping, ret);
2213                 end_extent_writepage(page, ret, page_start, page_end);
2214                 ClearPageChecked(page);
2215                 goto out;
2216         }
2217
2218         ClearPageChecked(page);
2219         set_page_dirty(page);
2220         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2221 out:
2222         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2223                              &cached_state);
2224 out_page:
2225         unlock_page(page);
2226         put_page(page);
2227         kfree(fixup);
2228         extent_changeset_free(data_reserved);
2229 }
2230
2231 /*
2232  * There are a few paths in the higher layers of the kernel that directly
2233  * set the page dirty bit without asking the filesystem if it is a
2234  * good idea.  This causes problems because we want to make sure COW
2235  * properly happens and the data=ordered rules are followed.
2236  *
2237  * In our case any range that doesn't have the ORDERED bit set
2238  * hasn't been properly setup for IO.  We kick off an async process
2239  * to fix it up.  The async helper will wait for ordered extents, set
2240  * the delalloc bit and make it safe to write the page.
2241  */
2242 int btrfs_writepage_cow_fixup(struct page *page, u64 start, u64 end)
2243 {
2244         struct inode *inode = page->mapping->host;
2245         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2246         struct btrfs_writepage_fixup *fixup;
2247
2248         /* this page is properly in the ordered list */
2249         if (TestClearPagePrivate2(page))
2250                 return 0;
2251
2252         if (PageChecked(page))
2253                 return -EAGAIN;
2254
2255         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2256         if (!fixup)
2257                 return -EAGAIN;
2258
2259         SetPageChecked(page);
2260         get_page(page);
2261         btrfs_init_work(&fixup->work, btrfs_writepage_fixup_worker, NULL, NULL);
2262         fixup->page = page;
2263         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2264         return -EBUSY;
2265 }
2266
2267 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2268                                        struct inode *inode, u64 file_pos,
2269                                        u64 disk_bytenr, u64 disk_num_bytes,
2270                                        u64 num_bytes, u64 ram_bytes,
2271                                        u8 compression, u8 encryption,
2272                                        u16 other_encoding, int extent_type)
2273 {
2274         struct btrfs_root *root = BTRFS_I(inode)->root;
2275         struct btrfs_file_extent_item *fi;
2276         struct btrfs_path *path;
2277         struct extent_buffer *leaf;
2278         struct btrfs_key ins;
2279         u64 qg_released;
2280         int extent_inserted = 0;
2281         int ret;
2282
2283         path = btrfs_alloc_path();
2284         if (!path)
2285                 return -ENOMEM;
2286
2287         /*
2288          * we may be replacing one extent in the tree with another.
2289          * The new extent is pinned in the extent map, and we don't want
2290          * to drop it from the cache until it is completely in the btree.
2291          *
2292          * So, tell btrfs_drop_extents to leave this extent in the cache.
2293          * the caller is expected to unpin it and allow it to be merged
2294          * with the others.
2295          */
2296         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2297                                    file_pos + num_bytes, NULL, 0,
2298                                    1, sizeof(*fi), &extent_inserted);
2299         if (ret)
2300                 goto out;
2301
2302         if (!extent_inserted) {
2303                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2304                 ins.offset = file_pos;
2305                 ins.type = BTRFS_EXTENT_DATA_KEY;
2306
2307                 path->leave_spinning = 1;
2308                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2309                                               sizeof(*fi));
2310                 if (ret)
2311                         goto out;
2312         }
2313         leaf = path->nodes[0];
2314         fi = btrfs_item_ptr(leaf, path->slots[0],
2315                             struct btrfs_file_extent_item);
2316         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2317         btrfs_set_file_extent_type(leaf, fi, extent_type);
2318         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2319         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2320         btrfs_set_file_extent_offset(leaf, fi, 0);
2321         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2322         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2323         btrfs_set_file_extent_compression(leaf, fi, compression);
2324         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2325         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2326
2327         btrfs_mark_buffer_dirty(leaf);
2328         btrfs_release_path(path);
2329
2330         inode_add_bytes(inode, num_bytes);
2331
2332         ins.objectid = disk_bytenr;
2333         ins.offset = disk_num_bytes;
2334         ins.type = BTRFS_EXTENT_ITEM_KEY;
2335
2336         /*
2337          * Release the reserved range from inode dirty range map, as it is
2338          * already moved into delayed_ref_head
2339          */
2340         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2341         if (ret < 0)
2342                 goto out;
2343         qg_released = ret;
2344         ret = btrfs_alloc_reserved_file_extent(trans, root,
2345                                                btrfs_ino(BTRFS_I(inode)),
2346                                                file_pos, qg_released, &ins);
2347 out:
2348         btrfs_free_path(path);
2349
2350         return ret;
2351 }
2352
2353 /* snapshot-aware defrag */
2354 struct sa_defrag_extent_backref {
2355         struct rb_node node;
2356         struct old_sa_defrag_extent *old;
2357         u64 root_id;
2358         u64 inum;
2359         u64 file_pos;
2360         u64 extent_offset;
2361         u64 num_bytes;
2362         u64 generation;
2363 };
2364
2365 struct old_sa_defrag_extent {
2366         struct list_head list;
2367         struct new_sa_defrag_extent *new;
2368
2369         u64 extent_offset;
2370         u64 bytenr;
2371         u64 offset;
2372         u64 len;
2373         int count;
2374 };
2375
2376 struct new_sa_defrag_extent {
2377         struct rb_root root;
2378         struct list_head head;
2379         struct btrfs_path *path;
2380         struct inode *inode;
2381         u64 file_pos;
2382         u64 len;
2383         u64 bytenr;
2384         u64 disk_len;
2385         u8 compress_type;
2386 };
2387
2388 static int backref_comp(struct sa_defrag_extent_backref *b1,
2389                         struct sa_defrag_extent_backref *b2)
2390 {
2391         if (b1->root_id < b2->root_id)
2392                 return -1;
2393         else if (b1->root_id > b2->root_id)
2394                 return 1;
2395
2396         if (b1->inum < b2->inum)
2397                 return -1;
2398         else if (b1->inum > b2->inum)
2399                 return 1;
2400
2401         if (b1->file_pos < b2->file_pos)
2402                 return -1;
2403         else if (b1->file_pos > b2->file_pos)
2404                 return 1;
2405
2406         /*
2407          * [------------------------------] ===> (a range of space)
2408          *     |<--->|   |<---->| =============> (fs/file tree A)
2409          * |<---------------------------->| ===> (fs/file tree B)
2410          *
2411          * A range of space can refer to two file extents in one tree while
2412          * refer to only one file extent in another tree.
2413          *
2414          * So we may process a disk offset more than one time(two extents in A)
2415          * and locate at the same extent(one extent in B), then insert two same
2416          * backrefs(both refer to the extent in B).
2417          */
2418         return 0;
2419 }
2420
2421 static void backref_insert(struct rb_root *root,
2422                            struct sa_defrag_extent_backref *backref)
2423 {
2424         struct rb_node **p = &root->rb_node;
2425         struct rb_node *parent = NULL;
2426         struct sa_defrag_extent_backref *entry;
2427         int ret;
2428
2429         while (*p) {
2430                 parent = *p;
2431                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2432
2433                 ret = backref_comp(backref, entry);
2434                 if (ret < 0)
2435                         p = &(*p)->rb_left;
2436                 else
2437                         p = &(*p)->rb_right;
2438         }
2439
2440         rb_link_node(&backref->node, parent, p);
2441         rb_insert_color(&backref->node, root);
2442 }
2443
2444 /*
2445  * Note the backref might has changed, and in this case we just return 0.
2446  */
2447 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2448                                        void *ctx)
2449 {
2450         struct btrfs_file_extent_item *extent;
2451         struct old_sa_defrag_extent *old = ctx;
2452         struct new_sa_defrag_extent *new = old->new;
2453         struct btrfs_path *path = new->path;
2454         struct btrfs_key key;
2455         struct btrfs_root *root;
2456         struct sa_defrag_extent_backref *backref;
2457         struct extent_buffer *leaf;
2458         struct inode *inode = new->inode;
2459         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2460         int slot;
2461         int ret;
2462         u64 extent_offset;
2463         u64 num_bytes;
2464
2465         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2466             inum == btrfs_ino(BTRFS_I(inode)))
2467                 return 0;
2468
2469         key.objectid = root_id;
2470         key.type = BTRFS_ROOT_ITEM_KEY;
2471         key.offset = (u64)-1;
2472
2473         root = btrfs_read_fs_root_no_name(fs_info, &key);
2474         if (IS_ERR(root)) {
2475                 if (PTR_ERR(root) == -ENOENT)
2476                         return 0;
2477                 WARN_ON(1);
2478                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2479                          inum, offset, root_id);
2480                 return PTR_ERR(root);
2481         }
2482
2483         key.objectid = inum;
2484         key.type = BTRFS_EXTENT_DATA_KEY;
2485         if (offset > (u64)-1 << 32)
2486                 key.offset = 0;
2487         else
2488                 key.offset = offset;
2489
2490         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2491         if (WARN_ON(ret < 0))
2492                 return ret;
2493         ret = 0;
2494
2495         while (1) {
2496                 cond_resched();
2497
2498                 leaf = path->nodes[0];
2499                 slot = path->slots[0];
2500
2501                 if (slot >= btrfs_header_nritems(leaf)) {
2502                         ret = btrfs_next_leaf(root, path);
2503                         if (ret < 0) {
2504                                 goto out;
2505                         } else if (ret > 0) {
2506                                 ret = 0;
2507                                 goto out;
2508                         }
2509                         continue;
2510                 }
2511
2512                 path->slots[0]++;
2513
2514                 btrfs_item_key_to_cpu(leaf, &key, slot);
2515
2516                 if (key.objectid > inum)
2517                         goto out;
2518
2519                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2520                         continue;
2521
2522                 extent = btrfs_item_ptr(leaf, slot,
2523                                         struct btrfs_file_extent_item);
2524
2525                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2526                         continue;
2527
2528                 /*
2529                  * 'offset' refers to the exact key.offset,
2530                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2531                  * (key.offset - extent_offset).
2532                  */
2533                 if (key.offset != offset)
2534                         continue;
2535
2536                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2537                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2538
2539                 if (extent_offset >= old->extent_offset + old->offset +
2540                     old->len || extent_offset + num_bytes <=
2541                     old->extent_offset + old->offset)
2542                         continue;
2543                 break;
2544         }
2545
2546         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2547         if (!backref) {
2548                 ret = -ENOENT;
2549                 goto out;
2550         }
2551
2552         backref->root_id = root_id;
2553         backref->inum = inum;
2554         backref->file_pos = offset;
2555         backref->num_bytes = num_bytes;
2556         backref->extent_offset = extent_offset;
2557         backref->generation = btrfs_file_extent_generation(leaf, extent);
2558         backref->old = old;
2559         backref_insert(&new->root, backref);
2560         old->count++;
2561 out:
2562         btrfs_release_path(path);
2563         WARN_ON(ret);
2564         return ret;
2565 }
2566
2567 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2568                                    struct new_sa_defrag_extent *new)
2569 {
2570         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2571         struct old_sa_defrag_extent *old, *tmp;
2572         int ret;
2573
2574         new->path = path;
2575
2576         list_for_each_entry_safe(old, tmp, &new->head, list) {
2577                 ret = iterate_inodes_from_logical(old->bytenr +
2578                                                   old->extent_offset, fs_info,
2579                                                   path, record_one_backref,
2580                                                   old, false);
2581                 if (ret < 0 && ret != -ENOENT)
2582                         return false;
2583
2584                 /* no backref to be processed for this extent */
2585                 if (!old->count) {
2586                         list_del(&old->list);
2587                         kfree(old);
2588                 }
2589         }
2590
2591         if (list_empty(&new->head))
2592                 return false;
2593
2594         return true;
2595 }
2596
2597 static int relink_is_mergable(struct extent_buffer *leaf,
2598                               struct btrfs_file_extent_item *fi,
2599                               struct new_sa_defrag_extent *new)
2600 {
2601         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2602                 return 0;
2603
2604         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2605                 return 0;
2606
2607         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2608                 return 0;
2609
2610         if (btrfs_file_extent_encryption(leaf, fi) ||
2611             btrfs_file_extent_other_encoding(leaf, fi))
2612                 return 0;
2613
2614         return 1;
2615 }
2616
2617 /*
2618  * Note the backref might has changed, and in this case we just return 0.
2619  */
2620 static noinline int relink_extent_backref(struct btrfs_path *path,
2621                                  struct sa_defrag_extent_backref *prev,
2622                                  struct sa_defrag_extent_backref *backref)
2623 {
2624         struct btrfs_file_extent_item *extent;
2625         struct btrfs_file_extent_item *item;
2626         struct btrfs_ordered_extent *ordered;
2627         struct btrfs_trans_handle *trans;
2628         struct btrfs_ref ref = { 0 };
2629         struct btrfs_root *root;
2630         struct btrfs_key key;
2631         struct extent_buffer *leaf;
2632         struct old_sa_defrag_extent *old = backref->old;
2633         struct new_sa_defrag_extent *new = old->new;
2634         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2635         struct inode *inode;
2636         struct extent_state *cached = NULL;
2637         int ret = 0;
2638         u64 start;
2639         u64 len;
2640         u64 lock_start;
2641         u64 lock_end;
2642         bool merge = false;
2643         int index;
2644
2645         if (prev && prev->root_id == backref->root_id &&
2646             prev->inum == backref->inum &&
2647             prev->file_pos + prev->num_bytes == backref->file_pos)
2648                 merge = true;
2649
2650         /* step 1: get root */
2651         key.objectid = backref->root_id;
2652         key.type = BTRFS_ROOT_ITEM_KEY;
2653         key.offset = (u64)-1;
2654
2655         index = srcu_read_lock(&fs_info->subvol_srcu);
2656
2657         root = btrfs_read_fs_root_no_name(fs_info, &key);
2658         if (IS_ERR(root)) {
2659                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2660                 if (PTR_ERR(root) == -ENOENT)
2661                         return 0;
2662                 return PTR_ERR(root);
2663         }
2664
2665         if (btrfs_root_readonly(root)) {
2666                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2667                 return 0;
2668         }
2669
2670         /* step 2: get inode */
2671         key.objectid = backref->inum;
2672         key.type = BTRFS_INODE_ITEM_KEY;
2673         key.offset = 0;
2674
2675         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2676         if (IS_ERR(inode)) {
2677                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2678                 return 0;
2679         }
2680
2681         srcu_read_unlock(&fs_info->subvol_srcu, index);
2682
2683         /* step 3: relink backref */
2684         lock_start = backref->file_pos;
2685         lock_end = backref->file_pos + backref->num_bytes - 1;
2686         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2687                          &cached);
2688
2689         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2690         if (ordered) {
2691                 btrfs_put_ordered_extent(ordered);
2692                 goto out_unlock;
2693         }
2694
2695         trans = btrfs_join_transaction(root);
2696         if (IS_ERR(trans)) {
2697                 ret = PTR_ERR(trans);
2698                 goto out_unlock;
2699         }
2700
2701         key.objectid = backref->inum;
2702         key.type = BTRFS_EXTENT_DATA_KEY;
2703         key.offset = backref->file_pos;
2704
2705         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2706         if (ret < 0) {
2707                 goto out_free_path;
2708         } else if (ret > 0) {
2709                 ret = 0;
2710                 goto out_free_path;
2711         }
2712
2713         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2714                                 struct btrfs_file_extent_item);
2715
2716         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2717             backref->generation)
2718                 goto out_free_path;
2719
2720         btrfs_release_path(path);
2721
2722         start = backref->file_pos;
2723         if (backref->extent_offset < old->extent_offset + old->offset)
2724                 start += old->extent_offset + old->offset -
2725                          backref->extent_offset;
2726
2727         len = min(backref->extent_offset + backref->num_bytes,
2728                   old->extent_offset + old->offset + old->len);
2729         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2730
2731         ret = btrfs_drop_extents(trans, root, inode, start,
2732                                  start + len, 1);
2733         if (ret)
2734                 goto out_free_path;
2735 again:
2736         key.objectid = btrfs_ino(BTRFS_I(inode));
2737         key.type = BTRFS_EXTENT_DATA_KEY;
2738         key.offset = start;
2739
2740         path->leave_spinning = 1;
2741         if (merge) {
2742                 struct btrfs_file_extent_item *fi;
2743                 u64 extent_len;
2744                 struct btrfs_key found_key;
2745
2746                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2747                 if (ret < 0)
2748                         goto out_free_path;
2749
2750                 path->slots[0]--;
2751                 leaf = path->nodes[0];
2752                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2753
2754                 fi = btrfs_item_ptr(leaf, path->slots[0],
2755                                     struct btrfs_file_extent_item);
2756                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2757
2758                 if (extent_len + found_key.offset == start &&
2759                     relink_is_mergable(leaf, fi, new)) {
2760                         btrfs_set_file_extent_num_bytes(leaf, fi,
2761                                                         extent_len + len);
2762                         btrfs_mark_buffer_dirty(leaf);
2763                         inode_add_bytes(inode, len);
2764
2765                         ret = 1;
2766                         goto out_free_path;
2767                 } else {
2768                         merge = false;
2769                         btrfs_release_path(path);
2770                         goto again;
2771                 }
2772         }
2773
2774         ret = btrfs_insert_empty_item(trans, root, path, &key,
2775                                         sizeof(*extent));
2776         if (ret) {
2777                 btrfs_abort_transaction(trans, ret);
2778                 goto out_free_path;
2779         }
2780
2781         leaf = path->nodes[0];
2782         item = btrfs_item_ptr(leaf, path->slots[0],
2783                                 struct btrfs_file_extent_item);
2784         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2785         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2786         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2787         btrfs_set_file_extent_num_bytes(leaf, item, len);
2788         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2789         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2790         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2791         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2792         btrfs_set_file_extent_encryption(leaf, item, 0);
2793         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2794
2795         btrfs_mark_buffer_dirty(leaf);
2796         inode_add_bytes(inode, len);
2797         btrfs_release_path(path);
2798
2799         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new->bytenr,
2800                                new->disk_len, 0);
2801         btrfs_init_data_ref(&ref, backref->root_id, backref->inum,
2802                             new->file_pos);  /* start - extent_offset */
2803         ret = btrfs_inc_extent_ref(trans, &ref);
2804         if (ret) {
2805                 btrfs_abort_transaction(trans, ret);
2806                 goto out_free_path;
2807         }
2808
2809         ret = 1;
2810 out_free_path:
2811         btrfs_release_path(path);
2812         path->leave_spinning = 0;
2813         btrfs_end_transaction(trans);
2814 out_unlock:
2815         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2816                              &cached);
2817         iput(inode);
2818         return ret;
2819 }
2820
2821 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2822 {
2823         struct old_sa_defrag_extent *old, *tmp;
2824
2825         if (!new)
2826                 return;
2827
2828         list_for_each_entry_safe(old, tmp, &new->head, list) {
2829                 kfree(old);
2830         }
2831         kfree(new);
2832 }
2833
2834 static void relink_file_extents(struct new_sa_defrag_extent *new)
2835 {
2836         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2837         struct btrfs_path *path;
2838         struct sa_defrag_extent_backref *backref;
2839         struct sa_defrag_extent_backref *prev = NULL;
2840         struct rb_node *node;
2841         int ret;
2842
2843         path = btrfs_alloc_path();
2844         if (!path)
2845                 return;
2846
2847         if (!record_extent_backrefs(path, new)) {
2848                 btrfs_free_path(path);
2849                 goto out;
2850         }
2851         btrfs_release_path(path);
2852
2853         while (1) {
2854                 node = rb_first(&new->root);
2855                 if (!node)
2856                         break;
2857                 rb_erase(node, &new->root);
2858
2859                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2860
2861                 ret = relink_extent_backref(path, prev, backref);
2862                 WARN_ON(ret < 0);
2863
2864                 kfree(prev);
2865
2866                 if (ret == 1)
2867                         prev = backref;
2868                 else
2869                         prev = NULL;
2870                 cond_resched();
2871         }
2872         kfree(prev);
2873
2874         btrfs_free_path(path);
2875 out:
2876         free_sa_defrag_extent(new);
2877
2878         atomic_dec(&fs_info->defrag_running);
2879         wake_up(&fs_info->transaction_wait);
2880 }
2881
2882 static struct new_sa_defrag_extent *
2883 record_old_file_extents(struct inode *inode,
2884                         struct btrfs_ordered_extent *ordered)
2885 {
2886         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2887         struct btrfs_root *root = BTRFS_I(inode)->root;
2888         struct btrfs_path *path;
2889         struct btrfs_key key;
2890         struct old_sa_defrag_extent *old;
2891         struct new_sa_defrag_extent *new;
2892         int ret;
2893
2894         new = kmalloc(sizeof(*new), GFP_NOFS);
2895         if (!new)
2896                 return NULL;
2897
2898         new->inode = inode;
2899         new->file_pos = ordered->file_offset;
2900         new->len = ordered->len;
2901         new->bytenr = ordered->start;
2902         new->disk_len = ordered->disk_len;
2903         new->compress_type = ordered->compress_type;
2904         new->root = RB_ROOT;
2905         INIT_LIST_HEAD(&new->head);
2906
2907         path = btrfs_alloc_path();
2908         if (!path)
2909                 goto out_kfree;
2910
2911         key.objectid = btrfs_ino(BTRFS_I(inode));
2912         key.type = BTRFS_EXTENT_DATA_KEY;
2913         key.offset = new->file_pos;
2914
2915         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2916         if (ret < 0)
2917                 goto out_free_path;
2918         if (ret > 0 && path->slots[0] > 0)
2919                 path->slots[0]--;
2920
2921         /* find out all the old extents for the file range */
2922         while (1) {
2923                 struct btrfs_file_extent_item *extent;
2924                 struct extent_buffer *l;
2925                 int slot;
2926                 u64 num_bytes;
2927                 u64 offset;
2928                 u64 end;
2929                 u64 disk_bytenr;
2930                 u64 extent_offset;
2931
2932                 l = path->nodes[0];
2933                 slot = path->slots[0];
2934
2935                 if (slot >= btrfs_header_nritems(l)) {
2936                         ret = btrfs_next_leaf(root, path);
2937                         if (ret < 0)
2938                                 goto out_free_path;
2939                         else if (ret > 0)
2940                                 break;
2941                         continue;
2942                 }
2943
2944                 btrfs_item_key_to_cpu(l, &key, slot);
2945
2946                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2947                         break;
2948                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2949                         break;
2950                 if (key.offset >= new->file_pos + new->len)
2951                         break;
2952
2953                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2954
2955                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2956                 if (key.offset + num_bytes < new->file_pos)
2957                         goto next;
2958
2959                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2960                 if (!disk_bytenr)
2961                         goto next;
2962
2963                 extent_offset = btrfs_file_extent_offset(l, extent);
2964
2965                 old = kmalloc(sizeof(*old), GFP_NOFS);
2966                 if (!old)
2967                         goto out_free_path;
2968
2969                 offset = max(new->file_pos, key.offset);
2970                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2971
2972                 old->bytenr = disk_bytenr;
2973                 old->extent_offset = extent_offset;
2974                 old->offset = offset - key.offset;
2975                 old->len = end - offset;
2976                 old->new = new;
2977                 old->count = 0;
2978                 list_add_tail(&old->list, &new->head);
2979 next:
2980                 path->slots[0]++;
2981                 cond_resched();
2982         }
2983
2984         btrfs_free_path(path);
2985         atomic_inc(&fs_info->defrag_running);
2986
2987         return new;
2988
2989 out_free_path:
2990         btrfs_free_path(path);
2991 out_kfree:
2992         free_sa_defrag_extent(new);
2993         return NULL;
2994 }
2995
2996 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2997                                          u64 start, u64 len)
2998 {
2999         struct btrfs_block_group_cache *cache;
3000
3001         cache = btrfs_lookup_block_group(fs_info, start);
3002         ASSERT(cache);
3003
3004         spin_lock(&cache->lock);
3005         cache->delalloc_bytes -= len;
3006         spin_unlock(&cache->lock);
3007
3008         btrfs_put_block_group(cache);
3009 }
3010
3011 /* as ordered data IO finishes, this gets called so we can finish
3012  * an ordered extent if the range of bytes in the file it covers are
3013  * fully written.
3014  */
3015 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
3016 {
3017         struct inode *inode = ordered_extent->inode;
3018         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3019         struct btrfs_root *root = BTRFS_I(inode)->root;
3020         struct btrfs_trans_handle *trans = NULL;
3021         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3022         struct extent_state *cached_state = NULL;
3023         struct new_sa_defrag_extent *new = NULL;
3024         int compress_type = 0;
3025         int ret = 0;
3026         u64 logical_len = ordered_extent->len;
3027         bool nolock;
3028         bool truncated = false;
3029         bool range_locked = false;
3030         bool clear_new_delalloc_bytes = false;
3031         bool clear_reserved_extent = true;
3032
3033         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3034             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
3035             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
3036                 clear_new_delalloc_bytes = true;
3037
3038         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
3039
3040         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
3041                 ret = -EIO;
3042                 goto out;
3043         }
3044
3045         btrfs_free_io_failure_record(BTRFS_I(inode),
3046                         ordered_extent->file_offset,
3047                         ordered_extent->file_offset +
3048                         ordered_extent->len - 1);
3049
3050         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
3051                 truncated = true;
3052                 logical_len = ordered_extent->truncated_len;
3053                 /* Truncated the entire extent, don't bother adding */
3054                 if (!logical_len)
3055                         goto out;
3056         }
3057
3058         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
3059                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
3060
3061                 /*
3062                  * For mwrite(mmap + memset to write) case, we still reserve
3063                  * space for NOCOW range.
3064                  * As NOCOW won't cause a new delayed ref, just free the space
3065                  */
3066                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3067                                        ordered_extent->len);
3068                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3069                 if (nolock)
3070                         trans = btrfs_join_transaction_nolock(root);
3071                 else
3072                         trans = btrfs_join_transaction(root);
3073                 if (IS_ERR(trans)) {
3074                         ret = PTR_ERR(trans);
3075                         trans = NULL;
3076                         goto out;
3077                 }
3078                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3079                 ret = btrfs_update_inode_fallback(trans, root, inode);
3080                 if (ret) /* -ENOMEM or corruption */
3081                         btrfs_abort_transaction(trans, ret);
3082                 goto out;
3083         }
3084
3085         range_locked = true;
3086         lock_extent_bits(io_tree, ordered_extent->file_offset,
3087                          ordered_extent->file_offset + ordered_extent->len - 1,
3088                          &cached_state);
3089
3090         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3091                         ordered_extent->file_offset + ordered_extent->len - 1,
3092                         EXTENT_DEFRAG, 0, cached_state);
3093         if (ret) {
3094                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3095                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3096                         /* the inode is shared */
3097                         new = record_old_file_extents(inode, ordered_extent);
3098
3099                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3100                         ordered_extent->file_offset + ordered_extent->len - 1,
3101                         EXTENT_DEFRAG, 0, 0, &cached_state);
3102         }
3103
3104         if (nolock)
3105                 trans = btrfs_join_transaction_nolock(root);
3106         else
3107                 trans = btrfs_join_transaction(root);
3108         if (IS_ERR(trans)) {
3109                 ret = PTR_ERR(trans);
3110                 trans = NULL;
3111                 goto out;
3112         }
3113
3114         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3115
3116         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3117                 compress_type = ordered_extent->compress_type;
3118         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3119                 BUG_ON(compress_type);
3120                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3121                                        ordered_extent->len);
3122                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3123                                                 ordered_extent->file_offset,
3124                                                 ordered_extent->file_offset +
3125                                                 logical_len);
3126         } else {
3127                 BUG_ON(root == fs_info->tree_root);
3128                 ret = insert_reserved_file_extent(trans, inode,
3129                                                 ordered_extent->file_offset,
3130                                                 ordered_extent->start,
3131                                                 ordered_extent->disk_len,
3132                                                 logical_len, logical_len,
3133                                                 compress_type, 0, 0,
3134                                                 BTRFS_FILE_EXTENT_REG);
3135                 if (!ret) {
3136                         clear_reserved_extent = false;
3137                         btrfs_release_delalloc_bytes(fs_info,
3138                                                      ordered_extent->start,
3139                                                      ordered_extent->disk_len);
3140                 }
3141         }
3142         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3143                            ordered_extent->file_offset, ordered_extent->len,
3144                            trans->transid);
3145         if (ret < 0) {
3146                 btrfs_abort_transaction(trans, ret);
3147                 goto out;
3148         }
3149
3150         ret = add_pending_csums(trans, inode, &ordered_extent->list);
3151         if (ret) {
3152                 btrfs_abort_transaction(trans, ret);
3153                 goto out;
3154         }
3155
3156         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3157         ret = btrfs_update_inode_fallback(trans, root, inode);
3158         if (ret) { /* -ENOMEM or corruption */
3159                 btrfs_abort_transaction(trans, ret);
3160                 goto out;
3161         }
3162         ret = 0;
3163 out:
3164         if (range_locked || clear_new_delalloc_bytes) {
3165                 unsigned int clear_bits = 0;
3166
3167                 if (range_locked)
3168                         clear_bits |= EXTENT_LOCKED;
3169                 if (clear_new_delalloc_bytes)
3170                         clear_bits |= EXTENT_DELALLOC_NEW;
3171                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3172                                  ordered_extent->file_offset,
3173                                  ordered_extent->file_offset +
3174                                  ordered_extent->len - 1,
3175                                  clear_bits,
3176                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3177                                  0, &cached_state);
3178         }
3179
3180         if (trans)
3181                 btrfs_end_transaction(trans);
3182
3183         if (ret || truncated) {
3184                 u64 start, end;
3185
3186                 if (truncated)
3187                         start = ordered_extent->file_offset + logical_len;
3188                 else
3189                         start = ordered_extent->file_offset;
3190                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3191                 clear_extent_uptodate(io_tree, start, end, NULL);
3192
3193                 /* Drop the cache for the part of the extent we didn't write. */
3194                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3195
3196                 /*
3197                  * If the ordered extent had an IOERR or something else went
3198                  * wrong we need to return the space for this ordered extent
3199                  * back to the allocator.  We only free the extent in the
3200                  * truncated case if we didn't write out the extent at all.
3201                  *
3202                  * If we made it past insert_reserved_file_extent before we
3203                  * errored out then we don't need to do this as the accounting
3204                  * has already been done.
3205                  */
3206                 if ((ret || !logical_len) &&
3207                     clear_reserved_extent &&
3208                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3209                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3210                         btrfs_free_reserved_extent(fs_info,
3211                                                    ordered_extent->start,
3212                                                    ordered_extent->disk_len, 1);
3213         }
3214
3215
3216         /*
3217          * This needs to be done to make sure anybody waiting knows we are done
3218          * updating everything for this ordered extent.
3219          */
3220         btrfs_remove_ordered_extent(inode, ordered_extent);
3221
3222         /* for snapshot-aware defrag */
3223         if (new) {
3224                 if (ret) {
3225                         free_sa_defrag_extent(new);
3226                         atomic_dec(&fs_info->defrag_running);
3227                 } else {
3228                         relink_file_extents(new);
3229                 }
3230         }
3231
3232         /* once for us */
3233         btrfs_put_ordered_extent(ordered_extent);
3234         /* once for the tree */
3235         btrfs_put_ordered_extent(ordered_extent);
3236
3237         return ret;
3238 }
3239
3240 static void finish_ordered_fn(struct btrfs_work *work)
3241 {
3242         struct btrfs_ordered_extent *ordered_extent;
3243         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3244         btrfs_finish_ordered_io(ordered_extent);
3245 }
3246
3247 void btrfs_writepage_endio_finish_ordered(struct page *page, u64 start,
3248                                           u64 end, int uptodate)
3249 {
3250         struct inode *inode = page->mapping->host;
3251         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3252         struct btrfs_ordered_extent *ordered_extent = NULL;
3253         struct btrfs_workqueue *wq;
3254
3255         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3256
3257         ClearPagePrivate2(page);
3258         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3259                                             end - start + 1, uptodate))
3260                 return;
3261
3262         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
3263                 wq = fs_info->endio_freespace_worker;
3264         else
3265                 wq = fs_info->endio_write_workers;
3266
3267         btrfs_init_work(&ordered_extent->work, finish_ordered_fn, NULL, NULL);
3268         btrfs_queue_work(wq, &ordered_extent->work);
3269 }
3270
3271 static int __readpage_endio_check(struct inode *inode,
3272                                   struct btrfs_io_bio *io_bio,
3273                                   int icsum, struct page *page,
3274                                   int pgoff, u64 start, size_t len)
3275 {
3276         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3277         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
3278         char *kaddr;
3279         u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
3280         u8 *csum_expected;
3281         u8 csum[BTRFS_CSUM_SIZE];
3282
3283         csum_expected = ((u8 *)io_bio->csum) + icsum * csum_size;
3284
3285         kaddr = kmap_atomic(page);
3286         shash->tfm = fs_info->csum_shash;
3287
3288         crypto_shash_init(shash);
3289         crypto_shash_update(shash, kaddr + pgoff, len);
3290         crypto_shash_final(shash, csum);
3291
3292         if (memcmp(csum, csum_expected, csum_size))
3293                 goto zeroit;
3294
3295         kunmap_atomic(kaddr);
3296         return 0;
3297 zeroit:
3298         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3299                                     io_bio->mirror_num);
3300         memset(kaddr + pgoff, 1, len);
3301         flush_dcache_page(page);
3302         kunmap_atomic(kaddr);
3303         return -EIO;
3304 }
3305
3306 /*
3307  * when reads are done, we need to check csums to verify the data is correct
3308  * if there's a match, we allow the bio to finish.  If not, the code in
3309  * extent_io.c will try to find good copies for us.
3310  */
3311 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3312                                       u64 phy_offset, struct page *page,
3313                                       u64 start, u64 end, int mirror)
3314 {
3315         size_t offset = start - page_offset(page);
3316         struct inode *inode = page->mapping->host;
3317         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3318         struct btrfs_root *root = BTRFS_I(inode)->root;
3319
3320         if (PageChecked(page)) {
3321                 ClearPageChecked(page);
3322                 return 0;
3323         }
3324
3325         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3326                 return 0;
3327
3328         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3329             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3330                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3331                 return 0;
3332         }
3333
3334         phy_offset >>= inode->i_sb->s_blocksize_bits;
3335         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3336                                       start, (size_t)(end - start + 1));
3337 }
3338
3339 /*
3340  * btrfs_add_delayed_iput - perform a delayed iput on @inode
3341  *
3342  * @inode: The inode we want to perform iput on
3343  *
3344  * This function uses the generic vfs_inode::i_count to track whether we should
3345  * just decrement it (in case it's > 1) or if this is the last iput then link
3346  * the inode to the delayed iput machinery. Delayed iputs are processed at
3347  * transaction commit time/superblock commit/cleaner kthread.
