fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/compat.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/xattr.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/falloc.h>
  38 #include <linux/slab.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/mount.h>
  41 #include <linux/btrfs.h>
  42 #include <linux/blkdev.h>
  43 #include <linux/posix_acl_xattr.h>
  44 #include <linux/uio.h>
  45 #include <linux/magic.h>
  46 #include <linux/iversion.h>
  47 #include "ctree.h"
  48 #include "disk-io.h"
  49 #include "transaction.h"
  50 #include "btrfs_inode.h"
  51 #include "print-tree.h"
  52 #include "ordered-data.h"
  53 #include "xattr.h"
  54 #include "tree-log.h"
  55 #include "volumes.h"
  56 #include "compression.h"
  57 #include "locking.h"
  58 #include "free-space-cache.h"
  59 #include "inode-map.h"
  60 #include "backref.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63 #include "dedupe.h"
  64
  65 struct btrfs_iget_args {
  66         struct btrfs_key *location;
  67         struct btrfs_root *root;
  68 };
  69
  70 struct btrfs_dio_data {
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74         int overwrite;
  75 };
  76
  77 static const struct inode_operations btrfs_dir_inode_operations;
  78 static const struct inode_operations btrfs_symlink_inode_operations;
  79 static const struct inode_operations btrfs_dir_ro_inode_operations;
  80 static const struct inode_operations btrfs_special_inode_operations;
  81 static const struct inode_operations btrfs_file_inode_operations;
  82 static const struct address_space_operations btrfs_aops;
  83 static const struct address_space_operations btrfs_symlink_aops;
  84 static const struct file_operations btrfs_dir_file_operations;
  85 static const struct extent_io_ops btrfs_extent_io_ops;
  86
  87 static struct kmem_cache *btrfs_inode_cachep;
  88 struct kmem_cache *btrfs_trans_handle_cachep;
  89 struct kmem_cache *btrfs_path_cachep;
  90 struct kmem_cache *btrfs_free_space_cachep;
  91
  92 #define S_SHIFT 12
  93 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  94         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  95         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  96         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  97         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  98         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  99         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 100         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 101 };
 102
 103 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 104 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
 105 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 106 static noinline int cow_file_range(struct inode *inode,
 107                                    struct page *locked_page,
 108                                    u64 start, u64 end, u64 delalloc_end,
 109                                    int *page_started, unsigned long *nr_written,
 110                                    int unlock, struct btrfs_dedupe_hash *hash);
 111 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 112                                        u64 orig_start, u64 block_start,
 113                                        u64 block_len, u64 orig_block_len,
 114                                        u64 ram_bytes, int compress_type,
 115                                        int type);
 116
 117 static void __endio_write_update_ordered(struct inode *inode,
 118                                          const u64 offset, const u64 bytes,
 119                                          const bool uptodate);
 120
 121 /*
 122  * Cleanup all submitted ordered extents in specified range to handle errors
 123  * from the fill_dellaloc() callback.
 124  *
 125  * NOTE: caller must ensure that when an error happens, it can not call
 126  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 127  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 128  * to be released, which we want to happen only when finishing the ordered
 129  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 130  * fill_delalloc() callback already does proper cleanup for the first page of
 131  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 132  * range of the first page.
 133  */
 134 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 135                                                  const u64 offset,
 136                                                  const u64 bytes)
 137 {
 138         unsigned long index = offset >> PAGE_SHIFT;
 139         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 140         struct page *page;
 141
 142         while (index <= end_index) {
 143                 page = find_get_page(inode->i_mapping, index);
 144                 index++;
 145                 if (!page)
 146                         continue;
 147                 ClearPagePrivate2(page);
 148                 put_page(page);
 149         }
 150         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 151                                             bytes - PAGE_SIZE, false);
 152 }
 153
 154 static int btrfs_dirty_inode(struct inode *inode);
 155
 156 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 157 void btrfs_test_inode_set_ops(struct inode *inode)
 158 {
 159         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 160 }
 161 #endif
 162
 163 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 164                                      struct inode *inode,  struct inode *dir,
 165                                      const struct qstr *qstr)
 166 {
 167         int err;
 168
 169         err = btrfs_init_acl(trans, inode, dir);
 170         if (!err)
 171                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 172         return err;
 173 }
 174
 175 /*
 176  * this does all the hard work for inserting an inline extent into
 177  * the btree.  The caller should have done a btrfs_drop_extents so that
 178  * no overlapping inline items exist in the btree
 179  */
 180 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 181                                 struct btrfs_path *path, int extent_inserted,
 182                                 struct btrfs_root *root, struct inode *inode,
 183                                 u64 start, size_t size, size_t compressed_size,
 184                                 int compress_type,
 185                                 struct page **compressed_pages)
 186 {
 187         struct extent_buffer *leaf;
 188         struct page *page = NULL;
 189         char *kaddr;
 190         unsigned long ptr;
 191         struct btrfs_file_extent_item *ei;
 192         int ret;
 193         size_t cur_size = size;
 194         unsigned long offset;
 195
 196         if (compressed_size && compressed_pages)
 197                 cur_size = compressed_size;
 198
 199         inode_add_bytes(inode, size);
 200
 201         if (!extent_inserted) {
 202                 struct btrfs_key key;
 203                 size_t datasize;
 204
 205                 key.objectid = btrfs_ino(BTRFS_I(inode));
 206                 key.offset = start;
 207                 key.type = BTRFS_EXTENT_DATA_KEY;
 208
 209                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 210                 path->leave_spinning = 1;
 211                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 212                                               datasize);
 213                 if (ret)
 214                         goto fail;
 215         }
 216         leaf = path->nodes[0];
 217         ei = btrfs_item_ptr(leaf, path->slots[0],
 218                             struct btrfs_file_extent_item);
 219         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 220         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 221         btrfs_set_file_extent_encryption(leaf, ei, 0);
 222         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 223         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 224         ptr = btrfs_file_extent_inline_start(ei);
 225
 226         if (compress_type != BTRFS_COMPRESS_NONE) {
 227                 struct page *cpage;
 228                 int i = 0;
 229                 while (compressed_size > 0) {
 230                         cpage = compressed_pages[i];
 231                         cur_size = min_t(unsigned long, compressed_size,
 232                                        PAGE_SIZE);
 233
 234                         kaddr = kmap_atomic(cpage);
 235                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 236                         kunmap_atomic(kaddr);
 237
 238                         i++;
 239                         ptr += cur_size;
 240                         compressed_size -= cur_size;
 241                 }
 242                 btrfs_set_file_extent_compression(leaf, ei,
 243                                                   compress_type);
 244         } else {
 245                 page = find_get_page(inode->i_mapping,
 246                                      start >> PAGE_SHIFT);
 247                 btrfs_set_file_extent_compression(leaf, ei, 0);
 248                 kaddr = kmap_atomic(page);
 249                 offset = start & (PAGE_SIZE - 1);
 250                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 251                 kunmap_atomic(kaddr);
 252                 put_page(page);
 253         }
 254         btrfs_mark_buffer_dirty(leaf);
 255         btrfs_release_path(path);
 256
 257         /*
 258          * we're an inline extent, so nobody can
 259          * extend the file past i_size without locking
 260          * a page we already have locked.
 261          *
 262          * We must do any isize and inode updates
 263          * before we unlock the pages.  Otherwise we
 264          * could end up racing with unlink.
 265          */
 266         BTRFS_I(inode)->disk_i_size = inode->i_size;
 267         ret = btrfs_update_inode(trans, root, inode);
 268
 269 fail:
 270         return ret;
 271 }
 272
 273
 274 /*
 275  * conditionally insert an inline extent into the file.  This
 276  * does the checks required to make sure the data is small enough
 277  * to fit as an inline extent.
 278  */
 279 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 280                                           u64 end, size_t compressed_size,
 281                                           int compress_type,
 282                                           struct page **compressed_pages)
 283 {
 284         struct btrfs_root *root = BTRFS_I(inode)->root;
 285         struct btrfs_fs_info *fs_info = root->fs_info;
 286         struct btrfs_trans_handle *trans;
 287         u64 isize = i_size_read(inode);
 288         u64 actual_end = min(end + 1, isize);
 289         u64 inline_len = actual_end - start;
 290         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 291         u64 data_len = inline_len;
 292         int ret;
 293         struct btrfs_path *path;
 294         int extent_inserted = 0;
 295         u32 extent_item_size;
 296
 297         if (compressed_size)
 298                 data_len = compressed_size;
 299
 300         if (start > 0 ||
 301             actual_end > fs_info->sectorsize ||
 302             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 303             (!compressed_size &&
 304             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 305             end + 1 < isize ||
 306             data_len > fs_info->max_inline) {
 307                 return 1;
 308         }
 309
 310         path = btrfs_alloc_path();
 311         if (!path)
 312                 return -ENOMEM;
 313
 314         trans = btrfs_join_transaction(root);
 315         if (IS_ERR(trans)) {
 316                 btrfs_free_path(path);
 317                 return PTR_ERR(trans);
 318         }
 319         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 320
 321         if (compressed_size && compressed_pages)
 322                 extent_item_size = btrfs_file_extent_calc_inline_size(
 323                    compressed_size);
 324         else
 325                 extent_item_size = btrfs_file_extent_calc_inline_size(
 326                     inline_len);
 327
 328         ret = __btrfs_drop_extents(trans, root, inode, path,
 329                                    start, aligned_end, NULL,
 330                                    1, 1, extent_item_size, &extent_inserted);
 331         if (ret) {
 332                 btrfs_abort_transaction(trans, ret);
 333                 goto out;
 334         }
 335
 336         if (isize > actual_end)
 337                 inline_len = min_t(u64, isize, actual_end);
 338         ret = insert_inline_extent(trans, path, extent_inserted,
 339                                    root, inode, start,
 340                                    inline_len, compressed_size,
 341                                    compress_type, compressed_pages);
 342         if (ret && ret != -ENOSPC) {
 343                 btrfs_abort_transaction(trans, ret);
 344                 goto out;
 345         } else if (ret == -ENOSPC) {
 346                 ret = 1;
 347                 goto out;
 348         }
 349
 350         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 351         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 352 out:
 353         /*
 354          * Don't forget to free the reserved space, as for inlined extent
 355          * it won't count as data extent, free them directly here.
 356          * And at reserve time, it's always aligned to page size, so
 357          * just free one page here.
 358          */
 359         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 360         btrfs_free_path(path);
 361         btrfs_end_transaction(trans);
 362         return ret;
 363 }
 364
 365 struct async_extent {
 366         u64 start;
 367         u64 ram_size;
 368         u64 compressed_size;
 369         struct page **pages;
 370         unsigned long nr_pages;
 371         int compress_type;
 372         struct list_head list;
 373 };
 374
 375 struct async_cow {
 376         struct inode *inode;
 377         struct btrfs_root *root;
 378         struct page *locked_page;
 379         u64 start;
 380         u64 end;
 381         unsigned int write_flags;
 382         struct list_head extents;
 383         struct btrfs_work work;
 384 };
 385
 386 static noinline int add_async_extent(struct async_cow *cow,
 387                                      u64 start, u64 ram_size,
 388                                      u64 compressed_size,
 389                                      struct page **pages,
 390                                      unsigned long nr_pages,
 391                                      int compress_type)
 392 {
 393         struct async_extent *async_extent;
 394
 395         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 396         BUG_ON(!async_extent); /* -ENOMEM */
 397         async_extent->start = start;
 398         async_extent->ram_size = ram_size;
 399         async_extent->compressed_size = compressed_size;
 400         async_extent->pages = pages;
 401         async_extent->nr_pages = nr_pages;
 402         async_extent->compress_type = compress_type;
 403         list_add_tail(&async_extent->list, &cow->extents);
 404         return 0;
 405 }
 406
 407 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 408 {
 409         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 410
 411         /* force compress */
 412         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 413                 return 1;
 414         /* defrag ioctl */
 415         if (BTRFS_I(inode)->defrag_compress)
 416                 return 1;
 417         /* bad compression ratios */
 418         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 419                 return 0;
 420         if (btrfs_test_opt(fs_info, COMPRESS) ||
 421             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 422             BTRFS_I(inode)->prop_compress)
 423                 return btrfs_compress_heuristic(inode, start, end);
 424         return 0;
 425 }
 426
 427 static inline void inode_should_defrag(struct btrfs_inode *inode,
 428                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 429 {
 430         /* If this is a small write inside eof, kick off a defrag */
 431         if (num_bytes < small_write &&
 432             (start > 0 || end + 1 < inode->disk_i_size))
 433                 btrfs_add_inode_defrag(NULL, inode);
 434 }
 435
 436 /*
 437  * we create compressed extents in two phases.  The first
 438  * phase compresses a range of pages that have already been
 439  * locked (both pages and state bits are locked).
 440  *
 441  * This is done inside an ordered work queue, and the compression
 442  * is spread across many cpus.  The actual IO submission is step
 443  * two, and the ordered work queue takes care of making sure that
 444  * happens in the same order things were put onto the queue by
 445  * writepages and friends.
 446  *
 447  * If this code finds it can't get good compression, it puts an
 448  * entry onto the work queue to write the uncompressed bytes.  This
 449  * makes sure that both compressed inodes and uncompressed inodes
 450  * are written in the same order that the flusher thread sent them
 451  * down.
 452  */
 453 static noinline void compress_file_range(struct inode *inode,
 454                                         struct page *locked_page,
 455                                         u64 start, u64 end,
 456                                         struct async_cow *async_cow,
 457                                         int *num_added)
 458 {
 459         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 460         u64 blocksize = fs_info->sectorsize;
 461         u64 actual_end;
 462         u64 isize = i_size_read(inode);
 463         int ret = 0;
 464         struct page **pages = NULL;
 465         unsigned long nr_pages;
 466         unsigned long total_compressed = 0;
 467         unsigned long total_in = 0;
 468         int i;
 469         int will_compress;
 470         int compress_type = fs_info->compress_type;
 471         int redirty = 0;
 472
 473         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 474                         SZ_16K);
 475
 476         actual_end = min_t(u64, isize, end + 1);
 477 again:
 478         will_compress = 0;
 479         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 480         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 481         nr_pages = min_t(unsigned long, nr_pages,
 482                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 483
 484         /*
 485          * we don't want to send crud past the end of i_size through
 486          * compression, that's just a waste of CPU time.  So, if the
 487          * end of the file is before the start of our current
 488          * requested range of bytes, we bail out to the uncompressed
 489          * cleanup code that can deal with all of this.
 490          *
 491          * It isn't really the fastest way to fix things, but this is a
 492          * very uncommon corner.
 493          */
 494         if (actual_end <= start)
 495                 goto cleanup_and_bail_uncompressed;
 496
 497         total_compressed = actual_end - start;
 498
 499         /*
 500          * skip compression for a small file range(<=blocksize) that
 501          * isn't an inline extent, since it doesn't save disk space at all.
 502          */
 503         if (total_compressed <= blocksize &&
 504            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 505                 goto cleanup_and_bail_uncompressed;
 506
 507         total_compressed = min_t(unsigned long, total_compressed,
 508                         BTRFS_MAX_UNCOMPRESSED);
 509         total_in = 0;
 510         ret = 0;
 511
 512         /*
 513          * we do compression for mount -o compress and when the
 514          * inode has not been flagged as nocompress.  This flag can
 515          * change at any time if we discover bad compression ratios.
