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