3348  */
3349 void btrfs_add_delayed_iput(struct inode *inode)
3350 {
3351         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3352         struct btrfs_inode *binode = BTRFS_I(inode);
3353
3354         if (atomic_add_unless(&inode->i_count, -1, 1))
3355                 return;
3356
3357         atomic_inc(&fs_info->nr_delayed_iputs);
3358         spin_lock(&fs_info->delayed_iput_lock);
3359         ASSERT(list_empty(&binode->delayed_iput));
3360         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3361         spin_unlock(&fs_info->delayed_iput_lock);
3362         if (!test_bit(BTRFS_FS_CLEANER_RUNNING, &fs_info->flags))
3363                 wake_up_process(fs_info->cleaner_kthread);
3364 }
3365
3366 static void run_delayed_iput_locked(struct btrfs_fs_info *fs_info,
3367                                     struct btrfs_inode *inode)
3368 {
3369         list_del_init(&inode->delayed_iput);
3370         spin_unlock(&fs_info->delayed_iput_lock);
3371         iput(&inode->vfs_inode);
3372         if (atomic_dec_and_test(&fs_info->nr_delayed_iputs))
3373                 wake_up(&fs_info->delayed_iputs_wait);
3374         spin_lock(&fs_info->delayed_iput_lock);
3375 }
3376
3377 static void btrfs_run_delayed_iput(struct btrfs_fs_info *fs_info,
3378                                    struct btrfs_inode *inode)
3379 {
3380         if (!list_empty(&inode->delayed_iput)) {
3381                 spin_lock(&fs_info->delayed_iput_lock);
3382                 if (!list_empty(&inode->delayed_iput))
3383                         run_delayed_iput_locked(fs_info, inode);
3384                 spin_unlock(&fs_info->delayed_iput_lock);
3385         }
3386 }
3387
3388 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3389 {
3390
3391         spin_lock(&fs_info->delayed_iput_lock);
3392         while (!list_empty(&fs_info->delayed_iputs)) {
3393                 struct btrfs_inode *inode;
3394
3395                 inode = list_first_entry(&fs_info->delayed_iputs,
3396                                 struct btrfs_inode, delayed_iput);
3397                 run_delayed_iput_locked(fs_info, inode);
3398         }
3399         spin_unlock(&fs_info->delayed_iput_lock);
3400 }
3401
3402 /**
3403  * btrfs_wait_on_delayed_iputs - wait on the delayed iputs to be done running
3404  * @fs_info - the fs_info for this fs
3405  * @return - EINTR if we were killed, 0 if nothing's pending
3406  *
3407  * This will wait on any delayed iputs that are currently running with KILLABLE
3408  * set.  Once they are all done running we will return, unless we are killed in
3409  * which case we return EINTR. This helps in user operations like fallocate etc
3410  * that might get blocked on the iputs.
3411  */
3412 int btrfs_wait_on_delayed_iputs(struct btrfs_fs_info *fs_info)
3413 {
3414         int ret = wait_event_killable(fs_info->delayed_iputs_wait,
3415                         atomic_read(&fs_info->nr_delayed_iputs) == 0);
3416         if (ret)
3417                 return -EINTR;
3418         return 0;
3419 }
3420
3421 /*
3422  * This creates an orphan entry for the given inode in case something goes wrong
3423  * in the middle of an unlink.
3424  */
3425 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3426                      struct btrfs_inode *inode)
3427 {
3428         int ret;
3429
3430         ret = btrfs_insert_orphan_item(trans, inode->root, btrfs_ino(inode));
3431         if (ret && ret != -EEXIST) {
3432                 btrfs_abort_transaction(trans, ret);
3433                 return ret;
3434         }
3435
3436         return 0;
3437 }
3438
3439 /*
3440  * We have done the delete so we can go ahead and remove the orphan item for
3441  * this particular inode.
3442  */
3443 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3444                             struct btrfs_inode *inode)
3445 {
3446         return btrfs_del_orphan_item(trans, inode->root, btrfs_ino(inode));
3447 }
3448
3449 /*
3450  * this cleans up any orphans that may be left on the list from the last use
3451  * of this root.
3452  */
3453 int btrfs_orphan_cleanup(struct btrfs_root *root)
3454 {
3455         struct btrfs_fs_info *fs_info = root->fs_info;
3456         struct btrfs_path *path;
3457         struct extent_buffer *leaf;
3458         struct btrfs_key key, found_key;
3459         struct btrfs_trans_handle *trans;
3460         struct inode *inode;
3461         u64 last_objectid = 0;
3462         int ret = 0, nr_unlink = 0;
3463
3464         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3465                 return 0;
3466
3467         path = btrfs_alloc_path();
3468         if (!path) {
3469                 ret = -ENOMEM;
3470                 goto out;
3471         }
3472         path->reada = READA_BACK;
3473
3474         key.objectid = BTRFS_ORPHAN_OBJECTID;
3475         key.type = BTRFS_ORPHAN_ITEM_KEY;
3476         key.offset = (u64)-1;
3477
3478         while (1) {
3479                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3480                 if (ret < 0)
3481                         goto out;
3482
3483                 /*
3484                  * if ret == 0 means we found what we were searching for, which
3485                  * is weird, but possible, so only screw with path if we didn't
3486                  * find the key and see if we have stuff that matches
3487                  */
3488                 if (ret > 0) {
3489                         ret = 0;
3490                         if (path->slots[0] == 0)
3491                                 break;
3492                         path->slots[0]--;
3493                 }
3494
3495                 /* pull out the item */
3496                 leaf = path->nodes[0];
3497                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3498
3499                 /* make sure the item matches what we want */
3500                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3501                         break;
3502                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3503                         break;
3504
3505                 /* release the path since we're done with it */
3506                 btrfs_release_path(path);
3507
3508                 /*
3509                  * this is where we are basically btrfs_lookup, without the
3510                  * crossing root thing.  we store the inode number in the
3511                  * offset of the orphan item.
3512                  */
3513
3514                 if (found_key.offset == last_objectid) {
3515                         btrfs_err(fs_info,
3516                                   "Error removing orphan entry, stopping orphan cleanup");
3517                         ret = -EINVAL;
3518                         goto out;
3519                 }
3520
3521                 last_objectid = found_key.offset;
3522
3523                 found_key.objectid = found_key.offset;
3524                 found_key.type = BTRFS_INODE_ITEM_KEY;
3525                 found_key.offset = 0;
3526                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3527                 ret = PTR_ERR_OR_ZERO(inode);
3528                 if (ret && ret != -ENOENT)
3529                         goto out;
3530
3531                 if (ret == -ENOENT && root == fs_info->tree_root) {
3532                         struct btrfs_root *dead_root;
3533                         struct btrfs_fs_info *fs_info = root->fs_info;
3534                         int is_dead_root = 0;
3535
3536                         /*
3537                          * this is an orphan in the tree root. Currently these
3538                          * could come from 2 sources:
3539                          *  a) a snapshot deletion in progress
3540                          *  b) a free space cache inode
3541                          * We need to distinguish those two, as the snapshot
3542                          * orphan must not get deleted.
3543                          * find_dead_roots already ran before us, so if this
3544                          * is a snapshot deletion, we should find the root
3545                          * in the dead_roots list
3546                          */
3547                         spin_lock(&fs_info->trans_lock);
3548                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3549                                             root_list) {
3550                                 if (dead_root->root_key.objectid ==
3551                                     found_key.objectid) {
3552                                         is_dead_root = 1;
3553                                         break;
3554                                 }
3555                         }
3556                         spin_unlock(&fs_info->trans_lock);
3557                         if (is_dead_root) {
3558                                 /* prevent this orphan from being found again */
3559                                 key.offset = found_key.objectid - 1;
3560                                 continue;
3561                         }
3562
3563                 }
3564
3565                 /*
3566                  * If we have an inode with links, there are a couple of
3567                  * possibilities. Old kernels (before v3.12) used to create an
3568                  * orphan item for truncate indicating that there were possibly
3569                  * extent items past i_size that needed to be deleted. In v3.12,
3570                  * truncate was changed to update i_size in sync with the extent
3571                  * items, but the (useless) orphan item was still created. Since
3572                  * v4.18, we don't create the orphan item for truncate at all.
3573                  *
3574                  * So, this item could mean that we need to do a truncate, but
3575                  * only if this filesystem was last used on a pre-v3.12 kernel
3576                  * and was not cleanly unmounted. The odds of that are quite
3577                  * slim, and it's a pain to do the truncate now, so just delete
3578                  * the orphan item.
3579                  *
3580                  * It's also possible that this orphan item was supposed to be
3581                  * deleted but wasn't. The inode number may have been reused,
3582                  * but either way, we can delete the orphan item.
3583                  */
3584                 if (ret == -ENOENT || inode->i_nlink) {
3585                         if (!ret)
3586                                 iput(inode);
3587                         trans = btrfs_start_transaction(root, 1);
3588                         if (IS_ERR(trans)) {
3589                                 ret = PTR_ERR(trans);
3590                                 goto out;
3591                         }
3592                         btrfs_debug(fs_info, "auto deleting %Lu",
3593                                     found_key.objectid);
3594                         ret = btrfs_del_orphan_item(trans, root,
3595                                                     found_key.objectid);
3596                         btrfs_end_transaction(trans);
3597                         if (ret)
3598                                 goto out;
3599                         continue;
3600                 }
3601
3602                 nr_unlink++;
3603
3604                 /* this will do delete_inode and everything for us */
3605                 iput(inode);
3606         }
3607         /* release the path since we're done with it */
3608         btrfs_release_path(path);
3609
3610         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3611
3612         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3613                 trans = btrfs_join_transaction(root);
3614                 if (!IS_ERR(trans))
3615                         btrfs_end_transaction(trans);
3616         }
3617
3618         if (nr_unlink)
3619                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3620
3621 out:
3622         if (ret)
3623                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3624         btrfs_free_path(path);
3625         return ret;
3626 }
3627
3628 /*
3629  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3630  * don't find any xattrs, we know there can't be any acls.
3631  *
3632  * slot is the slot the inode is in, objectid is the objectid of the inode
3633  */
3634 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3635                                           int slot, u64 objectid,
3636                                           int *first_xattr_slot)
3637 {
3638         u32 nritems = btrfs_header_nritems(leaf);
3639         struct btrfs_key found_key;
3640         static u64 xattr_access = 0;
3641         static u64 xattr_default = 0;
3642         int scanned = 0;
3643
3644         if (!xattr_access) {
3645                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3646                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3647                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3648                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3649         }
3650
3651         slot++;
3652         *first_xattr_slot = -1;
3653         while (slot < nritems) {
3654                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3655
3656                 /* we found a different objectid, there must not be acls */
3657                 if (found_key.objectid != objectid)
3658                         return 0;
3659
3660                 /* we found an xattr, assume we've got an acl */
3661                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3662                         if (*first_xattr_slot == -1)
3663                                 *first_xattr_slot = slot;
3664                         if (found_key.offset == xattr_access ||
3665                             found_key.offset == xattr_default)
3666                                 return 1;
3667                 }
3668
3669                 /*
3670                  * we found a key greater than an xattr key, there can't
3671                  * be any acls later on
3672                  */
3673                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3674                         return 0;
3675
3676                 slot++;
3677                 scanned++;
3678
3679                 /*
3680                  * it goes inode, inode backrefs, xattrs, extents,
3681                  * so if there are a ton of hard links to an inode there can
3682                  * be a lot of backrefs.  Don't waste time searching too hard,
3683                  * this is just an optimization
3684                  */
3685                 if (scanned >= 8)
3686                         break;
3687         }
3688         /* we hit the end of the leaf before we found an xattr or
3689          * something larger than an xattr.  We have to assume the inode
3690          * has acls
3691          */
3692         if (*first_xattr_slot == -1)
3693                 *first_xattr_slot = slot;
3694         return 1;
3695 }
3696
3697 /*
3698  * read an inode from the btree into the in-memory inode
3699  */
3700 static int btrfs_read_locked_inode(struct inode *inode,
3701                                    struct btrfs_path *in_path)
3702 {
3703         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3704         struct btrfs_path *path = in_path;
3705         struct extent_buffer *leaf;
3706         struct btrfs_inode_item *inode_item;
3707         struct btrfs_root *root = BTRFS_I(inode)->root;
3708         struct btrfs_key location;
3709         unsigned long ptr;
3710         int maybe_acls;
3711         u32 rdev;
3712         int ret;
3713         bool filled = false;
3714         int first_xattr_slot;
3715
3716         ret = btrfs_fill_inode(inode, &rdev);
3717         if (!ret)
3718                 filled = true;
3719
3720         if (!path) {
3721                 path = btrfs_alloc_path();
3722                 if (!path)
3723                         return -ENOMEM;
3724         }
3725
3726         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3727
3728         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3729         if (ret) {
3730                 if (path != in_path)
3731                         btrfs_free_path(path);
3732                 return ret;
3733         }
3734
3735         leaf = path->nodes[0];
3736
3737         if (filled)
3738                 goto cache_index;
3739
3740         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3741                                     struct btrfs_inode_item);
3742         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3743         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3744         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3745         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3746         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3747
3748         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3749         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3750
3751         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3752         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3753
3754         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3755         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3756
3757         BTRFS_I(inode)->i_otime.tv_sec =
3758                 btrfs_timespec_sec(leaf, &inode_item->otime);
3759         BTRFS_I(inode)->i_otime.tv_nsec =
3760                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3761
3762         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3763         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3764         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3765
3766         inode_set_iversion_queried(inode,
3767                                    btrfs_inode_sequence(leaf, inode_item));
3768         inode->i_generation = BTRFS_I(inode)->generation;
3769         inode->i_rdev = 0;
3770         rdev = btrfs_inode_rdev(leaf, inode_item);
3771
3772         BTRFS_I(inode)->index_cnt = (u64)-1;
3773         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3774
3775 cache_index:
3776         /*
3777          * If we were modified in the current generation and evicted from memory
3778          * and then re-read we need to do a full sync since we don't have any
3779          * idea about which extents were modified before we were evicted from
3780          * cache.
3781          *
3782          * This is required for both inode re-read from disk and delayed inode
3783          * in delayed_nodes_tree.
3784          */
3785         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3786                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3787                         &BTRFS_I(inode)->runtime_flags);
3788
3789         /*
3790          * We don't persist the id of the transaction where an unlink operation
3791          * against the inode was last made. So here we assume the inode might
3792          * have been evicted, and therefore the exact value of last_unlink_trans
3793          * lost, and set it to last_trans to avoid metadata inconsistencies
3794          * between the inode and its parent if the inode is fsync'ed and the log
3795          * replayed. For example, in the scenario:
3796          *
3797          * touch mydir/foo
3798          * ln mydir/foo mydir/bar
3799          * sync
3800          * unlink mydir/bar
3801          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3802          * xfs_io -c fsync mydir/foo
3803          * <power failure>
3804          * mount fs, triggers fsync log replay
3805          *
3806          * We must make sure that when we fsync our inode foo we also log its
3807          * parent inode, otherwise after log replay the parent still has the
3808          * dentry with the "bar" name but our inode foo has a link count of 1
3809          * and doesn't have an inode ref with the name "bar" anymore.
3810          *
3811          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3812          * but it guarantees correctness at the expense of occasional full
3813          * transaction commits on fsync if our inode is a directory, or if our
3814          * inode is not a directory, logging its parent unnecessarily.
3815          */
3816         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3817
3818         path->slots[0]++;
3819         if (inode->i_nlink != 1 ||
3820             path->slots[0] >= btrfs_header_nritems(leaf))
3821                 goto cache_acl;
3822
3823         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3824         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3825                 goto cache_acl;
3826
3827         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3828         if (location.type == BTRFS_INODE_REF_KEY) {
3829                 struct btrfs_inode_ref *ref;
3830
3831                 ref = (struct btrfs_inode_ref *)ptr;
3832                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3833         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3834                 struct btrfs_inode_extref *extref;
3835
3836                 extref = (struct btrfs_inode_extref *)ptr;
3837                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3838                                                                      extref);
3839         }
3840 cache_acl:
3841         /*
3842          * try to precache a NULL acl entry for files that don't have
3843          * any xattrs or acls
3844          */
3845         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3846                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3847         if (first_xattr_slot != -1) {
3848                 path->slots[0] = first_xattr_slot;
3849                 ret = btrfs_load_inode_props(inode, path);
3850                 if (ret)
3851                         btrfs_err(fs_info,
3852                                   "error loading props for ino %llu (root %llu): %d",
3853                                   btrfs_ino(BTRFS_I(inode)),
3854                                   root->root_key.objectid, ret);
3855         }
3856         if (path != in_path)
3857                 btrfs_free_path(path);
3858
3859         if (!maybe_acls)
3860                 cache_no_acl(inode);
3861
3862         switch (inode->i_mode & S_IFMT) {
3863         case S_IFREG:
3864                 inode->i_mapping->a_ops = &btrfs_aops;
3865                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3866                 inode->i_fop = &btrfs_file_operations;
3867                 inode->i_op = &btrfs_file_inode_operations;
3868                 break;
3869         case S_IFDIR:
3870                 inode->i_fop = &btrfs_dir_file_operations;
3871                 inode->i_op = &btrfs_dir_inode_operations;
3872                 break;
3873         case S_IFLNK:
3874                 inode->i_op = &btrfs_symlink_inode_operations;
3875                 inode_nohighmem(inode);
3876                 inode->i_mapping->a_ops = &btrfs_aops;
3877                 break;
3878         default:
3879                 inode->i_op = &btrfs_special_inode_operations;
3880                 init_special_inode(inode, inode->i_mode, rdev);
3881                 break;
3882         }
3883
3884         btrfs_sync_inode_flags_to_i_flags(inode);
3885         return 0;
3886 }
3887
3888 /*
3889  * given a leaf and an inode, copy the inode fields into the leaf
3890  */
3891 static void fill_inode_item(struct btrfs_trans_handle *trans,
3892                             struct extent_buffer *leaf,
3893                             struct btrfs_inode_item *item,
3894                             struct inode *inode)
3895 {
3896         struct btrfs_map_token token;
3897
3898         btrfs_init_map_token(&token, leaf);
3899
3900         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3901         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3902         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3903                                    &token);
3904         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3905         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3906
3907         btrfs_set_token_timespec_sec(leaf, &item->atime,
3908                                      inode->i_atime.tv_sec, &token);
3909         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3910                                       inode->i_atime.tv_nsec, &token);
3911
3912         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3913                                      inode->i_mtime.tv_sec, &token);
3914         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3915                                       inode->i_mtime.tv_nsec, &token);
3916
3917         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3918                                      inode->i_ctime.tv_sec, &token);
3919         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3920                                       inode->i_ctime.tv_nsec, &token);
3921
3922         btrfs_set_token_timespec_sec(leaf, &item->otime,
3923                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3924         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3925                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3926
3927         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3928                                      &token);
3929         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3930                                          &token);
3931         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
3932                                        &token);
3933         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3934         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3935         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3936         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3937 }
3938
3939 /*
3940  * copy everything in the in-memory inode into the btree.
3941  */
3942 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3943                                 struct btrfs_root *root, struct inode *inode)
3944 {
3945         struct btrfs_inode_item *inode_item;
3946         struct btrfs_path *path;
3947         struct extent_buffer *leaf;
3948         int ret;
3949
3950         path = btrfs_alloc_path();
3951         if (!path)
3952                 return -ENOMEM;
3953
3954         path->leave_spinning = 1;
3955         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3956                                  1);
3957         if (ret) {
3958                 if (ret > 0)
3959                         ret = -ENOENT;
3960                 goto failed;
3961         }
3962
3963         leaf = path->nodes[0];
3964         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3965                                     struct btrfs_inode_item);
3966
3967         fill_inode_item(trans, leaf, inode_item, inode);
3968         btrfs_mark_buffer_dirty(leaf);
3969         btrfs_set_inode_last_trans(trans, inode);
3970         ret = 0;
3971 failed:
3972         btrfs_free_path(path);
3973         return ret;
3974 }
3975
3976 /*
3977  * copy everything in the in-memory inode into the btree.
3978  */
3979 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3980                                 struct btrfs_root *root, struct inode *inode)
3981 {
3982         struct btrfs_fs_info *fs_info = root->fs_info;
3983         int ret;
3984
3985         /*
3986          * If the inode is a free space inode, we can deadlock during commit
3987          * if we put it into the delayed code.
3988          *
3989          * The data relocation inode should also be directly updated
3990          * without delay
3991          */
3992         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
3993             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3994             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
3995                 btrfs_update_root_times(trans, root);
3996
3997                 ret = btrfs_delayed_update_inode(trans, root, inode);
3998                 if (!ret)
3999                         btrfs_set_inode_last_trans(trans, inode);
4000                 return ret;
4001         }
4002
4003         return btrfs_update_inode_item(trans, root, inode);
4004 }
4005
4006 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4007                                          struct btrfs_root *root,
4008                                          struct inode *inode)
4009 {
4010         int ret;
4011
4012         ret = btrfs_update_inode(trans, root, inode);
4013         if (ret == -ENOSPC)
4014                 return btrfs_update_inode_item(trans, root, inode);
4015         return ret;
4016 }
4017
4018 /*
4019  * unlink helper that gets used here in inode.c and in the tree logging
4020  * recovery code.  It remove a link in a directory with a given name, and
4021  * also drops the back refs in the inode to the directory
4022  */
4023 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4024                                 struct btrfs_root *root,
4025                                 struct btrfs_inode *dir,
4026                                 struct btrfs_inode *inode,
4027                                 const char *name, int name_len)
4028 {
4029         struct btrfs_fs_info *fs_info = root->fs_info;
4030         struct btrfs_path *path;
4031         int ret = 0;
4032         struct btrfs_dir_item *di;
4033         u64 index;
4034         u64 ino = btrfs_ino(inode);
4035         u64 dir_ino = btrfs_ino(dir);
4036
4037         path = btrfs_alloc_path();
4038         if (!path) {
4039                 ret = -ENOMEM;
4040                 goto out;
4041         }
4042
4043         path->leave_spinning = 1;
4044         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4045                                     name, name_len, -1);
4046         if (IS_ERR_OR_NULL(di)) {
4047                 ret = di ? PTR_ERR(di) : -ENOENT;
4048                 goto err;
4049         }
4050         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4051         if (ret)
4052                 goto err;
4053         btrfs_release_path(path);
4054
4055         /*
4056          * If we don't have dir index, we have to get it by looking up
4057          * the inode ref, since we get the inode ref, remove it directly,
4058          * it is unnecessary to do delayed deletion.
4059          *
4060          * But if we have dir index, needn't search inode ref to get it.
4061          * Since the inode ref is close to the inode item, it is better
4062          * that we delay to delete it, and just do this deletion when
4063          * we update the inode item.
4064          */
4065         if (inode->dir_index) {
4066                 ret = btrfs_delayed_delete_inode_ref(inode);
4067                 if (!ret) {
4068                         index = inode->dir_index;
4069                         goto skip_backref;
4070                 }
4071         }
4072
4073         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4074                                   dir_ino, &index);
4075         if (ret) {
4076                 btrfs_info(fs_info,
4077                         "failed to delete reference to %.*s, inode %llu parent %llu",
4078                         name_len, name, ino, dir_ino);
4079                 btrfs_abort_transaction(trans, ret);
4080                 goto err;
4081         }
4082 skip_backref:
4083         ret = btrfs_delete_delayed_dir_index(trans, dir, index);
4084         if (ret) {
4085                 btrfs_abort_transaction(trans, ret);
4086                 goto err;
4087         }
4088
4089         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4090                         dir_ino);
4091         if (ret != 0 && ret != -ENOENT) {
4092                 btrfs_abort_transaction(trans, ret);
4093                 goto err;
4094         }
4095
4096         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4097                         index);
4098         if (ret == -ENOENT)
4099                 ret = 0;
4100         else if (ret)
4101                 btrfs_abort_transaction(trans, ret);
4102
4103         /*
4104          * If we have a pending delayed iput we could end up with the final iput
4105          * being run in btrfs-cleaner context.  If we have enough of these built
4106          * up we can end up burning a lot of time in btrfs-cleaner without any
4107          * way to throttle the unlinks.  Since we're currently holding a ref on
4108          * the inode we can run the delayed iput here without any issues as the
4109          * final iput won't be done until after we drop the ref we're currently
4110          * holding.
4111          */
4112         btrfs_run_delayed_iput(fs_info, inode);
4113 err:
4114         btrfs_free_path(path);
4115         if (ret)
4116                 goto out;
4117
4118         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4119         inode_inc_iversion(&inode->vfs_inode);
4120         inode_inc_iversion(&dir->vfs_inode);
4121         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4122                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4123         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4124 out:
4125         return ret;
4126 }
4127
4128 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4129                        struct btrfs_root *root,
4130                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4131                        const char *name, int name_len)
4132 {
4133         int ret;
4134         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4135         if (!ret) {
4136                 drop_nlink(&inode->vfs_inode);
4137                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4138         }
4139         return ret;
4140 }
4141
4142 /*
4143  * helper to start transaction for unlink and rmdir.
4144  *
4145  * unlink and rmdir are special in btrfs, they do not always free space, so
4146  * if we cannot make our reservations the normal way try and see if there is
4147  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4148  * allow the unlink to occur.
4149  */
4150 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4151 {
4152         struct btrfs_root *root = BTRFS_I(dir)->root;
4153
4154         /*
4155          * 1 for the possible orphan item
4156          * 1 for the dir item
4157          * 1 for the dir index
4158          * 1 for the inode ref
4159          * 1 for the inode
4160          */
4161         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4162 }
4163
4164 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4165 {
4166         struct btrfs_root *root = BTRFS_I(dir)->root;
4167         struct btrfs_trans_handle *trans;
4168         struct inode *inode = d_inode(dentry);
4169         int ret;
4170
4171         trans = __unlink_start_trans(dir);
4172         if (IS_ERR(trans))
4173                 return PTR_ERR(trans);
4174
4175         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4176                         0);
4177
4178         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4179                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4180                         dentry->d_name.len);
4181         if (ret)
4182                 goto out;
4183
4184         if (inode->i_nlink == 0) {
4185                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4186                 if (ret)
4187                         goto out;
4188         }
4189
4190 out:
4191         btrfs_end_transaction(trans);
4192         btrfs_btree_balance_dirty(root->fs_info);
4193         return ret;
4194 }
4195
4196 static int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4197                                struct inode *dir, u64 objectid,
4198                                const char *name, int name_len)
4199 {
4200         struct btrfs_root *root = BTRFS_I(dir)->root;
4201         struct btrfs_path *path;
4202         struct extent_buffer *leaf;
4203         struct btrfs_dir_item *di;
4204         struct btrfs_key key;
4205         u64 index;
4206         int ret;
4207         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4208
4209         path = btrfs_alloc_path();
4210         if (!path)
4211                 return -ENOMEM;
4212
4213         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4214                                    name, name_len, -1);
4215         if (IS_ERR_OR_NULL(di)) {
4216                 ret = di ? PTR_ERR(di) : -ENOENT;
4217                 goto out;
4218         }
4219
4220         leaf = path->nodes[0];
4221         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4222         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4223         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4224         if (ret) {
4225                 btrfs_abort_transaction(trans, ret);
4226                 goto out;
4227         }
4228         btrfs_release_path(path);
4229
4230         ret = btrfs_del_root_ref(trans, objectid, root->root_key.objectid,
4231                                  dir_ino, &index, name, name_len);
4232         if (ret < 0) {
4233                 if (ret != -ENOENT) {
4234                         btrfs_abort_transaction(trans, ret);
4235                         goto out;
4236                 }
4237                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4238                                                  name, name_len);
4239                 if (IS_ERR_OR_NULL(di)) {
4240                         if (!di)
4241                                 ret = -ENOENT;
4242                         else
4243                                 ret = PTR_ERR(di);
4244                         btrfs_abort_transaction(trans, ret);
4245                         goto out;
4246                 }
4247
4248                 leaf = path->nodes[0];
4249                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4250                 index = key.offset;
4251         }
4252         btrfs_release_path(path);
4253
4254         ret = btrfs_delete_delayed_dir_index(trans, BTRFS_I(dir), index);
4255         if (ret) {
4256                 btrfs_abort_transaction(trans, ret);
4257                 goto out;
4258         }
4259
4260         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4261         inode_inc_iversion(dir);
4262         dir->i_mtime = dir->i_ctime = current_time(dir);
4263         ret = btrfs_update_inode_fallback(trans, root, dir);
4264         if (ret)
4265                 btrfs_abort_transaction(trans, ret);
4266 out:
4267         btrfs_free_path(path);
4268         return ret;
4269 }
4270
4271 /*
4272  * Helper to check if the subvolume references other subvolumes or if it's
4273  * default.
4274  */
4275 static noinline int may_destroy_subvol(struct btrfs_root *root)
4276 {
4277         struct btrfs_fs_info *fs_info = root->fs_info;
4278         struct btrfs_path *path;
4279         struct btrfs_dir_item *di;
4280         struct btrfs_key key;
4281         u64 dir_id;
4282         int ret;
4283
4284         path = btrfs_alloc_path();
4285         if (!path)
4286                 return -ENOMEM;
4287
4288         /* Make sure this root isn't set as the default subvol */
4289         dir_id = btrfs_super_root_dir(fs_info->super_copy);
4290         di = btrfs_lookup_dir_item(NULL, fs_info->tree_root, path,
4291                                    dir_id, "default", 7, 0);
4292         if (di && !IS_ERR(di)) {
4293                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
4294                 if (key.objectid == root->root_key.objectid) {
4295                         ret = -EPERM;
4296                         btrfs_err(fs_info,
4297                                   "deleting default subvolume %llu is not allowed",
4298                                   key.objectid);
4299                         goto out;
4300                 }
4301                 btrfs_release_path(path);
4302         }
4303
4304         key.objectid = root->root_key.objectid;
4305         key.type = BTRFS_ROOT_REF_KEY;
4306         key.offset = (u64)-1;
4307
4308         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
4309         if (ret < 0)
4310                 goto out;
4311         BUG_ON(ret == 0);
4312
4313         ret = 0;
4314         if (path->slots[0] > 0) {
4315                 path->slots[0]--;
4316                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
4317                 if (key.objectid == root->root_key.objectid &&
4318                     key.type == BTRFS_ROOT_REF_KEY)
4319                         ret = -ENOTEMPTY;
4320         }
4321 out:
4322         btrfs_free_path(path);
4323         return ret;
4324 }
4325
4326 /* Delete all dentries for inodes belonging to the root */
4327 static void btrfs_prune_dentries(struct btrfs_root *root)
4328 {
4329         struct btrfs_fs_info *fs_info = root->fs_info;
4330         struct rb_node *node;
4331         struct rb_node *prev;
4332         struct btrfs_inode *entry;
4333         struct inode *inode;
4334         u64 objectid = 0;
4335
4336         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4337                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
4338
4339         spin_lock(&root->inode_lock);
4340 again:
4341         node = root->inode_tree.rb_node;
4342         prev = NULL;
4343         while (node) {
4344                 prev = node;
4345                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4346
4347                 if (objectid < btrfs_ino(entry))
4348                         node = node->rb_left;
4349                 else if (objectid > btrfs_ino(entry))
4350                         node = node->rb_right;
4351                 else
4352                         break;
4353         }
4354         if (!node) {
4355                 while (prev) {
4356                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
4357                         if (objectid <= btrfs_ino(entry)) {
4358                                 node = prev;
4359                                 break;
4360                         }
4361                         prev = rb_next(prev);
4362                 }
4363         }
4364         while (node) {
4365                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4366                 objectid = btrfs_ino(entry) + 1;
4367                 inode = igrab(&entry->vfs_inode);
4368                 if (inode) {
4369                         spin_unlock(&root->inode_lock);
4370                         if (atomic_read(&inode->i_count) > 1)
4371                                 d_prune_aliases(inode);
4372                         /*
4373                          * btrfs_drop_inode will have it removed from the inode
4374                          * cache when its usage count hits zero.
4375                          */
4376                         iput(inode);
4377                         cond_resched();
4378                         spin_lock(&root->inode_lock);
4379                         goto again;
4380                 }
4381
4382                 if (cond_resched_lock(&root->inode_lock))
4383                         goto again;
4384
4385                 node = rb_next(node);
4386         }
4387         spin_unlock(&root->inode_lock);
4388 }
4389
4390 int btrfs_delete_subvolume(struct inode *dir, struct dentry *dentry)
4391 {
4392         struct btrfs_fs_info *fs_info = btrfs_sb(dentry->d_sb);
4393         struct btrfs_root *root = BTRFS_I(dir)->root;
4394         struct inode *inode = d_inode(dentry);
4395         struct btrfs_root *dest = BTRFS_I(inode)->root;
4396         struct btrfs_trans_handle *trans;
4397         struct btrfs_block_rsv block_rsv;
4398         u64 root_flags;
4399         int ret;
4400         int err;
4401
4402         /*
4403          * Don't allow to delete a subvolume with send in progress. This is
4404          * inside the inode lock so the error handling that has to drop the bit
4405          * again is not run concurrently.
4406          */
4407         spin_lock(&dest->root_item_lock);
4408         if (dest->send_in_progress) {
4409                 spin_unlock(&dest->root_item_lock);
4410                 btrfs_warn(fs_info,
4411                            "attempt to delete subvolume %llu during send",
4412                            dest->root_key.objectid);
4413                 return -EPERM;
4414         }
4415         root_flags = btrfs_root_flags(&dest->root_item);
4416         btrfs_set_root_flags(&dest->root_item,
4417                              root_flags | BTRFS_ROOT_SUBVOL_DEAD);
4418         spin_unlock(&dest->root_item_lock);
4419
4420         down_write(&fs_info->subvol_sem);
4421
4422         err = may_destroy_subvol(dest);
4423         if (err)
4424                 goto out_up_write;
4425
4426         btrfs_init_block_rsv(&block_rsv, BTRFS_BLOCK_RSV_TEMP);
4427         /*
4428          * One for dir inode,
4429          * two for dir entries,
4430          * two for root ref/backref.