 516          */
 517         if (inode_need_compress(inode, start, end)) {
 518                 WARN_ON(pages);
 519                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 520                 if (!pages) {
 521                         /* just bail out to the uncompressed code */
 522                         goto cont;
 523                 }
 524
 525                 if (BTRFS_I(inode)->defrag_compress)
 526                         compress_type = BTRFS_I(inode)->defrag_compress;
 527                 else if (BTRFS_I(inode)->prop_compress)
 528                         compress_type = BTRFS_I(inode)->prop_compress;
 529
 530                 /*
 531                  * we need to call clear_page_dirty_for_io on each
 532                  * page in the range.  Otherwise applications with the file
 533                  * mmap'd can wander in and change the page contents while
 534                  * we are compressing them.
 535                  *
 536                  * If the compression fails for any reason, we set the pages
 537                  * dirty again later on.
 538                  *
 539                  * Note that the remaining part is redirtied, the start pointer
 540                  * has moved, the end is the original one.
 541                  */
 542                 if (!redirty) {
 543                         extent_range_clear_dirty_for_io(inode, start, end);
 544                         redirty = 1;
 545                 }
 546
 547                 /* Compression level is applied here and only here */
 548                 ret = btrfs_compress_pages(
 549                         compress_type | (fs_info->compress_level << 4),
 550                                            inode->i_mapping, start,
 551                                            pages,
 552                                            &nr_pages,
 553                                            &total_in,
 554                                            &total_compressed);
 555
 556                 if (!ret) {
 557                         unsigned long offset = total_compressed &
 558                                 (PAGE_SIZE - 1);
 559                         struct page *page = pages[nr_pages - 1];
 560                         char *kaddr;
 561
 562                         /* zero the tail end of the last page, we might be
 563                          * sending it down to disk
 564                          */
 565                         if (offset) {
 566                                 kaddr = kmap_atomic(page);
 567                                 memset(kaddr + offset, 0,
 568                                        PAGE_SIZE - offset);
 569                                 kunmap_atomic(kaddr);
 570                         }
 571                         will_compress = 1;
 572                 }
 573         }
 574 cont:
 575         if (start == 0) {
 576                 /* lets try to make an inline extent */
 577                 if (ret || total_in < actual_end) {
 578                         /* we didn't compress the entire range, try
 579                          * to make an uncompressed inline extent.
 580                          */
 581                         ret = cow_file_range_inline(inode, start, end, 0,
 582                                                     BTRFS_COMPRESS_NONE, NULL);
 583                 } else {
 584                         /* try making a compressed inline extent */
 585                         ret = cow_file_range_inline(inode, start, end,
 586                                                     total_compressed,
 587                                                     compress_type, pages);
 588                 }
 589                 if (ret <= 0) {
 590                         unsigned long clear_flags = EXTENT_DELALLOC |
 591                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 592                                 EXTENT_DO_ACCOUNTING;
 593                         unsigned long page_error_op;
 594
 595                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 596
 597                         /*
 598                          * inline extent creation worked or returned error,
 599                          * we don't need to create any more async work items.
 600                          * Unlock and free up our temp pages.
 601                          *
 602                          * We use DO_ACCOUNTING here because we need the
 603                          * delalloc_release_metadata to be done _after_ we drop
 604                          * our outstanding extent for clearing delalloc for this
 605                          * range.
 606                          */
 607                         extent_clear_unlock_delalloc(inode, start, end, end,
 608                                                      NULL, clear_flags,
 609                                                      PAGE_UNLOCK |
 610                                                      PAGE_CLEAR_DIRTY |
 611                                                      PAGE_SET_WRITEBACK |
 612                                                      page_error_op |
 613                                                      PAGE_END_WRITEBACK);
 614                         goto free_pages_out;
 615                 }
 616         }
 617
 618         if (will_compress) {
 619                 /*
 620                  * we aren't doing an inline extent round the compressed size
 621                  * up to a block size boundary so the allocator does sane
 622                  * things
 623                  */
 624                 total_compressed = ALIGN(total_compressed, blocksize);
 625
 626                 /*
 627                  * one last check to make sure the compression is really a
 628                  * win, compare the page count read with the blocks on disk,
 629                  * compression must free at least one sector size
 630                  */
 631                 total_in = ALIGN(total_in, PAGE_SIZE);
 632                 if (total_compressed + blocksize <= total_in) {
 633                         *num_added += 1;
 634
 635                         /*
 636                          * The async work queues will take care of doing actual
 637                          * allocation on disk for these compressed pages, and
 638                          * will submit them to the elevator.
 639                          */
 640                         add_async_extent(async_cow, start, total_in,
 641                                         total_compressed, pages, nr_pages,
 642                                         compress_type);
 643
 644                         if (start + total_in < end) {
 645                                 start += total_in;
 646                                 pages = NULL;
 647                                 cond_resched();
 648                                 goto again;
 649                         }
 650                         return;
 651                 }
 652         }
 653         if (pages) {
 654                 /*
 655                  * the compression code ran but failed to make things smaller,
 656                  * free any pages it allocated and our page pointer array
 657                  */
 658                 for (i = 0; i < nr_pages; i++) {
 659                         WARN_ON(pages[i]->mapping);
 660                         put_page(pages[i]);
 661                 }
 662                 kfree(pages);
 663                 pages = NULL;
 664                 total_compressed = 0;
 665                 nr_pages = 0;
 666
 667                 /* flag the file so we don't compress in the future */
 668                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 669                     !(BTRFS_I(inode)->prop_compress)) {
 670                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 671                 }
 672         }
 673 cleanup_and_bail_uncompressed:
 674         /*
 675          * No compression, but we still need to write the pages in the file
 676          * we've been given so far.  redirty the locked page if it corresponds
 677          * to our extent and set things up for the async work queue to run
 678          * cow_file_range to do the normal delalloc dance.
 679          */
 680         if (page_offset(locked_page) >= start &&
 681             page_offset(locked_page) <= end)
 682                 __set_page_dirty_nobuffers(locked_page);
 683                 /* unlocked later on in the async handlers */
 684
 685         if (redirty)
 686                 extent_range_redirty_for_io(inode, start, end);
 687         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 688                          BTRFS_COMPRESS_NONE);
 689         *num_added += 1;
 690
 691         return;
 692
 693 free_pages_out:
 694         for (i = 0; i < nr_pages; i++) {
 695                 WARN_ON(pages[i]->mapping);
 696                 put_page(pages[i]);
 697         }
 698         kfree(pages);
 699 }
 700
 701 static void free_async_extent_pages(struct async_extent *async_extent)
 702 {
 703         int i;
 704
 705         if (!async_extent->pages)
 706                 return;
 707
 708         for (i = 0; i < async_extent->nr_pages; i++) {
 709                 WARN_ON(async_extent->pages[i]->mapping);
 710                 put_page(async_extent->pages[i]);
 711         }
 712         kfree(async_extent->pages);
 713         async_extent->nr_pages = 0;
 714         async_extent->pages = NULL;
 715 }
 716
 717 /*
 718  * phase two of compressed writeback.  This is the ordered portion
 719  * of the code, which only gets called in the order the work was
 720  * queued.  We walk all the async extents created by compress_file_range
 721  * and send them down to the disk.
 722  */
 723 static noinline void submit_compressed_extents(struct inode *inode,
 724                                               struct async_cow *async_cow)
 725 {
 726         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 727         struct async_extent *async_extent;
 728         u64 alloc_hint = 0;
 729         struct btrfs_key ins;
 730         struct extent_map *em;
 731         struct btrfs_root *root = BTRFS_I(inode)->root;
 732         struct extent_io_tree *io_tree;
 733         int ret = 0;
 734
 735 again:
 736         while (!list_empty(&async_cow->extents)) {
 737                 async_extent = list_entry(async_cow->extents.next,
 738                                           struct async_extent, list);
 739                 list_del(&async_extent->list);
 740
 741                 io_tree = &BTRFS_I(inode)->io_tree;
 742
 743 retry:
 744                 /* did the compression code fall back to uncompressed IO? */
 745                 if (!async_extent->pages) {
 746                         int page_started = 0;
 747                         unsigned long nr_written = 0;
 748
 749                         lock_extent(io_tree, async_extent->start,
 750                                          async_extent->start +
 751                                          async_extent->ram_size - 1);
 752
 753                         /* allocate blocks */
 754                         ret = cow_file_range(inode, async_cow->locked_page,
 755                                              async_extent->start,
 756                                              async_extent->start +
 757                                              async_extent->ram_size - 1,
 758                                              async_extent->start +
 759                                              async_extent->ram_size - 1,
 760                                              &page_started, &nr_written, 0,
 761                                              NULL);
 762
 763                         /* JDM XXX */
 764
 765                         /*
 766                          * if page_started, cow_file_range inserted an
 767                          * inline extent and took care of all the unlocking
 768                          * and IO for us.  Otherwise, we need to submit
 769                          * all those pages down to the drive.
 770                          */
 771                         if (!page_started && !ret)
 772                                 extent_write_locked_range(inode,
 773                                                   async_extent->start,
 774                                                   async_extent->start +
 775                                                   async_extent->ram_size - 1,
 776                                                   WB_SYNC_ALL);
 777                         else if (ret)
 778                                 unlock_page(async_cow->locked_page);
 779                         kfree(async_extent);
 780                         cond_resched();
 781                         continue;
 782                 }
 783
 784                 lock_extent(io_tree, async_extent->start,
 785                             async_extent->start + async_extent->ram_size - 1);
 786
 787                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 788                                            async_extent->compressed_size,
 789                                            async_extent->compressed_size,
 790                                            0, alloc_hint, &ins, 1, 1);
 791                 if (ret) {
 792                         free_async_extent_pages(async_extent);
 793
 794                         if (ret == -ENOSPC) {
 795                                 unlock_extent(io_tree, async_extent->start,
 796                                               async_extent->start +
 797                                               async_extent->ram_size - 1);
 798
 799                                 /*
 800                                  * we need to redirty the pages if we decide to
 801                                  * fallback to uncompressed IO, otherwise we
 802                                  * will not submit these pages down to lower
 803                                  * layers.
 804                                  */
 805                                 extent_range_redirty_for_io(inode,
 806                                                 async_extent->start,
 807                                                 async_extent->start +
 808                                                 async_extent->ram_size - 1);
 809
 810                                 goto retry;
 811                         }
 812                         goto out_free;
 813                 }
 814                 /*
 815                  * here we're doing allocation and writeback of the
 816                  * compressed pages
 817                  */
 818                 em = create_io_em(inode, async_extent->start,
 819                                   async_extent->ram_size, /* len */
 820                                   async_extent->start, /* orig_start */
 821                                   ins.objectid, /* block_start */
 822                                   ins.offset, /* block_len */
 823                                   ins.offset, /* orig_block_len */
 824                                   async_extent->ram_size, /* ram_bytes */
 825                                   async_extent->compress_type,
 826                                   BTRFS_ORDERED_COMPRESSED);
 827                 if (IS_ERR(em))
 828                         /* ret value is not necessary due to void function */
 829                         goto out_free_reserve;
 830                 free_extent_map(em);
 831
 832                 ret = btrfs_add_ordered_extent_compress(inode,
 833                                                 async_extent->start,
 834                                                 ins.objectid,
 835                                                 async_extent->ram_size,
 836                                                 ins.offset,
 837                                                 BTRFS_ORDERED_COMPRESSED,
 838                                                 async_extent->compress_type);
 839                 if (ret) {
 840                         btrfs_drop_extent_cache(BTRFS_I(inode),
 841                                                 async_extent->start,
 842                                                 async_extent->start +
 843                                                 async_extent->ram_size - 1, 0);
 844                         goto out_free_reserve;
 845                 }
 846                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 847
 848                 /*
 849                  * clear dirty, set writeback and unlock the pages.
 850                  */
 851                 extent_clear_unlock_delalloc(inode, async_extent->start,
 852                                 async_extent->start +
 853                                 async_extent->ram_size - 1,
 854                                 async_extent->start +
 855                                 async_extent->ram_size - 1,
 856                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 857                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 858                                 PAGE_SET_WRITEBACK);
 859                 if (btrfs_submit_compressed_write(inode,
 860                                     async_extent->start,
 861                                     async_extent->ram_size,
 862                                     ins.objectid,
 863                                     ins.offset, async_extent->pages,
 864                                     async_extent->nr_pages,
 865                                     async_cow->write_flags)) {
 866                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 867                         struct page *p = async_extent->pages[0];
 868                         const u64 start = async_extent->start;
 869                         const u64 end = start + async_extent->ram_size - 1;
 870
 871                         p->mapping = inode->i_mapping;
 872                         tree->ops->writepage_end_io_hook(p, start, end,
 873                                                          NULL, 0);
 874                         p->mapping = NULL;
 875                         extent_clear_unlock_delalloc(inode, start, end, end,
 876                                                      NULL, 0,
 877                                                      PAGE_END_WRITEBACK |
 878                                                      PAGE_SET_ERROR);
 879                         free_async_extent_pages(async_extent);
 880                 }
 881                 alloc_hint = ins.objectid + ins.offset;
 882                 kfree(async_extent);
 883                 cond_resched();
 884         }
 885         return;
 886 out_free_reserve:
 887         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 888         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 889 out_free:
 890         extent_clear_unlock_delalloc(inode, async_extent->start,
 891                                      async_extent->start +
 892                                      async_extent->ram_size - 1,
 893                                      async_extent->start +
 894                                      async_extent->ram_size - 1,
 895                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 896                                      EXTENT_DELALLOC_NEW |
 897                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 898                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 899                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 900                                      PAGE_SET_ERROR);
 901         free_async_extent_pages(async_extent);
 902         kfree(async_extent);
 903         goto again;
 904 }
 905
 906 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 907                                       u64 num_bytes)
 908 {
 909         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 910         struct extent_map *em;
 911         u64 alloc_hint = 0;
 912
 913         read_lock(&em_tree->lock);
 914         em = search_extent_mapping(em_tree, start, num_bytes);
 915         if (em) {
 916                 /*
 917                  * if block start isn't an actual block number then find the
 918                  * first block in this inode and use that as a hint.  If that
 919                  * block is also bogus then just don't worry about it.
 920                  */
 921                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 922                         free_extent_map(em);
 923                         em = search_extent_mapping(em_tree, 0, 0);
 924                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 925                                 alloc_hint = em->block_start;
 926                         if (em)
 927                                 free_extent_map(em);
 928                 } else {
 929                         alloc_hint = em->block_start;
 930                         free_extent_map(em);
 931                 }
 932         }
 933         read_unlock(&em_tree->lock);
 934
 935         return alloc_hint;
 936 }
 937
 938 /*
 939  * when extent_io.c finds a delayed allocation range in the file,
 940  * the call backs end up in this code.  The basic idea is to
 941  * allocate extents on disk for the range, and create ordered data structs
 942  * in ram to track those extents.
 943  *
 944  * locked_page is the page that writepage had locked already.  We use
 945  * it to make sure we don't do extra locks or unlocks.
 946  *
 947  * *page_started is set to one if we unlock locked_page and do everything
 948  * required to start IO on it.  It may be clean and already done with
 949  * IO when we return.
 950  */
 951 static noinline int cow_file_range(struct inode *inode,
 952                                    struct page *locked_page,
 953                                    u64 start, u64 end, u64 delalloc_end,
 954                                    int *page_started, unsigned long *nr_written,
 955                                    int unlock, struct btrfs_dedupe_hash *hash)
 956 {
 957         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 958         struct btrfs_root *root = BTRFS_I(inode)->root;
 959         u64 alloc_hint = 0;
 960         u64 num_bytes;
 961         unsigned long ram_size;
 962         u64 cur_alloc_size = 0;
 963         u64 blocksize = fs_info->sectorsize;
 964         struct btrfs_key ins;
 965         struct extent_map *em;
 966         unsigned clear_bits;
 967         unsigned long page_ops;
 968         bool extent_reserved = false;
 969         int ret = 0;
 970
 971         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 972                 WARN_ON_ONCE(1);
 973                 ret = -EINVAL;
 974                 goto out_unlock;
 975         }
 976
 977         num_bytes = ALIGN(end - start + 1, blocksize);
 978         num_bytes = max(blocksize,  num_bytes);
 979         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
 980
 981         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 982
 983         if (start == 0) {
 984                 /* lets try to make an inline extent */
 985                 ret = cow_file_range_inline(inode, start, end, 0,
 986                                             BTRFS_COMPRESS_NONE, NULL);
 987                 if (ret == 0) {
 988                         /*
 989                          * We use DO_ACCOUNTING here because we need the
 990                          * delalloc_release_metadata to be run _after_ we drop
 991                          * our outstanding extent for clearing delalloc for this
 992                          * range.
 993                          */
 994                         extent_clear_unlock_delalloc(inode, start, end,
 995                                      delalloc_end, NULL,
 996                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 997                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 998                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
 999                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1000                                      PAGE_END_WRITEBACK);
1001                         *nr_written = *nr_written +
1002                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1003                         *page_started = 1;
1004                         goto out;
1005                 } else if (ret < 0) {
1006                         goto out_unlock;
1007                 }
1008         }
1009
1010         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1011         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1012                         start + num_bytes - 1, 0);
1013
1014         while (num_bytes > 0) {
1015                 cur_alloc_size = num_bytes;
1016                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1017                                            fs_info->sectorsize, 0, alloc_hint,
1018                                            &ins, 1, 1);
1019                 if (ret < 0)
1020                         goto out_unlock;
1021                 cur_alloc_size = ins.offset;
1022                 extent_reserved = true;
1023
1024                 ram_size = ins.offset;
1025                 em = create_io_em(inode, start, ins.offset, /* len */
1026                                   start, /* orig_start */
1027                                   ins.objectid, /* block_start */
1028                                   ins.offset, /* block_len */
1029                                   ins.offset, /* orig_block_len */
1030                                   ram_size, /* ram_bytes */
1031                                   BTRFS_COMPRESS_NONE, /* compress_type */
1032                                   BTRFS_ORDERED_REGULAR /* type */);
1033                 if (IS_ERR(em))
1034                         goto out_reserve;
1035                 free_extent_map(em);
1036
1037                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1038                                                ram_size, cur_alloc_size, 0);
1039                 if (ret)
1040                         goto out_drop_extent_cache;
1041
1042                 if (root->root_key.objectid ==
1043                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1044                         ret = btrfs_reloc_clone_csums(inode, start,
1045                                                       cur_alloc_size);
1046                         /*
1047                          * Only drop cache here, and process as normal.
1048                          *
1049                          * We must not allow extent_clear_unlock_delalloc()
1050                          * at out_unlock label to free meta of this ordered
1051                          * extent, as its meta should be freed by
1052                          * btrfs_finish_ordered_io().
1053                          *
1054                          * So we must continue until @start is increased to
1055                          * skip current ordered extent.
1056                          */
1057                         if (ret)
1058                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1059                                                 start + ram_size - 1, 0);
1060                 }
1061
1062                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1063
1064                 /* we're not doing compressed IO, don't unlock the first
1065                  * page (which the caller expects to stay locked), don't
1066                  * clear any dirty bits and don't set any writeback bits
1067                  *
1068                  * Do set the Private2 bit so we know this page was properly
1069                  * setup for writepage
1070                  */
1071                 page_ops = unlock ? PAGE_UNLOCK : 0;
1072                 page_ops |= PAGE_SET_PRIVATE2;
1073
1074                 extent_clear_unlock_delalloc(inode, start,
1075                                              start + ram_size - 1,
1076                                              delalloc_end, locked_page,
1077                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1078                                              page_ops);
1079                 if (num_bytes < cur_alloc_size)
1080                         num_bytes = 0;
1081                 else
1082                         num_bytes -= cur_alloc_size;
1083                 alloc_hint = ins.objectid + ins.offset;
1084                 start += cur_alloc_size;
1085                 extent_reserved = false;
1086
1087                 /*
1088                  * btrfs_reloc_clone_csums() error, since start is increased
1089                  * extent_clear_unlock_delalloc() at out_unlock label won't
1090                  * free metadata of current ordered extent, we're OK to exit.
1091                  */
1092                 if (ret)
1093                         goto out_unlock;
1094         }
1095 out:
1096         return ret;
1097
1098 out_drop_extent_cache:
1099         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1100 out_reserve:
1101         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1102         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1103 out_unlock:
1104         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1105                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1106         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1107                 PAGE_END_WRITEBACK;
1108         /*
1109          * If we reserved an extent for our delalloc range (or a subrange) and
1110          * failed to create the respective ordered extent, then it means that
1111          * when we reserved the extent we decremented the extent's size from
1112          * the data space_info's bytes_may_use counter and incremented the
1113          * space_info's bytes_reserved counter by the same amount. We must make
1114          * sure extent_clear_unlock_delalloc() does not try to decrement again
1115          * the data space_info's bytes_may_use counter, therefore we do not pass
1116          * it the flag EXTENT_CLEAR_DATA_RESV.
1117          */
1118         if (extent_reserved) {
1119                 extent_clear_unlock_delalloc(inode, start,
1120                                              start + cur_alloc_size,
1121                                              start + cur_alloc_size,
1122                                              locked_page,
1123                                              clear_bits,
1124                                              page_ops);
1125                 start += cur_alloc_size;
1126                 if (start >= end)
1127                         goto out;
1128         }
1129         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1130                                      locked_page,
1131                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1132                                      page_ops);
1133         goto out;
1134 }
1135
1136 /*
1137  * work queue call back to started compression on a file and pages
1138  */
1139 static noinline void async_cow_start(struct btrfs_work *work)
1140 {
1141         struct async_cow *async_cow;
1142         int num_added = 0;
1143         async_cow = container_of(work, struct async_cow, work);
1144
1145         compress_file_range(async_cow->inode, async_cow->locked_page,
1146                             async_cow->start, async_cow->end, async_cow,
1147                             &num_added);
1148         if (num_added == 0) {
1149                 btrfs_add_delayed_iput(async_cow->inode);
1150                 async_cow->inode = NULL;
1151         }
1152 }
1153
1154 /*
1155  * work queue call back to submit previously compressed pages
1156  */
1157 static noinline void async_cow_submit(struct btrfs_work *work)
1158 {
1159         struct btrfs_fs_info *fs_info;
1160         struct async_cow *async_cow;
1161         struct btrfs_root *root;
1162         unsigned long nr_pages;
1163
1164         async_cow = container_of(work, struct async_cow, work);
1165
1166         root = async_cow->root;
1167         fs_info = root->fs_info;
1168         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1169                 PAGE_SHIFT;
1170
1171         /*
1172          * atomic_sub_return implies a barrier for waitqueue_active
1173          */
1174         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1175             5 * SZ_1M &&
1176             waitqueue_active(&fs_info->async_submit_wait))
1177                 wake_up(&fs_info->async_submit_wait);
1178
1179         if (async_cow->inode)
1180                 submit_compressed_extents(async_cow->inode, async_cow);
1181 }
1182
1183 static noinline void async_cow_free(struct btrfs_work *work)
1184 {
1185         struct async_cow *async_cow;
1186         async_cow = container_of(work, struct async_cow, work);
1187         if (async_cow->inode)
1188                 btrfs_add_delayed_iput(async_cow->inode);
1189         kfree(async_cow);
1190 }
1191
1192 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1193                                 u64 start, u64 end, int *page_started,
1194                                 unsigned long *nr_written,
1195                                 unsigned int write_flags)
1196 {
1197         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1198         struct async_cow *async_cow;
1199         struct btrfs_root *root = BTRFS_I(inode)->root;
1200         unsigned long nr_pages;
1201         u64 cur_end;
1202
1203         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1204                          1, 0, NULL);
1205         while (start < end) {
1206                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1207                 BUG_ON(!async_cow); /* -ENOMEM */
1208                 async_cow->inode = igrab(inode);
1209                 async_cow->root = root;
1210                 async_cow->locked_page = locked_page;
1211                 async_cow->start = start;
1212                 async_cow->write_flags = write_flags;
1213
1214                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1215                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1216                         cur_end = end;
1217                 else
1218                         cur_end = min(end, start + SZ_512K - 1);
1219
1220                 async_cow->end = cur_end;
1221                 INIT_LIST_HEAD(&async_cow->extents);
1222
1223                 btrfs_init_work(&async_cow->work,
1224                                 btrfs_delalloc_helper,
1225                                 async_cow_start, async_cow_submit,
1226                                 async_cow_free);
1227
1228                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1229                         PAGE_SHIFT;
1230                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1231
1232                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1233
1234                 *nr_written += nr_pages;
1235                 start = cur_end + 1;
1236         }
1237         *page_started = 1;
1238         return 0;
1239 }
1240
1241 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1242                                         u64 bytenr, u64 num_bytes)
1243 {
1244         int ret;
1245         struct btrfs_ordered_sum *sums;
1246         LIST_HEAD(list);
1247
1248         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1249                                        bytenr + num_bytes - 1, &list, 0);
1250         if (ret == 0 && list_empty(&list))
1251                 return 0;
1252
1253         while (!list_empty(&list)) {
1254                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1255                 list_del(&sums->list);
1256                 kfree(sums);
1257         }
1258         if (ret < 0)
1259                 return ret;
1260         return 1;
1261 }
1262
1263 /*
1264  * when nowcow writeback call back.  This checks for snapshots or COW copies
1265  * of the extents that exist in the file, and COWs the file as required.
1266  *
1267  * If no cow copies or snapshots exist, we write directly to the existing
1268  * blocks on disk
1269  */
1270 static noinline int run_delalloc_nocow(struct inode *inode,
1271                                        struct page *locked_page,
1272                               u64 start, u64 end, int *page_started, int force,
1273                               unsigned long *nr_written)
1274 {
1275         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1276         struct btrfs_root *root = BTRFS_I(inode)->root;
1277         struct extent_buffer *leaf;
1278         struct btrfs_path *path;
1279         struct btrfs_file_extent_item *fi;
1280         struct btrfs_key found_key;
1281         struct extent_map *em;
1282         u64 cow_start;
1283         u64 cur_offset;
1284         u64 extent_end;
1285         u64 extent_offset;
1286         u64 disk_bytenr;
1287         u64 num_bytes;
1288         u64 disk_num_bytes;
1289         u64 ram_bytes;
1290         int extent_type;
1291         int ret, err;
1292         int type;
1293         int nocow;
1294         int check_prev = 1;
1295         bool nolock;
1296         u64 ino = btrfs_ino(BTRFS_I(inode));
1297
1298         path = btrfs_alloc_path();
1299         if (!path) {
1300                 extent_clear_unlock_delalloc(inode, start, end, end,
1301                                              locked_page,
1302                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1303                                              EXTENT_DO_ACCOUNTING |
1304                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1305                                              PAGE_CLEAR_DIRTY |
1306                                              PAGE_SET_WRITEBACK |
1307                                              PAGE_END_WRITEBACK);
1308                 return -ENOMEM;
1309         }
1310
1311         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1312
1313         cow_start = (u64)-1;
1314         cur_offset = start;
1315         while (1) {
1316                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1317                                                cur_offset, 0);
1318                 if (ret < 0)
1319                         goto error;
1320                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1321                         leaf = path->nodes[0];
1322                         btrfs_item_key_to_cpu(leaf, &found_key,
1323                                               path->slots[0] - 1);
1324                         if (found_key.objectid == ino &&
1325                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1326                                 path->slots[0]--;
1327                 }
1328                 check_prev = 0;
1329 next_slot:
1330                 leaf = path->nodes[0];
1331                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1332                         ret = btrfs_next_leaf(root, path);
1333                         if (ret < 0) {
1334                                 if (cow_start != (u64)-1)
1335                                         cur_offset = cow_start;
1336                                 goto error;
1337                         }
1338                         if (ret > 0)
1339                                 break;
1340                         leaf = path->nodes[0];
1341                 }
1342
1343                 nocow = 0;
1344                 disk_bytenr = 0;
1345                 num_bytes = 0;
1346                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1347
1348                 if (found_key.objectid > ino)
1349                         break;
1350                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1351                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1352                         path->slots[0]++;
1353                         goto next_slot;
1354                 }
1355                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1356                     found_key.offset > end)
1357                         break;
1358
1359                 if (found_key.offset > cur_offset) {
1360                         extent_end = found_key.offset;
1361                         extent_type = 0;
1362                         goto out_check;
1363                 }
1364
1365                 fi = btrfs_item_ptr(leaf, path->slots[0],
1366                                     struct btrfs_file_extent_item);
1367                 extent_type = btrfs_file_extent_type(leaf, fi);
1368
1369                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1370                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1371                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1372                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1373                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1374                         extent_end = found_key.offset +
1375                                 btrfs_file_extent_num_bytes(leaf, fi);
1376                         disk_num_bytes =
1377                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1378                         if (extent_end <= start) {
1379                                 path->slots[0]++;
1380                                 goto next_slot;
1381                         }
1382                         if (disk_bytenr == 0)
1383                                 goto out_check;
1384                         if (btrfs_file_extent_compression(leaf, fi) ||
1385                             btrfs_file_extent_encryption(leaf, fi) ||
1386                             btrfs_file_extent_other_encoding(leaf, fi))
1387                                 goto out_check;
1388                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1389                                 goto out_check;
1390                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1391                                 goto out_check;
1392                         ret = btrfs_cross_ref_exist(root, ino,
1393                                                     found_key.offset -
1394                                                     extent_offset, disk_bytenr);
1395                         if (ret) {
1396                                 /*
1397                                  * ret could be -EIO if the above fails to read
1398                                  * metadata.
1399                                  */
1400                                 if (ret < 0) {
1401                                         if (cow_start != (u64)-1)
1402                                                 cur_offset = cow_start;
1403                                         goto error;
1404                                 }
1405
1406                                 WARN_ON_ONCE(nolock);
1407                                 goto out_check;
1408                         }
1409                         disk_bytenr += extent_offset;
1410                         disk_bytenr += cur_offset - found_key.offset;
1411                         num_bytes = min(end + 1, extent_end) - cur_offset;
1412                         /*
1413                          * if there are pending snapshots for this root,
1414                          * we fall into common COW way.