4431          */
4432         err = btrfs_subvolume_reserve_metadata(root, &block_rsv, 5, true);
4433         if (err)
4434                 goto out_up_write;
4435
4436         trans = btrfs_start_transaction(root, 0);
4437         if (IS_ERR(trans)) {
4438                 err = PTR_ERR(trans);
4439                 goto out_release;
4440         }
4441         trans->block_rsv = &block_rsv;
4442         trans->bytes_reserved = block_rsv.size;
4443
4444         btrfs_record_snapshot_destroy(trans, BTRFS_I(dir));
4445
4446         ret = btrfs_unlink_subvol(trans, dir, dest->root_key.objectid,
4447                                   dentry->d_name.name, dentry->d_name.len);
4448         if (ret) {
4449                 err = ret;
4450                 btrfs_abort_transaction(trans, ret);
4451                 goto out_end_trans;
4452         }
4453
4454         btrfs_record_root_in_trans(trans, dest);
4455
4456         memset(&dest->root_item.drop_progress, 0,
4457                 sizeof(dest->root_item.drop_progress));
4458         dest->root_item.drop_level = 0;
4459         btrfs_set_root_refs(&dest->root_item, 0);
4460
4461         if (!test_and_set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &dest->state)) {
4462                 ret = btrfs_insert_orphan_item(trans,
4463                                         fs_info->tree_root,
4464                                         dest->root_key.objectid);
4465                 if (ret) {
4466                         btrfs_abort_transaction(trans, ret);
4467                         err = ret;
4468                         goto out_end_trans;
4469                 }
4470         }
4471
4472         ret = btrfs_uuid_tree_remove(trans, dest->root_item.uuid,
4473                                   BTRFS_UUID_KEY_SUBVOL,
4474                                   dest->root_key.objectid);
4475         if (ret && ret != -ENOENT) {
4476                 btrfs_abort_transaction(trans, ret);
4477                 err = ret;
4478                 goto out_end_trans;
4479         }
4480         if (!btrfs_is_empty_uuid(dest->root_item.received_uuid)) {
4481                 ret = btrfs_uuid_tree_remove(trans,
4482                                           dest->root_item.received_uuid,
4483                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
4484                                           dest->root_key.objectid);
4485                 if (ret && ret != -ENOENT) {
4486                         btrfs_abort_transaction(trans, ret);
4487                         err = ret;
4488                         goto out_end_trans;
4489                 }
4490         }
4491
4492 out_end_trans:
4493         trans->block_rsv = NULL;
4494         trans->bytes_reserved = 0;
4495         ret = btrfs_end_transaction(trans);
4496         if (ret && !err)
4497                 err = ret;
4498         inode->i_flags |= S_DEAD;
4499 out_release:
4500         btrfs_subvolume_release_metadata(fs_info, &block_rsv);
4501 out_up_write:
4502         up_write(&fs_info->subvol_sem);
4503         if (err) {
4504                 spin_lock(&dest->root_item_lock);
4505                 root_flags = btrfs_root_flags(&dest->root_item);
4506                 btrfs_set_root_flags(&dest->root_item,
4507                                 root_flags & ~BTRFS_ROOT_SUBVOL_DEAD);
4508                 spin_unlock(&dest->root_item_lock);
4509         } else {
4510                 d_invalidate(dentry);
4511                 btrfs_prune_dentries(dest);
4512                 ASSERT(dest->send_in_progress == 0);
4513
4514                 /* the last ref */
4515                 if (dest->ino_cache_inode) {
4516                         iput(dest->ino_cache_inode);
4517                         dest->ino_cache_inode = NULL;
4518                 }
4519         }
4520
4521         return err;
4522 }
4523
4524 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4525 {
4526         struct inode *inode = d_inode(dentry);
4527         int err = 0;
4528         struct btrfs_root *root = BTRFS_I(dir)->root;
4529         struct btrfs_trans_handle *trans;
4530         u64 last_unlink_trans;
4531
4532         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4533                 return -ENOTEMPTY;
4534         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4535                 return btrfs_delete_subvolume(dir, dentry);
4536
4537         trans = __unlink_start_trans(dir);
4538         if (IS_ERR(trans))
4539                 return PTR_ERR(trans);
4540
4541         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4542                 err = btrfs_unlink_subvol(trans, dir,
4543                                           BTRFS_I(inode)->location.objectid,
4544                                           dentry->d_name.name,
4545                                           dentry->d_name.len);
4546                 goto out;
4547         }
4548
4549         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4550         if (err)
4551                 goto out;
4552
4553         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4554
4555         /* now the directory is empty */
4556         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4557                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4558                         dentry->d_name.len);
4559         if (!err) {
4560                 btrfs_i_size_write(BTRFS_I(inode), 0);
4561                 /*
4562                  * Propagate the last_unlink_trans value of the deleted dir to
4563                  * its parent directory. This is to prevent an unrecoverable
4564                  * log tree in the case we do something like this:
4565                  * 1) create dir foo
4566                  * 2) create snapshot under dir foo
4567                  * 3) delete the snapshot
4568                  * 4) rmdir foo
4569                  * 5) mkdir foo
4570                  * 6) fsync foo or some file inside foo
4571                  */
4572                 if (last_unlink_trans >= trans->transid)
4573                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4574         }
4575 out:
4576         btrfs_end_transaction(trans);
4577         btrfs_btree_balance_dirty(root->fs_info);
4578
4579         return err;
4580 }
4581
4582 /*
4583  * Return this if we need to call truncate_block for the last bit of the
4584  * truncate.
4585  */
4586 #define NEED_TRUNCATE_BLOCK 1
4587
4588 /*
4589  * this can truncate away extent items, csum items and directory items.
4590  * It starts at a high offset and removes keys until it can't find
4591  * any higher than new_size
4592  *
4593  * csum items that cross the new i_size are truncated to the new size
4594  * as well.
4595  *
4596  * min_type is the minimum key type to truncate down to.  If set to 0, this
4597  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4598  */
4599 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4600                                struct btrfs_root *root,
4601                                struct inode *inode,
4602                                u64 new_size, u32 min_type)
4603 {
4604         struct btrfs_fs_info *fs_info = root->fs_info;
4605         struct btrfs_path *path;
4606         struct extent_buffer *leaf;
4607         struct btrfs_file_extent_item *fi;
4608         struct btrfs_key key;
4609         struct btrfs_key found_key;
4610         u64 extent_start = 0;
4611         u64 extent_num_bytes = 0;
4612         u64 extent_offset = 0;
4613         u64 item_end = 0;
4614         u64 last_size = new_size;
4615         u32 found_type = (u8)-1;
4616         int found_extent;
4617         int del_item;
4618         int pending_del_nr = 0;
4619         int pending_del_slot = 0;
4620         int extent_type = -1;
4621         int ret;
4622         u64 ino = btrfs_ino(BTRFS_I(inode));
4623         u64 bytes_deleted = 0;
4624         bool be_nice = false;
4625         bool should_throttle = false;
4626
4627         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4628
4629         /*
4630          * for non-free space inodes and ref cows, we want to back off from
4631          * time to time
4632          */
4633         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4634             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4635                 be_nice = true;
4636
4637         path = btrfs_alloc_path();
4638         if (!path)
4639                 return -ENOMEM;
4640         path->reada = READA_BACK;
4641
4642         /*
4643          * We want to drop from the next block forward in case this new size is
4644          * not block aligned since we will be keeping the last block of the
4645          * extent just the way it is.
4646          */
4647         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4648             root == fs_info->tree_root)
4649                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4650                                         fs_info->sectorsize),
4651                                         (u64)-1, 0);
4652
4653         /*
4654          * This function is also used to drop the items in the log tree before
4655          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4656          * it is used to drop the logged items. So we shouldn't kill the delayed
4657          * items.
4658          */
4659         if (min_type == 0 && root == BTRFS_I(inode)->root)
4660                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4661
4662         key.objectid = ino;
4663         key.offset = (u64)-1;
4664         key.type = (u8)-1;
4665
4666 search_again:
4667         /*
4668          * with a 16K leaf size and 128MB extents, you can actually queue
4669          * up a huge file in a single leaf.  Most of the time that
4670          * bytes_deleted is > 0, it will be huge by the time we get here
4671          */
4672         if (be_nice && bytes_deleted > SZ_32M &&
4673             btrfs_should_end_transaction(trans)) {
4674                 ret = -EAGAIN;
4675                 goto out;
4676         }
4677
4678         path->leave_spinning = 1;
4679         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4680         if (ret < 0)
4681                 goto out;
4682
4683         if (ret > 0) {
4684                 ret = 0;
4685                 /* there are no items in the tree for us to truncate, we're
4686                  * done
4687                  */
4688                 if (path->slots[0] == 0)
4689                         goto out;
4690                 path->slots[0]--;
4691         }
4692
4693         while (1) {
4694                 fi = NULL;
4695                 leaf = path->nodes[0];
4696                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4697                 found_type = found_key.type;
4698
4699                 if (found_key.objectid != ino)
4700                         break;
4701
4702                 if (found_type < min_type)
4703                         break;
4704
4705                 item_end = found_key.offset;
4706                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4707                         fi = btrfs_item_ptr(leaf, path->slots[0],
4708                                             struct btrfs_file_extent_item);
4709                         extent_type = btrfs_file_extent_type(leaf, fi);
4710                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4711                                 item_end +=
4712                                     btrfs_file_extent_num_bytes(leaf, fi);
4713
4714                                 trace_btrfs_truncate_show_fi_regular(
4715                                         BTRFS_I(inode), leaf, fi,
4716                                         found_key.offset);
4717                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4718                                 item_end += btrfs_file_extent_ram_bytes(leaf,
4719                                                                         fi);
4720
4721                                 trace_btrfs_truncate_show_fi_inline(
4722                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4723                                         found_key.offset);
4724                         }
4725                         item_end--;
4726                 }
4727                 if (found_type > min_type) {
4728                         del_item = 1;
4729                 } else {
4730                         if (item_end < new_size)
4731                                 break;
4732                         if (found_key.offset >= new_size)
4733                                 del_item = 1;
4734                         else
4735                                 del_item = 0;
4736                 }
4737                 found_extent = 0;
4738                 /* FIXME, shrink the extent if the ref count is only 1 */
4739                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4740                         goto delete;
4741
4742                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4743                         u64 num_dec;
4744                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4745                         if (!del_item) {
4746                                 u64 orig_num_bytes =
4747                                         btrfs_file_extent_num_bytes(leaf, fi);
4748                                 extent_num_bytes = ALIGN(new_size -
4749                                                 found_key.offset,
4750                                                 fs_info->sectorsize);
4751                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4752                                                          extent_num_bytes);
4753                                 num_dec = (orig_num_bytes -
4754                                            extent_num_bytes);
4755                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4756                                              &root->state) &&
4757                                     extent_start != 0)
4758                                         inode_sub_bytes(inode, num_dec);
4759                                 btrfs_mark_buffer_dirty(leaf);
4760                         } else {
4761                                 extent_num_bytes =
4762                                         btrfs_file_extent_disk_num_bytes(leaf,
4763                                                                          fi);
4764                                 extent_offset = found_key.offset -
4765                                         btrfs_file_extent_offset(leaf, fi);
4766
4767                                 /* FIXME blocksize != 4096 */
4768                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4769                                 if (extent_start != 0) {
4770                                         found_extent = 1;
4771                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4772                                                      &root->state))
4773                                                 inode_sub_bytes(inode, num_dec);
4774                                 }
4775                         }
4776                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4777                         /*
4778                          * we can't truncate inline items that have had
4779                          * special encodings
4780                          */
4781                         if (!del_item &&
4782                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4783                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4784                             btrfs_file_extent_compression(leaf, fi) == 0) {
4785                                 u32 size = (u32)(new_size - found_key.offset);
4786
4787                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4788                                 size = btrfs_file_extent_calc_inline_size(size);
4789                                 btrfs_truncate_item(path, size, 1);
4790                         } else if (!del_item) {
4791                                 /*
4792                                  * We have to bail so the last_size is set to
4793                                  * just before this extent.
4794                                  */
4795                                 ret = NEED_TRUNCATE_BLOCK;
4796                                 break;
4797                         }
4798
4799                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4800                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4801                 }
4802 delete:
4803                 if (del_item)
4804                         last_size = found_key.offset;
4805                 else
4806                         last_size = new_size;
4807                 if (del_item) {
4808                         if (!pending_del_nr) {
4809                                 /* no pending yet, add ourselves */
4810                                 pending_del_slot = path->slots[0];
4811                                 pending_del_nr = 1;
4812                         } else if (pending_del_nr &&
4813                                    path->slots[0] + 1 == pending_del_slot) {
4814                                 /* hop on the pending chunk */
4815                                 pending_del_nr++;
4816                                 pending_del_slot = path->slots[0];
4817                         } else {
4818                                 BUG();
4819                         }
4820                 } else {
4821                         break;
4822                 }
4823                 should_throttle = false;
4824
4825                 if (found_extent &&
4826                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4827                      root == fs_info->tree_root)) {
4828                         struct btrfs_ref ref = { 0 };
4829
4830                         btrfs_set_path_blocking(path);
4831                         bytes_deleted += extent_num_bytes;
4832
4833                         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF,
4834                                         extent_start, extent_num_bytes, 0);
4835                         ref.real_root = root->root_key.objectid;
4836                         btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
4837                                         ino, extent_offset);
4838                         ret = btrfs_free_extent(trans, &ref);
4839                         if (ret) {
4840                                 btrfs_abort_transaction(trans, ret);
4841                                 break;
4842                         }
4843                         if (be_nice) {
4844                                 if (btrfs_should_throttle_delayed_refs(trans))
4845                                         should_throttle = true;
4846                         }
4847                 }
4848
4849                 if (found_type == BTRFS_INODE_ITEM_KEY)
4850                         break;
4851
4852                 if (path->slots[0] == 0 ||
4853                     path->slots[0] != pending_del_slot ||
4854                     should_throttle) {
4855                         if (pending_del_nr) {
4856                                 ret = btrfs_del_items(trans, root, path,
4857                                                 pending_del_slot,
4858                                                 pending_del_nr);
4859                                 if (ret) {
4860                                         btrfs_abort_transaction(trans, ret);
4861                                         break;
4862                                 }
4863                                 pending_del_nr = 0;
4864                         }
4865                         btrfs_release_path(path);
4866
4867                         /*
4868                          * We can generate a lot of delayed refs, so we need to
4869                          * throttle every once and a while and make sure we're
4870                          * adding enough space to keep up with the work we are
4871                          * generating.  Since we hold a transaction here we
4872                          * can't flush, and we don't want to FLUSH_LIMIT because
4873                          * we could have generated too many delayed refs to
4874                          * actually allocate, so just bail if we're short and
4875                          * let the normal reservation dance happen higher up.
4876                          */
4877                         if (should_throttle) {
4878                                 ret = btrfs_delayed_refs_rsv_refill(fs_info,
4879                                                         BTRFS_RESERVE_NO_FLUSH);
4880                                 if (ret) {
4881                                         ret = -EAGAIN;
4882                                         break;
4883                                 }
4884                         }
4885                         goto search_again;
4886                 } else {
4887                         path->slots[0]--;
4888                 }
4889         }
4890 out:
4891         if (ret >= 0 && pending_del_nr) {
4892                 int err;
4893
4894                 err = btrfs_del_items(trans, root, path, pending_del_slot,
4895                                       pending_del_nr);
4896                 if (err) {
4897                         btrfs_abort_transaction(trans, err);
4898                         ret = err;
4899                 }
4900         }
4901         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4902                 ASSERT(last_size >= new_size);
4903                 if (!ret && last_size > new_size)
4904                         last_size = new_size;
4905                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4906         }
4907
4908         btrfs_free_path(path);
4909         return ret;
4910 }
4911
4912 /*
4913  * btrfs_truncate_block - read, zero a chunk and write a block
4914  * @inode - inode that we're zeroing
4915  * @from - the offset to start zeroing
4916  * @len - the length to zero, 0 to zero the entire range respective to the
4917  *      offset
4918  * @front - zero up to the offset instead of from the offset on
4919  *
4920  * This will find the block for the "from" offset and cow the block and zero the
4921  * part we want to zero.  This is used with truncate and hole punching.
4922  */
4923 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4924                         int front)
4925 {
4926         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4927         struct address_space *mapping = inode->i_mapping;
4928         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4929         struct btrfs_ordered_extent *ordered;
4930         struct extent_state *cached_state = NULL;
4931         struct extent_changeset *data_reserved = NULL;
4932         char *kaddr;
4933         u32 blocksize = fs_info->sectorsize;
4934         pgoff_t index = from >> PAGE_SHIFT;
4935         unsigned offset = from & (blocksize - 1);
4936         struct page *page;
4937         gfp_t mask = btrfs_alloc_write_mask(mapping);
4938         int ret = 0;
4939         u64 block_start;
4940         u64 block_end;
4941
4942         if (IS_ALIGNED(offset, blocksize) &&
4943             (!len || IS_ALIGNED(len, blocksize)))
4944                 goto out;
4945
4946         block_start = round_down(from, blocksize);
4947         block_end = block_start + blocksize - 1;
4948
4949         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4950                                            block_start, blocksize);
4951         if (ret)
4952                 goto out;
4953
4954 again:
4955         page = find_or_create_page(mapping, index, mask);
4956         if (!page) {
4957                 btrfs_delalloc_release_space(inode, data_reserved,
4958                                              block_start, blocksize, true);
4959                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4960                 ret = -ENOMEM;
4961                 goto out;
4962         }
4963
4964         if (!PageUptodate(page)) {
4965                 ret = btrfs_readpage(NULL, page);
4966                 lock_page(page);
4967                 if (page->mapping != mapping) {
4968                         unlock_page(page);
4969                         put_page(page);
4970                         goto again;
4971                 }
4972                 if (!PageUptodate(page)) {
4973                         ret = -EIO;
4974                         goto out_unlock;
4975                 }
4976         }
4977         wait_on_page_writeback(page);
4978
4979         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4980         set_page_extent_mapped(page);
4981
4982         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4983         if (ordered) {
4984                 unlock_extent_cached(io_tree, block_start, block_end,
4985                                      &cached_state);
4986                 unlock_page(page);
4987                 put_page(page);
4988                 btrfs_start_ordered_extent(inode, ordered, 1);
4989                 btrfs_put_ordered_extent(ordered);
4990                 goto again;
4991         }
4992
4993         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4994                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4995                          0, 0, &cached_state);
4996
4997         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4998                                         &cached_state);
4999         if (ret) {
5000                 unlock_extent_cached(io_tree, block_start, block_end,
5001                                      &cached_state);
5002                 goto out_unlock;
5003         }
5004
5005         if (offset != blocksize) {
5006                 if (!len)
5007                         len = blocksize - offset;
5008                 kaddr = kmap(page);
5009                 if (front)
5010                         memset(kaddr + (block_start - page_offset(page)),
5011                                 0, offset);
5012                 else
5013                         memset(kaddr + (block_start - page_offset(page)) +  offset,
5014                                 0, len);
5015                 flush_dcache_page(page);
5016                 kunmap(page);
5017         }
5018         ClearPageChecked(page);
5019         set_page_dirty(page);
5020         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
5021
5022 out_unlock:
5023         if (ret)
5024                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
5025                                              blocksize, true);
5026         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
5027         unlock_page(page);
5028         put_page(page);
5029 out:
5030         extent_changeset_free(data_reserved);
5031         return ret;
5032 }
5033
5034 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
5035                              u64 offset, u64 len)
5036 {
5037         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5038         struct btrfs_trans_handle *trans;
5039         int ret;
5040
5041         /*
5042          * Still need to make sure the inode looks like it's been updated so
5043          * that any holes get logged if we fsync.
5044          */
5045         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
5046                 BTRFS_I(inode)->last_trans = fs_info->generation;
5047                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
5048                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
5049                 return 0;
5050         }
5051
5052         /*
5053          * 1 - for the one we're dropping
5054          * 1 - for the one we're adding
5055          * 1 - for updating the inode.
5056          */
5057         trans = btrfs_start_transaction(root, 3);
5058         if (IS_ERR(trans))
5059                 return PTR_ERR(trans);
5060
5061         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
5062         if (ret) {
5063                 btrfs_abort_transaction(trans, ret);
5064                 btrfs_end_transaction(trans);
5065                 return ret;
5066         }
5067
5068         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
5069                         offset, 0, 0, len, 0, len, 0, 0, 0);
5070         if (ret)
5071                 btrfs_abort_transaction(trans, ret);
5072         else
5073                 btrfs_update_inode(trans, root, inode);
5074         btrfs_end_transaction(trans);
5075         return ret;
5076 }
5077
5078 /*
5079  * This function puts in dummy file extents for the area we're creating a hole
5080  * for.  So if we are truncating this file to a larger size we need to insert
5081  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
5082  * the range between oldsize and size
5083  */
5084 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
5085 {
5086         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5087         struct btrfs_root *root = BTRFS_I(inode)->root;
5088         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5089         struct extent_map *em = NULL;
5090         struct extent_state *cached_state = NULL;
5091         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
5092         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
5093         u64 block_end = ALIGN(size, fs_info->sectorsize);
5094         u64 last_byte;
5095         u64 cur_offset;
5096         u64 hole_size;
5097         int err = 0;
5098
5099         /*
5100          * If our size started in the middle of a block we need to zero out the
5101          * rest of the block before we expand the i_size, otherwise we could
5102          * expose stale data.
5103          */
5104         err = btrfs_truncate_block(inode, oldsize, 0, 0);
5105         if (err)
5106                 return err;
5107
5108         if (size <= hole_start)
5109                 return 0;
5110
5111         btrfs_lock_and_flush_ordered_range(io_tree, BTRFS_I(inode), hole_start,
5112                                            block_end - 1, &cached_state);
5113         cur_offset = hole_start;
5114         while (1) {
5115                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
5116                                 block_end - cur_offset, 0);
5117                 if (IS_ERR(em)) {
5118                         err = PTR_ERR(em);
5119                         em = NULL;
5120                         break;
5121                 }
5122                 last_byte = min(extent_map_end(em), block_end);
5123                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
5124                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
5125                         struct extent_map *hole_em;
5126                         hole_size = last_byte - cur_offset;
5127
5128                         err = maybe_insert_hole(root, inode, cur_offset,
5129                                                 hole_size);
5130                         if (err)
5131                                 break;
5132                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
5133                                                 cur_offset + hole_size - 1, 0);
5134                         hole_em = alloc_extent_map();
5135                         if (!hole_em) {
5136                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5137                                         &BTRFS_I(inode)->runtime_flags);
5138                                 goto next;
5139                         }
5140                         hole_em->start = cur_offset;
5141                         hole_em->len = hole_size;
5142                         hole_em->orig_start = cur_offset;
5143
5144                         hole_em->block_start = EXTENT_MAP_HOLE;
5145                         hole_em->block_len = 0;
5146                         hole_em->orig_block_len = 0;
5147                         hole_em->ram_bytes = hole_size;
5148                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
5149                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5150                         hole_em->generation = fs_info->generation;
5151
5152                         while (1) {
5153                                 write_lock(&em_tree->lock);
5154                                 err = add_extent_mapping(em_tree, hole_em, 1);
5155                                 write_unlock(&em_tree->lock);
5156                                 if (err != -EEXIST)
5157                                         break;
5158                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5159                                                         cur_offset,
5160                                                         cur_offset +
5161                                                         hole_size - 1, 0);
5162                         }
5163                         free_extent_map(hole_em);
5164                 }
5165 next:
5166                 free_extent_map(em);
5167                 em = NULL;
5168                 cur_offset = last_byte;
5169                 if (cur_offset >= block_end)
5170                         break;
5171         }
5172         free_extent_map(em);
5173         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5174         return err;
5175 }
5176
5177 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5178 {
5179         struct btrfs_root *root = BTRFS_I(inode)->root;
5180         struct btrfs_trans_handle *trans;
5181         loff_t oldsize = i_size_read(inode);
5182         loff_t newsize = attr->ia_size;
5183         int mask = attr->ia_valid;
5184         int ret;
5185
5186         /*
5187          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5188          * special case where we need to update the times despite not having
5189          * these flags set.  For all other operations the VFS set these flags
5190          * explicitly if it wants a timestamp update.
5191          */
5192         if (newsize != oldsize) {
5193                 inode_inc_iversion(inode);
5194                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5195                         inode->i_ctime = inode->i_mtime =
5196                                 current_time(inode);
5197         }
5198
5199         if (newsize > oldsize) {
5200                 /*
5201                  * Don't do an expanding truncate while snapshotting is ongoing.
5202                  * This is to ensure the snapshot captures a fully consistent
5203                  * state of this file - if the snapshot captures this expanding
5204                  * truncation, it must capture all writes that happened before
5205                  * this truncation.
5206                  */
5207                 btrfs_wait_for_snapshot_creation(root);
5208                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5209                 if (ret) {
5210                         btrfs_end_write_no_snapshotting(root);
5211                         return ret;
5212                 }
5213
5214                 trans = btrfs_start_transaction(root, 1);
5215                 if (IS_ERR(trans)) {
5216                         btrfs_end_write_no_snapshotting(root);
5217                         return PTR_ERR(trans);
5218                 }
5219
5220                 i_size_write(inode, newsize);
5221                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5222                 pagecache_isize_extended(inode, oldsize, newsize);
5223                 ret = btrfs_update_inode(trans, root, inode);
5224                 btrfs_end_write_no_snapshotting(root);
5225                 btrfs_end_transaction(trans);
5226         } else {
5227
5228                 /*
5229                  * We're truncating a file that used to have good data down to
5230                  * zero. Make sure it gets into the ordered flush list so that
5231                  * any new writes get down to disk quickly.
5232                  */
5233                 if (newsize == 0)
5234                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5235                                 &BTRFS_I(inode)->runtime_flags);
5236
5237                 truncate_setsize(inode, newsize);
5238
5239                 /* Disable nonlocked read DIO to avoid the endless truncate */
5240                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5241                 inode_dio_wait(inode);
5242                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5243
5244                 ret = btrfs_truncate(inode, newsize == oldsize);
5245                 if (ret && inode->i_nlink) {
5246                         int err;
5247
5248                         /*
5249                          * Truncate failed, so fix up the in-memory size. We
5250                          * adjusted disk_i_size down as we removed extents, so
5251                          * wait for disk_i_size to be stable and then update the
5252                          * in-memory size to match.
5253                          */
5254                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5255                         if (err)
5256                                 return err;
5257                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5258                 }
5259         }
5260
5261         return ret;
5262 }
5263
5264 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5265 {
5266         struct inode *inode = d_inode(dentry);
5267         struct btrfs_root *root = BTRFS_I(inode)->root;
5268         int err;
5269
5270         if (btrfs_root_readonly(root))
5271                 return -EROFS;
5272
5273         err = setattr_prepare(dentry, attr);
5274         if (err)
5275                 return err;
5276
5277         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5278                 err = btrfs_setsize(inode, attr);
5279                 if (err)
5280                         return err;
5281         }
5282
5283         if (attr->ia_valid) {
5284                 setattr_copy(inode, attr);
5285                 inode_inc_iversion(inode);
5286                 err = btrfs_dirty_inode(inode);
5287
5288                 if (!err && attr->ia_valid & ATTR_MODE)
5289                         err = posix_acl_chmod(inode, inode->i_mode);
5290         }
5291
5292         return err;
5293 }
5294
5295 /*
5296  * While truncating the inode pages during eviction, we get the VFS calling
5297  * btrfs_invalidatepage() against each page of the inode. This is slow because
5298  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5299  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5300  * extent_state structures over and over, wasting lots of time.
5301  *
5302  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5303  * those expensive operations on a per page basis and do only the ordered io
5304  * finishing, while we release here the extent_map and extent_state structures,
5305  * without the excessive merging and splitting.
5306  */
5307 static void evict_inode_truncate_pages(struct inode *inode)
5308 {
5309         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5310         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5311         struct rb_node *node;
5312
5313         ASSERT(inode->i_state & I_FREEING);
5314         truncate_inode_pages_final(&inode->i_data);
5315
5316         write_lock(&map_tree->lock);
5317         while (!RB_EMPTY_ROOT(&map_tree->map.rb_root)) {
5318                 struct extent_map *em;
5319
5320                 node = rb_first_cached(&map_tree->map);
5321                 em = rb_entry(node, struct extent_map, rb_node);
5322                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5323                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5324                 remove_extent_mapping(map_tree, em);
5325                 free_extent_map(em);
5326                 if (need_resched()) {
5327                         write_unlock(&map_tree->lock);
5328                         cond_resched();
5329                         write_lock(&map_tree->lock);
5330                 }
5331         }
5332         write_unlock(&map_tree->lock);
5333
5334         /*
5335          * Keep looping until we have no more ranges in the io tree.
5336          * We can have ongoing bios started by readpages (called from readahead)
5337          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5338          * still in progress (unlocked the pages in the bio but did not yet
5339          * unlocked the ranges in the io tree). Therefore this means some
5340          * ranges can still be locked and eviction started because before
5341          * submitting those bios, which are executed by a separate task (work
5342          * queue kthread), inode references (inode->i_count) were not taken
5343          * (which would be dropped in the end io callback of each bio).
5344          * Therefore here we effectively end up waiting for those bios and
5345          * anyone else holding locked ranges without having bumped the inode's
5346          * reference count - if we don't do it, when they access the inode's
5347          * io_tree to unlock a range it may be too late, leading to an
5348          * use-after-free issue.
5349          */
5350         spin_lock(&io_tree->lock);
5351         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5352                 struct extent_state *state;
5353                 struct extent_state *cached_state = NULL;
5354                 u64 start;
5355                 u64 end;
5356                 unsigned state_flags;
5357
5358                 node = rb_first(&io_tree->state);
5359                 state = rb_entry(node, struct extent_state, rb_node);
5360                 start = state->start;
5361                 end = state->end;
5362                 state_flags = state->state;
5363                 spin_unlock(&io_tree->lock);
5364
5365                 lock_extent_bits(io_tree, start, end, &cached_state);
5366
5367                 /*
5368                  * If still has DELALLOC flag, the extent didn't reach disk,
5369                  * and its reserved space won't be freed by delayed_ref.
5370                  * So we need to free its reserved space here.
5371                  * (Refer to comment in btrfs_invalidatepage, case 2)
5372                  *
5373                  * Note, end is the bytenr of last byte, so we need + 1 here.
5374                  */
5375                 if (state_flags & EXTENT_DELALLOC)
5376                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5377
5378                 clear_extent_bit(io_tree, start, end,
5379                                  EXTENT_LOCKED | EXTENT_DELALLOC |
5380                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
5381                                  &cached_state);
5382
5383                 cond_resched();
5384                 spin_lock(&io_tree->lock);
5385         }
5386         spin_unlock(&io_tree->lock);
5387 }
5388
5389 static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
5390                                                         struct btrfs_block_rsv *rsv)
5391 {
5392         struct btrfs_fs_info *fs_info = root->fs_info;
5393         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5394         struct btrfs_trans_handle *trans;
5395         u64 delayed_refs_extra = btrfs_calc_insert_metadata_size(fs_info, 1);
5396         int ret;
5397
5398         /*
5399          * Eviction should be taking place at some place safe because of our
5400          * delayed iputs.  However the normal flushing code will run delayed
5401          * iputs, so we cannot use FLUSH_ALL otherwise we'll deadlock.
5402          *
5403          * We reserve the delayed_refs_extra here again because we can't use
5404          * btrfs_start_transaction(root, 0) for the same deadlocky reason as
5405          * above.  We reserve our extra bit here because we generate a ton of
5406          * delayed refs activity by truncating.
5407          *
5408          * If we cannot make our reservation we'll attempt to steal from the
5409          * global reserve, because we really want to be able to free up space.
5410          */
5411         ret = btrfs_block_rsv_refill(root, rsv, rsv->size + delayed_refs_extra,
5412                                      BTRFS_RESERVE_FLUSH_EVICT);
5413         if (ret) {
5414                 /*
5415                  * Try to steal from the global reserve if there is space for
5416                  * it.
5417                  */
5418                 if (btrfs_check_space_for_delayed_refs(fs_info) ||
5419                     btrfs_block_rsv_migrate(global_rsv, rsv, rsv->size, 0)) {
5420                         btrfs_warn(fs_info,
5421                                    "could not allocate space for delete; will truncate on mount");
5422                         return ERR_PTR(-ENOSPC);
5423                 }
5424                 delayed_refs_extra = 0;
5425         }
5426
5427         trans = btrfs_join_transaction(root);
5428         if (IS_ERR(trans))
5429                 return trans;
5430
5431         if (delayed_refs_extra) {
5432                 trans->block_rsv = &fs_info->trans_block_rsv;
5433                 trans->bytes_reserved = delayed_refs_extra;
5434                 btrfs_block_rsv_migrate(rsv, trans->block_rsv,
5435                                         delayed_refs_extra, 1);
5436         }
5437         return trans;
5438 }
5439
5440 void btrfs_evict_inode(struct inode *inode)
5441 {
5442         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5443         struct btrfs_trans_handle *trans;
5444         struct btrfs_root *root = BTRFS_I(inode)->root;
5445         struct btrfs_block_rsv *rsv;
5446         int ret;
5447
5448         trace_btrfs_inode_evict(inode);
5449
5450         if (!root) {
5451                 clear_inode(inode);
5452                 return;
5453         }
5454
5455         evict_inode_truncate_pages(inode);
5456
5457         if (inode->i_nlink &&
5458             ((btrfs_root_refs(&root->root_item) != 0 &&
5459               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5460              btrfs_is_free_space_inode(BTRFS_I(inode))))
5461                 goto no_delete;
5462
5463         if (is_bad_inode(inode))
5464                 goto no_delete;
5465
5466         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5467
5468         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
5469                 goto no_delete;
5470
5471         if (inode->i_nlink > 0) {
5472                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5473                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5474                 goto no_delete;
5475         }
5476
5477         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5478         if (ret)
5479                 goto no_delete;
5480
5481         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5482         if (!rsv)
5483                 goto no_delete;
5484         rsv->size = btrfs_calc_metadata_size(fs_info, 1);
5485         rsv->failfast = 1;
5486
5487         btrfs_i_size_write(BTRFS_I(inode), 0);
5488
5489         while (1) {
5490                 trans = evict_refill_and_join(root, rsv);
5491                 if (IS_ERR(trans))
5492                         goto free_rsv;
5493
5494                 trans->block_rsv = rsv;
5495
5496                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5497                 trans->block_rsv = &fs_info->trans_block_rsv;
5498                 btrfs_end_transaction(trans);
5499                 btrfs_btree_balance_dirty(fs_info);
5500                 if (ret && ret != -ENOSPC && ret != -EAGAIN)
5501                         goto free_rsv;
5502                 else if (!ret)
5503                         break;
5504         }
5505
5506         /*
5507          * Errors here aren't a big deal, it just means we leave orphan items in
5508          * the tree. They will be cleaned up on the next mount. If the inode
5509          * number gets reused, cleanup deletes the orphan item without doing
5510          * anything, and unlink reuses the existing orphan item.
5511          *
5512          * If it turns out that we are dropping too many of these, we might want
5513          * to add a mechanism for retrying these after a commit.
5514          */
5515         trans = evict_refill_and_join(root, rsv);
5516         if (!IS_ERR(trans)) {
5517                 trans->block_rsv = rsv;
5518                 btrfs_orphan_del(trans, BTRFS_I(inode));
5519                 trans->block_rsv = &fs_info->trans_block_rsv;
5520                 btrfs_end_transaction(trans);
5521         }
5522
5523         if (!(root == fs_info->tree_root ||
5524               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5525                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5526
5527 free_rsv:
5528         btrfs_free_block_rsv(fs_info, rsv);
5529 no_delete:
5530         /*
5531          * If we didn't successfully delete, the orphan item will still be in
5532          * the tree and we'll retry on the next mount. Again, we might also want
5533          * to retry these periodically in the future.
5534          */
5535         btrfs_remove_delayed_node(BTRFS_I(inode));
5536         clear_inode(inode);
5537 }
5538
5539 /*
5540  * Return the key found in the dir entry in the location pointer, fill @type
5541  * with BTRFS_FT_*, and return 0.
5542  *
5543  * If no dir entries were found, returns -ENOENT.
5544  * If found a corrupted location in dir entry, returns -EUCLEAN.