1415                          */
1416                         if (!nolock) {
1417                                 err = btrfs_start_write_no_snapshotting(root);
1418                                 if (!err)
1419                                         goto out_check;
1420                         }
1421                         /*
1422                          * force cow if csum exists in the range.
1423                          * this ensure that csum for a given extent are
1424                          * either valid or do not exist.
1425                          */
1426                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1427                                                   num_bytes);
1428                         if (ret) {
1429                                 if (!nolock)
1430                                         btrfs_end_write_no_snapshotting(root);
1431
1432                                 /*
1433                                  * ret could be -EIO if the above fails to read
1434                                  * metadata.
1435                                  */
1436                                 if (ret < 0) {
1437                                         if (cow_start != (u64)-1)
1438                                                 cur_offset = cow_start;
1439                                         goto error;
1440                                 }
1441                                 WARN_ON_ONCE(nolock);
1442                                 goto out_check;
1443                         }
1444                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1445                                 if (!nolock)
1446                                         btrfs_end_write_no_snapshotting(root);
1447                                 goto out_check;
1448                         }
1449                         nocow = 1;
1450                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1451                         extent_end = found_key.offset +
1452                                 btrfs_file_extent_inline_len(leaf,
1453                                                      path->slots[0], fi);
1454                         extent_end = ALIGN(extent_end,
1455                                            fs_info->sectorsize);
1456                 } else {
1457                         BUG_ON(1);
1458                 }
1459 out_check:
1460                 if (extent_end <= start) {
1461                         path->slots[0]++;
1462                         if (!nolock && nocow)
1463                                 btrfs_end_write_no_snapshotting(root);
1464                         if (nocow)
1465                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1466                         goto next_slot;
1467                 }
1468                 if (!nocow) {
1469                         if (cow_start == (u64)-1)
1470                                 cow_start = cur_offset;
1471                         cur_offset = extent_end;
1472                         if (cur_offset > end)
1473                                 break;
1474                         path->slots[0]++;
1475                         goto next_slot;
1476                 }
1477
1478                 btrfs_release_path(path);
1479                 if (cow_start != (u64)-1) {
1480                         ret = cow_file_range(inode, locked_page,
1481                                              cow_start, found_key.offset - 1,
1482                                              end, page_started, nr_written, 1,
1483                                              NULL);
1484                         if (ret) {
1485                                 if (!nolock && nocow)
1486                                         btrfs_end_write_no_snapshotting(root);
1487                                 if (nocow)
1488                                         btrfs_dec_nocow_writers(fs_info,
1489                                                                 disk_bytenr);
1490                                 goto error;
1491                         }
1492                         cow_start = (u64)-1;
1493                 }
1494
1495                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1496                         u64 orig_start = found_key.offset - extent_offset;
1497
1498                         em = create_io_em(inode, cur_offset, num_bytes,
1499                                           orig_start,
1500                                           disk_bytenr, /* block_start */
1501                                           num_bytes, /* block_len */
1502                                           disk_num_bytes, /* orig_block_len */
1503                                           ram_bytes, BTRFS_COMPRESS_NONE,
1504                                           BTRFS_ORDERED_PREALLOC);
1505                         if (IS_ERR(em)) {
1506                                 if (!nolock && nocow)
1507                                         btrfs_end_write_no_snapshotting(root);
1508                                 if (nocow)
1509                                         btrfs_dec_nocow_writers(fs_info,
1510                                                                 disk_bytenr);
1511                                 ret = PTR_ERR(em);
1512                                 goto error;
1513                         }
1514                         free_extent_map(em);
1515                 }
1516
1517                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1518                         type = BTRFS_ORDERED_PREALLOC;
1519                 } else {
1520                         type = BTRFS_ORDERED_NOCOW;
1521                 }
1522
1523                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1524                                                num_bytes, num_bytes, type);
1525                 if (nocow)
1526                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1527                 BUG_ON(ret); /* -ENOMEM */
1528
1529                 if (root->root_key.objectid ==
1530                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1531                         /*
1532                          * Error handled later, as we must prevent
1533                          * extent_clear_unlock_delalloc() in error handler
1534                          * from freeing metadata of created ordered extent.
1535                          */
1536                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1537                                                       num_bytes);
1538
1539                 extent_clear_unlock_delalloc(inode, cur_offset,
1540                                              cur_offset + num_bytes - 1, end,
1541                                              locked_page, EXTENT_LOCKED |
1542                                              EXTENT_DELALLOC |
1543                                              EXTENT_CLEAR_DATA_RESV,
1544                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1545
1546                 if (!nolock && nocow)
1547                         btrfs_end_write_no_snapshotting(root);
1548                 cur_offset = extent_end;
1549
1550                 /*
1551                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1552                  * handler, as metadata for created ordered extent will only
1553                  * be freed by btrfs_finish_ordered_io().
1554                  */
1555                 if (ret)
1556                         goto error;
1557                 if (cur_offset > end)
1558                         break;
1559         }
1560         btrfs_release_path(path);
1561
1562         if (cur_offset <= end && cow_start == (u64)-1) {
1563                 cow_start = cur_offset;
1564                 cur_offset = end;
1565         }
1566
1567         if (cow_start != (u64)-1) {
1568                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1569                                      page_started, nr_written, 1, NULL);
1570                 if (ret)
1571                         goto error;
1572         }
1573
1574 error:
1575         if (ret && cur_offset < end)
1576                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1577                                              locked_page, EXTENT_LOCKED |
1578                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1579                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1580                                              PAGE_CLEAR_DIRTY |
1581                                              PAGE_SET_WRITEBACK |
1582                                              PAGE_END_WRITEBACK);
1583         btrfs_free_path(path);
1584         return ret;
1585 }
1586
1587 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1588 {
1589
1590         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1591             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1592                 return 0;
1593
1594         /*
1595          * @defrag_bytes is a hint value, no spinlock held here,
1596          * if is not zero, it means the file is defragging.
1597          * Force cow if given extent needs to be defragged.
1598          */
1599         if (BTRFS_I(inode)->defrag_bytes &&
1600             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1601                            EXTENT_DEFRAG, 0, NULL))
1602                 return 1;
1603
1604         return 0;
1605 }
1606
1607 /*
1608  * extent_io.c call back to do delayed allocation processing
1609  */
1610 static int run_delalloc_range(void *private_data, struct page *locked_page,
1611                               u64 start, u64 end, int *page_started,
1612                               unsigned long *nr_written,
1613                               struct writeback_control *wbc)
1614 {
1615         struct inode *inode = private_data;
1616         int ret;
1617         int force_cow = need_force_cow(inode, start, end);
1618         unsigned int write_flags = wbc_to_write_flags(wbc);
1619
1620         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1621                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1622                                          page_started, 1, nr_written);
1623         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1624                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1625                                          page_started, 0, nr_written);
1626         } else if (!inode_need_compress(inode, start, end)) {
1627                 ret = cow_file_range(inode, locked_page, start, end, end,
1628                                       page_started, nr_written, 1, NULL);
1629         } else {
1630                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1631                         &BTRFS_I(inode)->runtime_flags);
1632                 ret = cow_file_range_async(inode, locked_page, start, end,
1633                                            page_started, nr_written,
1634                                            write_flags);
1635         }
1636         if (ret)
1637                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1638         return ret;
1639 }
1640
1641 static void btrfs_split_extent_hook(void *private_data,
1642                                     struct extent_state *orig, u64 split)
1643 {
1644         struct inode *inode = private_data;
1645         u64 size;
1646
1647         /* not delalloc, ignore it */
1648         if (!(orig->state & EXTENT_DELALLOC))
1649                 return;
1650
1651         size = orig->end - orig->start + 1;
1652         if (size > BTRFS_MAX_EXTENT_SIZE) {
1653                 u32 num_extents;
1654                 u64 new_size;
1655
1656                 /*
1657                  * See the explanation in btrfs_merge_extent_hook, the same
1658                  * applies here, just in reverse.
1659                  */
1660                 new_size = orig->end - split + 1;
1661                 num_extents = count_max_extents(new_size);
1662                 new_size = split - orig->start;
1663                 num_extents += count_max_extents(new_size);
1664                 if (count_max_extents(size) >= num_extents)
1665                         return;
1666         }
1667
1668         spin_lock(&BTRFS_I(inode)->lock);
1669         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1670         spin_unlock(&BTRFS_I(inode)->lock);
1671 }
1672
1673 /*
1674  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1675  * extents so we can keep track of new extents that are just merged onto old
1676  * extents, such as when we are doing sequential writes, so we can properly
1677  * account for the metadata space we'll need.
1678  */
1679 static void btrfs_merge_extent_hook(void *private_data,
1680                                     struct extent_state *new,
1681                                     struct extent_state *other)
1682 {
1683         struct inode *inode = private_data;
1684         u64 new_size, old_size;
1685         u32 num_extents;
1686
1687         /* not delalloc, ignore it */
1688         if (!(other->state & EXTENT_DELALLOC))
1689                 return;
1690
1691         if (new->start > other->start)
1692                 new_size = new->end - other->start + 1;
1693         else
1694                 new_size = other->end - new->start + 1;
1695
1696         /* we're not bigger than the max, unreserve the space and go */
1697         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1698                 spin_lock(&BTRFS_I(inode)->lock);
1699                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1700                 spin_unlock(&BTRFS_I(inode)->lock);
1701                 return;
1702         }
1703
1704         /*
1705          * We have to add up either side to figure out how many extents were
1706          * accounted for before we merged into one big extent.  If the number of
1707          * extents we accounted for is <= the amount we need for the new range
1708          * then we can return, otherwise drop.  Think of it like this
1709          *
1710          * [ 4k][MAX_SIZE]
1711          *
1712          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1713          * need 2 outstanding extents, on one side we have 1 and the other side
1714          * we have 1 so they are == and we can return.  But in this case
1715          *
1716          * [MAX_SIZE+4k][MAX_SIZE+4k]
1717          *
1718          * Each range on their own accounts for 2 extents, but merged together
1719          * they are only 3 extents worth of accounting, so we need to drop in
1720          * this case.
1721          */
1722         old_size = other->end - other->start + 1;
1723         num_extents = count_max_extents(old_size);
1724         old_size = new->end - new->start + 1;
1725         num_extents += count_max_extents(old_size);
1726         if (count_max_extents(new_size) >= num_extents)
1727                 return;
1728
1729         spin_lock(&BTRFS_I(inode)->lock);
1730         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1731         spin_unlock(&BTRFS_I(inode)->lock);
1732 }
1733
1734 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1735                                       struct inode *inode)
1736 {
1737         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1738
1739         spin_lock(&root->delalloc_lock);
1740         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1741                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1742                               &root->delalloc_inodes);
1743                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1744                         &BTRFS_I(inode)->runtime_flags);
1745                 root->nr_delalloc_inodes++;
1746                 if (root->nr_delalloc_inodes == 1) {
1747                         spin_lock(&fs_info->delalloc_root_lock);
1748                         BUG_ON(!list_empty(&root->delalloc_root));
1749                         list_add_tail(&root->delalloc_root,
1750                                       &fs_info->delalloc_roots);
1751                         spin_unlock(&fs_info->delalloc_root_lock);
1752                 }
1753         }
1754         spin_unlock(&root->delalloc_lock);
1755 }
1756
1757 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1758                                      struct btrfs_inode *inode)
1759 {
1760         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1761
1762         spin_lock(&root->delalloc_lock);
1763         if (!list_empty(&inode->delalloc_inodes)) {
1764                 list_del_init(&inode->delalloc_inodes);
1765                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1766                           &inode->runtime_flags);
1767                 root->nr_delalloc_inodes--;
1768                 if (!root->nr_delalloc_inodes) {
1769                         spin_lock(&fs_info->delalloc_root_lock);
1770                         BUG_ON(list_empty(&root->delalloc_root));
1771                         list_del_init(&root->delalloc_root);
1772                         spin_unlock(&fs_info->delalloc_root_lock);
1773                 }
1774         }
1775         spin_unlock(&root->delalloc_lock);
1776 }
1777
1778 /*
1779  * extent_io.c set_bit_hook, used to track delayed allocation
1780  * bytes in this file, and to maintain the list of inodes that
1781  * have pending delalloc work to be done.
1782  */
1783 static void btrfs_set_bit_hook(void *private_data,
1784                                struct extent_state *state, unsigned *bits)
1785 {
1786         struct inode *inode = private_data;
1787
1788         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1789
1790         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1791                 WARN_ON(1);
1792         /*
1793          * set_bit and clear bit hooks normally require _irqsave/restore
1794          * but in this case, we are only testing for the DELALLOC
1795          * bit, which is only set or cleared with irqs on
1796          */
1797         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1798                 struct btrfs_root *root = BTRFS_I(inode)->root;
1799                 u64 len = state->end + 1 - state->start;
1800                 u32 num_extents = count_max_extents(len);
1801                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1802
1803                 spin_lock(&BTRFS_I(inode)->lock);
1804                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1805                 spin_unlock(&BTRFS_I(inode)->lock);
1806
1807                 /* For sanity tests */
1808                 if (btrfs_is_testing(fs_info))
1809                         return;
1810
1811                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1812                                          fs_info->delalloc_batch);
1813                 spin_lock(&BTRFS_I(inode)->lock);
1814                 BTRFS_I(inode)->delalloc_bytes += len;
1815                 if (*bits & EXTENT_DEFRAG)
1816                         BTRFS_I(inode)->defrag_bytes += len;
1817                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1818                                          &BTRFS_I(inode)->runtime_flags))
1819                         btrfs_add_delalloc_inodes(root, inode);
1820                 spin_unlock(&BTRFS_I(inode)->lock);
1821         }
1822
1823         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1824             (*bits & EXTENT_DELALLOC_NEW)) {
1825                 spin_lock(&BTRFS_I(inode)->lock);
1826                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1827                         state->start;
1828                 spin_unlock(&BTRFS_I(inode)->lock);
1829         }
1830 }
1831
1832 /*
1833  * extent_io.c clear_bit_hook, see set_bit_hook for why
1834  */
1835 static void btrfs_clear_bit_hook(void *private_data,
1836                                  struct extent_state *state,
1837                                  unsigned *bits)
1838 {
1839         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1840         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1841         u64 len = state->end + 1 - state->start;
1842         u32 num_extents = count_max_extents(len);
1843
1844         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1845                 spin_lock(&inode->lock);
1846                 inode->defrag_bytes -= len;
1847                 spin_unlock(&inode->lock);
1848         }
1849
1850         /*
1851          * set_bit and clear bit hooks normally require _irqsave/restore
1852          * but in this case, we are only testing for the DELALLOC
1853          * bit, which is only set or cleared with irqs on
1854          */
1855         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1856                 struct btrfs_root *root = inode->root;
1857                 bool do_list = !btrfs_is_free_space_inode(inode);
1858
1859                 spin_lock(&inode->lock);
1860                 btrfs_mod_outstanding_extents(inode, -num_extents);
1861                 spin_unlock(&inode->lock);
1862
1863                 /*
1864                  * We don't reserve metadata space for space cache inodes so we
1865                  * don't need to call dellalloc_release_metadata if there is an
1866                  * error.
1867                  */
1868                 if (*bits & EXTENT_CLEAR_META_RESV &&
1869                     root != fs_info->tree_root)
1870                         btrfs_delalloc_release_metadata(inode, len);
1871
1872                 /* For sanity tests. */
1873                 if (btrfs_is_testing(fs_info))
1874                         return;
1875
1876                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1877                     do_list && !(state->state & EXTENT_NORESERVE) &&
1878                     (*bits & EXTENT_CLEAR_DATA_RESV))
1879                         btrfs_free_reserved_data_space_noquota(
1880                                         &inode->vfs_inode,
1881                                         state->start, len);
1882
1883                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1884                                          fs_info->delalloc_batch);
1885                 spin_lock(&inode->lock);
1886                 inode->delalloc_bytes -= len;
1887                 if (do_list && inode->delalloc_bytes == 0 &&
1888                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1889                                         &inode->runtime_flags))
1890                         btrfs_del_delalloc_inode(root, inode);
1891                 spin_unlock(&inode->lock);
1892         }
1893
1894         if ((state->state & EXTENT_DELALLOC_NEW) &&
1895             (*bits & EXTENT_DELALLOC_NEW)) {
1896                 spin_lock(&inode->lock);
1897                 ASSERT(inode->new_delalloc_bytes >= len);
1898                 inode->new_delalloc_bytes -= len;
1899                 spin_unlock(&inode->lock);
1900         }
1901 }
1902
1903 /*
1904  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1905  * we don't create bios that span stripes or chunks
1906  *
1907  * return 1 if page cannot be merged to bio
1908  * return 0 if page can be merged to bio
1909  * return error otherwise
1910  */
1911 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1912                          size_t size, struct bio *bio,
1913                          unsigned long bio_flags)
1914 {
1915         struct inode *inode = page->mapping->host;
1916         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1917         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1918         u64 length = 0;
1919         u64 map_length;
1920         int ret;
1921
1922         if (bio_flags & EXTENT_BIO_COMPRESSED)
1923                 return 0;
1924
1925         length = bio->bi_iter.bi_size;
1926         map_length = length;
1927         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1928                               NULL, 0);
1929         if (ret < 0)
1930                 return ret;
1931         if (map_length < length + size)
1932                 return 1;
1933         return 0;
1934 }
1935
1936 /*
1937  * in order to insert checksums into the metadata in large chunks,
1938  * we wait until bio submission time.   All the pages in the bio are
1939  * checksummed and sums are attached onto the ordered extent record.
1940  *
1941  * At IO completion time the cums attached on the ordered extent record
1942  * are inserted into the btree
1943  */
1944 static blk_status_t __btrfs_submit_bio_start(void *private_data, struct bio *bio,
1945                                     int mirror_num, unsigned long bio_flags,
1946                                     u64 bio_offset)
1947 {
1948         struct inode *inode = private_data;
1949         blk_status_t ret = 0;
1950
1951         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1952         BUG_ON(ret); /* -ENOMEM */
1953         return 0;
1954 }
1955
1956 /*
1957  * in order to insert checksums into the metadata in large chunks,
1958  * we wait until bio submission time.   All the pages in the bio are
1959  * checksummed and sums are attached onto the ordered extent record.
1960  *
1961  * At IO completion time the cums attached on the ordered extent record
1962  * are inserted into the btree
1963  */
1964 static blk_status_t __btrfs_submit_bio_done(void *private_data, struct bio *bio,
1965                           int mirror_num, unsigned long bio_flags,
1966                           u64 bio_offset)
1967 {
1968         struct inode *inode = private_data;
1969         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1970         blk_status_t ret;
1971
1972         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1973         if (ret) {
1974                 bio->bi_status = ret;
1975                 bio_endio(bio);
1976         }
1977         return ret;
1978 }
1979
1980 /*
1981  * extent_io.c submission hook. This does the right thing for csum calculation
1982  * on write, or reading the csums from the tree before a read.
1983  *
1984  * Rules about async/sync submit,
1985  * a) read:                             sync submit
1986  *
1987  * b) write without checksum:           sync submit
1988  *
1989  * c) write with checksum:
1990  *    c-1) if bio is issued by fsync:   sync submit
1991  *         (sync_writers != 0)
1992  *
1993  *    c-2) if root is reloc root:       sync submit
1994  *         (only in case of buffered IO)
1995  *
1996  *    c-3) otherwise:                   async submit
1997  */
1998 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1999                                  int mirror_num, unsigned long bio_flags,
2000                                  u64 bio_offset)
2001 {
2002         struct inode *inode = private_data;
2003         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2004         struct btrfs_root *root = BTRFS_I(inode)->root;
2005         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2006         blk_status_t ret = 0;
2007         int skip_sum;
2008         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2009
2010         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2011
2012         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2013                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2014
2015         if (bio_op(bio) != REQ_OP_WRITE) {
2016                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2017                 if (ret)
2018                         goto out;
2019
2020                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2021                         ret = btrfs_submit_compressed_read(inode, bio,
2022                                                            mirror_num,
2023                                                            bio_flags);
2024                         goto out;
2025                 } else if (!skip_sum) {
2026                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2027                         if (ret)
2028                                 goto out;
2029                 }
2030                 goto mapit;
2031         } else if (async && !skip_sum) {
2032                 /* csum items have already been cloned */
2033                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2034                         goto mapit;
2035                 /* we're doing a write, do the async checksumming */
2036                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2037                                           bio_offset, inode,
2038                                           __btrfs_submit_bio_start,
2039                                           __btrfs_submit_bio_done);
2040                 goto out;
2041         } else if (!skip_sum) {
2042                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2043                 if (ret)
2044                         goto out;
2045         }
2046
2047 mapit:
2048         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2049
2050 out:
2051         if (ret) {
2052                 bio->bi_status = ret;
2053                 bio_endio(bio);
2054         }
2055         return ret;
2056 }
2057
2058 /*
2059  * given a list of ordered sums record them in the inode.  This happens
2060  * at IO completion time based on sums calculated at bio submission time.
2061  */
2062 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2063                              struct inode *inode, struct list_head *list)
2064 {
2065         struct btrfs_ordered_sum *sum;
2066         int ret;
2067
2068         list_for_each_entry(sum, list, list) {
2069                 trans->adding_csums = true;
2070                 ret = btrfs_csum_file_blocks(trans,
2071                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2072                 trans->adding_csums = false;
2073                 if (ret)
2074                         return ret;
2075         }
2076         return 0;
2077 }
2078
2079 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2080                               unsigned int extra_bits,
2081                               struct extent_state **cached_state, int dedupe)
2082 {
2083         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2084         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2085                                    extra_bits, cached_state);
2086 }
2087
2088 /* see btrfs_writepage_start_hook for details on why this is required */
2089 struct btrfs_writepage_fixup {
2090         struct page *page;
2091         struct btrfs_work work;
2092 };
2093
2094 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2095 {
2096         struct btrfs_writepage_fixup *fixup;
2097         struct btrfs_ordered_extent *ordered;
2098         struct extent_state *cached_state = NULL;
2099         struct extent_changeset *data_reserved = NULL;
2100         struct page *page;
2101         struct inode *inode;
2102         u64 page_start;
2103         u64 page_end;
2104         int ret;
2105
2106         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2107         page = fixup->page;
2108 again:
2109         lock_page(page);
2110         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2111                 ClearPageChecked(page);
2112                 goto out_page;
2113         }
2114
2115         inode = page->mapping->host;
2116         page_start = page_offset(page);
2117         page_end = page_offset(page) + PAGE_SIZE - 1;
2118
2119         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2120                          &cached_state);
2121
2122         /* already ordered? We're done */
2123         if (PagePrivate2(page))
2124                 goto out;
2125
2126         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2127                                         PAGE_SIZE);
2128         if (ordered) {
2129                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2130                                      page_end, &cached_state);
2131                 unlock_page(page);
2132                 btrfs_start_ordered_extent(inode, ordered, 1);
2133                 btrfs_put_ordered_extent(ordered);
2134                 goto again;
2135         }
2136
2137         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2138                                            PAGE_SIZE);
2139         if (ret) {
2140                 mapping_set_error(page->mapping, ret);
2141                 end_extent_writepage(page, ret, page_start, page_end);
2142                 ClearPageChecked(page);
2143                 goto out;
2144          }
2145
2146         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2147                                         &cached_state, 0);
2148         if (ret) {
2149                 mapping_set_error(page->mapping, ret);
2150                 end_extent_writepage(page, ret, page_start, page_end);
2151                 ClearPageChecked(page);
2152                 goto out;
2153         }
2154
2155         ClearPageChecked(page);
2156         set_page_dirty(page);
2157         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2158 out:
2159         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2160                              &cached_state);
2161 out_page:
2162         unlock_page(page);
2163         put_page(page);
2164         kfree(fixup);
2165         extent_changeset_free(data_reserved);
2166 }
2167
2168 /*
2169  * There are a few paths in the higher layers of the kernel that directly
2170  * set the page dirty bit without asking the filesystem if it is a
2171  * good idea.  This causes problems because we want to make sure COW
2172  * properly happens and the data=ordered rules are followed.
2173  *
2174  * In our case any range that doesn't have the ORDERED bit set
2175  * hasn't been properly setup for IO.  We kick off an async process
2176  * to fix it up.  The async helper will wait for ordered extents, set
2177  * the delalloc bit and make it safe to write the page.
2178  */
2179 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2180 {
2181         struct inode *inode = page->mapping->host;
2182         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2183         struct btrfs_writepage_fixup *fixup;
2184
2185         /* this page is properly in the ordered list */
2186         if (TestClearPagePrivate2(page))
2187                 return 0;
2188
2189         if (PageChecked(page))
2190                 return -EAGAIN;
2191
2192         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2193         if (!fixup)
2194                 return -EAGAIN;
2195
2196         SetPageChecked(page);
2197         get_page(page);
2198         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2199                         btrfs_writepage_fixup_worker, NULL, NULL);
2200         fixup->page = page;
2201         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2202         return -EBUSY;
2203 }
2204
2205 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2206                                        struct inode *inode, u64 file_pos,
2207                                        u64 disk_bytenr, u64 disk_num_bytes,
2208                                        u64 num_bytes, u64 ram_bytes,
2209                                        u8 compression, u8 encryption,
2210                                        u16 other_encoding, int extent_type)
2211 {
2212         struct btrfs_root *root = BTRFS_I(inode)->root;
2213         struct btrfs_file_extent_item *fi;
2214         struct btrfs_path *path;
2215         struct extent_buffer *leaf;
2216         struct btrfs_key ins;
2217         u64 qg_released;
2218         int extent_inserted = 0;
2219         int ret;
2220
2221         path = btrfs_alloc_path();
2222         if (!path)
2223                 return -ENOMEM;
2224
2225         /*
2226          * we may be replacing one extent in the tree with another.
2227          * The new extent is pinned in the extent map, and we don't want
2228          * to drop it from the cache until it is completely in the btree.
2229          *
2230          * So, tell btrfs_drop_extents to leave this extent in the cache.
2231          * the caller is expected to unpin it and allow it to be merged
2232          * with the others.
2233          */
2234         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2235                                    file_pos + num_bytes, NULL, 0,
2236                                    1, sizeof(*fi), &extent_inserted);
2237         if (ret)
2238                 goto out;
2239
2240         if (!extent_inserted) {
2241                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2242                 ins.offset = file_pos;
2243                 ins.type = BTRFS_EXTENT_DATA_KEY;
2244
2245                 path->leave_spinning = 1;
2246                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2247                                               sizeof(*fi));
2248                 if (ret)
2249                         goto out;
2250         }
2251         leaf = path->nodes[0];
2252         fi = btrfs_item_ptr(leaf, path->slots[0],
2253                             struct btrfs_file_extent_item);
2254         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2255         btrfs_set_file_extent_type(leaf, fi, extent_type);
2256         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2257         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2258         btrfs_set_file_extent_offset(leaf, fi, 0);
2259         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2260         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2261         btrfs_set_file_extent_compression(leaf, fi, compression);
2262         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2263         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2264
2265         btrfs_mark_buffer_dirty(leaf);
2266         btrfs_release_path(path);
2267
2268         inode_add_bytes(inode, num_bytes);
2269
2270         ins.objectid = disk_bytenr;
2271         ins.offset = disk_num_bytes;
2272         ins.type = BTRFS_EXTENT_ITEM_KEY;
2273
2274         /*
2275          * Release the reserved range from inode dirty range map, as it is
2276          * already moved into delayed_ref_head
2277          */
2278         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2279         if (ret < 0)
2280                 goto out;
2281         qg_released = ret;
2282         ret = btrfs_alloc_reserved_file_extent(trans, root,
2283                                                btrfs_ino(BTRFS_I(inode)),
2284                                                file_pos, qg_released, &ins);
2285 out:
2286         btrfs_free_path(path);
2287
2288         return ret;
2289 }
2290
2291 /* snapshot-aware defrag */
2292 struct sa_defrag_extent_backref {
2293         struct rb_node node;
2294         struct old_sa_defrag_extent *old;
2295         u64 root_id;
2296         u64 inum;
2297         u64 file_pos;
2298         u64 extent_offset;
2299         u64 num_bytes;
2300         u64 generation;
2301 };
2302
2303 struct old_sa_defrag_extent {
2304         struct list_head list;
2305         struct new_sa_defrag_extent *new;
2306
2307         u64 extent_offset;
2308         u64 bytenr;
2309         u64 offset;
2310         u64 len;
2311         int count;
2312 };
2313
2314 struct new_sa_defrag_extent {
2315         struct rb_root root;
2316         struct list_head head;
2317         struct btrfs_path *path;
2318         struct inode *inode;
2319         u64 file_pos;
2320         u64 len;
2321         u64 bytenr;
2322         u64 disk_len;
2323         u8 compress_type;
2324 };
2325
2326 static int backref_comp(struct sa_defrag_extent_backref *b1,
2327                         struct sa_defrag_extent_backref *b2)
2328 {
2329         if (b1->root_id < b2->root_id)
2330                 return -1;
2331         else if (b1->root_id > b2->root_id)
2332                 return 1;
2333
2334         if (b1->inum < b2->inum)
2335                 return -1;
2336         else if (b1->inum > b2->inum)
2337                 return 1;
2338
2339         if (b1->file_pos < b2->file_pos)
2340                 return -1;
2341         else if (b1->file_pos > b2->file_pos)
2342                 return 1;
2343
2344         /*
2345          * [------------------------------] ===> (a range of space)
2346          *     |<--->|   |<---->| =============> (fs/file tree A)
2347          * |<---------------------------->| ===> (fs/file tree B)
2348          *
2349          * A range of space can refer to two file extents in one tree while
2350          * refer to only one file extent in another tree.
2351          *
2352          * So we may process a disk offset more than one time(two extents in A)
2353          * and locate at the same extent(one extent in B), then insert two same
2354          * backrefs(both refer to the extent in B).
2355          */
2356         return 0;
2357 }
2358
2359 static void backref_insert(struct rb_root *root,
2360                            struct sa_defrag_extent_backref *backref)
2361 {
2362         struct rb_node **p = &root->rb_node;
2363         struct rb_node *parent = NULL;
2364         struct sa_defrag_extent_backref *entry;
2365         int ret;
2366
2367         while (*p) {
2368                 parent = *p;
2369                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2370
2371                 ret = backref_comp(backref, entry);
2372                 if (ret < 0)
2373                         p = &(*p)->rb_left;
2374                 else
2375                         p = &(*p)->rb_right;
2376         }
2377
2378         rb_link_node(&backref->node, parent, p);
2379         rb_insert_color(&backref->node, root);
2380 }
2381
2382 /*
2383  * Note the backref might has changed, and in this case we just return 0.