5545  */
5546 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5547                                struct btrfs_key *location, u8 *type)
5548 {
5549         const char *name = dentry->d_name.name;
5550         int namelen = dentry->d_name.len;
5551         struct btrfs_dir_item *di;
5552         struct btrfs_path *path;
5553         struct btrfs_root *root = BTRFS_I(dir)->root;
5554         int ret = 0;
5555
5556         path = btrfs_alloc_path();
5557         if (!path)
5558                 return -ENOMEM;
5559
5560         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5561                         name, namelen, 0);
5562         if (IS_ERR_OR_NULL(di)) {
5563                 ret = di ? PTR_ERR(di) : -ENOENT;
5564                 goto out;
5565         }
5566
5567         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5568         if (location->type != BTRFS_INODE_ITEM_KEY &&
5569             location->type != BTRFS_ROOT_ITEM_KEY) {
5570                 ret = -EUCLEAN;
5571                 btrfs_warn(root->fs_info,
5572 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5573                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5574                            location->objectid, location->type, location->offset);
5575         }
5576         if (!ret)
5577                 *type = btrfs_dir_type(path->nodes[0], di);
5578 out:
5579         btrfs_free_path(path);
5580         return ret;
5581 }
5582
5583 /*
5584  * when we hit a tree root in a directory, the btrfs part of the inode
5585  * needs to be changed to reflect the root directory of the tree root.  This
5586  * is kind of like crossing a mount point.
5587  */
5588 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5589                                     struct inode *dir,
5590                                     struct dentry *dentry,
5591                                     struct btrfs_key *location,
5592                                     struct btrfs_root **sub_root)
5593 {
5594         struct btrfs_path *path;
5595         struct btrfs_root *new_root;
5596         struct btrfs_root_ref *ref;
5597         struct extent_buffer *leaf;
5598         struct btrfs_key key;
5599         int ret;
5600         int err = 0;
5601
5602         path = btrfs_alloc_path();
5603         if (!path) {
5604                 err = -ENOMEM;
5605                 goto out;
5606         }
5607
5608         err = -ENOENT;
5609         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5610         key.type = BTRFS_ROOT_REF_KEY;
5611         key.offset = location->objectid;
5612
5613         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5614         if (ret) {
5615                 if (ret < 0)
5616                         err = ret;
5617                 goto out;
5618         }
5619
5620         leaf = path->nodes[0];
5621         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5622         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5623             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5624                 goto out;
5625
5626         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5627                                    (unsigned long)(ref + 1),
5628                                    dentry->d_name.len);
5629         if (ret)
5630                 goto out;
5631
5632         btrfs_release_path(path);
5633
5634         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5635         if (IS_ERR(new_root)) {
5636                 err = PTR_ERR(new_root);
5637                 goto out;
5638         }
5639
5640         *sub_root = new_root;
5641         location->objectid = btrfs_root_dirid(&new_root->root_item);
5642         location->type = BTRFS_INODE_ITEM_KEY;
5643         location->offset = 0;
5644         err = 0;
5645 out:
5646         btrfs_free_path(path);
5647         return err;
5648 }
5649
5650 static void inode_tree_add(struct inode *inode)
5651 {
5652         struct btrfs_root *root = BTRFS_I(inode)->root;
5653         struct btrfs_inode *entry;
5654         struct rb_node **p;
5655         struct rb_node *parent;
5656         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5657         u64 ino = btrfs_ino(BTRFS_I(inode));
5658
5659         if (inode_unhashed(inode))
5660                 return;
5661         parent = NULL;
5662         spin_lock(&root->inode_lock);
5663         p = &root->inode_tree.rb_node;
5664         while (*p) {
5665                 parent = *p;
5666                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5667
5668                 if (ino < btrfs_ino(entry))
5669                         p = &parent->rb_left;
5670                 else if (ino > btrfs_ino(entry))
5671                         p = &parent->rb_right;
5672                 else {
5673                         WARN_ON(!(entry->vfs_inode.i_state &
5674                                   (I_WILL_FREE | I_FREEING)));
5675                         rb_replace_node(parent, new, &root->inode_tree);
5676                         RB_CLEAR_NODE(parent);
5677                         spin_unlock(&root->inode_lock);
5678                         return;
5679                 }
5680         }
5681         rb_link_node(new, parent, p);
5682         rb_insert_color(new, &root->inode_tree);
5683         spin_unlock(&root->inode_lock);
5684 }
5685
5686 static void inode_tree_del(struct inode *inode)
5687 {
5688         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5689         struct btrfs_root *root = BTRFS_I(inode)->root;
5690         int empty = 0;
5691
5692         spin_lock(&root->inode_lock);
5693         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5694                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5695                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5696                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5697         }
5698         spin_unlock(&root->inode_lock);
5699
5700         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5701                 synchronize_srcu(&fs_info->subvol_srcu);
5702                 spin_lock(&root->inode_lock);
5703                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5704                 spin_unlock(&root->inode_lock);
5705                 if (empty)
5706                         btrfs_add_dead_root(root);
5707         }
5708 }
5709
5710
5711 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5712 {
5713         struct btrfs_iget_args *args = p;
5714         inode->i_ino = args->location->objectid;
5715         memcpy(&BTRFS_I(inode)->location, args->location,
5716                sizeof(*args->location));
5717         BTRFS_I(inode)->root = args->root;
5718         return 0;
5719 }
5720
5721 static int btrfs_find_actor(struct inode *inode, void *opaque)
5722 {
5723         struct btrfs_iget_args *args = opaque;
5724         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5725                 args->root == BTRFS_I(inode)->root;
5726 }
5727
5728 static struct inode *btrfs_iget_locked(struct super_block *s,
5729                                        struct btrfs_key *location,
5730                                        struct btrfs_root *root)
5731 {
5732         struct inode *inode;
5733         struct btrfs_iget_args args;
5734         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5735
5736         args.location = location;
5737         args.root = root;
5738
5739         inode = iget5_locked(s, hashval, btrfs_find_actor,
5740                              btrfs_init_locked_inode,
5741                              (void *)&args);
5742         return inode;
5743 }
5744
5745 /* Get an inode object given its location and corresponding root.
5746  * Returns in *is_new if the inode was read from disk
5747  */
5748 struct inode *btrfs_iget_path(struct super_block *s, struct btrfs_key *location,
5749                               struct btrfs_root *root, int *new,
5750                               struct btrfs_path *path)
5751 {
5752         struct inode *inode;
5753
5754         inode = btrfs_iget_locked(s, location, root);
5755         if (!inode)
5756                 return ERR_PTR(-ENOMEM);
5757
5758         if (inode->i_state & I_NEW) {
5759                 int ret;
5760
5761                 ret = btrfs_read_locked_inode(inode, path);
5762                 if (!ret) {
5763                         inode_tree_add(inode);
5764                         unlock_new_inode(inode);
5765                         if (new)
5766                                 *new = 1;
5767                 } else {
5768                         iget_failed(inode);
5769                         /*
5770                          * ret > 0 can come from btrfs_search_slot called by
5771                          * btrfs_read_locked_inode, this means the inode item
5772                          * was not found.
5773                          */
5774                         if (ret > 0)
5775                                 ret = -ENOENT;
5776                         inode = ERR_PTR(ret);
5777                 }
5778         }
5779
5780         return inode;
5781 }
5782
5783 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5784                          struct btrfs_root *root, int *new)
5785 {
5786         return btrfs_iget_path(s, location, root, new, NULL);
5787 }
5788
5789 static struct inode *new_simple_dir(struct super_block *s,
5790                                     struct btrfs_key *key,
5791                                     struct btrfs_root *root)
5792 {
5793         struct inode *inode = new_inode(s);
5794
5795         if (!inode)
5796                 return ERR_PTR(-ENOMEM);
5797
5798         BTRFS_I(inode)->root = root;
5799         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5800         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5801
5802         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5803         inode->i_op = &btrfs_dir_ro_inode_operations;
5804         inode->i_opflags &= ~IOP_XATTR;
5805         inode->i_fop = &simple_dir_operations;
5806         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5807         inode->i_mtime = current_time(inode);
5808         inode->i_atime = inode->i_mtime;
5809         inode->i_ctime = inode->i_mtime;
5810         BTRFS_I(inode)->i_otime = inode->i_mtime;
5811
5812         return inode;
5813 }
5814
5815 static inline u8 btrfs_inode_type(struct inode *inode)
5816 {
5817         /*
5818          * Compile-time asserts that generic FT_* types still match
5819          * BTRFS_FT_* types
5820          */
5821         BUILD_BUG_ON(BTRFS_FT_UNKNOWN != FT_UNKNOWN);
5822         BUILD_BUG_ON(BTRFS_FT_REG_FILE != FT_REG_FILE);
5823         BUILD_BUG_ON(BTRFS_FT_DIR != FT_DIR);
5824         BUILD_BUG_ON(BTRFS_FT_CHRDEV != FT_CHRDEV);
5825         BUILD_BUG_ON(BTRFS_FT_BLKDEV != FT_BLKDEV);
5826         BUILD_BUG_ON(BTRFS_FT_FIFO != FT_FIFO);
5827         BUILD_BUG_ON(BTRFS_FT_SOCK != FT_SOCK);
5828         BUILD_BUG_ON(BTRFS_FT_SYMLINK != FT_SYMLINK);
5829
5830         return fs_umode_to_ftype(inode->i_mode);
5831 }
5832
5833 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5834 {
5835         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5836         struct inode *inode;
5837         struct btrfs_root *root = BTRFS_I(dir)->root;
5838         struct btrfs_root *sub_root = root;
5839         struct btrfs_key location;
5840         u8 di_type = 0;
5841         int index;
5842         int ret = 0;
5843
5844         if (dentry->d_name.len > BTRFS_NAME_LEN)
5845                 return ERR_PTR(-ENAMETOOLONG);
5846
5847         ret = btrfs_inode_by_name(dir, dentry, &location, &di_type);
5848         if (ret < 0)
5849                 return ERR_PTR(ret);
5850
5851         if (location.type == BTRFS_INODE_ITEM_KEY) {
5852                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5853                 if (IS_ERR(inode))
5854                         return inode;
5855
5856                 /* Do extra check against inode mode with di_type */
5857                 if (btrfs_inode_type(inode) != di_type) {
5858                         btrfs_crit(fs_info,
5859 "inode mode mismatch with dir: inode mode=0%o btrfs type=%u dir type=%u",
5860                                   inode->i_mode, btrfs_inode_type(inode),
5861                                   di_type);
5862                         iput(inode);
5863                         return ERR_PTR(-EUCLEAN);
5864                 }
5865                 return inode;
5866         }
5867
5868         index = srcu_read_lock(&fs_info->subvol_srcu);
5869         ret = fixup_tree_root_location(fs_info, dir, dentry,
5870                                        &location, &sub_root);
5871         if (ret < 0) {
5872                 if (ret != -ENOENT)
5873                         inode = ERR_PTR(ret);
5874                 else
5875                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5876         } else {
5877                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5878         }
5879         srcu_read_unlock(&fs_info->subvol_srcu, index);
5880
5881         if (!IS_ERR(inode) && root != sub_root) {
5882                 down_read(&fs_info->cleanup_work_sem);
5883                 if (!sb_rdonly(inode->i_sb))
5884                         ret = btrfs_orphan_cleanup(sub_root);
5885                 up_read(&fs_info->cleanup_work_sem);
5886                 if (ret) {
5887                         iput(inode);
5888                         inode = ERR_PTR(ret);
5889                 }
5890         }
5891
5892         return inode;
5893 }
5894
5895 static int btrfs_dentry_delete(const struct dentry *dentry)
5896 {
5897         struct btrfs_root *root;
5898         struct inode *inode = d_inode(dentry);
5899
5900         if (!inode && !IS_ROOT(dentry))
5901                 inode = d_inode(dentry->d_parent);
5902
5903         if (inode) {
5904                 root = BTRFS_I(inode)->root;
5905                 if (btrfs_root_refs(&root->root_item) == 0)
5906                         return 1;
5907
5908                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5909                         return 1;
5910         }
5911         return 0;
5912 }
5913
5914 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5915                                    unsigned int flags)
5916 {
5917         struct inode *inode = btrfs_lookup_dentry(dir, dentry);
5918
5919         if (inode == ERR_PTR(-ENOENT))
5920                 inode = NULL;
5921         return d_splice_alias(inode, dentry);
5922 }
5923
5924 /*
5925  * All this infrastructure exists because dir_emit can fault, and we are holding
5926  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5927  * our information into that, and then dir_emit from the buffer.  This is
5928  * similar to what NFS does, only we don't keep the buffer around in pagecache
5929  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5930  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5931  * tree lock.
5932  */
5933 static int btrfs_opendir(struct inode *inode, struct file *file)
5934 {
5935         struct btrfs_file_private *private;
5936
5937         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5938         if (!private)
5939                 return -ENOMEM;
5940         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5941         if (!private->filldir_buf) {
5942                 kfree(private);
5943                 return -ENOMEM;
5944         }
5945         file->private_data = private;
5946         return 0;
5947 }
5948
5949 struct dir_entry {
5950         u64 ino;
5951         u64 offset;
5952         unsigned type;
5953         int name_len;
5954 };
5955
5956 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5957 {
5958         while (entries--) {
5959                 struct dir_entry *entry = addr;
5960                 char *name = (char *)(entry + 1);
5961
5962                 ctx->pos = get_unaligned(&entry->offset);
5963                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
5964                                          get_unaligned(&entry->ino),
5965                                          get_unaligned(&entry->type)))
5966                         return 1;
5967                 addr += sizeof(struct dir_entry) +
5968                         get_unaligned(&entry->name_len);
5969                 ctx->pos++;
5970         }
5971         return 0;
5972 }
5973
5974 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5975 {
5976         struct inode *inode = file_inode(file);
5977         struct btrfs_root *root = BTRFS_I(inode)->root;
5978         struct btrfs_file_private *private = file->private_data;
5979         struct btrfs_dir_item *di;
5980         struct btrfs_key key;
5981         struct btrfs_key found_key;
5982         struct btrfs_path *path;
5983         void *addr;
5984         struct list_head ins_list;
5985         struct list_head del_list;
5986         int ret;
5987         struct extent_buffer *leaf;
5988         int slot;
5989         char *name_ptr;
5990         int name_len;
5991         int entries = 0;
5992         int total_len = 0;
5993         bool put = false;
5994         struct btrfs_key location;
5995
5996         if (!dir_emit_dots(file, ctx))
5997                 return 0;
5998
5999         path = btrfs_alloc_path();
6000         if (!path)
6001                 return -ENOMEM;
6002
6003         addr = private->filldir_buf;
6004         path->reada = READA_FORWARD;
6005
6006         INIT_LIST_HEAD(&ins_list);
6007         INIT_LIST_HEAD(&del_list);
6008         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
6009
6010 again:
6011         key.type = BTRFS_DIR_INDEX_KEY;
6012         key.offset = ctx->pos;
6013         key.objectid = btrfs_ino(BTRFS_I(inode));
6014
6015         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6016         if (ret < 0)
6017                 goto err;
6018
6019         while (1) {
6020                 struct dir_entry *entry;
6021
6022                 leaf = path->nodes[0];
6023                 slot = path->slots[0];
6024                 if (slot >= btrfs_header_nritems(leaf)) {
6025                         ret = btrfs_next_leaf(root, path);
6026                         if (ret < 0)
6027                                 goto err;
6028                         else if (ret > 0)
6029                                 break;
6030                         continue;
6031                 }
6032
6033                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
6034
6035                 if (found_key.objectid != key.objectid)
6036                         break;
6037                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
6038                         break;
6039                 if (found_key.offset < ctx->pos)
6040                         goto next;
6041                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
6042                         goto next;
6043                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
6044                 name_len = btrfs_dir_name_len(leaf, di);
6045                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
6046                     PAGE_SIZE) {
6047                         btrfs_release_path(path);
6048                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6049                         if (ret)
6050                                 goto nopos;
6051                         addr = private->filldir_buf;
6052                         entries = 0;
6053                         total_len = 0;
6054                         goto again;
6055                 }
6056
6057                 entry = addr;
6058                 put_unaligned(name_len, &entry->name_len);
6059                 name_ptr = (char *)(entry + 1);
6060                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
6061                                    name_len);
6062                 put_unaligned(fs_ftype_to_dtype(btrfs_dir_type(leaf, di)),
6063                                 &entry->type);
6064                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
6065                 put_unaligned(location.objectid, &entry->ino);
6066                 put_unaligned(found_key.offset, &entry->offset);
6067                 entries++;
6068                 addr += sizeof(struct dir_entry) + name_len;
6069                 total_len += sizeof(struct dir_entry) + name_len;
6070 next:
6071                 path->slots[0]++;
6072         }
6073         btrfs_release_path(path);
6074
6075         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6076         if (ret)
6077                 goto nopos;
6078
6079         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6080         if (ret)
6081                 goto nopos;
6082
6083         /*
6084          * Stop new entries from being returned after we return the last
6085          * entry.
6086          *
6087          * New directory entries are assigned a strictly increasing
6088          * offset.  This means that new entries created during readdir
6089          * are *guaranteed* to be seen in the future by that readdir.
6090          * This has broken buggy programs which operate on names as
6091          * they're returned by readdir.  Until we re-use freed offsets
6092          * we have this hack to stop new entries from being returned
6093          * under the assumption that they'll never reach this huge
6094          * offset.
6095          *
6096          * This is being careful not to overflow 32bit loff_t unless the
6097          * last entry requires it because doing so has broken 32bit apps
6098          * in the past.
6099          */
6100         if (ctx->pos >= INT_MAX)
6101                 ctx->pos = LLONG_MAX;
6102         else
6103                 ctx->pos = INT_MAX;
6104 nopos:
6105         ret = 0;
6106 err:
6107         if (put)
6108                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6109         btrfs_free_path(path);
6110         return ret;
6111 }
6112
6113 /*
6114  * This is somewhat expensive, updating the tree every time the
6115  * inode changes.  But, it is most likely to find the inode in cache.
6116  * FIXME, needs more benchmarking...there are no reasons other than performance
6117  * to keep or drop this code.
6118  */
6119 static int btrfs_dirty_inode(struct inode *inode)
6120 {
6121         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6122         struct btrfs_root *root = BTRFS_I(inode)->root;
6123         struct btrfs_trans_handle *trans;
6124         int ret;
6125
6126         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6127                 return 0;
6128
6129         trans = btrfs_join_transaction(root);
6130         if (IS_ERR(trans))
6131                 return PTR_ERR(trans);
6132
6133         ret = btrfs_update_inode(trans, root, inode);
6134         if (ret && ret == -ENOSPC) {
6135                 /* whoops, lets try again with the full transaction */
6136                 btrfs_end_transaction(trans);
6137                 trans = btrfs_start_transaction(root, 1);
6138                 if (IS_ERR(trans))
6139                         return PTR_ERR(trans);
6140
6141                 ret = btrfs_update_inode(trans, root, inode);
6142         }
6143         btrfs_end_transaction(trans);
6144         if (BTRFS_I(inode)->delayed_node)
6145                 btrfs_balance_delayed_items(fs_info);
6146
6147         return ret;
6148 }
6149
6150 /*
6151  * This is a copy of file_update_time.  We need this so we can return error on
6152  * ENOSPC for updating the inode in the case of file write and mmap writes.
6153  */
6154 static int btrfs_update_time(struct inode *inode, struct timespec64 *now,
6155                              int flags)
6156 {
6157         struct btrfs_root *root = BTRFS_I(inode)->root;
6158         bool dirty = flags & ~S_VERSION;
6159
6160         if (btrfs_root_readonly(root))
6161                 return -EROFS;
6162
6163         if (flags & S_VERSION)
6164                 dirty |= inode_maybe_inc_iversion(inode, dirty);
6165         if (flags & S_CTIME)
6166                 inode->i_ctime = *now;
6167         if (flags & S_MTIME)
6168                 inode->i_mtime = *now;
6169         if (flags & S_ATIME)
6170                 inode->i_atime = *now;
6171         return dirty ? btrfs_dirty_inode(inode) : 0;
6172 }
6173
6174 /*
6175  * find the highest existing sequence number in a directory
6176  * and then set the in-memory index_cnt variable to reflect
6177  * free sequence numbers
6178  */
6179 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6180 {
6181         struct btrfs_root *root = inode->root;
6182         struct btrfs_key key, found_key;
6183         struct btrfs_path *path;
6184         struct extent_buffer *leaf;
6185         int ret;
6186
6187         key.objectid = btrfs_ino(inode);
6188         key.type = BTRFS_DIR_INDEX_KEY;
6189         key.offset = (u64)-1;
6190
6191         path = btrfs_alloc_path();
6192         if (!path)
6193                 return -ENOMEM;
6194
6195         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6196         if (ret < 0)
6197                 goto out;
6198         /* FIXME: we should be able to handle this */
6199         if (ret == 0)
6200                 goto out;
6201         ret = 0;
6202
6203         /*
6204          * MAGIC NUMBER EXPLANATION:
6205          * since we search a directory based on f_pos we have to start at 2
6206          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6207          * else has to start at 2
6208          */
6209         if (path->slots[0] == 0) {
6210                 inode->index_cnt = 2;
6211                 goto out;
6212         }
6213
6214         path->slots[0]--;
6215
6216         leaf = path->nodes[0];
6217         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6218
6219         if (found_key.objectid != btrfs_ino(inode) ||
6220             found_key.type != BTRFS_DIR_INDEX_KEY) {
6221                 inode->index_cnt = 2;
6222                 goto out;
6223         }
6224
6225         inode->index_cnt = found_key.offset + 1;
6226 out:
6227         btrfs_free_path(path);
6228         return ret;
6229 }
6230
6231 /*
6232  * helper to find a free sequence number in a given directory.  This current
6233  * code is very simple, later versions will do smarter things in the btree
6234  */
6235 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6236 {
6237         int ret = 0;
6238
6239         if (dir->index_cnt == (u64)-1) {
6240                 ret = btrfs_inode_delayed_dir_index_count(dir);
6241                 if (ret) {
6242                         ret = btrfs_set_inode_index_count(dir);
6243                         if (ret)
6244                                 return ret;
6245                 }
6246         }
6247
6248         *index = dir->index_cnt;
6249         dir->index_cnt++;
6250
6251         return ret;
6252 }
6253
6254 static int btrfs_insert_inode_locked(struct inode *inode)
6255 {
6256         struct btrfs_iget_args args;
6257         args.location = &BTRFS_I(inode)->location;
6258         args.root = BTRFS_I(inode)->root;
6259
6260         return insert_inode_locked4(inode,
6261                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6262                    btrfs_find_actor, &args);
6263 }
6264
6265 /*
6266  * Inherit flags from the parent inode.
6267  *
6268  * Currently only the compression flags and the cow flags are inherited.
6269  */
6270 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6271 {
6272         unsigned int flags;
6273
6274         if (!dir)
6275                 return;
6276
6277         flags = BTRFS_I(dir)->flags;
6278
6279         if (flags & BTRFS_INODE_NOCOMPRESS) {
6280                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6281                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6282         } else if (flags & BTRFS_INODE_COMPRESS) {
6283                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6284                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6285         }
6286
6287         if (flags & BTRFS_INODE_NODATACOW) {
6288                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6289                 if (S_ISREG(inode->i_mode))
6290                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6291         }
6292
6293         btrfs_sync_inode_flags_to_i_flags(inode);
6294 }
6295
6296 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6297                                      struct btrfs_root *root,
6298                                      struct inode *dir,
6299                                      const char *name, int name_len,
6300                                      u64 ref_objectid, u64 objectid,
6301                                      umode_t mode, u64 *index)
6302 {
6303         struct btrfs_fs_info *fs_info = root->fs_info;
6304         struct inode *inode;
6305         struct btrfs_inode_item *inode_item;
6306         struct btrfs_key *location;
6307         struct btrfs_path *path;
6308         struct btrfs_inode_ref *ref;
6309         struct btrfs_key key[2];
6310         u32 sizes[2];
6311         int nitems = name ? 2 : 1;
6312         unsigned long ptr;
6313         unsigned int nofs_flag;
6314         int ret;
6315
6316         path = btrfs_alloc_path();
6317         if (!path)
6318                 return ERR_PTR(-ENOMEM);
6319
6320         nofs_flag = memalloc_nofs_save();
6321         inode = new_inode(fs_info->sb);
6322         memalloc_nofs_restore(nofs_flag);
6323         if (!inode) {
6324                 btrfs_free_path(path);
6325                 return ERR_PTR(-ENOMEM);
6326         }
6327
6328         /*
6329          * O_TMPFILE, set link count to 0, so that after this point,
6330          * we fill in an inode item with the correct link count.
6331          */
6332         if (!name)
6333                 set_nlink(inode, 0);
6334
6335         /*
6336          * we have to initialize this early, so we can reclaim the inode
6337          * number if we fail afterwards in this function.
6338          */
6339         inode->i_ino = objectid;
6340
6341         if (dir && name) {
6342                 trace_btrfs_inode_request(dir);
6343
6344                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6345                 if (ret) {
6346                         btrfs_free_path(path);
6347                         iput(inode);
6348                         return ERR_PTR(ret);
6349                 }
6350         } else if (dir) {
6351                 *index = 0;
6352         }
6353         /*
6354          * index_cnt is ignored for everything but a dir,
6355          * btrfs_set_inode_index_count has an explanation for the magic
6356          * number
6357          */
6358         BTRFS_I(inode)->index_cnt = 2;
6359         BTRFS_I(inode)->dir_index = *index;
6360         BTRFS_I(inode)->root = root;
6361         BTRFS_I(inode)->generation = trans->transid;
6362         inode->i_generation = BTRFS_I(inode)->generation;
6363
6364         /*
6365          * We could have gotten an inode number from somebody who was fsynced
6366          * and then removed in this same transaction, so let's just set full
6367          * sync since it will be a full sync anyway and this will blow away the
6368          * old info in the log.
6369          */
6370         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6371
6372         key[0].objectid = objectid;
6373         key[0].type = BTRFS_INODE_ITEM_KEY;
6374         key[0].offset = 0;
6375
6376         sizes[0] = sizeof(struct btrfs_inode_item);
6377
6378         if (name) {
6379                 /*
6380                  * Start new inodes with an inode_ref. This is slightly more
6381                  * efficient for small numbers of hard links since they will
6382                  * be packed into one item. Extended refs will kick in if we
6383                  * add more hard links than can fit in the ref item.
6384                  */
6385                 key[1].objectid = objectid;
6386                 key[1].type = BTRFS_INODE_REF_KEY;
6387                 key[1].offset = ref_objectid;
6388
6389                 sizes[1] = name_len + sizeof(*ref);
6390         }
6391
6392         location = &BTRFS_I(inode)->location;
6393         location->objectid = objectid;
6394         location->offset = 0;
6395         location->type = BTRFS_INODE_ITEM_KEY;
6396
6397         ret = btrfs_insert_inode_locked(inode);
6398         if (ret < 0) {
6399                 iput(inode);
6400                 goto fail;
6401         }
6402
6403         path->leave_spinning = 1;
6404         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6405         if (ret != 0)
6406                 goto fail_unlock;
6407
6408         inode_init_owner(inode, dir, mode);
6409         inode_set_bytes(inode, 0);
6410
6411         inode->i_mtime = current_time(inode);
6412         inode->i_atime = inode->i_mtime;
6413         inode->i_ctime = inode->i_mtime;
6414         BTRFS_I(inode)->i_otime = inode->i_mtime;
6415
6416         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6417                                   struct btrfs_inode_item);
6418         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6419                              sizeof(*inode_item));
6420         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6421
6422         if (name) {
6423                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6424                                      struct btrfs_inode_ref);
6425                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6426                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6427                 ptr = (unsigned long)(ref + 1);
6428                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6429         }
6430
6431         btrfs_mark_buffer_dirty(path->nodes[0]);
6432         btrfs_free_path(path);
6433
6434         btrfs_inherit_iflags(inode, dir);
6435
6436         if (S_ISREG(mode)) {
6437                 if (btrfs_test_opt(fs_info, NODATASUM))
6438                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6439                 if (btrfs_test_opt(fs_info, NODATACOW))
6440                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6441                                 BTRFS_INODE_NODATASUM;
6442         }
6443
6444         inode_tree_add(inode);
6445
6446         trace_btrfs_inode_new(inode);
6447         btrfs_set_inode_last_trans(trans, inode);
6448
6449         btrfs_update_root_times(trans, root);
6450
6451         ret = btrfs_inode_inherit_props(trans, inode, dir);
6452         if (ret)
6453                 btrfs_err(fs_info,
6454                           "error inheriting props for ino %llu (root %llu): %d",
6455                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6456
6457         return inode;
6458
6459 fail_unlock:
6460         discard_new_inode(inode);
6461 fail:
6462         if (dir && name)
6463                 BTRFS_I(dir)->index_cnt--;
6464         btrfs_free_path(path);
6465         return ERR_PTR(ret);
6466 }
6467
6468 /*
6469  * utility function to add 'inode' into 'parent_inode' with
6470  * a give name and a given sequence number.
6471  * if 'add_backref' is true, also insert a backref from the
6472  * inode to the parent directory.
6473  */
6474 int btrfs_add_link(struct btrfs_trans_handle *trans,
6475                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6476                    const char *name, int name_len, int add_backref, u64 index)
6477 {
6478         int ret = 0;
6479         struct btrfs_key key;
6480         struct btrfs_root *root = parent_inode->root;
6481         u64 ino = btrfs_ino(inode);
6482         u64 parent_ino = btrfs_ino(parent_inode);
6483
6484         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6485                 memcpy(&key, &inode->root->root_key, sizeof(key));
6486         } else {
6487                 key.objectid = ino;
6488                 key.type = BTRFS_INODE_ITEM_KEY;
6489                 key.offset = 0;
6490         }
6491
6492         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6493                 ret = btrfs_add_root_ref(trans, key.objectid,
6494                                          root->root_key.objectid, parent_ino,
6495                                          index, name, name_len);
6496         } else if (add_backref) {
6497                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6498                                              parent_ino, index);
6499         }
6500
6501         /* Nothing to clean up yet */
6502         if (ret)
6503                 return ret;
6504
6505         ret = btrfs_insert_dir_item(trans, name, name_len, parent_inode, &key,
6506                                     btrfs_inode_type(&inode->vfs_inode), index);
6507         if (ret == -EEXIST || ret == -EOVERFLOW)
6508                 goto fail_dir_item;
6509         else if (ret) {
6510                 btrfs_abort_transaction(trans, ret);
6511                 return ret;
6512         }
6513
6514         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6515                            name_len * 2);
6516         inode_inc_iversion(&parent_inode->vfs_inode);
6517         /*
6518          * If we are replaying a log tree, we do not want to update the mtime
6519          * and ctime of the parent directory with the current time, since the
6520          * log replay procedure is responsible for setting them to their correct
6521          * values (the ones it had when the fsync was done).
6522          */
6523         if (!test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags)) {
6524                 struct timespec64 now = current_time(&parent_inode->vfs_inode);
6525
6526                 parent_inode->vfs_inode.i_mtime = now;
6527                 parent_inode->vfs_inode.i_ctime = now;
6528         }
6529         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6530         if (ret)
6531                 btrfs_abort_transaction(trans, ret);
6532         return ret;
6533
6534 fail_dir_item:
6535         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6536                 u64 local_index;
6537                 int err;
6538                 err = btrfs_del_root_ref(trans, key.objectid,
6539                                          root->root_key.objectid, parent_ino,
6540                                          &local_index, name, name_len);
6541                 if (err)
6542                         btrfs_abort_transaction(trans, err);
6543         } else if (add_backref) {
6544                 u64 local_index;
6545                 int err;
6546
6547                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6548                                           ino, parent_ino, &local_index);
6549                 if (err)
6550                         btrfs_abort_transaction(trans, err);
6551         }
6552
6553         /* Return the original error code */
6554         return ret;
6555 }
6556
6557 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6558                             struct btrfs_inode *dir, struct dentry *dentry,
6559                             struct btrfs_inode *inode, int backref, u64 index)
6560 {
6561         int err = btrfs_add_link(trans, dir, inode,
6562                                  dentry->d_name.name, dentry->d_name.len,
6563                                  backref, index);
6564         if (err > 0)
6565                 err = -EEXIST;
6566         return err;
6567 }
6568
6569 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6570                         umode_t mode, dev_t rdev)
6571 {
6572         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6573         struct btrfs_trans_handle *trans;
6574         struct btrfs_root *root = BTRFS_I(dir)->root;
6575         struct inode *inode = NULL;
6576         int err;
6577         u64 objectid;
6578         u64 index = 0;
6579
6580         /*
6581          * 2 for inode item and ref
6582          * 2 for dir items
6583          * 1 for xattr if selinux is on
6584          */
6585         trans = btrfs_start_transaction(root, 5);
6586         if (IS_ERR(trans))
6587                 return PTR_ERR(trans);
6588
6589         err = btrfs_find_free_ino(root, &objectid);
6590         if (err)
6591                 goto out_unlock;
6592
6593         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6594                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6595                         mode, &index);
6596         if (IS_ERR(inode)) {
6597                 err = PTR_ERR(inode);
6598                 inode = NULL;
6599                 goto out_unlock;
6600         }
6601
6602         /*
6603         * If the active LSM wants to access the inode during
6604         * d_instantiate it needs these. Smack checks to see
6605         * if the filesystem supports xattrs by looking at the
6606         * ops vector.
6607         */
6608         inode->i_op = &btrfs_special_inode_operations;
6609         init_special_inode(inode, inode->i_mode, rdev);
6610
6611         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6612         if (err)
6613                 goto out_unlock;
6614
6615         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6616                         0, index);
6617         if (err)
6618                 goto out_unlock;
6619
6620         btrfs_update_inode(trans, root, inode);
6621         d_instantiate_new(dentry, inode);
6622
6623 out_unlock:
6624         btrfs_end_transaction(trans);
6625         btrfs_btree_balance_dirty(fs_info);
6626         if (err && inode) {
6627                 inode_dec_link_count(inode);
6628                 discard_new_inode(inode);
6629         }
6630         return err;
6631 }
6632
6633 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6634                         umode_t mode, bool excl)
6635 {
6636         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6637         struct btrfs_trans_handle *trans;
6638         struct btrfs_root *root = BTRFS_I(dir)->root;
6639         struct inode *inode = NULL;
6640         int err;
6641         u64 objectid;
6642         u64 index = 0;
6643
6644         /*
6645          * 2 for inode item and ref
6646          * 2 for dir items
6647          * 1 for xattr if selinux is on
6648          */
6649         trans = btrfs_start_transaction(root, 5);
6650         if (IS_ERR(trans))
6651                 return PTR_ERR(trans);
6652
6653         err = btrfs_find_free_ino(root, &objectid);
6654         if (err)
6655                 goto out_unlock;
6656
6657         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6658                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6659                         mode, &index);
6660         if (IS_ERR(inode)) {
6661                 err = PTR_ERR(inode);
6662                 inode = NULL;
6663                 goto out_unlock;
6664         }
6665         /*
6666         * If the active LSM wants to access the inode during
6667         * d_instantiate it needs these. Smack checks to see
6668         * if the filesystem supports xattrs by looking at the
6669         * ops vector.