2384  */
2385 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2386                                        void *ctx)
2387 {
2388         struct btrfs_file_extent_item *extent;
2389         struct old_sa_defrag_extent *old = ctx;
2390         struct new_sa_defrag_extent *new = old->new;
2391         struct btrfs_path *path = new->path;
2392         struct btrfs_key key;
2393         struct btrfs_root *root;
2394         struct sa_defrag_extent_backref *backref;
2395         struct extent_buffer *leaf;
2396         struct inode *inode = new->inode;
2397         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2398         int slot;
2399         int ret;
2400         u64 extent_offset;
2401         u64 num_bytes;
2402
2403         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2404             inum == btrfs_ino(BTRFS_I(inode)))
2405                 return 0;
2406
2407         key.objectid = root_id;
2408         key.type = BTRFS_ROOT_ITEM_KEY;
2409         key.offset = (u64)-1;
2410
2411         root = btrfs_read_fs_root_no_name(fs_info, &key);
2412         if (IS_ERR(root)) {
2413                 if (PTR_ERR(root) == -ENOENT)
2414                         return 0;
2415                 WARN_ON(1);
2416                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2417                          inum, offset, root_id);
2418                 return PTR_ERR(root);
2419         }
2420
2421         key.objectid = inum;
2422         key.type = BTRFS_EXTENT_DATA_KEY;
2423         if (offset > (u64)-1 << 32)
2424                 key.offset = 0;
2425         else
2426                 key.offset = offset;
2427
2428         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2429         if (WARN_ON(ret < 0))
2430                 return ret;
2431         ret = 0;
2432
2433         while (1) {
2434                 cond_resched();
2435
2436                 leaf = path->nodes[0];
2437                 slot = path->slots[0];
2438
2439                 if (slot >= btrfs_header_nritems(leaf)) {
2440                         ret = btrfs_next_leaf(root, path);
2441                         if (ret < 0) {
2442                                 goto out;
2443                         } else if (ret > 0) {
2444                                 ret = 0;
2445                                 goto out;
2446                         }
2447                         continue;
2448                 }
2449
2450                 path->slots[0]++;
2451
2452                 btrfs_item_key_to_cpu(leaf, &key, slot);
2453
2454                 if (key.objectid > inum)
2455                         goto out;
2456
2457                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2458                         continue;
2459
2460                 extent = btrfs_item_ptr(leaf, slot,
2461                                         struct btrfs_file_extent_item);
2462
2463                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2464                         continue;
2465
2466                 /*
2467                  * 'offset' refers to the exact key.offset,
2468                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2469                  * (key.offset - extent_offset).
2470                  */
2471                 if (key.offset != offset)
2472                         continue;
2473
2474                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2475                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2476
2477                 if (extent_offset >= old->extent_offset + old->offset +
2478                     old->len || extent_offset + num_bytes <=
2479                     old->extent_offset + old->offset)
2480                         continue;
2481                 break;
2482         }
2483
2484         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2485         if (!backref) {
2486                 ret = -ENOENT;
2487                 goto out;
2488         }
2489
2490         backref->root_id = root_id;
2491         backref->inum = inum;
2492         backref->file_pos = offset;
2493         backref->num_bytes = num_bytes;
2494         backref->extent_offset = extent_offset;
2495         backref->generation = btrfs_file_extent_generation(leaf, extent);
2496         backref->old = old;
2497         backref_insert(&new->root, backref);
2498         old->count++;
2499 out:
2500         btrfs_release_path(path);
2501         WARN_ON(ret);
2502         return ret;
2503 }
2504
2505 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2506                                    struct new_sa_defrag_extent *new)
2507 {
2508         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2509         struct old_sa_defrag_extent *old, *tmp;
2510         int ret;
2511
2512         new->path = path;
2513
2514         list_for_each_entry_safe(old, tmp, &new->head, list) {
2515                 ret = iterate_inodes_from_logical(old->bytenr +
2516                                                   old->extent_offset, fs_info,
2517                                                   path, record_one_backref,
2518                                                   old, false);
2519                 if (ret < 0 && ret != -ENOENT)
2520                         return false;
2521
2522                 /* no backref to be processed for this extent */
2523                 if (!old->count) {
2524                         list_del(&old->list);
2525                         kfree(old);
2526                 }
2527         }
2528
2529         if (list_empty(&new->head))
2530                 return false;
2531
2532         return true;
2533 }
2534
2535 static int relink_is_mergable(struct extent_buffer *leaf,
2536                               struct btrfs_file_extent_item *fi,
2537                               struct new_sa_defrag_extent *new)
2538 {
2539         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2540                 return 0;
2541
2542         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2543                 return 0;
2544
2545         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2546                 return 0;
2547
2548         if (btrfs_file_extent_encryption(leaf, fi) ||
2549             btrfs_file_extent_other_encoding(leaf, fi))
2550                 return 0;
2551
2552         return 1;
2553 }
2554
2555 /*
2556  * Note the backref might has changed, and in this case we just return 0.
2557  */
2558 static noinline int relink_extent_backref(struct btrfs_path *path,
2559                                  struct sa_defrag_extent_backref *prev,
2560                                  struct sa_defrag_extent_backref *backref)
2561 {
2562         struct btrfs_file_extent_item *extent;
2563         struct btrfs_file_extent_item *item;
2564         struct btrfs_ordered_extent *ordered;
2565         struct btrfs_trans_handle *trans;
2566         struct btrfs_root *root;
2567         struct btrfs_key key;
2568         struct extent_buffer *leaf;
2569         struct old_sa_defrag_extent *old = backref->old;
2570         struct new_sa_defrag_extent *new = old->new;
2571         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2572         struct inode *inode;
2573         struct extent_state *cached = NULL;
2574         int ret = 0;
2575         u64 start;
2576         u64 len;
2577         u64 lock_start;
2578         u64 lock_end;
2579         bool merge = false;
2580         int index;
2581
2582         if (prev && prev->root_id == backref->root_id &&
2583             prev->inum == backref->inum &&
2584             prev->file_pos + prev->num_bytes == backref->file_pos)
2585                 merge = true;
2586
2587         /* step 1: get root */
2588         key.objectid = backref->root_id;
2589         key.type = BTRFS_ROOT_ITEM_KEY;
2590         key.offset = (u64)-1;
2591
2592         index = srcu_read_lock(&fs_info->subvol_srcu);
2593
2594         root = btrfs_read_fs_root_no_name(fs_info, &key);
2595         if (IS_ERR(root)) {
2596                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2597                 if (PTR_ERR(root) == -ENOENT)
2598                         return 0;
2599                 return PTR_ERR(root);
2600         }
2601
2602         if (btrfs_root_readonly(root)) {
2603                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2604                 return 0;
2605         }
2606
2607         /* step 2: get inode */
2608         key.objectid = backref->inum;
2609         key.type = BTRFS_INODE_ITEM_KEY;
2610         key.offset = 0;
2611
2612         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2613         if (IS_ERR(inode)) {
2614                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2615                 return 0;
2616         }
2617
2618         srcu_read_unlock(&fs_info->subvol_srcu, index);
2619
2620         /* step 3: relink backref */
2621         lock_start = backref->file_pos;
2622         lock_end = backref->file_pos + backref->num_bytes - 1;
2623         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2624                          &cached);
2625
2626         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2627         if (ordered) {
2628                 btrfs_put_ordered_extent(ordered);
2629                 goto out_unlock;
2630         }
2631
2632         trans = btrfs_join_transaction(root);
2633         if (IS_ERR(trans)) {
2634                 ret = PTR_ERR(trans);
2635                 goto out_unlock;
2636         }
2637
2638         key.objectid = backref->inum;
2639         key.type = BTRFS_EXTENT_DATA_KEY;
2640         key.offset = backref->file_pos;
2641
2642         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2643         if (ret < 0) {
2644                 goto out_free_path;
2645         } else if (ret > 0) {
2646                 ret = 0;
2647                 goto out_free_path;
2648         }
2649
2650         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2651                                 struct btrfs_file_extent_item);
2652
2653         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2654             backref->generation)
2655                 goto out_free_path;
2656
2657         btrfs_release_path(path);
2658
2659         start = backref->file_pos;
2660         if (backref->extent_offset < old->extent_offset + old->offset)
2661                 start += old->extent_offset + old->offset -
2662                          backref->extent_offset;
2663
2664         len = min(backref->extent_offset + backref->num_bytes,
2665                   old->extent_offset + old->offset + old->len);
2666         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2667
2668         ret = btrfs_drop_extents(trans, root, inode, start,
2669                                  start + len, 1);
2670         if (ret)
2671                 goto out_free_path;
2672 again:
2673         key.objectid = btrfs_ino(BTRFS_I(inode));
2674         key.type = BTRFS_EXTENT_DATA_KEY;
2675         key.offset = start;
2676
2677         path->leave_spinning = 1;
2678         if (merge) {
2679                 struct btrfs_file_extent_item *fi;
2680                 u64 extent_len;
2681                 struct btrfs_key found_key;
2682
2683                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2684                 if (ret < 0)
2685                         goto out_free_path;
2686
2687                 path->slots[0]--;
2688                 leaf = path->nodes[0];
2689                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2690
2691                 fi = btrfs_item_ptr(leaf, path->slots[0],
2692                                     struct btrfs_file_extent_item);
2693                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2694
2695                 if (extent_len + found_key.offset == start &&
2696                     relink_is_mergable(leaf, fi, new)) {
2697                         btrfs_set_file_extent_num_bytes(leaf, fi,
2698                                                         extent_len + len);
2699                         btrfs_mark_buffer_dirty(leaf);
2700                         inode_add_bytes(inode, len);
2701
2702                         ret = 1;
2703                         goto out_free_path;
2704                 } else {
2705                         merge = false;
2706                         btrfs_release_path(path);
2707                         goto again;
2708                 }
2709         }
2710
2711         ret = btrfs_insert_empty_item(trans, root, path, &key,
2712                                         sizeof(*extent));
2713         if (ret) {
2714                 btrfs_abort_transaction(trans, ret);
2715                 goto out_free_path;
2716         }
2717
2718         leaf = path->nodes[0];
2719         item = btrfs_item_ptr(leaf, path->slots[0],
2720                                 struct btrfs_file_extent_item);
2721         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2722         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2723         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2724         btrfs_set_file_extent_num_bytes(leaf, item, len);
2725         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2726         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2727         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2728         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2729         btrfs_set_file_extent_encryption(leaf, item, 0);
2730         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2731
2732         btrfs_mark_buffer_dirty(leaf);
2733         inode_add_bytes(inode, len);
2734         btrfs_release_path(path);
2735
2736         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2737                         new->disk_len, 0,
2738                         backref->root_id, backref->inum,
2739                         new->file_pos); /* start - extent_offset */
2740         if (ret) {
2741                 btrfs_abort_transaction(trans, ret);
2742                 goto out_free_path;
2743         }
2744
2745         ret = 1;
2746 out_free_path:
2747         btrfs_release_path(path);
2748         path->leave_spinning = 0;
2749         btrfs_end_transaction(trans);
2750 out_unlock:
2751         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2752                              &cached);
2753         iput(inode);
2754         return ret;
2755 }
2756
2757 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2758 {
2759         struct old_sa_defrag_extent *old, *tmp;
2760
2761         if (!new)
2762                 return;
2763
2764         list_for_each_entry_safe(old, tmp, &new->head, list) {
2765                 kfree(old);
2766         }
2767         kfree(new);
2768 }
2769
2770 static void relink_file_extents(struct new_sa_defrag_extent *new)
2771 {
2772         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2773         struct btrfs_path *path;
2774         struct sa_defrag_extent_backref *backref;
2775         struct sa_defrag_extent_backref *prev = NULL;
2776         struct inode *inode;
2777         struct btrfs_root *root;
2778         struct rb_node *node;
2779         int ret;
2780
2781         inode = new->inode;
2782         root = BTRFS_I(inode)->root;
2783
2784         path = btrfs_alloc_path();
2785         if (!path)
2786                 return;
2787
2788         if (!record_extent_backrefs(path, new)) {
2789                 btrfs_free_path(path);
2790                 goto out;
2791         }
2792         btrfs_release_path(path);
2793
2794         while (1) {
2795                 node = rb_first(&new->root);
2796                 if (!node)
2797                         break;
2798                 rb_erase(node, &new->root);
2799
2800                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2801
2802                 ret = relink_extent_backref(path, prev, backref);
2803                 WARN_ON(ret < 0);
2804
2805                 kfree(prev);
2806
2807                 if (ret == 1)
2808                         prev = backref;
2809                 else
2810                         prev = NULL;
2811                 cond_resched();
2812         }
2813         kfree(prev);
2814
2815         btrfs_free_path(path);
2816 out:
2817         free_sa_defrag_extent(new);
2818
2819         atomic_dec(&fs_info->defrag_running);
2820         wake_up(&fs_info->transaction_wait);
2821 }
2822
2823 static struct new_sa_defrag_extent *
2824 record_old_file_extents(struct inode *inode,
2825                         struct btrfs_ordered_extent *ordered)
2826 {
2827         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2828         struct btrfs_root *root = BTRFS_I(inode)->root;
2829         struct btrfs_path *path;
2830         struct btrfs_key key;
2831         struct old_sa_defrag_extent *old;
2832         struct new_sa_defrag_extent *new;
2833         int ret;
2834
2835         new = kmalloc(sizeof(*new), GFP_NOFS);
2836         if (!new)
2837                 return NULL;
2838
2839         new->inode = inode;
2840         new->file_pos = ordered->file_offset;
2841         new->len = ordered->len;
2842         new->bytenr = ordered->start;
2843         new->disk_len = ordered->disk_len;
2844         new->compress_type = ordered->compress_type;
2845         new->root = RB_ROOT;
2846         INIT_LIST_HEAD(&new->head);
2847
2848         path = btrfs_alloc_path();
2849         if (!path)
2850                 goto out_kfree;
2851
2852         key.objectid = btrfs_ino(BTRFS_I(inode));
2853         key.type = BTRFS_EXTENT_DATA_KEY;
2854         key.offset = new->file_pos;
2855
2856         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2857         if (ret < 0)
2858                 goto out_free_path;
2859         if (ret > 0 && path->slots[0] > 0)
2860                 path->slots[0]--;
2861
2862         /* find out all the old extents for the file range */
2863         while (1) {
2864                 struct btrfs_file_extent_item *extent;
2865                 struct extent_buffer *l;
2866                 int slot;
2867                 u64 num_bytes;
2868                 u64 offset;
2869                 u64 end;
2870                 u64 disk_bytenr;
2871                 u64 extent_offset;
2872
2873                 l = path->nodes[0];
2874                 slot = path->slots[0];
2875
2876                 if (slot >= btrfs_header_nritems(l)) {
2877                         ret = btrfs_next_leaf(root, path);
2878                         if (ret < 0)
2879                                 goto out_free_path;
2880                         else if (ret > 0)
2881                                 break;
2882                         continue;
2883                 }
2884
2885                 btrfs_item_key_to_cpu(l, &key, slot);
2886
2887                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2888                         break;
2889                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2890                         break;
2891                 if (key.offset >= new->file_pos + new->len)
2892                         break;
2893
2894                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2895
2896                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2897                 if (key.offset + num_bytes < new->file_pos)
2898                         goto next;
2899
2900                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2901                 if (!disk_bytenr)
2902                         goto next;
2903
2904                 extent_offset = btrfs_file_extent_offset(l, extent);
2905
2906                 old = kmalloc(sizeof(*old), GFP_NOFS);
2907                 if (!old)
2908                         goto out_free_path;
2909
2910                 offset = max(new->file_pos, key.offset);
2911                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2912
2913                 old->bytenr = disk_bytenr;
2914                 old->extent_offset = extent_offset;
2915                 old->offset = offset - key.offset;
2916                 old->len = end - offset;
2917                 old->new = new;
2918                 old->count = 0;
2919                 list_add_tail(&old->list, &new->head);
2920 next:
2921                 path->slots[0]++;
2922                 cond_resched();
2923         }
2924
2925         btrfs_free_path(path);
2926         atomic_inc(&fs_info->defrag_running);
2927
2928         return new;
2929
2930 out_free_path:
2931         btrfs_free_path(path);
2932 out_kfree:
2933         free_sa_defrag_extent(new);
2934         return NULL;
2935 }
2936
2937 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2938                                          u64 start, u64 len)
2939 {
2940         struct btrfs_block_group_cache *cache;
2941
2942         cache = btrfs_lookup_block_group(fs_info, start);
2943         ASSERT(cache);
2944
2945         spin_lock(&cache->lock);
2946         cache->delalloc_bytes -= len;
2947         spin_unlock(&cache->lock);
2948
2949         btrfs_put_block_group(cache);
2950 }
2951
2952 /* as ordered data IO finishes, this gets called so we can finish
2953  * an ordered extent if the range of bytes in the file it covers are
2954  * fully written.
2955  */
2956 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2957 {
2958         struct inode *inode = ordered_extent->inode;
2959         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2960         struct btrfs_root *root = BTRFS_I(inode)->root;
2961         struct btrfs_trans_handle *trans = NULL;
2962         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2963         struct extent_state *cached_state = NULL;
2964         struct new_sa_defrag_extent *new = NULL;
2965         int compress_type = 0;
2966         int ret = 0;
2967         u64 logical_len = ordered_extent->len;
2968         bool nolock;
2969         bool truncated = false;
2970         bool range_locked = false;
2971         bool clear_new_delalloc_bytes = false;
2972
2973         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2974             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2975             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2976                 clear_new_delalloc_bytes = true;
2977
2978         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2979
2980         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2981                 ret = -EIO;
2982                 goto out;
2983         }
2984
2985         btrfs_free_io_failure_record(BTRFS_I(inode),
2986                         ordered_extent->file_offset,
2987                         ordered_extent->file_offset +
2988                         ordered_extent->len - 1);
2989
2990         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2991                 truncated = true;
2992                 logical_len = ordered_extent->truncated_len;
2993                 /* Truncated the entire extent, don't bother adding */
2994                 if (!logical_len)
2995                         goto out;
2996         }
2997
2998         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2999                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
3000
3001                 /*
3002                  * For mwrite(mmap + memset to write) case, we still reserve
3003                  * space for NOCOW range.
3004                  * As NOCOW won't cause a new delayed ref, just free the space
3005                  */
3006                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3007                                        ordered_extent->len);
3008                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3009                 if (nolock)
3010                         trans = btrfs_join_transaction_nolock(root);
3011                 else
3012                         trans = btrfs_join_transaction(root);
3013                 if (IS_ERR(trans)) {
3014                         ret = PTR_ERR(trans);
3015                         trans = NULL;
3016                         goto out;
3017                 }
3018                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3019                 ret = btrfs_update_inode_fallback(trans, root, inode);
3020                 if (ret) /* -ENOMEM or corruption */
3021                         btrfs_abort_transaction(trans, ret);
3022                 goto out;
3023         }
3024
3025         range_locked = true;
3026         lock_extent_bits(io_tree, ordered_extent->file_offset,
3027                          ordered_extent->file_offset + ordered_extent->len - 1,
3028                          &cached_state);
3029
3030         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3031                         ordered_extent->file_offset + ordered_extent->len - 1,
3032                         EXTENT_DEFRAG, 0, cached_state);
3033         if (ret) {
3034                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3035                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3036                         /* the inode is shared */
3037                         new = record_old_file_extents(inode, ordered_extent);
3038
3039                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3040                         ordered_extent->file_offset + ordered_extent->len - 1,
3041                         EXTENT_DEFRAG, 0, 0, &cached_state);
3042         }
3043
3044         if (nolock)
3045                 trans = btrfs_join_transaction_nolock(root);
3046         else
3047                 trans = btrfs_join_transaction(root);
3048         if (IS_ERR(trans)) {
3049                 ret = PTR_ERR(trans);
3050                 trans = NULL;
3051                 goto out;
3052         }
3053
3054         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3055
3056         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3057                 compress_type = ordered_extent->compress_type;
3058         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3059                 BUG_ON(compress_type);
3060                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3061                                        ordered_extent->len);
3062                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3063                                                 ordered_extent->file_offset,
3064                                                 ordered_extent->file_offset +
3065                                                 logical_len);
3066         } else {
3067                 BUG_ON(root == fs_info->tree_root);
3068                 ret = insert_reserved_file_extent(trans, inode,
3069                                                 ordered_extent->file_offset,
3070                                                 ordered_extent->start,
3071                                                 ordered_extent->disk_len,
3072                                                 logical_len, logical_len,
3073                                                 compress_type, 0, 0,
3074                                                 BTRFS_FILE_EXTENT_REG);
3075                 if (!ret)
3076                         btrfs_release_delalloc_bytes(fs_info,
3077                                                      ordered_extent->start,
3078                                                      ordered_extent->disk_len);
3079         }
3080         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3081                            ordered_extent->file_offset, ordered_extent->len,
3082                            trans->transid);
3083         if (ret < 0) {
3084                 btrfs_abort_transaction(trans, ret);
3085                 goto out;
3086         }
3087
3088         ret = add_pending_csums(trans, inode, &ordered_extent->list);
3089         if (ret) {
3090                 btrfs_abort_transaction(trans, ret);
3091                 goto out;
3092         }
3093
3094         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3095         ret = btrfs_update_inode_fallback(trans, root, inode);
3096         if (ret) { /* -ENOMEM or corruption */
3097                 btrfs_abort_transaction(trans, ret);
3098                 goto out;
3099         }
3100         ret = 0;
3101 out:
3102         if (range_locked || clear_new_delalloc_bytes) {
3103                 unsigned int clear_bits = 0;
3104
3105                 if (range_locked)
3106                         clear_bits |= EXTENT_LOCKED;
3107                 if (clear_new_delalloc_bytes)
3108                         clear_bits |= EXTENT_DELALLOC_NEW;
3109                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3110                                  ordered_extent->file_offset,
3111                                  ordered_extent->file_offset +
3112                                  ordered_extent->len - 1,
3113                                  clear_bits,
3114                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3115                                  0, &cached_state);
3116         }
3117
3118         if (trans)
3119                 btrfs_end_transaction(trans);
3120
3121         if (ret || truncated) {
3122                 u64 start, end;
3123
3124                 if (truncated)
3125                         start = ordered_extent->file_offset + logical_len;
3126                 else
3127                         start = ordered_extent->file_offset;
3128                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3129                 clear_extent_uptodate(io_tree, start, end, NULL);
3130
3131                 /* Drop the cache for the part of the extent we didn't write. */
3132                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3133
3134                 /*
3135                  * If the ordered extent had an IOERR or something else went
3136                  * wrong we need to return the space for this ordered extent
3137                  * back to the allocator.  We only free the extent in the
3138                  * truncated case if we didn't write out the extent at all.
3139                  */
3140                 if ((ret || !logical_len) &&
3141                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3142                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3143                         btrfs_free_reserved_extent(fs_info,
3144                                                    ordered_extent->start,
3145                                                    ordered_extent->disk_len, 1);
3146         }
3147
3148
3149         /*
3150          * This needs to be done to make sure anybody waiting knows we are done
3151          * updating everything for this ordered extent.
3152          */
3153         btrfs_remove_ordered_extent(inode, ordered_extent);
3154
3155         /* for snapshot-aware defrag */
3156         if (new) {
3157                 if (ret) {
3158                         free_sa_defrag_extent(new);
3159                         atomic_dec(&fs_info->defrag_running);
3160                 } else {
3161                         relink_file_extents(new);
3162                 }
3163         }
3164
3165         /* once for us */
3166         btrfs_put_ordered_extent(ordered_extent);
3167         /* once for the tree */
3168         btrfs_put_ordered_extent(ordered_extent);
3169
3170         return ret;
3171 }
3172
3173 static void finish_ordered_fn(struct btrfs_work *work)
3174 {
3175         struct btrfs_ordered_extent *ordered_extent;
3176         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3177         btrfs_finish_ordered_io(ordered_extent);
3178 }
3179
3180 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3181                                 struct extent_state *state, int uptodate)
3182 {
3183         struct inode *inode = page->mapping->host;
3184         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3185         struct btrfs_ordered_extent *ordered_extent = NULL;
3186         struct btrfs_workqueue *wq;
3187         btrfs_work_func_t func;
3188
3189         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3190
3191         ClearPagePrivate2(page);
3192         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3193                                             end - start + 1, uptodate))
3194                 return;
3195
3196         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3197                 wq = fs_info->endio_freespace_worker;
3198                 func = btrfs_freespace_write_helper;
3199         } else {
3200                 wq = fs_info->endio_write_workers;
3201                 func = btrfs_endio_write_helper;
3202         }
3203
3204         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3205                         NULL);
3206         btrfs_queue_work(wq, &ordered_extent->work);
3207 }
3208
3209 static int __readpage_endio_check(struct inode *inode,
3210                                   struct btrfs_io_bio *io_bio,
3211                                   int icsum, struct page *page,
3212                                   int pgoff, u64 start, size_t len)
3213 {
3214         char *kaddr;
3215         u32 csum_expected;
3216         u32 csum = ~(u32)0;
3217
3218         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3219
3220         kaddr = kmap_atomic(page);
3221         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3222         btrfs_csum_final(csum, (u8 *)&csum);
3223         if (csum != csum_expected)
3224                 goto zeroit;
3225
3226         kunmap_atomic(kaddr);
3227         return 0;
3228 zeroit:
3229         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3230                                     io_bio->mirror_num);
3231         memset(kaddr + pgoff, 1, len);
3232         flush_dcache_page(page);
3233         kunmap_atomic(kaddr);
3234         return -EIO;
3235 }
3236
3237 /*
3238  * when reads are done, we need to check csums to verify the data is correct
3239  * if there's a match, we allow the bio to finish.  If not, the code in
3240  * extent_io.c will try to find good copies for us.
3241  */
3242 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3243                                       u64 phy_offset, struct page *page,
3244                                       u64 start, u64 end, int mirror)
3245 {
3246         size_t offset = start - page_offset(page);
3247         struct inode *inode = page->mapping->host;
3248         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3249         struct btrfs_root *root = BTRFS_I(inode)->root;
3250
3251         if (PageChecked(page)) {
3252                 ClearPageChecked(page);
3253                 return 0;
3254         }
3255
3256         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3257                 return 0;
3258
3259         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3260             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3261                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3262                 return 0;
3263         }
3264
3265         phy_offset >>= inode->i_sb->s_blocksize_bits;
3266         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3267                                       start, (size_t)(end - start + 1));
3268 }
3269
3270 /*
3271  * btrfs_add_delayed_iput - perform a delayed iput on @inode
3272  *
3273  * @inode: The inode we want to perform iput on
3274  *
3275  * This function uses the generic vfs_inode::i_count to track whether we should
3276  * just decrement it (in case it's > 1) or if this is the last iput then link
3277  * the inode to the delayed iput machinery. Delayed iputs are processed at
3278  * transaction commit time/superblock commit/cleaner kthread.
3279  */
3280 void btrfs_add_delayed_iput(struct inode *inode)
3281 {
3282         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3283         struct btrfs_inode *binode = BTRFS_I(inode);
3284
3285         if (atomic_add_unless(&inode->i_count, -1, 1))
3286                 return;
3287
3288         spin_lock(&fs_info->delayed_iput_lock);
3289         ASSERT(list_empty(&binode->delayed_iput));
3290         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3291         spin_unlock(&fs_info->delayed_iput_lock);
3292 }
3293
3294 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3295 {
3296
3297         spin_lock(&fs_info->delayed_iput_lock);
3298         while (!list_empty(&fs_info->delayed_iputs)) {
3299                 struct btrfs_inode *inode;
3300
3301                 inode = list_first_entry(&fs_info->delayed_iputs,
3302                                 struct btrfs_inode, delayed_iput);
3303                 list_del_init(&inode->delayed_iput);
3304                 spin_unlock(&fs_info->delayed_iput_lock);
3305                 iput(&inode->vfs_inode);
3306                 spin_lock(&fs_info->delayed_iput_lock);
3307         }
3308         spin_unlock(&fs_info->delayed_iput_lock);
3309 }
3310
3311 /*
3312  * This is called in transaction commit time. If there are no orphan
3313  * files in the subvolume, it removes orphan item and frees block_rsv
3314  * structure.
3315  */
3316 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3317                               struct btrfs_root *root)
3318 {
3319         struct btrfs_fs_info *fs_info = root->fs_info;
3320         struct btrfs_block_rsv *block_rsv;
3321         int ret;
3322
3323         if (atomic_read(&root->orphan_inodes) ||
3324             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3325                 return;
3326
3327         spin_lock(&root->orphan_lock);
3328         if (atomic_read(&root->orphan_inodes)) {
3329                 spin_unlock(&root->orphan_lock);
3330                 return;
3331         }
3332
3333         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3334                 spin_unlock(&root->orphan_lock);
3335                 return;
3336         }
3337
3338         block_rsv = root->orphan_block_rsv;
3339         root->orphan_block_rsv = NULL;
3340         spin_unlock(&root->orphan_lock);
3341
3342         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3343             btrfs_root_refs(&root->root_item) > 0) {
3344                 ret = btrfs_del_orphan_item(trans, fs_info->tree_root,
3345                                             root->root_key.objectid);
3346                 if (ret)
3347                         btrfs_abort_transaction(trans, ret);
3348                 else
3349                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3350                                   &root->state);
3351         }
3352
3353         if (block_rsv) {
3354                 WARN_ON(block_rsv->size > 0);
3355                 btrfs_free_block_rsv(fs_info, block_rsv);
3356         }
3357 }
3358
3359 /*
3360  * This creates an orphan entry for the given inode in case something goes
3361  * wrong in the middle of an unlink/truncate.
3362  *
3363  * NOTE: caller of this function should reserve 5 units of metadata for
3364  *       this function.