6670         */
6671         inode->i_fop = &btrfs_file_operations;
6672         inode->i_op = &btrfs_file_inode_operations;
6673         inode->i_mapping->a_ops = &btrfs_aops;
6674
6675         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6676         if (err)
6677                 goto out_unlock;
6678
6679         err = btrfs_update_inode(trans, root, inode);
6680         if (err)
6681                 goto out_unlock;
6682
6683         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6684                         0, index);
6685         if (err)
6686                 goto out_unlock;
6687
6688         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6689         d_instantiate_new(dentry, inode);
6690
6691 out_unlock:
6692         btrfs_end_transaction(trans);
6693         if (err && inode) {
6694                 inode_dec_link_count(inode);
6695                 discard_new_inode(inode);
6696         }
6697         btrfs_btree_balance_dirty(fs_info);
6698         return err;
6699 }
6700
6701 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6702                       struct dentry *dentry)
6703 {
6704         struct btrfs_trans_handle *trans = NULL;
6705         struct btrfs_root *root = BTRFS_I(dir)->root;
6706         struct inode *inode = d_inode(old_dentry);
6707         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6708         u64 index;
6709         int err;
6710         int drop_inode = 0;
6711
6712         /* do not allow sys_link's with other subvols of the same device */
6713         if (root->root_key.objectid != BTRFS_I(inode)->root->root_key.objectid)
6714                 return -EXDEV;
6715
6716         if (inode->i_nlink >= BTRFS_LINK_MAX)
6717                 return -EMLINK;
6718
6719         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6720         if (err)
6721                 goto fail;
6722
6723         /*
6724          * 2 items for inode and inode ref
6725          * 2 items for dir items
6726          * 1 item for parent inode
6727          * 1 item for orphan item deletion if O_TMPFILE
6728          */
6729         trans = btrfs_start_transaction(root, inode->i_nlink ? 5 : 6);
6730         if (IS_ERR(trans)) {
6731                 err = PTR_ERR(trans);
6732                 trans = NULL;
6733                 goto fail;
6734         }
6735
6736         /* There are several dir indexes for this inode, clear the cache. */
6737         BTRFS_I(inode)->dir_index = 0ULL;
6738         inc_nlink(inode);
6739         inode_inc_iversion(inode);
6740         inode->i_ctime = current_time(inode);
6741         ihold(inode);
6742         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6743
6744         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6745                         1, index);
6746
6747         if (err) {
6748                 drop_inode = 1;
6749         } else {
6750                 struct dentry *parent = dentry->d_parent;
6751                 int ret;
6752
6753                 err = btrfs_update_inode(trans, root, inode);
6754                 if (err)
6755                         goto fail;
6756                 if (inode->i_nlink == 1) {
6757                         /*
6758                          * If new hard link count is 1, it's a file created
6759                          * with open(2) O_TMPFILE flag.
6760                          */
6761                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6762                         if (err)
6763                                 goto fail;
6764                 }
6765                 d_instantiate(dentry, inode);
6766                 ret = btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent,
6767                                          true, NULL);
6768                 if (ret == BTRFS_NEED_TRANS_COMMIT) {
6769                         err = btrfs_commit_transaction(trans);
6770                         trans = NULL;
6771                 }
6772         }
6773
6774 fail:
6775         if (trans)
6776                 btrfs_end_transaction(trans);
6777         if (drop_inode) {
6778                 inode_dec_link_count(inode);
6779                 iput(inode);
6780         }
6781         btrfs_btree_balance_dirty(fs_info);
6782         return err;
6783 }
6784
6785 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6786 {
6787         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6788         struct inode *inode = NULL;
6789         struct btrfs_trans_handle *trans;
6790         struct btrfs_root *root = BTRFS_I(dir)->root;
6791         int err = 0;
6792         u64 objectid = 0;
6793         u64 index = 0;
6794
6795         /*
6796          * 2 items for inode and ref
6797          * 2 items for dir items
6798          * 1 for xattr if selinux is on
6799          */
6800         trans = btrfs_start_transaction(root, 5);
6801         if (IS_ERR(trans))
6802                 return PTR_ERR(trans);
6803
6804         err = btrfs_find_free_ino(root, &objectid);
6805         if (err)
6806                 goto out_fail;
6807
6808         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6809                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6810                         S_IFDIR | mode, &index);
6811         if (IS_ERR(inode)) {
6812                 err = PTR_ERR(inode);
6813                 inode = NULL;
6814                 goto out_fail;
6815         }
6816
6817         /* these must be set before we unlock the inode */
6818         inode->i_op = &btrfs_dir_inode_operations;
6819         inode->i_fop = &btrfs_dir_file_operations;
6820
6821         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6822         if (err)
6823                 goto out_fail;
6824
6825         btrfs_i_size_write(BTRFS_I(inode), 0);
6826         err = btrfs_update_inode(trans, root, inode);
6827         if (err)
6828                 goto out_fail;
6829
6830         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6831                         dentry->d_name.name,
6832                         dentry->d_name.len, 0, index);
6833         if (err)
6834                 goto out_fail;
6835
6836         d_instantiate_new(dentry, inode);
6837
6838 out_fail:
6839         btrfs_end_transaction(trans);
6840         if (err && inode) {
6841                 inode_dec_link_count(inode);
6842                 discard_new_inode(inode);
6843         }
6844         btrfs_btree_balance_dirty(fs_info);
6845         return err;
6846 }
6847
6848 static noinline int uncompress_inline(struct btrfs_path *path,
6849                                       struct page *page,
6850                                       size_t pg_offset, u64 extent_offset,
6851                                       struct btrfs_file_extent_item *item)
6852 {
6853         int ret;
6854         struct extent_buffer *leaf = path->nodes[0];
6855         char *tmp;
6856         size_t max_size;
6857         unsigned long inline_size;
6858         unsigned long ptr;
6859         int compress_type;
6860
6861         WARN_ON(pg_offset != 0);
6862         compress_type = btrfs_file_extent_compression(leaf, item);
6863         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6864         inline_size = btrfs_file_extent_inline_item_len(leaf,
6865                                         btrfs_item_nr(path->slots[0]));
6866         tmp = kmalloc(inline_size, GFP_NOFS);
6867         if (!tmp)
6868                 return -ENOMEM;
6869         ptr = btrfs_file_extent_inline_start(item);
6870
6871         read_extent_buffer(leaf, tmp, ptr, inline_size);
6872
6873         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6874         ret = btrfs_decompress(compress_type, tmp, page,
6875                                extent_offset, inline_size, max_size);
6876
6877         /*
6878          * decompression code contains a memset to fill in any space between the end
6879          * of the uncompressed data and the end of max_size in case the decompressed
6880          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6881          * the end of an inline extent and the beginning of the next block, so we
6882          * cover that region here.
6883          */
6884
6885         if (max_size + pg_offset < PAGE_SIZE) {
6886                 char *map = kmap(page);
6887                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6888                 kunmap(page);
6889         }
6890         kfree(tmp);
6891         return ret;
6892 }
6893
6894 /*
6895  * a bit scary, this does extent mapping from logical file offset to the disk.
6896  * the ugly parts come from merging extents from the disk with the in-ram
6897  * representation.  This gets more complex because of the data=ordered code,
6898  * where the in-ram extents might be locked pending data=ordered completion.
6899  *
6900  * This also copies inline extents directly into the page.
6901  */
6902 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6903                                     struct page *page,
6904                                     size_t pg_offset, u64 start, u64 len,
6905                                     int create)
6906 {
6907         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6908         int ret;
6909         int err = 0;
6910         u64 extent_start = 0;
6911         u64 extent_end = 0;
6912         u64 objectid = btrfs_ino(inode);
6913         int extent_type = -1;
6914         struct btrfs_path *path = NULL;
6915         struct btrfs_root *root = inode->root;
6916         struct btrfs_file_extent_item *item;
6917         struct extent_buffer *leaf;
6918         struct btrfs_key found_key;
6919         struct extent_map *em = NULL;
6920         struct extent_map_tree *em_tree = &inode->extent_tree;
6921         struct extent_io_tree *io_tree = &inode->io_tree;
6922         const bool new_inline = !page || create;
6923
6924         read_lock(&em_tree->lock);
6925         em = lookup_extent_mapping(em_tree, start, len);
6926         if (em)
6927                 em->bdev = fs_info->fs_devices->latest_bdev;
6928         read_unlock(&em_tree->lock);
6929
6930         if (em) {
6931                 if (em->start > start || em->start + em->len <= start)
6932                         free_extent_map(em);
6933                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6934                         free_extent_map(em);
6935                 else
6936                         goto out;
6937         }
6938         em = alloc_extent_map();
6939         if (!em) {
6940                 err = -ENOMEM;
6941                 goto out;
6942         }
6943         em->bdev = fs_info->fs_devices->latest_bdev;
6944         em->start = EXTENT_MAP_HOLE;
6945         em->orig_start = EXTENT_MAP_HOLE;
6946         em->len = (u64)-1;
6947         em->block_len = (u64)-1;
6948
6949         path = btrfs_alloc_path();
6950         if (!path) {
6951                 err = -ENOMEM;
6952                 goto out;
6953         }
6954
6955         /* Chances are we'll be called again, so go ahead and do readahead */
6956         path->reada = READA_FORWARD;
6957
6958         /*
6959          * Unless we're going to uncompress the inline extent, no sleep would
6960          * happen.
6961          */
6962         path->leave_spinning = 1;
6963
6964         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6965         if (ret < 0) {
6966                 err = ret;
6967                 goto out;
6968         } else if (ret > 0) {
6969                 if (path->slots[0] == 0)
6970                         goto not_found;
6971                 path->slots[0]--;
6972         }
6973
6974         leaf = path->nodes[0];
6975         item = btrfs_item_ptr(leaf, path->slots[0],
6976                               struct btrfs_file_extent_item);
6977         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6978         if (found_key.objectid != objectid ||
6979             found_key.type != BTRFS_EXTENT_DATA_KEY) {
6980                 /*
6981                  * If we backup past the first extent we want to move forward
6982                  * and see if there is an extent in front of us, otherwise we'll
6983                  * say there is a hole for our whole search range which can
6984                  * cause problems.
6985                  */
6986                 extent_end = start;
6987                 goto next;
6988         }
6989
6990         extent_type = btrfs_file_extent_type(leaf, item);
6991         extent_start = found_key.offset;
6992         if (extent_type == BTRFS_FILE_EXTENT_REG ||
6993             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
6994                 /* Only regular file could have regular/prealloc extent */
6995                 if (!S_ISREG(inode->vfs_inode.i_mode)) {
6996                         ret = -EUCLEAN;
6997                         btrfs_crit(fs_info,
6998                 "regular/prealloc extent found for non-regular inode %llu",
6999                                    btrfs_ino(inode));
7000                         goto out;
7001                 }
7002                 extent_end = extent_start +
7003                        btrfs_file_extent_num_bytes(leaf, item);
7004
7005                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
7006                                                        extent_start);
7007         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
7008                 size_t size;
7009
7010                 size = btrfs_file_extent_ram_bytes(leaf, item);
7011                 extent_end = ALIGN(extent_start + size,
7012                                    fs_info->sectorsize);
7013
7014                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
7015                                                       path->slots[0],
7016                                                       extent_start);
7017         }
7018 next:
7019         if (start >= extent_end) {
7020                 path->slots[0]++;
7021                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
7022                         ret = btrfs_next_leaf(root, path);
7023                         if (ret < 0) {
7024                                 err = ret;
7025                                 goto out;
7026                         } else if (ret > 0) {
7027                                 goto not_found;
7028                         }
7029                         leaf = path->nodes[0];
7030                 }
7031                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7032                 if (found_key.objectid != objectid ||
7033                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7034                         goto not_found;
7035                 if (start + len <= found_key.offset)
7036                         goto not_found;
7037                 if (start > found_key.offset)
7038                         goto next;
7039
7040                 /* New extent overlaps with existing one */
7041                 em->start = start;
7042                 em->orig_start = start;
7043                 em->len = found_key.offset - start;
7044                 em->block_start = EXTENT_MAP_HOLE;
7045                 goto insert;
7046         }
7047
7048         btrfs_extent_item_to_extent_map(inode, path, item,
7049                         new_inline, em);
7050
7051         if (extent_type == BTRFS_FILE_EXTENT_REG ||
7052             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
7053                 goto insert;
7054         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
7055                 unsigned long ptr;
7056                 char *map;
7057                 size_t size;
7058                 size_t extent_offset;
7059                 size_t copy_size;
7060
7061                 if (new_inline)
7062                         goto out;
7063
7064                 size = btrfs_file_extent_ram_bytes(leaf, item);
7065                 extent_offset = page_offset(page) + pg_offset - extent_start;
7066                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7067                                   size - extent_offset);
7068                 em->start = extent_start + extent_offset;
7069                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7070                 em->orig_block_len = em->len;
7071                 em->orig_start = em->start;
7072                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7073
7074                 btrfs_set_path_blocking(path);
7075                 if (!PageUptodate(page)) {
7076                         if (btrfs_file_extent_compression(leaf, item) !=
7077                             BTRFS_COMPRESS_NONE) {
7078                                 ret = uncompress_inline(path, page, pg_offset,
7079                                                         extent_offset, item);
7080                                 if (ret) {
7081                                         err = ret;
7082                                         goto out;
7083                                 }
7084                         } else {
7085                                 map = kmap(page);
7086                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7087                                                    copy_size);
7088                                 if (pg_offset + copy_size < PAGE_SIZE) {
7089                                         memset(map + pg_offset + copy_size, 0,
7090                                                PAGE_SIZE - pg_offset -
7091                                                copy_size);
7092                                 }
7093                                 kunmap(page);
7094                         }
7095                         flush_dcache_page(page);
7096                 }
7097                 set_extent_uptodate(io_tree, em->start,
7098                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7099                 goto insert;
7100         }
7101 not_found:
7102         em->start = start;
7103         em->orig_start = start;
7104         em->len = len;
7105         em->block_start = EXTENT_MAP_HOLE;
7106 insert:
7107         btrfs_release_path(path);
7108         if (em->start > start || extent_map_end(em) <= start) {
7109                 btrfs_err(fs_info,
7110                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7111                           em->start, em->len, start, len);
7112                 err = -EIO;
7113                 goto out;
7114         }
7115
7116         err = 0;
7117         write_lock(&em_tree->lock);
7118         err = btrfs_add_extent_mapping(fs_info, em_tree, &em, start, len);
7119         write_unlock(&em_tree->lock);
7120 out:
7121         btrfs_free_path(path);
7122
7123         trace_btrfs_get_extent(root, inode, em);
7124
7125         if (err) {
7126                 free_extent_map(em);
7127                 return ERR_PTR(err);
7128         }
7129         BUG_ON(!em); /* Error is always set */
7130         return em;
7131 }
7132
7133 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7134                                            u64 start, u64 len)
7135 {
7136         struct extent_map *em;
7137         struct extent_map *hole_em = NULL;
7138         u64 delalloc_start = start;
7139         u64 end;
7140         u64 delalloc_len;
7141         u64 delalloc_end;
7142         int err = 0;
7143
7144         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7145         if (IS_ERR(em))
7146                 return em;
7147         /*
7148          * If our em maps to:
7149          * - a hole or
7150          * - a pre-alloc extent,
7151          * there might actually be delalloc bytes behind it.
7152          */
7153         if (em->block_start != EXTENT_MAP_HOLE &&
7154             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7155                 return em;
7156         else
7157                 hole_em = em;
7158
7159         /* check to see if we've wrapped (len == -1 or similar) */
7160         end = start + len;
7161         if (end < start)
7162                 end = (u64)-1;
7163         else
7164                 end -= 1;
7165
7166         em = NULL;
7167
7168         /* ok, we didn't find anything, lets look for delalloc */
7169         delalloc_len = count_range_bits(&inode->io_tree, &delalloc_start,
7170                                  end, len, EXTENT_DELALLOC, 1);
7171         delalloc_end = delalloc_start + delalloc_len;
7172         if (delalloc_end < delalloc_start)
7173                 delalloc_end = (u64)-1;
7174
7175         /*
7176          * We didn't find anything useful, return the original results from
7177          * get_extent()
7178          */
7179         if (delalloc_start > end || delalloc_end <= start) {
7180                 em = hole_em;
7181                 hole_em = NULL;
7182                 goto out;
7183         }
7184
7185         /*
7186          * Adjust the delalloc_start to make sure it doesn't go backwards from
7187          * the start they passed in
7188          */
7189         delalloc_start = max(start, delalloc_start);
7190         delalloc_len = delalloc_end - delalloc_start;
7191
7192         if (delalloc_len > 0) {
7193                 u64 hole_start;
7194                 u64 hole_len;
7195                 const u64 hole_end = extent_map_end(hole_em);
7196
7197                 em = alloc_extent_map();
7198                 if (!em) {
7199                         err = -ENOMEM;
7200                         goto out;
7201                 }
7202                 em->bdev = NULL;
7203
7204                 ASSERT(hole_em);
7205                 /*
7206                  * When btrfs_get_extent can't find anything it returns one
7207                  * huge hole
7208                  *
7209                  * Make sure what it found really fits our range, and adjust to
7210                  * make sure it is based on the start from the caller
7211                  */
7212                 if (hole_end <= start || hole_em->start > end) {
7213                        free_extent_map(hole_em);
7214                        hole_em = NULL;
7215                 } else {
7216                        hole_start = max(hole_em->start, start);
7217                        hole_len = hole_end - hole_start;
7218                 }
7219
7220                 if (hole_em && delalloc_start > hole_start) {
7221                         /*
7222                          * Our hole starts before our delalloc, so we have to
7223                          * return just the parts of the hole that go until the
7224                          * delalloc starts
7225                          */
7226                         em->len = min(hole_len, delalloc_start - hole_start);
7227                         em->start = hole_start;
7228                         em->orig_start = hole_start;
7229                         /*
7230                          * Don't adjust block start at all, it is fixed at
7231                          * EXTENT_MAP_HOLE
7232                          */
7233                         em->block_start = hole_em->block_start;
7234                         em->block_len = hole_len;
7235                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7236                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7237                 } else {
7238                         /*
7239                          * Hole is out of passed range or it starts after
7240                          * delalloc range
7241                          */
7242                         em->start = delalloc_start;
7243                         em->len = delalloc_len;
7244                         em->orig_start = delalloc_start;
7245                         em->block_start = EXTENT_MAP_DELALLOC;
7246                         em->block_len = delalloc_len;
7247                 }
7248         } else {
7249                 return hole_em;
7250         }
7251 out:
7252
7253         free_extent_map(hole_em);
7254         if (err) {
7255                 free_extent_map(em);
7256                 return ERR_PTR(err);
7257         }
7258         return em;
7259 }
7260
7261 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7262                                                   const u64 start,
7263                                                   const u64 len,
7264                                                   const u64 orig_start,
7265                                                   const u64 block_start,
7266                                                   const u64 block_len,
7267                                                   const u64 orig_block_len,
7268                                                   const u64 ram_bytes,
7269                                                   const int type)
7270 {
7271         struct extent_map *em = NULL;
7272         int ret;
7273
7274         if (type != BTRFS_ORDERED_NOCOW) {
7275                 em = create_io_em(inode, start, len, orig_start,
7276                                   block_start, block_len, orig_block_len,
7277                                   ram_bytes,
7278                                   BTRFS_COMPRESS_NONE, /* compress_type */
7279                                   type);
7280                 if (IS_ERR(em))
7281                         goto out;
7282         }
7283         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7284                                            len, block_len, type);
7285         if (ret) {
7286                 if (em) {
7287                         free_extent_map(em);
7288                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7289                                                 start + len - 1, 0);
7290                 }
7291                 em = ERR_PTR(ret);
7292         }
7293  out:
7294
7295         return em;
7296 }
7297
7298 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7299                                                   u64 start, u64 len)
7300 {
7301         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7302         struct btrfs_root *root = BTRFS_I(inode)->root;
7303         struct extent_map *em;
7304         struct btrfs_key ins;
7305         u64 alloc_hint;
7306         int ret;
7307
7308         alloc_hint = get_extent_allocation_hint(inode, start, len);
7309         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7310                                    0, alloc_hint, &ins, 1, 1);
7311         if (ret)
7312                 return ERR_PTR(ret);
7313
7314         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7315                                      ins.objectid, ins.offset, ins.offset,
7316                                      ins.offset, BTRFS_ORDERED_REGULAR);
7317         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7318         if (IS_ERR(em))
7319                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7320                                            ins.offset, 1);
7321
7322         return em;
7323 }
7324
7325 /*
7326  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7327  * block must be cow'd
7328  */
7329 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7330                               u64 *orig_start, u64 *orig_block_len,
7331                               u64 *ram_bytes)
7332 {
7333         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7334         struct btrfs_path *path;
7335         int ret;
7336         struct extent_buffer *leaf;
7337         struct btrfs_root *root = BTRFS_I(inode)->root;
7338         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7339         struct btrfs_file_extent_item *fi;
7340         struct btrfs_key key;
7341         u64 disk_bytenr;
7342         u64 backref_offset;
7343         u64 extent_end;
7344         u64 num_bytes;
7345         int slot;
7346         int found_type;
7347         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7348
7349         path = btrfs_alloc_path();
7350         if (!path)
7351                 return -ENOMEM;
7352
7353         ret = btrfs_lookup_file_extent(NULL, root, path,
7354                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7355         if (ret < 0)
7356                 goto out;
7357
7358         slot = path->slots[0];
7359         if (ret == 1) {
7360                 if (slot == 0) {
7361                         /* can't find the item, must cow */
7362                         ret = 0;
7363                         goto out;
7364                 }
7365                 slot--;
7366         }
7367         ret = 0;
7368         leaf = path->nodes[0];
7369         btrfs_item_key_to_cpu(leaf, &key, slot);
7370         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7371             key.type != BTRFS_EXTENT_DATA_KEY) {
7372                 /* not our file or wrong item type, must cow */
7373                 goto out;
7374         }
7375
7376         if (key.offset > offset) {
7377                 /* Wrong offset, must cow */
7378                 goto out;
7379         }
7380
7381         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7382         found_type = btrfs_file_extent_type(leaf, fi);
7383         if (found_type != BTRFS_FILE_EXTENT_REG &&
7384             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7385                 /* not a regular extent, must cow */
7386                 goto out;
7387         }
7388
7389         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7390                 goto out;
7391
7392         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7393         if (extent_end <= offset)
7394                 goto out;
7395
7396         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7397         if (disk_bytenr == 0)
7398                 goto out;
7399
7400         if (btrfs_file_extent_compression(leaf, fi) ||
7401             btrfs_file_extent_encryption(leaf, fi) ||
7402             btrfs_file_extent_other_encoding(leaf, fi))
7403                 goto out;
7404
7405         /*
7406          * Do the same check as in btrfs_cross_ref_exist but without the
7407          * unnecessary search.
7408          */
7409         if (btrfs_file_extent_generation(leaf, fi) <=
7410             btrfs_root_last_snapshot(&root->root_item))
7411                 goto out;
7412
7413         backref_offset = btrfs_file_extent_offset(leaf, fi);
7414
7415         if (orig_start) {
7416                 *orig_start = key.offset - backref_offset;
7417                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7418                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7419         }
7420
7421         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7422                 goto out;
7423
7424         num_bytes = min(offset + *len, extent_end) - offset;
7425         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7426                 u64 range_end;
7427
7428                 range_end = round_up(offset + num_bytes,
7429                                      root->fs_info->sectorsize) - 1;
7430                 ret = test_range_bit(io_tree, offset, range_end,
7431                                      EXTENT_DELALLOC, 0, NULL);
7432                 if (ret) {
7433                         ret = -EAGAIN;
7434                         goto out;
7435                 }
7436         }
7437
7438         btrfs_release_path(path);
7439
7440         /*
7441          * look for other files referencing this extent, if we
7442          * find any we must cow
7443          */
7444
7445         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7446                                     key.offset - backref_offset, disk_bytenr);
7447         if (ret) {
7448                 ret = 0;
7449                 goto out;
7450         }
7451
7452         /*
7453          * adjust disk_bytenr and num_bytes to cover just the bytes
7454          * in this extent we are about to write.  If there
7455          * are any csums in that range we have to cow in order
7456          * to keep the csums correct
7457          */
7458         disk_bytenr += backref_offset;
7459         disk_bytenr += offset - key.offset;
7460         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7461                 goto out;
7462         /*
7463          * all of the above have passed, it is safe to overwrite this extent
7464          * without cow
7465          */
7466         *len = num_bytes;
7467         ret = 1;
7468 out:
7469         btrfs_free_path(path);
7470         return ret;
7471 }
7472
7473 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7474                               struct extent_state **cached_state, int writing)
7475 {
7476         struct btrfs_ordered_extent *ordered;
7477         int ret = 0;
7478
7479         while (1) {
7480                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7481                                  cached_state);
7482                 /*
7483                  * We're concerned with the entire range that we're going to be
7484                  * doing DIO to, so we need to make sure there's no ordered
7485                  * extents in this range.
7486                  */
7487                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7488                                                      lockend - lockstart + 1);
7489
7490                 /*
7491                  * We need to make sure there are no buffered pages in this
7492                  * range either, we could have raced between the invalidate in
7493                  * generic_file_direct_write and locking the extent.  The
7494                  * invalidate needs to happen so that reads after a write do not
7495                  * get stale data.
7496                  */
7497                 if (!ordered &&
7498                     (!writing || !filemap_range_has_page(inode->i_mapping,
7499                                                          lockstart, lockend)))
7500                         break;
7501
7502                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7503                                      cached_state);
7504
7505                 if (ordered) {
7506                         /*
7507                          * If we are doing a DIO read and the ordered extent we
7508                          * found is for a buffered write, we can not wait for it
7509                          * to complete and retry, because if we do so we can
7510                          * deadlock with concurrent buffered writes on page
7511                          * locks. This happens only if our DIO read covers more
7512                          * than one extent map, if at this point has already
7513                          * created an ordered extent for a previous extent map
7514                          * and locked its range in the inode's io tree, and a
7515                          * concurrent write against that previous extent map's
7516                          * range and this range started (we unlock the ranges
7517                          * in the io tree only when the bios complete and
7518                          * buffered writes always lock pages before attempting
7519                          * to lock range in the io tree).
7520                          */
7521                         if (writing ||
7522                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7523                                 btrfs_start_ordered_extent(inode, ordered, 1);
7524                         else
7525                                 ret = -ENOTBLK;
7526                         btrfs_put_ordered_extent(ordered);
7527                 } else {
7528                         /*
7529                          * We could trigger writeback for this range (and wait
7530                          * for it to complete) and then invalidate the pages for
7531                          * this range (through invalidate_inode_pages2_range()),
7532                          * but that can lead us to a deadlock with a concurrent
7533                          * call to readpages() (a buffered read or a defrag call
7534                          * triggered a readahead) on a page lock due to an
7535                          * ordered dio extent we created before but did not have
7536                          * yet a corresponding bio submitted (whence it can not
7537                          * complete), which makes readpages() wait for that
7538                          * ordered extent to complete while holding a lock on
7539                          * that page.
7540                          */
7541                         ret = -ENOTBLK;
7542                 }
7543
7544                 if (ret)
7545                         break;
7546
7547                 cond_resched();
7548         }
7549
7550         return ret;
7551 }
7552
7553 /* The callers of this must take lock_extent() */
7554 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7555                                        u64 orig_start, u64 block_start,
7556                                        u64 block_len, u64 orig_block_len,
7557                                        u64 ram_bytes, int compress_type,
7558                                        int type)
7559 {
7560         struct extent_map_tree *em_tree;
7561         struct extent_map *em;
7562         struct btrfs_root *root = BTRFS_I(inode)->root;
7563         int ret;
7564
7565         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7566                type == BTRFS_ORDERED_COMPRESSED ||
7567                type == BTRFS_ORDERED_NOCOW ||
7568                type == BTRFS_ORDERED_REGULAR);
7569
7570         em_tree = &BTRFS_I(inode)->extent_tree;
7571         em = alloc_extent_map();
7572         if (!em)
7573                 return ERR_PTR(-ENOMEM);
7574
7575         em->start = start;
7576         em->orig_start = orig_start;
7577         em->len = len;
7578         em->block_len = block_len;
7579         em->block_start = block_start;
7580         em->bdev = root->fs_info->fs_devices->latest_bdev;
7581         em->orig_block_len = orig_block_len;
7582         em->ram_bytes = ram_bytes;
7583         em->generation = -1;
7584         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7585         if (type == BTRFS_ORDERED_PREALLOC) {
7586                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7587         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7588                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7589                 em->compress_type = compress_type;
7590         }
7591
7592         do {
7593                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7594                                 em->start + em->len - 1, 0);
7595                 write_lock(&em_tree->lock);
7596                 ret = add_extent_mapping(em_tree, em, 1);
7597                 write_unlock(&em_tree->lock);
7598                 /*
7599                  * The caller has taken lock_extent(), who could race with us
7600                  * to add em?
7601                  */
7602         } while (ret == -EEXIST);
7603
7604         if (ret) {
7605                 free_extent_map(em);
7606                 return ERR_PTR(ret);
7607         }
7608
7609         /* em got 2 refs now, callers needs to do free_extent_map once. */
7610         return em;
7611 }
7612
7613
7614 static int btrfs_get_blocks_direct_read(struct extent_map *em,
7615                                         struct buffer_head *bh_result,
7616                                         struct inode *inode,
7617                                         u64 start, u64 len)
7618 {
7619         if (em->block_start == EXTENT_MAP_HOLE ||
7620                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7621                 return -ENOENT;
7622
7623         len = min(len, em->len - (start - em->start));
7624
7625         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7626                 inode->i_blkbits;
7627         bh_result->b_size = len;
7628         bh_result->b_bdev = em->bdev;
7629         set_buffer_mapped(bh_result);
7630
7631         return 0;
7632 }
7633
7634 static int btrfs_get_blocks_direct_write(struct extent_map **map,
7635                                          struct buffer_head *bh_result,
7636                                          struct inode *inode,
7637                                          struct btrfs_dio_data *dio_data,
7638                                          u64 start, u64 len)
7639 {
7640         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7641         struct extent_map *em = *map;
7642         int ret = 0;
7643
7644         /*
7645          * We don't allocate a new extent in the following cases
7646          *
7647          * 1) The inode is marked as NODATACOW. In this case we'll just use the
7648          * existing extent.
7649          * 2) The extent is marked as PREALLOC. We're good to go here and can
7650          * just use the extent.
7651          *
7652          */
7653         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7654             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7655              em->block_start != EXTENT_MAP_HOLE)) {
7656                 int type;
7657                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7658
7659                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7660                         type = BTRFS_ORDERED_PREALLOC;
7661                 else
7662                         type = BTRFS_ORDERED_NOCOW;
7663                 len = min(len, em->len - (start - em->start));
7664                 block_start = em->block_start + (start - em->start);
7665
7666                 if (can_nocow_extent(inode, start, &len, &orig_start,
7667                                      &orig_block_len, &ram_bytes) == 1 &&
7668                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7669                         struct extent_map *em2;
7670
7671                         em2 = btrfs_create_dio_extent(inode, start, len,
7672                                                       orig_start, block_start,
7673                                                       len, orig_block_len,
7674                                                       ram_bytes, type);
7675                         btrfs_dec_nocow_writers(fs_info, block_start);
7676                         if (type == BTRFS_ORDERED_PREALLOC) {
7677                                 free_extent_map(em);
7678                                 *map = em = em2;
7679                         }
7680
7681                         if (em2 && IS_ERR(em2)) {
7682                                 ret = PTR_ERR(em2);
7683                                 goto out;
7684                         }
7685                         /*
7686                          * For inode marked NODATACOW or extent marked PREALLOC,
7687                          * use the existing or preallocated extent, so does not
7688                          * need to adjust btrfs_space_info's bytes_may_use.
7689                          */
7690                         btrfs_free_reserved_data_space_noquota(inode, start,
7691                                                                len);
7692                         goto skip_cow;
7693                 }
7694         }
7695
7696         /* this will cow the extent */
7697         len = bh_result->b_size;
7698         free_extent_map(em);
7699         *map = em = btrfs_new_extent_direct(inode, start, len);
7700         if (IS_ERR(em)) {
7701                 ret = PTR_ERR(em);
7702                 goto out;
7703         }
7704
7705         len = min(len, em->len - (start - em->start));
7706
7707 skip_cow:
7708         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7709                 inode->i_blkbits;
7710         bh_result->b_size = len;
7711         bh_result->b_bdev = em->bdev;
7712         set_buffer_mapped(bh_result);
7713
7714         if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7715                 set_buffer_new(bh_result);
7716
7717         /*
7718          * Need to update the i_size under the extent lock so buffered
7719          * readers will get the updated i_size when we unlock.
7720          */
7721         if (!dio_data->overwrite && start + len > i_size_read(inode))
7722                 i_size_write(inode, start + len);
7723
7724         WARN_ON(dio_data->reserve < len);
7725         dio_data->reserve -= len;
7726         dio_data->unsubmitted_oe_range_end = start + len;
7727         current->journal_info = dio_data;
7728 out:
7729         return ret;
7730 }
7731
7732 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7733                                    struct buffer_head *bh_result, int create)
7734 {
7735         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7736         struct extent_map *em;
7737         struct extent_state *cached_state = NULL;
7738         struct btrfs_dio_data *dio_data = NULL;
7739         u64 start = iblock << inode->i_blkbits;
7740         u64 lockstart, lockend;
7741         u64 len = bh_result->b_size;
7742         int ret = 0;
7743
7744         if (!create)
7745                 len = min_t(u64, len, fs_info->sectorsize);
7746
7747         lockstart = start;
7748         lockend = start + len - 1;
7749
7750         if (current->journal_info) {
7751                 /*
7752                  * Need to pull our outstanding extents and set journal_info to NULL so
7753                  * that anything that needs to check if there's a transaction doesn't get
7754                  * confused.
7755                  */
7756                 dio_data = current->journal_info;
7757                 current->journal_info = NULL;
7758         }
7759
7760         /*
7761          * If this errors out it's because we couldn't invalidate pagecache for
7762          * this range and we need to fallback to buffered.
7763          */
7764         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7765                                create)) {
7766                 ret = -ENOTBLK;
7767                 goto err;
7768         }
7769
7770         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7771         if (IS_ERR(em)) {
7772                 ret = PTR_ERR(em);
7773                 goto unlock_err;
7774         }
7775
7776         /*
7777          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7778          * io.  INLINE is special, and we could probably kludge it in here, but
7779          * it's still buffered so for safety lets just fall back to the generic
7780          * buffered path.
7781          *
7782          * For COMPRESSED we _have_ to read the entire extent in so we can
7783          * decompress it, so there will be buffering required no matter what we
7784          * do, so go ahead and fallback to buffered.
7785          *
7786          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7787          * to buffered IO.  Don't blame me, this is the price we pay for using
7788          * the generic code.
7789          */
7790         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7791             em->block_start == EXTENT_MAP_INLINE) {
7792                 free_extent_map(em);
7793                 ret = -ENOTBLK;
7794                 goto unlock_err;
7795         }
7796
7797         if (create) {
7798                 ret = btrfs_get_blocks_direct_write(&em, bh_result, inode,
7799                                                     dio_data, start, len);
7800                 if (ret < 0)
7801                         goto unlock_err;
7802
7803                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
7804                                      lockend, &cached_state);
7805         } else {
7806                 ret = btrfs_get_blocks_direct_read(em, bh_result, inode,
7807                                                    start, len);
7808                 /* Can be negative only if we read from a hole */
7809                 if (ret < 0) {
7810                         ret = 0;
7811                         free_extent_map(em);
7812                         goto unlock_err;
7813                 }
7814                 /*
7815                  * We need to unlock only the end area that we aren't using.
7816                  * The rest is going to be unlocked by the endio routine.
7817                  */
7818                 lockstart = start + bh_result->b_size;
7819                 if (lockstart < lockend) {
7820                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
7821                                              lockstart, lockend, &cached_state);
7822                 } else {
7823                         free_extent_state(cached_state);
7824                 }
7825         }
7826
7827         free_extent_map(em);
7828
7829         return 0;
7830
7831 unlock_err:
7832         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7833                              &cached_state);
7834 err:
7835         if (dio_data)
7836                 current->journal_info = dio_data;
7837         return ret;
7838 }
7839
7840 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7841                                                  struct bio *bio,
7842                                                  int mirror_num)
7843 {
7844         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7845         blk_status_t ret;
7846
7847         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7848
7849         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7850         if (ret)
7851                 return ret;
7852
7853         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7854
7855         return ret;
7856 }
7857
7858 static int btrfs_check_dio_repairable(struct inode *inode,
7859                                       struct bio *failed_bio,
7860                                       struct io_failure_record *failrec,
7861                                       int failed_mirror)
7862 {
7863         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7864         int num_copies;
7865
7866         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7867         if (num_copies == 1) {
7868                 /*
7869                  * we only have a single copy of the data, so don't bother with
7870                  * all the retry and error correction code that follows. no
7871                  * matter what the error is, it is very likely to persist.
7872                  */
7873                 btrfs_debug(fs_info,
7874                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7875                         num_copies, failrec->this_mirror, failed_mirror);
7876                 return 0;
7877         }
7878
7879         failrec->failed_mirror = failed_mirror;
7880         failrec->this_mirror++;
7881         if (failrec->this_mirror == failed_mirror)
7882                 failrec->this_mirror++;
7883
7884         if (failrec->this_mirror > num_copies) {
7885                 btrfs_debug(fs_info,
7886                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7887                         num_copies, failrec->this_mirror, failed_mirror);
7888                 return 0;
7889         }
7890
7891         return 1;
7892 }
7893
7894 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7895                                    struct page *page, unsigned int pgoff,
7896                                    u64 start, u64 end, int failed_mirror,
7897                                    bio_end_io_t *repair_endio, void *repair_arg)
7898 {
7899         struct io_failure_record *failrec;
7900         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7901         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7902         struct bio *bio;
7903         int isector;
7904         unsigned int read_mode = 0;
7905         int segs;
7906         int ret;
7907         blk_status_t status;
7908         struct bio_vec bvec;
7909
7910         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7911
7912         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7913         if (ret)
7914                 return errno_to_blk_status(ret);
7915
7916         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7917                                          failed_mirror);
7918         if (!ret) {
7919                 free_io_failure(failure_tree, io_tree, failrec);
7920                 return BLK_STS_IOERR;
7921         }
7922
7923         segs = bio_segments(failed_bio);
7924         bio_get_first_bvec(failed_bio, &bvec);
7925         if (segs > 1 ||
7926             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7927                 read_mode |= REQ_FAILFAST_DEV;
7928
7929         isector = start - btrfs_io_bio(failed_bio)->logical;
7930         isector >>= inode->i_sb->s_blocksize_bits;
7931         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7932                                 pgoff, isector, repair_endio, repair_arg);
7933         bio->bi_opf = REQ_OP_READ | read_mode;
7934
7935         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7936                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7937                     read_mode, failrec->this_mirror, failrec->in_validation);
7938
7939         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7940         if (status) {
7941                 free_io_failure(failure_tree, io_tree, failrec);
7942                 bio_put(bio);
7943         }
7944
7945         return status;
7946 }
7947
7948 struct btrfs_retry_complete {
7949         struct completion done;
7950         struct inode *inode;
7951         u64 start;
7952         int uptodate;
7953 };
7954
7955 static void btrfs_retry_endio_nocsum(struct bio *bio)
7956 {
7957         struct btrfs_retry_complete *done = bio->bi_private;
7958         struct inode *inode = done->inode;
7959         struct bio_vec *bvec;
7960         struct extent_io_tree *io_tree, *failure_tree;
7961         struct bvec_iter_all iter_all;
7962
7963         if (bio->bi_status)
7964                 goto end;
7965
7966         ASSERT(bio->bi_vcnt == 1);
7967         io_tree = &BTRFS_I(inode)->io_tree;
7968         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7969         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
7970
7971         done->uptodate = 1;
7972         ASSERT(!bio_flagged(bio, BIO_CLONED));
7973         bio_for_each_segment_all(bvec, bio, iter_all)
7974                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
7975                                  io_tree, done->start, bvec->bv_page,
7976                                  btrfs_ino(BTRFS_I(inode)), 0);
7977 end:
7978         complete(&done->done);
7979         bio_put(bio);
7980 }
7981
7982 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
7983                                                 struct btrfs_io_bio *io_bio)
7984 {
7985         struct btrfs_fs_info *fs_info;
7986         struct bio_vec bvec;
7987         struct bvec_iter iter;
7988         struct btrfs_retry_complete done;
7989         u64 start;
7990         unsigned int pgoff;
7991         u32 sectorsize;
7992         int nr_sectors;
7993         blk_status_t ret;
7994         blk_status_t err = BLK_STS_OK;
7995
7996         fs_info = BTRFS_I(inode)->root->fs_info;
7997         sectorsize = fs_info->sectorsize;
7998
7999         start = io_bio->logical;
8000         done.inode = inode;
8001         io_bio->bio.bi_iter = io_bio->iter;
8002
8003         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8004                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8005                 pgoff = bvec.bv_offset;
8006
8007 next_block_or_try_again:
8008                 done.uptodate = 0;
8009                 done.start = start;
8010                 init_completion(&done.done);
8011
8012                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8013                                 pgoff, start, start + sectorsize - 1,
8014                                 io_bio->mirror_num,
8015                                 btrfs_retry_endio_nocsum, &done);
8016                 if (ret) {
8017                         err = ret;
8018                         goto next;
8019                 }
8020
8021                 wait_for_completion_io(&done.done);
8022
8023                 if (!done.uptodate) {
8024                         /* We might have another mirror, so try again */
8025                         goto next_block_or_try_again;
8026                 }
8027
8028 next:
8029                 start += sectorsize;
8030
8031                 nr_sectors--;
8032                 if (nr_sectors) {
8033                         pgoff += sectorsize;
8034                         ASSERT(pgoff < PAGE_SIZE);
8035                         goto next_block_or_try_again;
8036                 }
8037         }
8038
8039         return err;
8040 }
8041
8042 static void btrfs_retry_endio(struct bio *bio)
8043 {
8044         struct btrfs_retry_complete *done = bio->bi_private;
8045         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8046         struct extent_io_tree *io_tree, *failure_tree;
8047         struct inode *inode = done->inode;
8048         struct bio_vec *bvec;
8049         int uptodate;
8050         int ret;
8051         int i = 0;
8052         struct bvec_iter_all iter_all;
8053
8054         if (bio->bi_status)
8055                 goto end;
8056
8057         uptodate = 1;
8058
8059         ASSERT(bio->bi_vcnt == 1);
8060         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
8061
8062         io_tree = &BTRFS_I(inode)->io_tree;
8063         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8064
8065         ASSERT(!bio_flagged(bio, BIO_CLONED));
8066         bio_for_each_segment_all(bvec, bio, iter_all) {
8067                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8068                                              bvec->bv_offset, done->start,
8069                                              bvec->bv_len);
8070                 if (!ret)
8071                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8072                                          failure_tree, io_tree, done->start,
8073                                          bvec->bv_page,
8074                                          btrfs_ino(BTRFS_I(inode)),
8075                                          bvec->bv_offset);
8076                 else
8077                         uptodate = 0;
8078                 i++;
8079         }
8080
8081         done->uptodate = uptodate;
8082 end:
8083         complete(&done->done);
8084         bio_put(bio);
8085 }
8086
8087 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8088                 struct btrfs_io_bio *io_bio, blk_status_t err)
8089 {
8090         struct btrfs_fs_info *fs_info;
8091         struct bio_vec bvec;
8092         struct bvec_iter iter;
8093         struct btrfs_retry_complete done;
8094         u64 start;
8095         u64 offset = 0;
8096         u32 sectorsize;
8097         int nr_sectors;
8098         unsigned int pgoff;
8099         int csum_pos;
8100         bool uptodate = (err == 0);
8101         int ret;
8102         blk_status_t status;
8103
8104         fs_info = BTRFS_I(inode)->root->fs_info;
8105         sectorsize = fs_info->sectorsize;
8106
8107         err = BLK_STS_OK;
8108         start = io_bio->logical;
8109         done.inode = inode;
8110         io_bio->bio.bi_iter = io_bio->iter;
8111
8112         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8113                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8114
8115                 pgoff = bvec.bv_offset;
8116 next_block:
8117                 if (uptodate) {
8118                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8119                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8120                                         bvec.bv_page, pgoff, start, sectorsize);
8121                         if (likely(!ret))
8122                                 goto next;
8123                 }
8124 try_again:
8125                 done.uptodate = 0;
8126                 done.start = start;
8127                 init_completion(&done.done);
8128
8129                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8130                                         pgoff, start, start + sectorsize - 1,
8131                                         io_bio->mirror_num, btrfs_retry_endio,
8132                                         &done);
8133                 if (status) {
8134                         err = status;
8135                         goto next;
8136                 }
8137
8138                 wait_for_completion_io(&done.done);
8139
8140                 if (!done.uptodate) {
8141                         /* We might have another mirror, so try again */
8142                         goto try_again;
8143                 }
8144 next:
8145                 offset += sectorsize;
8146                 start += sectorsize;
8147
8148                 ASSERT(nr_sectors);
8149
8150                 nr_sectors--;
8151                 if (nr_sectors) {
8152                         pgoff += sectorsize;
8153                         ASSERT(pgoff < PAGE_SIZE);
8154                         goto next_block;
8155                 }
8156         }
8157
8158         return err;
8159 }
8160
8161 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8162                 struct btrfs_io_bio *io_bio, blk_status_t err)
8163 {
8164         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8165
8166         if (skip_csum) {
8167                 if (unlikely(err))
8168                         return __btrfs_correct_data_nocsum(inode, io_bio);
8169                 else
8170                         return BLK_STS_OK;
8171         } else {
8172                 return __btrfs_subio_endio_read(inode, io_bio, err);
8173         }
8174 }
8175
8176 static void btrfs_endio_direct_read(struct bio *bio)
8177 {
8178         struct btrfs_dio_private *dip = bio->bi_private;
8179         struct inode *inode = dip->inode;
8180         struct bio *dio_bio;
8181         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8182         blk_status_t err = bio->bi_status;
8183
8184         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8185                 err = btrfs_subio_endio_read(inode, io_bio, err);
8186
8187         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8188                       dip->logical_offset + dip->bytes - 1);
8189         dio_bio = dip->dio_bio;
8190
8191         kfree(dip);
8192
8193         dio_bio->bi_status = err;
8194         dio_end_io(dio_bio);
8195         btrfs_io_bio_free_csum(io_bio);
8196         bio_put(bio);
8197 }
8198
8199 static void __endio_write_update_ordered(struct inode *inode,
8200                                          const u64 offset, const u64 bytes,
8201                                          const bool uptodate)
8202 {
8203         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8204         struct btrfs_ordered_extent *ordered = NULL;
8205         struct btrfs_workqueue *wq;
8206         u64 ordered_offset = offset;
8207         u64 ordered_bytes = bytes;
8208         u64 last_offset;
8209
8210         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
8211                 wq = fs_info->endio_freespace_worker;
8212         else
8213                 wq = fs_info->endio_write_workers;
8214
8215         while (ordered_offset < offset + bytes) {
8216                 last_offset = ordered_offset;
8217                 if (btrfs_dec_test_first_ordered_pending(inode, &ordered,
8218                                                            &ordered_offset,
8219                                                            ordered_bytes,
8220                                                            uptodate)) {
8221                         btrfs_init_work(&ordered->work, finish_ordered_fn, NULL,
8222                                         NULL);
8223                         btrfs_queue_work(wq, &ordered->work);
8224                 }
8225                 /*
8226                  * If btrfs_dec_test_ordered_pending does not find any ordered
8227                  * extent in the range, we can exit.
8228                  */
8229                 if (ordered_offset == last_offset)
8230                         return;
8231                 /*
8232                  * Our bio might span multiple ordered extents. In this case
8233                  * we keep going until we have accounted the whole dio.
8234                  */
8235                 if (ordered_offset < offset + bytes) {
8236                         ordered_bytes = offset + bytes - ordered_offset;
8237                         ordered = NULL;
8238                 }
8239         }
8240 }
8241
8242 static void btrfs_endio_direct_write(struct bio *bio)
8243 {
8244         struct btrfs_dio_private *dip = bio->bi_private;
8245         struct bio *dio_bio = dip->dio_bio;
8246
8247         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8248                                      dip->bytes, !bio->bi_status);
8249
8250         kfree(dip);
8251
8252         dio_bio->bi_status = bio->bi_status;
8253         dio_end_io(dio_bio);
8254         bio_put(bio);
8255 }
8256
8257 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
8258                                     struct bio *bio, u64 offset)
8259 {
8260         struct inode *inode = private_data;
8261         blk_status_t ret;
8262         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8263         BUG_ON(ret); /* -ENOMEM */
8264         return 0;
8265 }
8266
8267 static void btrfs_end_dio_bio(struct bio *bio)
8268 {
8269         struct btrfs_dio_private *dip = bio->bi_private;
8270         blk_status_t err = bio->bi_status;
8271
8272         if (err)
8273                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8274                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8275                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8276                            bio->bi_opf,
8277                            (unsigned long long)bio->bi_iter.bi_sector,
8278                            bio->bi_iter.bi_size, err);
8279
8280         if (dip->subio_endio)
8281                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8282
8283         if (err) {
8284                 /*
8285                  * We want to perceive the errors flag being set before
8286                  * decrementing the reference count. We don't need a barrier
8287                  * since atomic operations with a return value are fully
8288                  * ordered as per atomic_t.txt
8289                  */
8290                 dip->errors = 1;
8291         }
8292
8293         /* if there are more bios still pending for this dio, just exit */
8294         if (!atomic_dec_and_test(&dip->pending_bios))
8295                 goto out;
8296
8297         if (dip->errors) {
8298                 bio_io_error(dip->orig_bio);
8299         } else {
8300                 dip->dio_bio->bi_status = BLK_STS_OK;
8301                 bio_endio(dip->orig_bio);
8302         }
8303 out:
8304         bio_put(bio);
8305 }
8306
8307 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8308                                                  struct btrfs_dio_private *dip,
8309                                                  struct bio *bio,
8310                                                  u64 file_offset)
8311 {
8312         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8313         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8314         blk_status_t ret;
8315
8316         /*
8317          * We load all the csum data we need when we submit
8318          * the first bio to reduce the csum tree search and
8319          * contention.
8320          */
8321         if (dip->logical_offset == file_offset) {
8322                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8323                                                 file_offset);
8324                 if (ret)
8325                         return ret;
8326         }
8327
8328         if (bio == dip->orig_bio)
8329                 return 0;
8330
8331         file_offset -= dip->logical_offset;
8332         file_offset >>= inode->i_sb->s_blocksize_bits;
8333         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8334
8335         return 0;
8336 }
8337
8338 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
8339                 struct inode *inode, u64 file_offset, int async_submit)
8340 {
8341         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8342         struct btrfs_dio_private *dip = bio->bi_private;
8343         bool write = bio_op(bio) == REQ_OP_WRITE;
8344         blk_status_t ret;
8345
8346         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8347         if (async_submit)
8348                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8349
8350         if (!write) {
8351                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8352                 if (ret)
8353                         goto err;
8354         }
8355
8356         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8357                 goto map;
8358
8359         if (write && async_submit) {
8360                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8361                                           file_offset, inode,
8362                                           btrfs_submit_bio_start_direct_io);
8363                 goto err;
8364         } else if (write) {
8365                 /*
8366                  * If we aren't doing async submit, calculate the csum of the
8367                  * bio now.
8368                  */
8369                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8370                 if (ret)
8371                         goto err;
8372         } else {
8373                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8374                                                      file_offset);
8375                 if (ret)
8376                         goto err;
8377         }
8378 map:
8379         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8380 err:
8381         return ret;
8382 }
8383
8384 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8385 {
8386         struct inode *inode = dip->inode;
8387         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8388         struct bio *bio;
8389         struct bio *orig_bio = dip->orig_bio;
8390         u64 start_sector = orig_bio->bi_iter.bi_sector;
8391         u64 file_offset = dip->logical_offset;
8392         int async_submit = 0;
8393         u64 submit_len;
8394         int clone_offset = 0;
8395         int clone_len;
8396         int ret;
8397         blk_status_t status;
8398         struct btrfs_io_geometry geom;
8399
8400         submit_len = orig_bio->bi_iter.bi_size;
8401         ret = btrfs_get_io_geometry(fs_info, btrfs_op(orig_bio),
8402                                     start_sector << 9, submit_len, &geom);
8403         if (ret)
8404                 return -EIO;
8405
8406         if (geom.len >= submit_len) {
8407                 bio = orig_bio;
8408                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8409                 goto submit;
8410         }
8411
8412         /* async crcs make it difficult to collect full stripe writes. */
8413         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8414                 async_submit = 0;
8415         else
8416                 async_submit = 1;
8417
8418         /* bio split */
8419         ASSERT(geom.len <= INT_MAX);
8420         atomic_inc(&dip->pending_bios);
8421         do {
8422                 clone_len = min_t(int, submit_len, geom.len);
8423
8424                 /*
8425                  * This will never fail as it's passing GPF_NOFS and
8426                  * the allocation is backed by btrfs_bioset.
8427                  */
8428                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8429                                               clone_len);
8430                 bio->bi_private = dip;
8431                 bio->bi_end_io = btrfs_end_dio_bio;
8432                 btrfs_io_bio(bio)->logical = file_offset;
8433
8434                 ASSERT(submit_len >= clone_len);
8435                 submit_len -= clone_len;
8436                 if (submit_len == 0)
8437                         break;
8438
8439                 /*
8440                  * Increase the count before we submit the bio so we know
8441                  * the end IO handler won't happen before we increase the
8442                  * count. Otherwise, the dip might get freed before we're
8443                  * done setting it up.
8444                  */
8445                 atomic_inc(&dip->pending_bios);
8446
8447                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
8448                                                 async_submit);
8449                 if (status) {
8450                         bio_put(bio);
8451                         atomic_dec(&dip->pending_bios);
8452                         goto out_err;
8453                 }
8454
8455                 clone_offset += clone_len;
8456                 start_sector += clone_len >> 9;
8457                 file_offset += clone_len;
8458
8459                 ret = btrfs_get_io_geometry(fs_info, btrfs_op(orig_bio),
8460                                       start_sector << 9, submit_len, &geom);
8461                 if (ret)
8462                         goto out_err;
8463         } while (submit_len > 0);
8464
8465 submit:
8466         status = btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8467         if (!status)
8468                 return 0;
8469
8470         bio_put(bio);
8471 out_err:
8472         dip->errors = 1;
8473         /*
8474          * Before atomic variable goto zero, we must  make sure dip->errors is
8475          * perceived to be set. This ordering is ensured by the fact that an
8476          * atomic operations with a return value are fully ordered as per
8477          * atomic_t.txt
8478          */
8479         if (atomic_dec_and_test(&dip->pending_bios))
8480                 bio_io_error(dip->orig_bio);
8481
8482         /* bio_end_io() will handle error, so we needn't return it */
8483         return 0;
8484 }
8485
8486 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8487                                 loff_t file_offset)
8488 {
8489         struct btrfs_dio_private *dip = NULL;
8490         struct bio *bio = NULL;
8491         struct btrfs_io_bio *io_bio;
8492         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8493         int ret = 0;
8494
8495         bio = btrfs_bio_clone(dio_bio);
8496
8497         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8498         if (!dip) {
8499                 ret = -ENOMEM;
8500                 goto free_ordered;
8501         }
8502
8503         dip->private = dio_bio->bi_private;
8504         dip->inode = inode;
8505         dip->logical_offset = file_offset;
8506         dip->bytes = dio_bio->bi_iter.bi_size;
8507         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8508         bio->bi_private = dip;
8509         dip->orig_bio = bio;
8510         dip->dio_bio = dio_bio;
8511         atomic_set(&dip->pending_bios, 0);
8512         io_bio = btrfs_io_bio(bio);
8513         io_bio->logical = file_offset;
8514
8515         if (write) {
8516                 bio->bi_end_io = btrfs_endio_direct_write;
8517         } else {
8518                 bio->bi_end_io = btrfs_endio_direct_read;
8519                 dip->subio_endio = btrfs_subio_endio_read;
8520         }
8521
8522         /*
8523          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8524          * even if we fail to submit a bio, because in such case we do the
8525          * corresponding error handling below and it must not be done a second
8526          * time by btrfs_direct_IO().
8527          */
8528         if (write) {
8529                 struct btrfs_dio_data *dio_data = current->journal_info;
8530
8531                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8532                         dip->bytes;
8533                 dio_data->unsubmitted_oe_range_start =
8534                         dio_data->unsubmitted_oe_range_end;
8535         }
8536
8537         ret = btrfs_submit_direct_hook(dip);
8538         if (!ret)
8539                 return;
8540
8541         btrfs_io_bio_free_csum(io_bio);
8542
8543 free_ordered:
8544         /*
8545          * If we arrived here it means either we failed to submit the dip
8546          * or we either failed to clone the dio_bio or failed to allocate the
8547          * dip. If we cloned the dio_bio and allocated the dip, we can just
8548          * call bio_endio against our io_bio so that we get proper resource
8549          * cleanup if we fail to submit the dip, otherwise, we must do the
8550          * same as btrfs_endio_direct_[write|read] because we can't call these
8551          * callbacks - they require an allocated dip and a clone of dio_bio.
8552          */
8553         if (bio && dip) {
8554                 bio_io_error(bio);
8555                 /*
8556                  * The end io callbacks free our dip, do the final put on bio
8557                  * and all the cleanup and final put for dio_bio (through
8558                  * dio_end_io()).
8559                  */
8560                 dip = NULL;
8561                 bio = NULL;
8562         } else {
8563                 if (write)
8564                         __endio_write_update_ordered(inode,
8565                                                 file_offset,
8566                                                 dio_bio->bi_iter.bi_size,
8567                                                 false);
8568                 else
8569                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8570                               file_offset + dio_bio->bi_iter.bi_size - 1);
8571
8572                 dio_bio->bi_status = BLK_STS_IOERR;
8573                 /*
8574                  * Releases and cleans up our dio_bio, no need to bio_put()
8575                  * nor bio_endio()/bio_io_error() against dio_bio.
8576                  */
8577                 dio_end_io(dio_bio);
8578         }
8579         if (bio)
8580                 bio_put(bio);
8581         kfree(dip);
8582 }
8583
8584 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8585                                const struct iov_iter *iter, loff_t offset)
8586 {
8587         int seg;
8588         int i;
8589         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8590         ssize_t retval = -EINVAL;
8591
8592         if (offset & blocksize_mask)
8593                 goto out;
8594
8595         if (iov_iter_alignment(iter) & blocksize_mask)
8596                 goto out;
8597
8598         /* If this is a write we don't need to check anymore */
8599         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8600                 return 0;
8601         /*
8602          * Check to make sure we don't have duplicate iov_base's in this
8603          * iovec, if so return EINVAL, otherwise we'll get csum errors
8604          * when reading back.
8605          */
8606         for (seg = 0; seg < iter->nr_segs; seg++) {
8607                 for (i = seg + 1; i < iter->nr_segs; i++) {
8608                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8609                                 goto out;
8610                 }
8611         }
8612         retval = 0;
8613 out:
8614         return retval;
8615 }
8616
8617 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8618 {
8619         struct file *file = iocb->ki_filp;
8620         struct inode *inode = file->f_mapping->host;
8621         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8622         struct btrfs_dio_data dio_data = { 0 };
8623         struct extent_changeset *data_reserved = NULL;
8624         loff_t offset = iocb->ki_pos;
8625         size_t count = 0;
8626         int flags = 0;
8627         bool wakeup = true;
8628         bool relock = false;
8629         ssize_t ret;
8630
8631         if (check_direct_IO(fs_info, iter, offset))
8632                 return 0;
8633
8634         inode_dio_begin(inode);
8635
8636         /*
8637          * The generic stuff only does filemap_write_and_wait_range, which
8638          * isn't enough if we've written compressed pages to this area, so
8639          * we need to flush the dirty pages again to make absolutely sure
8640          * that any outstanding dirty pages are on disk.
8641          */
8642         count = iov_iter_count(iter);
8643         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8644                      &BTRFS_I(inode)->runtime_flags))
8645                 filemap_fdatawrite_range(inode->i_mapping, offset,
8646                                          offset + count - 1);
8647
8648         if (iov_iter_rw(iter) == WRITE) {
8649                 /*
8650                  * If the write DIO is beyond the EOF, we need update
8651                  * the isize, but it is protected by i_mutex. So we can
8652                  * not unlock the i_mutex at this case.
8653                  */
8654                 if (offset + count <= inode->i_size) {
8655                         dio_data.overwrite = 1;
8656                         inode_unlock(inode);
8657                         relock = true;
8658                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8659                         ret = -EAGAIN;
8660                         goto out;
8661                 }
8662                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8663                                                    offset, count);
8664                 if (ret)
8665                         goto out;
8666
8667                 /*
8668                  * We need to know how many extents we reserved so that we can
8669                  * do the accounting properly if we go over the number we
8670                  * originally calculated.  Abuse current->journal_info for this.
8671                  */
8672                 dio_data.reserve = round_up(count,
8673                                             fs_info->sectorsize);
8674                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8675                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8676                 current->journal_info = &dio_data;
8677                 down_read(&BTRFS_I(inode)->dio_sem);
8678         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8679                                      &BTRFS_I(inode)->runtime_flags)) {
8680                 inode_dio_end(inode);
8681                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8682                 wakeup = false;
8683         }
8684
8685         ret = __blockdev_direct_IO(iocb, inode,
8686                                    fs_info->fs_devices->latest_bdev,
8687                                    iter, btrfs_get_blocks_direct, NULL,
8688                                    btrfs_submit_direct, flags);
8689         if (iov_iter_rw(iter) == WRITE) {
8690                 up_read(&BTRFS_I(inode)->dio_sem);
8691                 current->journal_info = NULL;
8692                 if (ret < 0 && ret != -EIOCBQUEUED) {
8693                         if (dio_data.reserve)
8694                                 btrfs_delalloc_release_space(inode, data_reserved,
8695                                         offset, dio_data.reserve, true);
8696                         /*
8697                          * On error we might have left some ordered extents
8698                          * without submitting corresponding bios for them, so
8699                          * cleanup them up to avoid other tasks getting them
8700                          * and waiting for them to complete forever.
8701                          */
8702                         if (dio_data.unsubmitted_oe_range_start <
8703                             dio_data.unsubmitted_oe_range_end)
8704                                 __endio_write_update_ordered(inode,
8705                                         dio_data.unsubmitted_oe_range_start,
8706                                         dio_data.unsubmitted_oe_range_end -
8707                                         dio_data.unsubmitted_oe_range_start,
8708                                         false);
8709                 } else if (ret >= 0 && (size_t)ret < count)
8710                         btrfs_delalloc_release_space(inode, data_reserved,
8711                                         offset, count - (size_t)ret, true);
8712                 btrfs_delalloc_release_extents(BTRFS_I(inode), count);
8713         }
8714 out:
8715         if (wakeup)
8716                 inode_dio_end(inode);
8717         if (relock)
8718                 inode_lock(inode);
8719
8720         extent_changeset_free(data_reserved);
8721         return ret;
8722 }
8723
8724 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8725
8726 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8727                 __u64 start, __u64 len)
8728 {
8729         int     ret;
8730
8731         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8732         if (ret)
8733                 return ret;
8734
8735         return extent_fiemap(inode, fieinfo, start, len);
8736 }
8737
8738 int btrfs_readpage(struct file *file, struct page *page)
8739 {
8740         struct extent_io_tree *tree;
8741         tree = &BTRFS_I(page->mapping->host)->io_tree;
8742         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8743 }
8744
8745 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8746 {
8747         struct inode *inode = page->mapping->host;
8748         int ret;
8749
8750         if (current->flags & PF_MEMALLOC) {
8751                 redirty_page_for_writepage(wbc, page);
8752                 unlock_page(page);
8753                 return 0;
8754         }
8755
8756         /*
8757          * If we are under memory pressure we will call this directly from the
8758          * VM, we need to make sure we have the inode referenced for the ordered
8759          * extent.  If not just return like we didn't do anything.
8760          */
8761         if (!igrab(inode)) {
8762                 redirty_page_for_writepage(wbc, page);
8763                 return AOP_WRITEPAGE_ACTIVATE;
8764         }
8765         ret = extent_write_full_page(page, wbc);
8766         btrfs_add_delayed_iput(inode);
8767         return ret;
8768 }
8769
8770 static int btrfs_writepages(struct address_space *mapping,
8771                             struct writeback_control *wbc)
8772 {
8773         return extent_writepages(mapping, wbc);
8774 }
8775
8776 static int
8777 btrfs_readpages(struct file *file, struct address_space *mapping,
8778                 struct list_head *pages, unsigned nr_pages)
8779 {
8780         return extent_readpages(mapping, pages, nr_pages);
8781 }
8782
8783 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8784 {
8785         int ret = try_release_extent_mapping(page, gfp_flags);
8786         if (ret == 1) {
8787                 ClearPagePrivate(page);
8788                 set_page_private(page, 0);
8789                 put_page(page);
8790         }
8791         return ret;
8792 }
8793
8794 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8795 {
8796         if (PageWriteback(page) || PageDirty(page))
8797                 return 0;
8798         return __btrfs_releasepage(page, gfp_flags);
8799 }
8800
8801 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8802                                  unsigned int length)
8803 {
8804         struct inode *inode = page->mapping->host;
8805         struct extent_io_tree *tree;
8806         struct btrfs_ordered_extent *ordered;
8807         struct extent_state *cached_state = NULL;
8808         u64 page_start = page_offset(page);
8809         u64 page_end = page_start + PAGE_SIZE - 1;
8810         u64 start;
8811         u64 end;
8812         int inode_evicting = inode->i_state & I_FREEING;
8813
8814         /*
8815          * we have the page locked, so new writeback can't start,
8816          * and the dirty bit won't be cleared while we are here.
8817          *
8818          * Wait for IO on this page so that we can safely clear
8819          * the PagePrivate2 bit and do ordered accounting
8820          */
8821         wait_on_page_writeback(page);
8822
8823         tree = &BTRFS_I(inode)->io_tree;
8824         if (offset) {
8825                 btrfs_releasepage(page, GFP_NOFS);
8826                 return;
8827         }
8828
8829         if (!inode_evicting)
8830                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8831 again:
8832         start = page_start;
8833         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8834                                         page_end - start + 1);
8835         if (ordered) {
8836                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8837                 /*
8838                  * IO on this page will never be started, so we need
8839                  * to account for any ordered extents now
8840                  */
8841                 if (!inode_evicting)
8842                         clear_extent_bit(tree, start, end,
8843                                          EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8844                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8845                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8846                 /*
8847                  * whoever cleared the private bit is responsible
8848                  * for the finish_ordered_io
8849                  */
8850                 if (TestClearPagePrivate2(page)) {
8851                         struct btrfs_ordered_inode_tree *tree;
8852                         u64 new_len;
8853
8854                         tree = &BTRFS_I(inode)->ordered_tree;
8855
8856                         spin_lock_irq(&tree->lock);
8857                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8858                         new_len = start - ordered->file_offset;
8859                         if (new_len < ordered->truncated_len)
8860                                 ordered->truncated_len = new_len;
8861                         spin_unlock_irq(&tree->lock);
8862
8863                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8864                                                            start,
8865                                                            end - start + 1, 1))
8866                                 btrfs_finish_ordered_io(ordered);
8867                 }
8868                 btrfs_put_ordered_extent(ordered);
8869                 if (!inode_evicting) {
8870                         cached_state = NULL;
8871                         lock_extent_bits(tree, start, end,
8872                                          &cached_state);
8873                 }
8874
8875                 start = end + 1;
8876                 if (start < page_end)
8877                         goto again;
8878         }
8879
8880         /*
8881          * Qgroup reserved space handler
8882          * Page here will be either
8883          * 1) Already written to disk
8884          *    In this case, its reserved space is released from data rsv map
8885          *    and will be freed by delayed_ref handler finally.
8886          *    So even we call qgroup_free_data(), it won't decrease reserved
8887          *    space.
8888          * 2) Not written to disk
8889          *    This means the reserved space should be freed here. However,
8890          *    if a truncate invalidates the page (by clearing PageDirty)
8891          *    and the page is accounted for while allocating extent
8892          *    in btrfs_check_data_free_space() we let delayed_ref to
8893          *    free the entire extent.
8894          */
8895         if (PageDirty(page))
8896                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8897         if (!inode_evicting) {
8898                 clear_extent_bit(tree, page_start, page_end, EXTENT_LOCKED |
8899                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8900                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8901                                  &cached_state);
8902
8903                 __btrfs_releasepage(page, GFP_NOFS);
8904         }
8905
8906         ClearPageChecked(page);
8907         if (PagePrivate(page)) {
8908                 ClearPagePrivate(page);
8909                 set_page_private(page, 0);
8910                 put_page(page);
8911         }
8912 }
8913
8914 /*
8915  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8916  * called from a page fault handler when a page is first dirtied. Hence we must
8917  * be careful to check for EOF conditions here. We set the page up correctly
8918  * for a written page which means we get ENOSPC checking when writing into
8919  * holes and correct delalloc and unwritten extent mapping on filesystems that
8920  * support these features.
8921  *
8922  * We are not allowed to take the i_mutex here so we have to play games to
8923  * protect against truncate races as the page could now be beyond EOF.  Because
8924  * truncate_setsize() writes the inode size before removing pages, once we have
8925  * the page lock we can determine safely if the page is beyond EOF. If it is not
8926  * beyond EOF, then the page is guaranteed safe against truncation until we
8927  * unlock the page.
8928  */
8929 vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
8930 {
8931         struct page *page = vmf->page;
8932         struct inode *inode = file_inode(vmf->vma->vm_file);
8933         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8934         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8935         struct btrfs_ordered_extent *ordered;
8936         struct extent_state *cached_state = NULL;
8937         struct extent_changeset *data_reserved = NULL;
8938         char *kaddr;
8939         unsigned long zero_start;
8940         loff_t size;
8941         vm_fault_t ret;
8942         int ret2;
8943         int reserved = 0;
8944         u64 reserved_space;
8945         u64 page_start;
8946         u64 page_end;
8947         u64 end;
8948
8949         reserved_space = PAGE_SIZE;
8950
8951         sb_start_pagefault(inode->i_sb);
8952         page_start = page_offset(page);
8953         page_end = page_start + PAGE_SIZE - 1;
8954         end = page_end;
8955
8956         /*
8957          * Reserving delalloc space after obtaining the page lock can lead to
8958          * deadlock. For example, if a dirty page is locked by this function
8959          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8960          * dirty page write out, then the btrfs_writepage() function could
8961          * end up waiting indefinitely to get a lock on the page currently
8962          * being processed by btrfs_page_mkwrite() function.
8963          */
8964         ret2 = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
8965                                            reserved_space);
8966         if (!ret2) {
8967                 ret2 = file_update_time(vmf->vma->vm_file);
8968                 reserved = 1;
8969         }
8970         if (ret2) {
8971                 ret = vmf_error(ret2);
8972                 if (reserved)
8973                         goto out;
8974                 goto out_noreserve;
8975         }
8976
8977         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8978 again:
8979         lock_page(page);
8980         size = i_size_read(inode);
8981
8982         if ((page->mapping != inode->i_mapping) ||
8983             (page_start >= size)) {
8984                 /* page got truncated out from underneath us */
8985                 goto out_unlock;
8986         }
8987         wait_on_page_writeback(page);
8988
8989         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8990         set_page_extent_mapped(page);
8991
8992         /*
8993          * we can't set the delalloc bits if there are pending ordered
8994          * extents.  Drop our locks and wait for them to finish
8995          */
8996         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
8997                         PAGE_SIZE);
8998         if (ordered) {
8999                 unlock_extent_cached(io_tree, page_start, page_end,
9000                                      &cached_state);
9001                 unlock_page(page);
9002                 btrfs_start_ordered_extent(inode, ordered, 1);
9003                 btrfs_put_ordered_extent(ordered);
9004                 goto again;
9005         }
9006
9007         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9008                 reserved_space = round_up(size - page_start,
9009                                           fs_info->sectorsize);
9010                 if (reserved_space < PAGE_SIZE) {
9011                         end = page_start + reserved_space - 1;
9012                         btrfs_delalloc_release_space(inode, data_reserved,
9013                                         page_start, PAGE_SIZE - reserved_space,
9014                                         true);
9015                 }
9016         }
9017
9018         /*
9019          * page_mkwrite gets called when the page is firstly dirtied after it's
9020          * faulted in, but write(2) could also dirty a page and set delalloc
9021          * bits, thus in this case for space account reason, we still need to
9022          * clear any delalloc bits within this page range since we have to
9023          * reserve data&meta space before lock_page() (see above comments).
9024          */
9025         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9026                           EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
9027                           EXTENT_DEFRAG, 0, 0, &cached_state);
9028
9029         ret2 = btrfs_set_extent_delalloc(inode, page_start, end, 0,
9030                                         &cached_state);
9031         if (ret2) {
9032                 unlock_extent_cached(io_tree, page_start, page_end,
9033                                      &cached_state);
9034                 ret = VM_FAULT_SIGBUS;
9035                 goto out_unlock;
9036         }
9037         ret2 = 0;
9038
9039         /* page is wholly or partially inside EOF */
9040         if (page_start + PAGE_SIZE > size)
9041                 zero_start = offset_in_page(size);
9042         else
9043                 zero_start = PAGE_SIZE;
9044
9045         if (zero_start != PAGE_SIZE) {
9046                 kaddr = kmap(page);
9047                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9048                 flush_dcache_page(page);
9049                 kunmap(page);
9050         }
9051         ClearPageChecked(page);
9052         set_page_dirty(page);
9053         SetPageUptodate(page);
9054
9055         BTRFS_I(inode)->last_trans = fs_info->generation;
9056         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9057         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9058
9059         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
9060
9061         if (!ret2) {
9062                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9063                 sb_end_pagefault(inode->i_sb);
9064                 extent_changeset_free(data_reserved);
9065                 return VM_FAULT_LOCKED;
9066         }
9067
9068 out_unlock:
9069         unlock_page(page);
9070 out:
9071         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9072         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9073                                      reserved_space, (ret != 0));
9074 out_noreserve:
9075         sb_end_pagefault(inode->i_sb);
9076         extent_changeset_free(data_reserved);
9077         return ret;
9078 }
9079
9080 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
9081 {
9082         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9083         struct btrfs_root *root = BTRFS_I(inode)->root;
9084         struct btrfs_block_rsv *rsv;
9085         int ret;
9086         struct btrfs_trans_handle *trans;
9087         u64 mask = fs_info->sectorsize - 1;
9088         u64 min_size = btrfs_calc_metadata_size(fs_info, 1);
9089
9090         if (!skip_writeback) {
9091                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9092                                                (u64)-1);
9093                 if (ret)
9094                         return ret;
9095         }
9096
9097         /*
9098          * Yes ladies and gentlemen, this is indeed ugly.  We have a couple of
9099          * things going on here:
9100          *
9101          * 1) We need to reserve space to update our inode.
9102          *
9103          * 2) We need to have something to cache all the space that is going to
9104          * be free'd up by the truncate operation, but also have some slack
9105          * space reserved in case it uses space during the truncate (thank you
9106          * very much snapshotting).
9107          *
9108          * And we need these to be separate.  The fact is we can use a lot of
9109          * space doing the truncate, and we have no earthly idea how much space
9110          * we will use, so we need the truncate reservation to be separate so it
9111          * doesn't end up using space reserved for updating the inode.  We also
9112          * need to be able to stop the transaction and start a new one, which
9113          * means we need to be able to update the inode several times, and we
9114          * have no idea of knowing how many times that will be, so we can't just
9115          * reserve 1 item for the entirety of the operation, so that has to be
9116          * done separately as well.
9117          *
9118          * So that leaves us with
9119          *
9120          * 1) rsv - for the truncate reservation, which we will steal from the
9121          * transaction reservation.
9122          * 2) fs_info->trans_block_rsv - this will have 1 items worth left for
9123          * updating the inode.
9124          */
9125         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9126         if (!rsv)
9127                 return -ENOMEM;
9128         rsv->size = min_size;
9129         rsv->failfast = 1;
9130
9131         /*
9132          * 1 for the truncate slack space
9133          * 1 for updating the inode.
9134          */
9135         trans = btrfs_start_transaction(root, 2);
9136         if (IS_ERR(trans)) {
9137                 ret = PTR_ERR(trans);
9138                 goto out;
9139         }
9140
9141         /* Migrate the slack space for the truncate to our reserve */
9142         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9143                                       min_size, false);
9144         BUG_ON(ret);
9145
9146         /*
9147          * So if we truncate and then write and fsync we normally would just
9148          * write the extents that changed, which is a problem if we need to
9149          * first truncate that entire inode.  So set this flag so we write out
9150          * all of the extents in the inode to the sync log so we're completely
9151          * safe.
9152          */
9153         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9154         trans->block_rsv = rsv;
9155
9156         while (1) {
9157                 ret = btrfs_truncate_inode_items(trans, root, inode,
9158                                                  inode->i_size,
9159                                                  BTRFS_EXTENT_DATA_KEY);
9160                 trans->block_rsv = &fs_info->trans_block_rsv;
9161                 if (ret != -ENOSPC && ret != -EAGAIN)
9162                         break;
9163
9164                 ret = btrfs_update_inode(trans, root, inode);
9165                 if (ret)
9166                         break;
9167
9168                 btrfs_end_transaction(trans);
9169                 btrfs_btree_balance_dirty(fs_info);
9170
9171                 trans = btrfs_start_transaction(root, 2);
9172                 if (IS_ERR(trans)) {
9173                         ret = PTR_ERR(trans);
9174                         trans = NULL;
9175                         break;
9176                 }
9177
9178                 btrfs_block_rsv_release(fs_info, rsv, -1);
9179                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9180                                               rsv, min_size, false);
9181                 BUG_ON(ret);    /* shouldn't happen */
9182                 trans->block_rsv = rsv;
9183         }
9184
9185         /*
9186          * We can't call btrfs_truncate_block inside a trans handle as we could
9187          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9188          * we've truncated everything except the last little bit, and can do
9189          * btrfs_truncate_block and then update the disk_i_size.
9190          */
9191         if (ret == NEED_TRUNCATE_BLOCK) {
9192                 btrfs_end_transaction(trans);
9193                 btrfs_btree_balance_dirty(fs_info);
9194
9195                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9196                 if (ret)
9197                         goto out;
9198                 trans = btrfs_start_transaction(root, 1);
9199                 if (IS_ERR(trans)) {
9200                         ret = PTR_ERR(trans);
9201                         goto out;
9202                 }
9203                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9204         }
9205
9206         if (trans) {
9207                 int ret2;
9208
9209                 trans->block_rsv = &fs_info->trans_block_rsv;
9210                 ret2 = btrfs_update_inode(trans, root, inode);
9211                 if (ret2 && !ret)
9212                         ret = ret2;
9213
9214                 ret2 = btrfs_end_transaction(trans);
9215                 if (ret2 && !ret)
9216                         ret = ret2;
9217                 btrfs_btree_balance_dirty(fs_info);
9218         }
9219 out:
9220         btrfs_free_block_rsv(fs_info, rsv);
9221
9222         return ret;
9223 }
9224
9225 /*
9226  * create a new subvolume directory/inode (helper for the ioctl).
9227  */
9228 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9229                              struct btrfs_root *new_root,
9230                              struct btrfs_root *parent_root,
9231                              u64 new_dirid)
9232 {
9233         struct inode *inode;
9234         int err;
9235         u64 index = 0;
9236
9237         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9238                                 new_dirid, new_dirid,
9239                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9240                                 &index);
9241         if (IS_ERR(inode))
9242                 return PTR_ERR(inode);
9243         inode->i_op = &btrfs_dir_inode_operations;
9244         inode->i_fop = &btrfs_dir_file_operations;
9245
9246         set_nlink(inode, 1);
9247         btrfs_i_size_write(BTRFS_I(inode), 0);
9248         unlock_new_inode(inode);
9249
9250         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9251         if (err)
9252                 btrfs_err(new_root->fs_info,
9253                           "error inheriting subvolume %llu properties: %d",
9254                           new_root->root_key.objectid, err);
9255
9256         err = btrfs_update_inode(trans, new_root, inode);
9257
9258         iput(inode);
9259         return err;
9260 }
9261
9262 struct inode *btrfs_alloc_inode(struct super_block *sb)
9263 {
9264         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9265         struct btrfs_inode *ei;
9266         struct inode *inode;
9267
9268         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9269         if (!ei)
9270                 return NULL;
9271
9272         ei->root = NULL;
9273         ei->generation = 0;
9274         ei->last_trans = 0;
9275         ei->last_sub_trans = 0;
9276         ei->logged_trans = 0;
9277         ei->delalloc_bytes = 0;
9278         ei->new_delalloc_bytes = 0;
9279         ei->defrag_bytes = 0;
9280         ei->disk_i_size = 0;
9281         ei->flags = 0;
9282         ei->csum_bytes = 0;
9283         ei->index_cnt = (u64)-1;
9284         ei->dir_index = 0;
9285         ei->last_unlink_trans = 0;
9286         ei->last_log_commit = 0;
9287
9288         spin_lock_init(&ei->lock);
9289         ei->outstanding_extents = 0;
9290         if (sb->s_magic != BTRFS_TEST_MAGIC)
9291                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9292                                               BTRFS_BLOCK_RSV_DELALLOC);
9293         ei->runtime_flags = 0;
9294         ei->prop_compress = BTRFS_COMPRESS_NONE;
9295         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9296
9297         ei->delayed_node = NULL;
9298
9299         ei->i_otime.tv_sec = 0;
9300         ei->i_otime.tv_nsec = 0;
9301
9302         inode = &ei->vfs_inode;
9303         extent_map_tree_init(&ei->extent_tree);
9304         extent_io_tree_init(fs_info, &ei->io_tree, IO_TREE_INODE_IO, inode);
9305         extent_io_tree_init(fs_info, &ei->io_failure_tree,
9306                             IO_TREE_INODE_IO_FAILURE, inode);
9307         ei->io_tree.track_uptodate = true;
9308         ei->io_failure_tree.track_uptodate = true;
9309         atomic_set(&ei->sync_writers, 0);
9310         mutex_init(&ei->log_mutex);
9311         mutex_init(&ei->delalloc_mutex);
9312         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9313         INIT_LIST_HEAD(&ei->delalloc_inodes);
9314         INIT_LIST_HEAD(&ei->delayed_iput);
9315         RB_CLEAR_NODE(&ei->rb_node);
9316         init_rwsem(&ei->dio_sem);
9317
9318         return inode;
9319 }
9320
9321 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9322 void btrfs_test_destroy_inode(struct inode *inode)
9323 {
9324         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9325         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9326 }
9327 #endif
9328
9329 void btrfs_free_inode(struct inode *inode)
9330 {
9331         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9332 }
9333
9334 void btrfs_destroy_inode(struct inode *inode)
9335 {
9336         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9337         struct btrfs_ordered_extent *ordered;
9338         struct btrfs_root *root = BTRFS_I(inode)->root;
9339
9340         WARN_ON(!hlist_empty(&inode->i_dentry));
9341         WARN_ON(inode->i_data.nrpages);
9342         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9343         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9344         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9345         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9346         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9347         WARN_ON(BTRFS_I(inode)->csum_bytes);
9348         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9349
9350         /*
9351          * This can happen where we create an inode, but somebody else also
9352          * created the same inode and we need to destroy the one we already
9353          * created.
9354          */
9355         if (!root)
9356                 return;
9357
9358         while (1) {
9359                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9360                 if (!ordered)
9361                         break;
9362                 else {
9363                         btrfs_err(fs_info,
9364                                   "found ordered extent %llu %llu on inode cleanup",
9365                                   ordered->file_offset, ordered->len);
9366                         btrfs_remove_ordered_extent(inode, ordered);
9367                         btrfs_put_ordered_extent(ordered);
9368                         btrfs_put_ordered_extent(ordered);
9369                 }
9370         }
9371         btrfs_qgroup_check_reserved_leak(inode);
9372         inode_tree_del(inode);
9373         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9374 }
9375
9376 int btrfs_drop_inode(struct inode *inode)
9377 {
9378         struct btrfs_root *root = BTRFS_I(inode)->root;
9379
9380         if (root == NULL)
9381                 return 1;
9382
9383         /* the snap/subvol tree is on deleting */
9384         if (btrfs_root_refs(&root->root_item) == 0)
9385                 return 1;
9386         else
9387                 return generic_drop_inode(inode);
9388 }
9389
9390 static void init_once(void *foo)
9391 {
9392         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9393
9394         inode_init_once(&ei->vfs_inode);
9395 }
9396
9397 void __cold btrfs_destroy_cachep(void)
9398 {
9399         /*
9400          * Make sure all delayed rcu free inodes are flushed before we
9401          * destroy cache.
9402          */
9403         rcu_barrier();
9404         kmem_cache_destroy(btrfs_inode_cachep);
9405         kmem_cache_destroy(btrfs_trans_handle_cachep);
9406         kmem_cache_destroy(btrfs_path_cachep);
9407         kmem_cache_destroy(btrfs_free_space_cachep);
9408         kmem_cache_destroy(btrfs_free_space_bitmap_cachep);
9409 }
9410
9411 int __init btrfs_init_cachep(void)
9412 {
9413         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9414                         sizeof(struct btrfs_inode), 0,
9415                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9416                         init_once);
9417         if (!btrfs_inode_cachep)
9418                 goto fail;
9419
9420         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9421                         sizeof(struct btrfs_trans_handle), 0,
9422                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9423         if (!btrfs_trans_handle_cachep)
9424                 goto fail;
9425
9426         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9427                         sizeof(struct btrfs_path), 0,
9428                         SLAB_MEM_SPREAD, NULL);
9429         if (!btrfs_path_cachep)
9430                 goto fail;
9431
9432         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9433                         sizeof(struct btrfs_free_space), 0,
9434                         SLAB_MEM_SPREAD, NULL);
9435         if (!btrfs_free_space_cachep)
9436                 goto fail;
9437
9438         btrfs_free_space_bitmap_cachep = kmem_cache_create("btrfs_free_space_bitmap",
9439                                                         PAGE_SIZE, PAGE_SIZE,
9440                                                         SLAB_RED_ZONE, NULL);
9441         if (!btrfs_free_space_bitmap_cachep)
9442                 goto fail;
9443
9444         return 0;
9445 fail:
9446         btrfs_destroy_cachep();
9447         return -ENOMEM;
9448 }
9449
9450 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9451                          u32 request_mask, unsigned int flags)
9452 {
9453         u64 delalloc_bytes;
9454         struct inode *inode = d_inode(path->dentry);
9455         u32 blocksize = inode->i_sb->s_blocksize;
9456         u32 bi_flags = BTRFS_I(inode)->flags;
9457
9458         stat->result_mask |= STATX_BTIME;
9459         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9460         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9461         if (bi_flags & BTRFS_INODE_APPEND)
9462                 stat->attributes |= STATX_ATTR_APPEND;
9463         if (bi_flags & BTRFS_INODE_COMPRESS)
9464                 stat->attributes |= STATX_ATTR_COMPRESSED;
9465         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9466                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9467         if (bi_flags & BTRFS_INODE_NODUMP)
9468                 stat->attributes |= STATX_ATTR_NODUMP;
9469
9470         stat->attributes_mask |= (STATX_ATTR_APPEND |
9471                                   STATX_ATTR_COMPRESSED |
9472                                   STATX_ATTR_IMMUTABLE |
9473                                   STATX_ATTR_NODUMP);
9474
9475         generic_fillattr(inode, stat);
9476         stat->dev = BTRFS_I(inode)->root->anon_dev;
9477
9478         spin_lock(&BTRFS_I(inode)->lock);
9479         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9480         spin_unlock(&BTRFS_I(inode)->lock);
9481         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9482                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9483         return 0;
9484 }
9485
9486 static int btrfs_rename_exchange(struct inode *old_dir,
9487                               struct dentry *old_dentry,
9488                               struct inode *new_dir,
9489                               struct dentry *new_dentry)
9490 {
9491         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9492         struct btrfs_trans_handle *trans;
9493         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9494         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9495         struct inode *new_inode = new_dentry->d_inode;
9496         struct inode *old_inode = old_dentry->d_inode;
9497         struct timespec64 ctime = current_time(old_inode);
9498         struct dentry *parent;
9499         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9500         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9501         u64 old_idx = 0;
9502         u64 new_idx = 0;
9503         u64 root_objectid;
9504         int ret;
9505         bool root_log_pinned = false;
9506         bool dest_log_pinned = false;
9507         struct btrfs_log_ctx ctx_root;
9508         struct btrfs_log_ctx ctx_dest;
9509         bool sync_log_root = false;
9510         bool sync_log_dest = false;
9511         bool commit_transaction = false;
9512
9513         /* we only allow rename subvolume link between subvolumes */
9514         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9515                 return -EXDEV;
9516
9517         btrfs_init_log_ctx(&ctx_root, old_inode);
9518         btrfs_init_log_ctx(&ctx_dest, new_inode);
9519
9520         /* close the race window with snapshot create/destroy ioctl */
9521         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9522                 down_read(&fs_info->subvol_sem);
9523         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9524                 down_read(&fs_info->subvol_sem);
9525
9526         /*
9527          * We want to reserve the absolute worst case amount of items.  So if
9528          * both inodes are subvols and we need to unlink them then that would
9529          * require 4 item modifications, but if they are both normal inodes it
9530          * would require 5 item modifications, so we'll assume their normal
9531          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9532          * should cover the worst case number of items we'll modify.
9533          */
9534         trans = btrfs_start_transaction(root, 12);
9535         if (IS_ERR(trans)) {
9536                 ret = PTR_ERR(trans);
9537                 goto out_notrans;
9538         }
9539
9540         /*
9541          * We need to find a free sequence number both in the source and
9542          * in the destination directory for the exchange.
9543          */
9544         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9545         if (ret)
9546                 goto out_fail;
9547         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9548         if (ret)
9549                 goto out_fail;
9550
9551         BTRFS_I(old_inode)->dir_index = 0ULL;
9552         BTRFS_I(new_inode)->dir_index = 0ULL;
9553
9554         /* Reference for the source. */
9555         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9556                 /* force full log commit if subvolume involved. */
9557                 btrfs_set_log_full_commit(trans);
9558         } else {
9559                 btrfs_pin_log_trans(root);
9560                 root_log_pinned = true;
9561                 ret = btrfs_insert_inode_ref(trans, dest,
9562                                              new_dentry->d_name.name,
9563                                              new_dentry->d_name.len,
9564                                              old_ino,
9565                                              btrfs_ino(BTRFS_I(new_dir)),
9566                                              old_idx);
9567                 if (ret)
9568                         goto out_fail;
9569         }
9570
9571         /* And now for the dest. */
9572         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9573                 /* force full log commit if subvolume involved. */
9574                 btrfs_set_log_full_commit(trans);
9575         } else {
9576                 btrfs_pin_log_trans(dest);
9577                 dest_log_pinned = true;
9578                 ret = btrfs_insert_inode_ref(trans, root,
9579                                              old_dentry->d_name.name,
9580                                              old_dentry->d_name.len,
9581                                              new_ino,
9582                                              btrfs_ino(BTRFS_I(old_dir)),
9583                                              new_idx);
9584                 if (ret)
9585                         goto out_fail;
9586         }
9587
9588         /* Update inode version and ctime/mtime. */
9589         inode_inc_iversion(old_dir);
9590         inode_inc_iversion(new_dir);
9591         inode_inc_iversion(old_inode);
9592         inode_inc_iversion(new_inode);
9593         old_dir->i_ctime = old_dir->i_mtime = ctime;
9594         new_dir->i_ctime = new_dir->i_mtime = ctime;
9595         old_inode->i_ctime = ctime;
9596         new_inode->i_ctime = ctime;
9597
9598         if (old_dentry->d_parent != new_dentry->d_parent) {
9599                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9600                                 BTRFS_I(old_inode), 1);
9601                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9602                                 BTRFS_I(new_inode), 1);
9603         }
9604
9605         /* src is a subvolume */
9606         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9607                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9608                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9609                                           old_dentry->d_name.name,
9610                                           old_dentry->d_name.len);
9611         } else { /* src is an inode */
9612                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9613                                            BTRFS_I(old_dentry->d_inode),
9614                                            old_dentry->d_name.name,
9615                                            old_dentry->d_name.len);
9616                 if (!ret)
9617                         ret = btrfs_update_inode(trans, root, old_inode);
9618         }
9619         if (ret) {
9620                 btrfs_abort_transaction(trans, ret);
9621                 goto out_fail;
9622         }
9623
9624         /* dest is a subvolume */
9625         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9626                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9627                 ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9628                                           new_dentry->d_name.name,
9629                                           new_dentry->d_name.len);
9630         } else { /* dest is an inode */
9631                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9632                                            BTRFS_I(new_dentry->d_inode),
9633                                            new_dentry->d_name.name,
9634                                            new_dentry->d_name.len);
9635                 if (!ret)
9636                         ret = btrfs_update_inode(trans, dest, new_inode);
9637         }
9638         if (ret) {
9639                 btrfs_abort_transaction(trans, ret);
9640                 goto out_fail;
9641         }
9642
9643         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9644                              new_dentry->d_name.name,
9645                              new_dentry->d_name.len, 0, old_idx);
9646         if (ret) {
9647                 btrfs_abort_transaction(trans, ret);
9648                 goto out_fail;
9649         }
9650
9651         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9652                              old_dentry->d_name.name,
9653                              old_dentry->d_name.len, 0, new_idx);
9654         if (ret) {
9655                 btrfs_abort_transaction(trans, ret);
9656                 goto out_fail;
9657         }
9658
9659         if (old_inode->i_nlink == 1)
9660                 BTRFS_I(old_inode)->dir_index = old_idx;
9661         if (new_inode->i_nlink == 1)
9662                 BTRFS_I(new_inode)->dir_index = new_idx;
9663
9664         if (root_log_pinned) {
9665                 parent = new_dentry->d_parent;
9666                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9667                                          BTRFS_I(old_dir), parent,
9668                                          false, &ctx_root);
9669                 if (ret == BTRFS_NEED_LOG_SYNC)
9670                         sync_log_root = true;
9671                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9672                         commit_transaction = true;
9673                 ret = 0;
9674                 btrfs_end_log_trans(root);
9675                 root_log_pinned = false;
9676         }
9677         if (dest_log_pinned) {
9678                 if (!commit_transaction) {
9679                         parent = old_dentry->d_parent;
9680                         ret = btrfs_log_new_name(trans, BTRFS_I(new_inode),
9681                                                  BTRFS_I(new_dir), parent,
9682                                                  false, &ctx_dest);
9683                         if (ret == BTRFS_NEED_LOG_SYNC)
9684                                 sync_log_dest = true;
9685                         else if (ret == BTRFS_NEED_TRANS_COMMIT)
9686                                 commit_transaction = true;
9687                         ret = 0;
9688                 }
9689                 btrfs_end_log_trans(dest);
9690                 dest_log_pinned = false;
9691         }
9692 out_fail:
9693         /*
9694          * If we have pinned a log and an error happened, we unpin tasks
9695          * trying to sync the log and force them to fallback to a transaction
9696          * commit if the log currently contains any of the inodes involved in
9697          * this rename operation (to ensure we do not persist a log with an
9698          * inconsistent state for any of these inodes or leading to any
9699          * inconsistencies when replayed). If the transaction was aborted, the
9700          * abortion reason is propagated to userspace when attempting to commit
9701          * the transaction. If the log does not contain any of these inodes, we
9702          * allow the tasks to sync it.
9703          */
9704         if (ret && (root_log_pinned || dest_log_pinned)) {
9705                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9706                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9707                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9708                     (new_inode &&
9709                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9710                         btrfs_set_log_full_commit(trans);
9711
9712                 if (root_log_pinned) {
9713                         btrfs_end_log_trans(root);
9714                         root_log_pinned = false;
9715                 }
9716                 if (dest_log_pinned) {
9717                         btrfs_end_log_trans(dest);
9718                         dest_log_pinned = false;
9719                 }
9720         }
9721         if (!ret && sync_log_root && !commit_transaction) {
9722                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root,
9723                                      &ctx_root);
9724                 if (ret)
9725                         commit_transaction = true;
9726         }
9727         if (!ret && sync_log_dest && !commit_transaction) {
9728                 ret = btrfs_sync_log(trans, BTRFS_I(new_inode)->root,
9729                                      &ctx_dest);
9730                 if (ret)
9731                         commit_transaction = true;
9732         }
9733         if (commit_transaction) {
9734                 /*
9735                  * We may have set commit_transaction when logging the new name
9736                  * in the destination root, in which case we left the source
9737                  * root context in the list of log contextes. So make sure we
9738                  * remove it to avoid invalid memory accesses, since the context
9739                  * was allocated in our stack frame.
9740                  */
9741                 if (sync_log_root) {
9742                         mutex_lock(&root->log_mutex);
9743                         list_del_init(&ctx_root.list);
9744                         mutex_unlock(&root->log_mutex);
9745                 }
9746                 ret = btrfs_commit_transaction(trans);
9747         } else {
9748                 int ret2;
9749
9750                 ret2 = btrfs_end_transaction(trans);
9751                 ret = ret ? ret : ret2;
9752         }
9753 out_notrans:
9754         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9755                 up_read(&fs_info->subvol_sem);
9756         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9757                 up_read(&fs_info->subvol_sem);
9758
9759         ASSERT(list_empty(&ctx_root.list));
9760         ASSERT(list_empty(&ctx_dest.list));
9761
9762         return ret;
9763 }
9764
9765 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9766                                      struct btrfs_root *root,
9767                                      struct inode *dir,
9768                                      struct dentry *dentry)
9769 {
9770         int ret;
9771         struct inode *inode;
9772         u64 objectid;
9773         u64 index;
9774
9775         ret = btrfs_find_free_ino(root, &objectid);
9776         if (ret)
9777                 return ret;
9778
9779         inode = btrfs_new_inode(trans, root, dir,
9780                                 dentry->d_name.name,
9781                                 dentry->d_name.len,
9782                                 btrfs_ino(BTRFS_I(dir)),
9783                                 objectid,
9784                                 S_IFCHR | WHITEOUT_MODE,
9785                                 &index);
9786
9787         if (IS_ERR(inode)) {
9788                 ret = PTR_ERR(inode);
9789                 return ret;
9790         }
9791
9792         inode->i_op = &btrfs_special_inode_operations;
9793         init_special_inode(inode, inode->i_mode,
9794                 WHITEOUT_DEV);
9795
9796         ret = btrfs_init_inode_security(trans, inode, dir,
9797                                 &dentry->d_name);
9798         if (ret)
9799                 goto out;
9800
9801         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9802                                 BTRFS_I(inode), 0, index);
9803         if (ret)
9804                 goto out;
9805
9806         ret = btrfs_update_inode(trans, root, inode);
9807 out:
9808         unlock_new_inode(inode);
9809         if (ret)
9810                 inode_dec_link_count(inode);
9811         iput(inode);
9812
9813         return ret;
9814 }
9815
9816 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9817                            struct inode *new_dir, struct dentry *new_dentry,
9818                            unsigned int flags)
9819 {
9820         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9821         struct btrfs_trans_handle *trans;
9822         unsigned int trans_num_items;
9823         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9824         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9825         struct inode *new_inode = d_inode(new_dentry);
9826         struct inode *old_inode = d_inode(old_dentry);
9827         u64 index = 0;
9828         u64 root_objectid;
9829         int ret;
9830         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9831         bool log_pinned = false;
9832         struct btrfs_log_ctx ctx;
9833         bool sync_log = false;
9834         bool commit_transaction = false;
9835
9836         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9837                 return -EPERM;
9838
9839         /* we only allow rename subvolume link between subvolumes */
9840         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9841                 return -EXDEV;
9842
9843         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9844             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9845                 return -ENOTEMPTY;
9846
9847         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9848             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9849                 return -ENOTEMPTY;
9850
9851
9852         /* check for collisions, even if the  name isn't there */
9853         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9854                              new_dentry->d_name.name,
9855                              new_dentry->d_name.len);
9856
9857         if (ret) {
9858                 if (ret == -EEXIST) {
9859                         /* we shouldn't get
9860                          * eexist without a new_inode */
9861                         if (WARN_ON(!new_inode)) {
9862                                 return ret;
9863                         }
9864                 } else {
9865                         /* maybe -EOVERFLOW */
9866                         return ret;
9867                 }
9868         }
9869         ret = 0;
9870
9871         /*
9872          * we're using rename to replace one file with another.  Start IO on it
9873          * now so  we don't add too much work to the end of the transaction
9874          */
9875         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9876                 filemap_flush(old_inode->i_mapping);
9877
9878         /* close the racy window with snapshot create/destroy ioctl */
9879         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9880                 down_read(&fs_info->subvol_sem);
9881         /*
9882          * We want to reserve the absolute worst case amount of items.  So if
9883          * both inodes are subvols and we need to unlink them then that would
9884          * require 4 item modifications, but if they are both normal inodes it
9885          * would require 5 item modifications, so we'll assume they are normal
9886          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9887          * should cover the worst case number of items we'll modify.
9888          * If our rename has the whiteout flag, we need more 5 units for the
9889          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9890          * when selinux is enabled).
9891          */
9892         trans_num_items = 11;
9893         if (flags & RENAME_WHITEOUT)
9894                 trans_num_items += 5;
9895         trans = btrfs_start_transaction(root, trans_num_items);
9896         if (IS_ERR(trans)) {
9897                 ret = PTR_ERR(trans);
9898                 goto out_notrans;
9899         }
9900
9901         if (dest != root)
9902                 btrfs_record_root_in_trans(trans, dest);
9903
9904         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9905         if (ret)
9906                 goto out_fail;
9907
9908         BTRFS_I(old_inode)->dir_index = 0ULL;
9909         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9910                 /* force full log commit if subvolume involved. */
9911                 btrfs_set_log_full_commit(trans);
9912         } else {
9913                 btrfs_pin_log_trans(root);
9914                 log_pinned = true;
9915                 ret = btrfs_insert_inode_ref(trans, dest,
9916                                              new_dentry->d_name.name,
9917                                              new_dentry->d_name.len,
9918                                              old_ino,
9919                                              btrfs_ino(BTRFS_I(new_dir)), index);
9920                 if (ret)
9921                         goto out_fail;
9922         }
9923
9924         inode_inc_iversion(old_dir);
9925         inode_inc_iversion(new_dir);
9926         inode_inc_iversion(old_inode);
9927         old_dir->i_ctime = old_dir->i_mtime =
9928         new_dir->i_ctime = new_dir->i_mtime =
9929         old_inode->i_ctime = current_time(old_dir);
9930
9931         if (old_dentry->d_parent != new_dentry->d_parent)
9932                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9933                                 BTRFS_I(old_inode), 1);
9934
9935         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9936                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9937                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9938                                         old_dentry->d_name.name,
9939                                         old_dentry->d_name.len);
9940         } else {
9941                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9942                                         BTRFS_I(d_inode(old_dentry)),
9943                                         old_dentry->d_name.name,
9944                                         old_dentry->d_name.len);
9945                 if (!ret)
9946                         ret = btrfs_update_inode(trans, root, old_inode);
9947         }
9948         if (ret) {
9949                 btrfs_abort_transaction(trans, ret);
9950                 goto out_fail;
9951         }
9952
9953         if (new_inode) {
9954                 inode_inc_iversion(new_inode);
9955                 new_inode->i_ctime = current_time(new_inode);
9956                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9957                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9958                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9959                         ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9960                                                 new_dentry->d_name.name,
9961                                                 new_dentry->d_name.len);
9962                         BUG_ON(new_inode->i_nlink == 0);
9963                 } else {
9964                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9965                                                  BTRFS_I(d_inode(new_dentry)),
9966                                                  new_dentry->d_name.name,
9967                                                  new_dentry->d_name.len);
9968                 }
9969                 if (!ret && new_inode->i_nlink == 0)
9970                         ret = btrfs_orphan_add(trans,
9971                                         BTRFS_I(d_inode(new_dentry)));
9972                 if (ret) {
9973                         btrfs_abort_transaction(trans, ret);
9974                         goto out_fail;
9975                 }
9976         }
9977
9978         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9979                              new_dentry->d_name.name,
9980                              new_dentry->d_name.len, 0, index);
9981         if (ret) {
9982                 btrfs_abort_transaction(trans, ret);
9983                 goto out_fail;
9984         }
9985
9986         if (old_inode->i_nlink == 1)
9987                 BTRFS_I(old_inode)->dir_index = index;
9988
9989         if (log_pinned) {
9990                 struct dentry *parent = new_dentry->d_parent;
9991
9992                 btrfs_init_log_ctx(&ctx, old_inode);
9993                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9994                                          BTRFS_I(old_dir), parent,
9995                                          false, &ctx);
9996                 if (ret == BTRFS_NEED_LOG_SYNC)
9997                         sync_log = true;
9998                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9999                         commit_transaction = true;
10000                 ret = 0;
10001                 btrfs_end_log_trans(root);
10002                 log_pinned = false;
10003         }
10004
10005         if (flags & RENAME_WHITEOUT) {
10006                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
10007                                                 old_dentry);
10008
10009                 if (ret) {
10010                         btrfs_abort_transaction(trans, ret);
10011                         goto out_fail;
10012                 }
10013         }
10014 out_fail:
10015         /*
10016          * If we have pinned the log and an error happened, we unpin tasks
10017          * trying to sync the log and force them to fallback to a transaction
10018          * commit if the log currently contains any of the inodes involved in
10019          * this rename operation (to ensure we do not persist a log with an
10020          * inconsistent state for any of these inodes or leading to any
10021          * inconsistencies when replayed). If the transaction was aborted, the
10022          * abortion reason is propagated to userspace when attempting to commit
10023          * the transaction. If the log does not contain any of these inodes, we
10024          * allow the tasks to sync it.
10025          */
10026         if (ret && log_pinned) {
10027                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
10028                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
10029                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10030                     (new_inode &&
10031                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10032                         btrfs_set_log_full_commit(trans);
10033
10034                 btrfs_end_log_trans(root);
10035                 log_pinned = false;
10036         }
10037         if (!ret && sync_log) {
10038                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root, &ctx);
10039                 if (ret)
10040                         commit_transaction = true;
10041         }
10042         if (commit_transaction) {
10043                 ret = btrfs_commit_transaction(trans);
10044         } else {
10045                 int ret2;
10046
10047                 ret2 = btrfs_end_transaction(trans);
10048                 ret = ret ? ret : ret2;
10049         }
10050 out_notrans:
10051         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10052                 up_read(&fs_info->subvol_sem);
10053
10054         return ret;
10055 }
10056
10057 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10058                          struct inode *new_dir, struct dentry *new_dentry,
10059                          unsigned int flags)
10060 {
10061         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10062                 return -EINVAL;
10063
10064         if (flags & RENAME_EXCHANGE)
10065                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10066                                           new_dentry);
10067
10068         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10069 }
10070
10071 struct btrfs_delalloc_work {
10072         struct inode *inode;
10073         struct completion completion;
10074         struct list_head list;
10075         struct btrfs_work work;
10076 };
10077
10078 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10079 {
10080         struct btrfs_delalloc_work *delalloc_work;
10081         struct inode *inode;
10082
10083         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10084                                      work);
10085         inode = delalloc_work->inode;
10086         filemap_flush(inode->i_mapping);
10087         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10088                                 &BTRFS_I(inode)->runtime_flags))
10089                 filemap_flush(inode->i_mapping);
10090
10091         iput(inode);
10092         complete(&delalloc_work->completion);
10093 }
10094
10095 static struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode)
10096 {
10097         struct btrfs_delalloc_work *work;
10098
10099         work = kmalloc(sizeof(*work), GFP_NOFS);
10100         if (!work)
10101                 return NULL;
10102
10103         init_completion(&work->completion);
10104         INIT_LIST_HEAD(&work->list);
10105         work->inode = inode;
10106         btrfs_init_work(&work->work, btrfs_run_delalloc_work, NULL, NULL);
10107
10108         return work;
10109 }
10110
10111 /*
10112  * some fairly slow code that needs optimization. This walks the list
10113  * of all the inodes with pending delalloc and forces them to disk.
10114  */
10115 static int start_delalloc_inodes(struct btrfs_root *root, int nr, bool snapshot)
10116 {
10117         struct btrfs_inode *binode;
10118         struct inode *inode;
10119         struct btrfs_delalloc_work *work, *next;
10120         struct list_head works;
10121         struct list_head splice;
10122         int ret = 0;
10123
10124         INIT_LIST_HEAD(&works);
10125         INIT_LIST_HEAD(&splice);
10126
10127         mutex_lock(&root->delalloc_mutex);
10128         spin_lock(&root->delalloc_lock);
10129         list_splice_init(&root->delalloc_inodes, &splice);
10130         while (!list_empty(&splice)) {
10131                 binode = list_entry(splice.next, struct btrfs_inode,
10132                                     delalloc_inodes);
10133
10134                 list_move_tail(&binode->delalloc_inodes,
10135                                &root->delalloc_inodes);
10136                 inode = igrab(&binode->vfs_inode);
10137                 if (!inode) {
10138                         cond_resched_lock(&root->delalloc_lock);
10139                         continue;
10140                 }
10141                 spin_unlock(&root->delalloc_lock);
10142
10143                 if (snapshot)
10144                         set_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
10145                                 &binode->runtime_flags);
10146                 work = btrfs_alloc_delalloc_work(inode);
10147                 if (!work) {
10148                         iput(inode);
10149                         ret = -ENOMEM;
10150                         goto out;
10151                 }
10152                 list_add_tail(&work->list, &works);
10153                 btrfs_queue_work(root->fs_info->flush_workers,
10154                                  &work->work);
10155                 ret++;
10156                 if (nr != -1 && ret >= nr)
10157                         goto out;
10158                 cond_resched();
10159                 spin_lock(&root->delalloc_lock);
10160         }
10161         spin_unlock(&root->delalloc_lock);
10162
10163 out:
10164         list_for_each_entry_safe(work, next, &works, list) {
10165                 list_del_init(&work->list);
10166                 wait_for_completion(&work->completion);
10167                 kfree(work);
10168         }
10169
10170         if (!list_empty(&splice)) {
10171                 spin_lock(&root->delalloc_lock);
10172                 list_splice_tail(&splice, &root->delalloc_inodes);
10173                 spin_unlock(&root->delalloc_lock);
10174         }
10175         mutex_unlock(&root->delalloc_mutex);
10176         return ret;
10177 }
10178
10179 int btrfs_start_delalloc_snapshot(struct btrfs_root *root)
10180 {
10181         struct btrfs_fs_info *fs_info = root->fs_info;
10182         int ret;
10183
10184         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10185                 return -EROFS;
10186
10187         ret = start_delalloc_inodes(root, -1, true);
10188         if (ret > 0)
10189                 ret = 0;
10190         return ret;
10191 }
10192
10193 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int nr)
10194 {
10195         struct btrfs_root *root;
10196         struct list_head splice;
10197         int ret;
10198
10199         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10200                 return -EROFS;
10201
10202         INIT_LIST_HEAD(&splice);
10203
10204         mutex_lock(&fs_info->delalloc_root_mutex);
10205         spin_lock(&fs_info->delalloc_root_lock);
10206         list_splice_init(&fs_info->delalloc_roots, &splice);
10207         while (!list_empty(&splice) && nr) {
10208                 root = list_first_entry(&splice, struct btrfs_root,
10209                                         delalloc_root);
10210                 root = btrfs_grab_fs_root(root);
10211                 BUG_ON(!root);
10212                 list_move_tail(&root->delalloc_root,
10213                                &fs_info->delalloc_roots);
10214                 spin_unlock(&fs_info->delalloc_root_lock);
10215
10216                 ret = start_delalloc_inodes(root, nr, false);
10217                 btrfs_put_fs_root(root);
10218                 if (ret < 0)
10219                         goto out;
10220
10221                 if (nr != -1) {
10222                         nr -= ret;
10223                         WARN_ON(nr < 0);
10224                 }
10225                 spin_lock(&fs_info->delalloc_root_lock);
10226         }
10227         spin_unlock(&fs_info->delalloc_root_lock);
10228
10229         ret = 0;
10230 out:
10231         if (!list_empty(&splice)) {
10232                 spin_lock(&fs_info->delalloc_root_lock);
10233                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10234                 spin_unlock(&fs_info->delalloc_root_lock);
10235         }
10236         mutex_unlock(&fs_info->delalloc_root_mutex);
10237         return ret;
10238 }
10239
10240 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10241                          const char *symname)
10242 {
10243         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10244         struct btrfs_trans_handle *trans;
10245         struct btrfs_root *root = BTRFS_I(dir)->root;
10246         struct btrfs_path *path;
10247         struct btrfs_key key;
10248         struct inode *inode = NULL;
10249         int err;
10250         u64 objectid;
10251         u64 index = 0;
10252         int name_len;
10253         int datasize;
10254         unsigned long ptr;
10255         struct btrfs_file_extent_item *ei;
10256         struct extent_buffer *leaf;
10257
10258         name_len = strlen(symname);
10259         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10260                 return -ENAMETOOLONG;
10261
10262         /*
10263          * 2 items for inode item and ref
10264          * 2 items for dir items
10265          * 1 item for updating parent inode item
10266          * 1 item for the inline extent item
10267          * 1 item for xattr if selinux is on
10268          */
10269         trans = btrfs_start_transaction(root, 7);
10270         if (IS_ERR(trans))
10271                 return PTR_ERR(trans);
10272
10273         err = btrfs_find_free_ino(root, &objectid);
10274         if (err)
10275                 goto out_unlock;
10276
10277         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10278                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10279                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10280         if (IS_ERR(inode)) {
10281                 err = PTR_ERR(inode);
10282                 inode = NULL;
10283                 goto out_unlock;
10284         }
10285
10286         /*
10287         * If the active LSM wants to access the inode during
10288         * d_instantiate it needs these. Smack checks to see
10289         * if the filesystem supports xattrs by looking at the
10290         * ops vector.
10291         */
10292         inode->i_fop = &btrfs_file_operations;
10293         inode->i_op = &btrfs_file_inode_operations;
10294         inode->i_mapping->a_ops = &btrfs_aops;
10295         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10296
10297         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10298         if (err)
10299                 goto out_unlock;
10300
10301         path = btrfs_alloc_path();
10302         if (!path) {
10303                 err = -ENOMEM;
10304                 goto out_unlock;
10305         }
10306         key.objectid = btrfs_ino(BTRFS_I(inode));
10307         key.offset = 0;
10308         key.type = BTRFS_EXTENT_DATA_KEY;
10309         datasize = btrfs_file_extent_calc_inline_size(name_len);
10310         err = btrfs_insert_empty_item(trans, root, path, &key,
10311                                       datasize);
10312         if (err) {
10313                 btrfs_free_path(path);
10314                 goto out_unlock;
10315         }
10316         leaf = path->nodes[0];
10317         ei = btrfs_item_ptr(leaf, path->slots[0],
10318                             struct btrfs_file_extent_item);
10319         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10320         btrfs_set_file_extent_type(leaf, ei,
10321                                    BTRFS_FILE_EXTENT_INLINE);
10322         btrfs_set_file_extent_encryption(leaf, ei, 0);
10323         btrfs_set_file_extent_compression(leaf, ei, 0);
10324         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10325         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10326
10327         ptr = btrfs_file_extent_inline_start(ei);
10328         write_extent_buffer(leaf, symname, ptr, name_len);
10329         btrfs_mark_buffer_dirty(leaf);
10330         btrfs_free_path(path);
10331
10332         inode->i_op = &btrfs_symlink_inode_operations;
10333         inode_nohighmem(inode);
10334         inode_set_bytes(inode, name_len);
10335         btrfs_i_size_write(BTRFS_I(inode), name_len);
10336         err = btrfs_update_inode(trans, root, inode);
10337         /*
10338          * Last step, add directory indexes for our symlink inode. This is the
10339          * last step to avoid extra cleanup of these indexes if an error happens
10340          * elsewhere above.
10341          */
10342         if (!err)
10343                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10344                                 BTRFS_I(inode), 0, index);
10345         if (err)
10346                 goto out_unlock;
10347
10348         d_instantiate_new(dentry, inode);
10349
10350 out_unlock:
10351         btrfs_end_transaction(trans);
10352         if (err && inode) {
10353                 inode_dec_link_count(inode);
10354                 discard_new_inode(inode);
10355         }
10356         btrfs_btree_balance_dirty(fs_info);
10357         return err;
10358 }
10359
10360 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10361                                        u64 start, u64 num_bytes, u64 min_size,
10362                                        loff_t actual_len, u64 *alloc_hint,
10363                                        struct btrfs_trans_handle *trans)
10364 {
10365         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10366         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10367         struct extent_map *em;
10368         struct btrfs_root *root = BTRFS_I(inode)->root;
10369         struct btrfs_key ins;
10370         u64 cur_offset = start;
10371         u64 i_size;
10372         u64 cur_bytes;
10373         u64 last_alloc = (u64)-1;
10374         int ret = 0;
10375         bool own_trans = true;
10376         u64 end = start + num_bytes - 1;
10377
10378         if (trans)
10379                 own_trans = false;
10380         while (num_bytes > 0) {
10381                 if (own_trans) {
10382                         trans = btrfs_start_transaction(root, 3);
10383                         if (IS_ERR(trans)) {
10384                                 ret = PTR_ERR(trans);
10385                                 break;
10386                         }
10387                 }
10388
10389                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10390                 cur_bytes = max(cur_bytes, min_size);
10391                 /*
10392                  * If we are severely fragmented we could end up with really
10393                  * small allocations, so if the allocator is returning small
10394                  * chunks lets make its job easier by only searching for those
10395                  * sized chunks.
10396                  */
10397                 cur_bytes = min(cur_bytes, last_alloc);
10398                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10399                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10400                 if (ret) {
10401                         if (own_trans)
10402                                 btrfs_end_transaction(trans);
10403                         break;
10404                 }
10405                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10406
10407                 last_alloc = ins.offset;
10408                 ret = insert_reserved_file_extent(trans, inode,
10409                                                   cur_offset, ins.objectid,
10410                                                   ins.offset, ins.offset,
10411                                                   ins.offset, 0, 0, 0,
10412                                                   BTRFS_FILE_EXTENT_PREALLOC);
10413                 if (ret) {
10414                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10415                                                    ins.offset, 0);
10416                         btrfs_abort_transaction(trans, ret);
10417                         if (own_trans)
10418                                 btrfs_end_transaction(trans);
10419                         break;
10420                 }
10421
10422                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10423                                         cur_offset + ins.offset -1, 0);
10424
10425                 em = alloc_extent_map();
10426                 if (!em) {
10427                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10428                                 &BTRFS_I(inode)->runtime_flags);
10429                         goto next;
10430                 }
10431
10432                 em->start = cur_offset;
10433                 em->orig_start = cur_offset;
10434                 em->len = ins.offset;
10435                 em->block_start = ins.objectid;
10436                 em->block_len = ins.offset;
10437                 em->orig_block_len = ins.offset;
10438                 em->ram_bytes = ins.offset;
10439                 em->bdev = fs_info->fs_devices->latest_bdev;
10440                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10441                 em->generation = trans->transid;
10442
10443                 while (1) {
10444                         write_lock(&em_tree->lock);
10445                         ret = add_extent_mapping(em_tree, em, 1);
10446                         write_unlock(&em_tree->lock);
10447                         if (ret != -EEXIST)
10448                                 break;
10449                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10450                                                 cur_offset + ins.offset - 1,
10451                                                 0);
10452                 }
10453                 free_extent_map(em);
10454 next:
10455                 num_bytes -= ins.offset;
10456                 cur_offset += ins.offset;
10457                 *alloc_hint = ins.objectid + ins.offset;
10458
10459                 inode_inc_iversion(inode);
10460                 inode->i_ctime = current_time(inode);
10461                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10462                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10463                     (actual_len > inode->i_size) &&
10464                     (cur_offset > inode->i_size)) {
10465                         if (cur_offset > actual_len)
10466                                 i_size = actual_len;
10467                         else
10468                                 i_size = cur_offset;
10469                         i_size_write(inode, i_size);
10470                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10471                 }
10472
10473                 ret = btrfs_update_inode(trans, root, inode);
10474
10475                 if (ret) {
10476                         btrfs_abort_transaction(trans, ret);
10477                         if (own_trans)
10478                                 btrfs_end_transaction(trans);
10479                         break;
10480                 }
10481
10482                 if (own_trans)
10483                         btrfs_end_transaction(trans);
10484         }
10485         if (cur_offset < end)
10486                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10487                         end - cur_offset + 1);
10488         return ret;
10489 }
10490
10491 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10492                               u64 start, u64 num_bytes, u64 min_size,
10493                               loff_t actual_len, u64 *alloc_hint)
10494 {
10495         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10496                                            min_size, actual_len, alloc_hint,
10497                                            NULL);
10498 }
10499
10500 int btrfs_prealloc_file_range_trans(struct inode *inode,
10501                                     struct btrfs_trans_handle *trans, int mode,
10502                                     u64 start, u64 num_bytes, u64 min_size,
10503                                     loff_t actual_len, u64 *alloc_hint)
10504 {
10505         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10506                                            min_size, actual_len, alloc_hint, trans);
10507 }
10508
10509 static int btrfs_set_page_dirty(struct page *page)
10510 {
10511         return __set_page_dirty_nobuffers(page);
10512 }
10513
10514 static int btrfs_permission(struct inode *inode, int mask)
10515 {
10516         struct btrfs_root *root = BTRFS_I(inode)->root;
10517         umode_t mode = inode->i_mode;
10518
10519         if (mask & MAY_WRITE &&
10520             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10521                 if (btrfs_root_readonly(root))
10522                         return -EROFS;
10523                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10524                         return -EACCES;
10525         }
10526         return generic_permission(inode, mask);
10527 }
10528
10529 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10530 {
10531         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10532         struct btrfs_trans_handle *trans;
10533         struct btrfs_root *root = BTRFS_I(dir)->root;
10534         struct inode *inode = NULL;
10535         u64 objectid;
10536         u64 index;
10537         int ret = 0;
10538
10539         /*
10540          * 5 units required for adding orphan entry
10541          */
10542         trans = btrfs_start_transaction(root, 5);
10543         if (IS_ERR(trans))
10544                 return PTR_ERR(trans);
10545
10546         ret = btrfs_find_free_ino(root, &objectid);
10547         if (ret)
10548                 goto out;
10549
10550         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10551                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10552         if (IS_ERR(inode)) {
10553                 ret = PTR_ERR(inode);
10554                 inode = NULL;
10555                 goto out;
10556         }
10557
10558         inode->i_fop = &btrfs_file_operations;
10559         inode->i_op = &btrfs_file_inode_operations;
10560
10561         inode->i_mapping->a_ops = &btrfs_aops;
10562         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10563
10564         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10565         if (ret)
10566                 goto out;
10567
10568         ret = btrfs_update_inode(trans, root, inode);
10569         if (ret)
10570                 goto out;
10571         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10572         if (ret)
10573                 goto out;
10574
10575         /*
10576          * We set number of links to 0 in btrfs_new_inode(), and here we set
10577          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10578          * through:
10579          *
10580          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10581          */
10582         set_nlink(inode, 1);
10583         d_tmpfile(dentry, inode);
10584         unlock_new_inode(inode);
10585         mark_inode_dirty(inode);
10586 out:
10587         btrfs_end_transaction(trans);
10588         if (ret && inode)
10589                 discard_new_inode(inode);
10590         btrfs_btree_balance_dirty(fs_info);
10591         return ret;
10592 }
10593
10594 void btrfs_set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
10595 {
10596         struct inode *inode = tree->private_data;
10597         unsigned long index = start >> PAGE_SHIFT;
10598         unsigned long end_index = end >> PAGE_SHIFT;
10599         struct page *page;
10600
10601         while (index <= end_index) {
10602                 page = find_get_page(inode->i_mapping, index);
10603                 ASSERT(page); /* Pages should be in the extent_io_tree */
10604                 set_page_writeback(page);
10605                 put_page(page);
10606                 index++;
10607         }
10608 }
10609
10610 #ifdef CONFIG_SWAP
10611 /*
10612  * Add an entry indicating a block group or device which is pinned by a
10613  * swapfile. Returns 0 on success, 1 if there is already an entry for it, or a
10614  * negative errno on failure.
10615  */
10616 static int btrfs_add_swapfile_pin(struct inode *inode, void *ptr,
10617                                   bool is_block_group)
10618 {
10619         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10620         struct btrfs_swapfile_pin *sp, *entry;
10621         struct rb_node **p;
10622         struct rb_node *parent = NULL;
10623
10624         sp = kmalloc(sizeof(*sp), GFP_NOFS);
10625         if (!sp)
10626                 return -ENOMEM;
10627         sp->ptr = ptr;
10628         sp->inode = inode;
10629         sp->is_block_group = is_block_group;
10630
10631         spin_lock(&fs_info->swapfile_pins_lock);
10632         p = &fs_info->swapfile_pins.rb_node;
10633         while (*p) {
10634                 parent = *p;
10635                 entry = rb_entry(parent, struct btrfs_swapfile_pin, node);
10636                 if (sp->ptr < entry->ptr ||
10637                     (sp->ptr == entry->ptr && sp->inode < entry->inode)) {
10638                         p = &(*p)->rb_left;
10639                 } else if (sp->ptr > entry->ptr ||
10640                            (sp->ptr == entry->ptr && sp->inode > entry->inode)) {
10641                         p = &(*p)->rb_right;
10642                 } else {
10643                         spin_unlock(&fs_info->swapfile_pins_lock);
10644                         kfree(sp);
10645                         return 1;
10646                 }
10647         }
10648         rb_link_node(&sp->node, parent, p);
10649         rb_insert_color(&sp->node, &fs_info->swapfile_pins);
10650         spin_unlock(&fs_info->swapfile_pins_lock);
10651         return 0;
10652 }
10653
10654 /* Free all of the entries pinned by this swapfile. */
10655 static void btrfs_free_swapfile_pins(struct inode *inode)
10656 {
10657         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10658         struct btrfs_swapfile_pin *sp;
10659         struct rb_node *node, *next;
10660
10661         spin_lock(&fs_info->swapfile_pins_lock);
10662         node = rb_first(&fs_info->swapfile_pins);
10663         while (node) {
10664                 next = rb_next(node);
10665                 sp = rb_entry(node, struct btrfs_swapfile_pin, node);
10666                 if (sp->inode == inode) {
10667                         rb_erase(&sp->node, &fs_info->swapfile_pins);
10668                         if (sp->is_block_group)
10669                                 btrfs_put_block_group(sp->ptr);
10670                         kfree(sp);
10671                 }
10672                 node = next;
10673         }
10674         spin_unlock(&fs_info->swapfile_pins_lock);
10675 }
10676
10677 struct btrfs_swap_info {
10678         u64 start;
10679         u64 block_start;
10680         u64 block_len;
10681         u64 lowest_ppage;
10682         u64 highest_ppage;
10683         unsigned long nr_pages;
10684         int nr_extents;
10685 };
10686
10687 static int btrfs_add_swap_extent(struct swap_info_struct *sis,
10688                                  struct btrfs_swap_info *bsi)
10689 {
10690         unsigned long nr_pages;
10691         u64 first_ppage, first_ppage_reported, next_ppage;
10692         int ret;
10693
10694         first_ppage = ALIGN(bsi->block_start, PAGE_SIZE) >> PAGE_SHIFT;
10695         next_ppage = ALIGN_DOWN(bsi->block_start + bsi->block_len,
10696                                 PAGE_SIZE) >> PAGE_SHIFT;
10697
10698         if (first_ppage >= next_ppage)
10699                 return 0;
10700         nr_pages = next_ppage - first_ppage;
10701
10702         first_ppage_reported = first_ppage;
10703         if (bsi->start == 0)
10704                 first_ppage_reported++;
10705         if (bsi->lowest_ppage > first_ppage_reported)
10706                 bsi->lowest_ppage = first_ppage_reported;
10707         if (bsi->highest_ppage < (next_ppage - 1))
10708                 bsi->highest_ppage = next_ppage - 1;
10709
10710         ret = add_swap_extent(sis, bsi->nr_pages, nr_pages, first_ppage);
10711         if (ret < 0)
10712                 return ret;
10713         bsi->nr_extents += ret;
10714         bsi->nr_pages += nr_pages;
10715         return 0;
10716 }
10717
10718 static void btrfs_swap_deactivate(struct file *file)
10719 {
10720         struct inode *inode = file_inode(file);
10721
10722         btrfs_free_swapfile_pins(inode);
10723         atomic_dec(&BTRFS_I(inode)->root->nr_swapfiles);
10724 }
10725
10726 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10727                                sector_t *span)
10728 {
10729         struct inode *inode = file_inode(file);
10730         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10731         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
10732         struct extent_state *cached_state = NULL;
10733         struct extent_map *em = NULL;
10734         struct btrfs_device *device = NULL;
10735         struct btrfs_swap_info bsi = {
10736                 .lowest_ppage = (sector_t)-1ULL,
10737         };
10738         int ret = 0;
10739         u64 isize;
10740         u64 start;
10741
10742         /*
10743          * If the swap file was just created, make sure delalloc is done. If the
10744          * file changes again after this, the user is doing something stupid and
10745          * we don't really care.
10746          */
10747         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
10748         if (ret)
10749                 return ret;
10750
10751         /*
10752          * The inode is locked, so these flags won't change after we check them.
10753          */
10754         if (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS) {
10755                 btrfs_warn(fs_info, "swapfile must not be compressed");
10756                 return -EINVAL;
10757         }
10758         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)) {
10759                 btrfs_warn(fs_info, "swapfile must not be copy-on-write");
10760                 return -EINVAL;
10761         }
10762         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
10763                 btrfs_warn(fs_info, "swapfile must not be checksummed");
10764                 return -EINVAL;
10765         }
10766
10767         /*
10768          * Balance or device remove/replace/resize can move stuff around from
10769          * under us. The EXCL_OP flag makes sure they aren't running/won't run
10770          * concurrently while we are mapping the swap extents, and
10771          * fs_info->swapfile_pins prevents them from running while the swap file
10772          * is active and moving the extents. Note that this also prevents a
10773          * concurrent device add which isn't actually necessary, but it's not
10774          * really worth the trouble to allow it.
10775          */
10776         if (test_and_set_bit(BTRFS_FS_EXCL_OP, &fs_info->flags)) {
10777                 btrfs_warn(fs_info,
10778            "cannot activate swapfile while exclusive operation is running");
10779                 return -EBUSY;
10780         }
10781         /*
10782          * Snapshots can create extents which require COW even if NODATACOW is
10783          * set. We use this counter to prevent snapshots. We must increment it
10784          * before walking the extents because we don't want a concurrent
10785          * snapshot to run after we've already checked the extents.
10786          */
10787         atomic_inc(&BTRFS_I(inode)->root->nr_swapfiles);
10788
10789         isize = ALIGN_DOWN(inode->i_size, fs_info->sectorsize);
10790
10791         lock_extent_bits(io_tree, 0, isize - 1, &cached_state);
10792         start = 0;
10793         while (start < isize) {
10794                 u64 logical_block_start, physical_block_start;
10795                 struct btrfs_block_group_cache *bg;
10796                 u64 len = isize - start;
10797
10798                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
10799                 if (IS_ERR(em)) {
10800                         ret = PTR_ERR(em);
10801                         goto out;
10802                 }
10803
10804                 if (em->block_start == EXTENT_MAP_HOLE) {
10805                         btrfs_warn(fs_info, "swapfile must not have holes");
10806                         ret = -EINVAL;
10807                         goto out;
10808                 }
10809                 if (em->block_start == EXTENT_MAP_INLINE) {
10810                         /*
10811                          * It's unlikely we'll ever actually find ourselves
10812                          * here, as a file small enough to fit inline won't be
10813                          * big enough to store more than the swap header, but in
10814                          * case something changes in the future, let's catch it
10815                          * here rather than later.
10816                          */
10817                         btrfs_warn(fs_info, "swapfile must not be inline");
10818                         ret = -EINVAL;
10819                         goto out;
10820                 }
10821                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
10822                         btrfs_warn(fs_info, "swapfile must not be compressed");
10823                         ret = -EINVAL;
10824                         goto out;
10825                 }
10826
10827                 logical_block_start = em->block_start + (start - em->start);
10828                 len = min(len, em->len - (start - em->start));
10829                 free_extent_map(em);
10830                 em = NULL;
10831
10832                 ret = can_nocow_extent(inode, start, &len, NULL, NULL, NULL);
10833                 if (ret < 0) {
10834                         goto out;
10835                 } else if (ret) {
10836                         ret = 0;
10837                 } else {
10838                         btrfs_warn(fs_info,
10839                                    "swapfile must not be copy-on-write");
10840                         ret = -EINVAL;
10841                         goto out;
10842                 }
10843
10844                 em = btrfs_get_chunk_map(fs_info, logical_block_start, len);
10845                 if (IS_ERR(em)) {
10846                         ret = PTR_ERR(em);
10847                         goto out;
10848                 }
10849
10850                 if (em->map_lookup->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
10851                         btrfs_warn(fs_info,
10852                                    "swapfile must have single data profile");
10853                         ret = -EINVAL;
10854                         goto out;
10855                 }
10856
10857                 if (device == NULL) {
10858                         device = em->map_lookup->stripes[0].dev;
10859                         ret = btrfs_add_swapfile_pin(inode, device, false);
10860                         if (ret == 1)
10861                                 ret = 0;
10862                         else if (ret)
10863                                 goto out;
10864                 } else if (device != em->map_lookup->stripes[0].dev) {
10865                         btrfs_warn(fs_info, "swapfile must be on one device");
10866                         ret = -EINVAL;
10867                         goto out;
10868                 }
10869
10870                 physical_block_start = (em->map_lookup->stripes[0].physical +
10871                                         (logical_block_start - em->start));
10872                 len = min(len, em->len - (logical_block_start - em->start));
10873                 free_extent_map(em);
10874                 em = NULL;
10875
10876                 bg = btrfs_lookup_block_group(fs_info, logical_block_start);
10877                 if (!bg) {
10878                         btrfs_warn(fs_info,
10879                            "could not find block group containing swapfile");
10880                         ret = -EINVAL;
10881                         goto out;
10882                 }
10883
10884                 ret = btrfs_add_swapfile_pin(inode, bg, true);
10885                 if (ret) {
10886                         btrfs_put_block_group(bg);
10887                         if (ret == 1)
10888                                 ret = 0;
10889                         else
10890                                 goto out;
10891                 }
10892
10893                 if (bsi.block_len &&
10894                     bsi.block_start + bsi.block_len == physical_block_start) {
10895                         bsi.block_len += len;
10896                 } else {
10897                         if (bsi.block_len) {
10898                                 ret = btrfs_add_swap_extent(sis, &bsi);
10899                                 if (ret)
10900                                         goto out;
10901                         }
10902                         bsi.start = start;
10903                         bsi.block_start = physical_block_start;
10904                         bsi.block_len = len;
10905                 }
10906
10907                 start += len;
10908         }
10909
10910         if (bsi.block_len)
10911                 ret = btrfs_add_swap_extent(sis, &bsi);
10912
10913 out:
10914         if (!IS_ERR_OR_NULL(em))
10915                 free_extent_map(em);
10916
10917         unlock_extent_cached(io_tree, 0, isize - 1, &cached_state);
10918
10919         if (ret)
10920                 btrfs_swap_deactivate(file);
10921
10922         clear_bit(BTRFS_FS_EXCL_OP, &fs_info->flags);
10923
10924         if (ret)
10925                 return ret;
10926
10927         if (device)
10928                 sis->bdev = device->bdev;
10929         *span = bsi.highest_ppage - bsi.lowest_ppage + 1;
10930         sis->max = bsi.nr_pages;
10931         sis->pages = bsi.nr_pages - 1;
10932         sis->highest_bit = bsi.nr_pages - 1;
10933         return bsi.nr_extents;
10934 }
10935 #else
10936 static void btrfs_swap_deactivate(struct file *file)
10937 {
10938 }
10939
10940 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10941                                sector_t *span)
10942 {
10943         return -EOPNOTSUPP;
10944 }
10945 #endif
10946
10947 static const struct inode_operations btrfs_dir_inode_operations = {
10948         .getattr        = btrfs_getattr,
10949         .lookup         = btrfs_lookup,
10950         .create         = btrfs_create,
10951         .unlink         = btrfs_unlink,
10952         .link           = btrfs_link,
10953         .mkdir          = btrfs_mkdir,
10954         .rmdir          = btrfs_rmdir,
10955         .rename         = btrfs_rename2,
10956         .symlink        = btrfs_symlink,
10957         .setattr        = btrfs_setattr,
10958         .mknod          = btrfs_mknod,
10959         .listxattr      = btrfs_listxattr,
10960         .permission     = btrfs_permission,
10961         .get_acl        = btrfs_get_acl,
10962         .set_acl        = btrfs_set_acl,
10963         .update_time    = btrfs_update_time,
10964         .tmpfile        = btrfs_tmpfile,
10965 };
10966 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10967         .lookup         = btrfs_lookup,
10968         .permission     = btrfs_permission,
10969         .update_time    = btrfs_update_time,
10970 };
10971
10972 static const struct file_operations btrfs_dir_file_operations = {
10973         .llseek         = generic_file_llseek,
10974         .read           = generic_read_dir,
10975         .iterate_shared = btrfs_real_readdir,
10976         .open           = btrfs_opendir,
10977         .unlocked_ioctl = btrfs_ioctl,
10978 #ifdef CONFIG_COMPAT
10979         .compat_ioctl   = btrfs_compat_ioctl,
10980 #endif
10981         .release        = btrfs_release_file,
10982         .fsync          = btrfs_sync_file,
10983 };
10984
10985 static const struct extent_io_ops btrfs_extent_io_ops = {
10986         /* mandatory callbacks */
10987         .submit_bio_hook = btrfs_submit_bio_hook,
10988         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10989 };
10990
10991 /*
10992  * btrfs doesn't support the bmap operation because swapfiles
10993  * use bmap to make a mapping of extents in the file.  They assume
10994  * these extents won't change over the life of the file and they
10995  * use the bmap result to do IO directly to the drive.
10996  *
10997  * the btrfs bmap call would return logical addresses that aren't
10998  * suitable for IO and they also will change frequently as COW
10999  * operations happen.  So, swapfile + btrfs == corruption.
11000  *
11001  * For now we're avoiding this by dropping bmap.
11002  */
11003 static const struct address_space_operations btrfs_aops = {
11004         .readpage       = btrfs_readpage,
11005         .writepage      = btrfs_writepage,
11006         .writepages     = btrfs_writepages,
11007         .readpages      = btrfs_readpages,
11008         .direct_IO      = btrfs_direct_IO,
11009         .invalidatepage = btrfs_invalidatepage,
11010         .releasepage    = btrfs_releasepage,
11011         .set_page_dirty = btrfs_set_page_dirty,
11012         .error_remove_page = generic_error_remove_page,
11013         .swap_activate  = btrfs_swap_activate,
11014         .swap_deactivate = btrfs_swap_deactivate,
11015 };
11016
11017 static const struct inode_operations btrfs_file_inode_operations = {
11018         .getattr        = btrfs_getattr,
11019         .setattr        = btrfs_setattr,
11020         .listxattr      = btrfs_listxattr,
11021         .permission     = btrfs_permission,
11022         .fiemap         = btrfs_fiemap,
11023         .get_acl        = btrfs_get_acl,
11024         .set_acl        = btrfs_set_acl,
11025         .update_time    = btrfs_update_time,
11026 };
11027 static const struct inode_operations btrfs_special_inode_operations = {
11028         .getattr        = btrfs_getattr,
11029         .setattr        = btrfs_setattr,
11030         .permission     = btrfs_permission,
11031         .listxattr      = btrfs_listxattr,
11032         .get_acl        = btrfs_get_acl,
11033         .set_acl        = btrfs_set_acl,
11034         .update_time    = btrfs_update_time,
11035 };
11036 static const struct inode_operations btrfs_symlink_inode_operations = {
11037         .get_link       = page_get_link,
11038         .getattr        = btrfs_getattr,
11039         .setattr        = btrfs_setattr,
11040         .permission     = btrfs_permission,
11041         .listxattr      = btrfs_listxattr,
11042         .update_time    = btrfs_update_time,
11043 };
11044
11045 const struct dentry_operations btrfs_dentry_operations = {
11046         .d_delete       = btrfs_dentry_delete,
11047 };