3365  */
3366 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3367                 struct btrfs_inode *inode)
3368 {
3369         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
3370         struct btrfs_root *root = inode->root;
3371         struct btrfs_block_rsv *block_rsv = NULL;
3372         int reserve = 0;
3373         int insert = 0;
3374         int ret;
3375
3376         if (!root->orphan_block_rsv) {
3377                 block_rsv = btrfs_alloc_block_rsv(fs_info,
3378                                                   BTRFS_BLOCK_RSV_TEMP);
3379                 if (!block_rsv)
3380                         return -ENOMEM;
3381         }
3382
3383         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3384                               &inode->runtime_flags)) {
3385 #if 0
3386                 /*
3387                  * For proper ENOSPC handling, we should do orphan
3388                  * cleanup when mounting. But this introduces backward
3389                  * compatibility issue.
3390                  */
3391                 if (!xchg(&root->orphan_item_inserted, 1))
3392                         insert = 2;
3393                 else
3394                         insert = 1;
3395 #endif
3396                 insert = 1;
3397         }
3398
3399         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3400                               &inode->runtime_flags))
3401                 reserve = 1;
3402
3403         spin_lock(&root->orphan_lock);
3404         /* If someone has created ->orphan_block_rsv, be happy to use it. */
3405         if (!root->orphan_block_rsv) {
3406                 root->orphan_block_rsv = block_rsv;
3407         } else if (block_rsv) {
3408                 btrfs_free_block_rsv(fs_info, block_rsv);
3409                 block_rsv = NULL;
3410         }
3411
3412         if (insert)
3413                 atomic_inc(&root->orphan_inodes);
3414         spin_unlock(&root->orphan_lock);
3415
3416         /* grab metadata reservation from transaction handle */
3417         if (reserve) {
3418                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3419                 ASSERT(!ret);
3420                 if (ret) {
3421                         /*
3422                          * dec doesn't need spin_lock as ->orphan_block_rsv
3423                          * would be released only if ->orphan_inodes is
3424                          * zero.
3425                          */
3426                         atomic_dec(&root->orphan_inodes);
3427                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3428                                   &inode->runtime_flags);
3429                         if (insert)
3430                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3431                                           &inode->runtime_flags);
3432                         return ret;
3433                 }
3434         }
3435
3436         /* insert an orphan item to track this unlinked/truncated file */
3437         if (insert >= 1) {
3438                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3439                 if (ret) {
3440                         if (reserve) {
3441                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3442                                           &inode->runtime_flags);
3443                                 btrfs_orphan_release_metadata(inode);
3444                         }
3445                         /*
3446                          * btrfs_orphan_commit_root may race with us and set
3447                          * ->orphan_block_rsv to zero, in order to avoid that,
3448                          * decrease ->orphan_inodes after everything is done.
3449                          */
3450                         atomic_dec(&root->orphan_inodes);
3451                         if (ret != -EEXIST) {
3452                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3453                                           &inode->runtime_flags);
3454                                 btrfs_abort_transaction(trans, ret);
3455                                 return ret;
3456                         }
3457                 }
3458                 ret = 0;
3459         }
3460
3461         /* insert an orphan item to track subvolume contains orphan files */
3462         if (insert >= 2) {
3463                 ret = btrfs_insert_orphan_item(trans, fs_info->tree_root,
3464                                                root->root_key.objectid);
3465                 if (ret && ret != -EEXIST) {
3466                         btrfs_abort_transaction(trans, ret);
3467                         return ret;
3468                 }
3469         }
3470         return 0;
3471 }
3472
3473 /*
3474  * We have done the truncate/delete so we can go ahead and remove the orphan
3475  * item for this particular inode.
3476  */
3477 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3478                             struct btrfs_inode *inode)
3479 {
3480         struct btrfs_root *root = inode->root;
3481         int delete_item = 0;
3482         int ret = 0;
3483
3484         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3485                                &inode->runtime_flags))
3486                 delete_item = 1;
3487
3488         if (delete_item && trans)
3489                 ret = btrfs_del_orphan_item(trans, root, btrfs_ino(inode));
3490
3491         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3492                                &inode->runtime_flags))
3493                 btrfs_orphan_release_metadata(inode);
3494
3495         /*
3496          * btrfs_orphan_commit_root may race with us and set ->orphan_block_rsv
3497          * to zero, in order to avoid that, decrease ->orphan_inodes after
3498          * everything is done.
3499          */
3500         if (delete_item)
3501                 atomic_dec(&root->orphan_inodes);
3502
3503         return ret;
3504 }
3505
3506 /*
3507  * this cleans up any orphans that may be left on the list from the last use
3508  * of this root.
3509  */
3510 int btrfs_orphan_cleanup(struct btrfs_root *root)
3511 {
3512         struct btrfs_fs_info *fs_info = root->fs_info;
3513         struct btrfs_path *path;
3514         struct extent_buffer *leaf;
3515         struct btrfs_key key, found_key;
3516         struct btrfs_trans_handle *trans;
3517         struct inode *inode;
3518         u64 last_objectid = 0;
3519         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3520
3521         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3522                 return 0;
3523
3524         path = btrfs_alloc_path();
3525         if (!path) {
3526                 ret = -ENOMEM;
3527                 goto out;
3528         }
3529         path->reada = READA_BACK;
3530
3531         key.objectid = BTRFS_ORPHAN_OBJECTID;
3532         key.type = BTRFS_ORPHAN_ITEM_KEY;
3533         key.offset = (u64)-1;
3534
3535         while (1) {
3536                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3537                 if (ret < 0)
3538                         goto out;
3539
3540                 /*
3541                  * if ret == 0 means we found what we were searching for, which
3542                  * is weird, but possible, so only screw with path if we didn't
3543                  * find the key and see if we have stuff that matches
3544                  */
3545                 if (ret > 0) {
3546                         ret = 0;
3547                         if (path->slots[0] == 0)
3548                                 break;
3549                         path->slots[0]--;
3550                 }
3551
3552                 /* pull out the item */
3553                 leaf = path->nodes[0];
3554                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3555
3556                 /* make sure the item matches what we want */
3557                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3558                         break;
3559                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3560                         break;
3561
3562                 /* release the path since we're done with it */
3563                 btrfs_release_path(path);
3564
3565                 /*
3566                  * this is where we are basically btrfs_lookup, without the
3567                  * crossing root thing.  we store the inode number in the
3568                  * offset of the orphan item.
3569                  */
3570
3571                 if (found_key.offset == last_objectid) {
3572                         btrfs_err(fs_info,
3573                                   "Error removing orphan entry, stopping orphan cleanup");
3574                         ret = -EINVAL;
3575                         goto out;
3576                 }
3577
3578                 last_objectid = found_key.offset;
3579
3580                 found_key.objectid = found_key.offset;
3581                 found_key.type = BTRFS_INODE_ITEM_KEY;
3582                 found_key.offset = 0;
3583                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3584                 ret = PTR_ERR_OR_ZERO(inode);
3585                 if (ret && ret != -ENOENT)
3586                         goto out;
3587
3588                 if (ret == -ENOENT && root == fs_info->tree_root) {
3589                         struct btrfs_root *dead_root;
3590                         struct btrfs_fs_info *fs_info = root->fs_info;
3591                         int is_dead_root = 0;
3592
3593                         /*
3594                          * this is an orphan in the tree root. Currently these
3595                          * could come from 2 sources:
3596                          *  a) a snapshot deletion in progress
3597                          *  b) a free space cache inode
3598                          * We need to distinguish those two, as the snapshot
3599                          * orphan must not get deleted.
3600                          * find_dead_roots already ran before us, so if this
3601                          * is a snapshot deletion, we should find the root
3602                          * in the dead_roots list
3603                          */
3604                         spin_lock(&fs_info->trans_lock);
3605                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3606                                             root_list) {
3607                                 if (dead_root->root_key.objectid ==
3608                                     found_key.objectid) {
3609                                         is_dead_root = 1;
3610                                         break;
3611                                 }
3612                         }
3613                         spin_unlock(&fs_info->trans_lock);
3614                         if (is_dead_root) {
3615                                 /* prevent this orphan from being found again */
3616                                 key.offset = found_key.objectid - 1;
3617                                 continue;
3618                         }
3619                 }
3620                 /*
3621                  * Inode is already gone but the orphan item is still there,
3622                  * kill the orphan item.
3623                  */
3624                 if (ret == -ENOENT) {
3625                         trans = btrfs_start_transaction(root, 1);
3626                         if (IS_ERR(trans)) {
3627                                 ret = PTR_ERR(trans);
3628                                 goto out;
3629                         }
3630                         btrfs_debug(fs_info, "auto deleting %Lu",
3631                                     found_key.objectid);
3632                         ret = btrfs_del_orphan_item(trans, root,
3633                                                     found_key.objectid);
3634                         btrfs_end_transaction(trans);
3635                         if (ret)
3636                                 goto out;
3637                         continue;
3638                 }
3639
3640                 /*
3641                  * add this inode to the orphan list so btrfs_orphan_del does
3642                  * the proper thing when we hit it
3643                  */
3644                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3645                         &BTRFS_I(inode)->runtime_flags);
3646                 atomic_inc(&root->orphan_inodes);
3647
3648                 /* if we have links, this was a truncate, lets do that */
3649                 if (inode->i_nlink) {
3650                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3651                                 iput(inode);
3652                                 continue;
3653                         }
3654                         nr_truncate++;
3655
3656                         /* 1 for the orphan item deletion. */
3657                         trans = btrfs_start_transaction(root, 1);
3658                         if (IS_ERR(trans)) {
3659                                 iput(inode);
3660                                 ret = PTR_ERR(trans);
3661                                 goto out;
3662                         }
3663                         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3664                         btrfs_end_transaction(trans);
3665                         if (ret) {
3666                                 iput(inode);
3667                                 goto out;
3668                         }
3669
3670                         ret = btrfs_truncate(inode, false);
3671                         if (ret)
3672                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
3673                 } else {
3674                         nr_unlink++;
3675                 }
3676
3677                 /* this will do delete_inode and everything for us */
3678                 iput(inode);
3679                 if (ret)
3680                         goto out;
3681         }
3682         /* release the path since we're done with it */
3683         btrfs_release_path(path);
3684
3685         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3686
3687         if (root->orphan_block_rsv)
3688                 btrfs_block_rsv_release(fs_info, root->orphan_block_rsv,
3689                                         (u64)-1);
3690
3691         if (root->orphan_block_rsv ||
3692             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3693                 trans = btrfs_join_transaction(root);
3694                 if (!IS_ERR(trans))
3695                         btrfs_end_transaction(trans);
3696         }
3697
3698         if (nr_unlink)
3699                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3700         if (nr_truncate)
3701                 btrfs_debug(fs_info, "truncated %d orphans", nr_truncate);
3702
3703 out:
3704         if (ret)
3705                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3706         btrfs_free_path(path);
3707         return ret;
3708 }
3709
3710 /*
3711  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3712  * don't find any xattrs, we know there can't be any acls.
3713  *
3714  * slot is the slot the inode is in, objectid is the objectid of the inode
3715  */
3716 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3717                                           int slot, u64 objectid,
3718                                           int *first_xattr_slot)
3719 {
3720         u32 nritems = btrfs_header_nritems(leaf);
3721         struct btrfs_key found_key;
3722         static u64 xattr_access = 0;
3723         static u64 xattr_default = 0;
3724         int scanned = 0;
3725
3726         if (!xattr_access) {
3727                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3728                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3729                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3730                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3731         }
3732
3733         slot++;
3734         *first_xattr_slot = -1;
3735         while (slot < nritems) {
3736                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3737
3738                 /* we found a different objectid, there must not be acls */
3739                 if (found_key.objectid != objectid)
3740                         return 0;
3741
3742                 /* we found an xattr, assume we've got an acl */
3743                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3744                         if (*first_xattr_slot == -1)
3745                                 *first_xattr_slot = slot;
3746                         if (found_key.offset == xattr_access ||
3747                             found_key.offset == xattr_default)
3748                                 return 1;
3749                 }
3750
3751                 /*
3752                  * we found a key greater than an xattr key, there can't
3753                  * be any acls later on
3754                  */
3755                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3756                         return 0;
3757
3758                 slot++;
3759                 scanned++;
3760
3761                 /*
3762                  * it goes inode, inode backrefs, xattrs, extents,
3763                  * so if there are a ton of hard links to an inode there can
3764                  * be a lot of backrefs.  Don't waste time searching too hard,
3765                  * this is just an optimization
3766                  */
3767                 if (scanned >= 8)
3768                         break;
3769         }
3770         /* we hit the end of the leaf before we found an xattr or
3771          * something larger than an xattr.  We have to assume the inode
3772          * has acls
3773          */
3774         if (*first_xattr_slot == -1)
3775                 *first_xattr_slot = slot;
3776         return 1;
3777 }
3778
3779 /*
3780  * read an inode from the btree into the in-memory inode
3781  */
3782 static int btrfs_read_locked_inode(struct inode *inode)
3783 {
3784         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3785         struct btrfs_path *path;
3786         struct extent_buffer *leaf;
3787         struct btrfs_inode_item *inode_item;
3788         struct btrfs_root *root = BTRFS_I(inode)->root;
3789         struct btrfs_key location;
3790         unsigned long ptr;
3791         int maybe_acls;
3792         u32 rdev;
3793         int ret;
3794         bool filled = false;
3795         int first_xattr_slot;
3796
3797         ret = btrfs_fill_inode(inode, &rdev);
3798         if (!ret)
3799                 filled = true;
3800
3801         path = btrfs_alloc_path();
3802         if (!path) {
3803                 ret = -ENOMEM;
3804                 goto make_bad;
3805         }
3806
3807         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3808
3809         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3810         if (ret) {
3811                 if (ret > 0)
3812                         ret = -ENOENT;
3813                 goto make_bad;
3814         }
3815
3816         leaf = path->nodes[0];
3817
3818         if (filled)
3819                 goto cache_index;
3820
3821         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3822                                     struct btrfs_inode_item);
3823         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3824         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3825         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3826         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3827         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3828
3829         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3830         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3831
3832         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3833         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3834
3835         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3836         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3837
3838         BTRFS_I(inode)->i_otime.tv_sec =
3839                 btrfs_timespec_sec(leaf, &inode_item->otime);
3840         BTRFS_I(inode)->i_otime.tv_nsec =
3841                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3842
3843         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3844         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3845         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3846
3847         inode_set_iversion_queried(inode,
3848                                    btrfs_inode_sequence(leaf, inode_item));
3849         inode->i_generation = BTRFS_I(inode)->generation;
3850         inode->i_rdev = 0;
3851         rdev = btrfs_inode_rdev(leaf, inode_item);
3852
3853         BTRFS_I(inode)->index_cnt = (u64)-1;
3854         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3855
3856 cache_index:
3857         /*
3858          * If we were modified in the current generation and evicted from memory
3859          * and then re-read we need to do a full sync since we don't have any
3860          * idea about which extents were modified before we were evicted from
3861          * cache.
3862          *
3863          * This is required for both inode re-read from disk and delayed inode
3864          * in delayed_nodes_tree.
3865          */
3866         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3867                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3868                         &BTRFS_I(inode)->runtime_flags);
3869
3870         /*
3871          * We don't persist the id of the transaction where an unlink operation
3872          * against the inode was last made. So here we assume the inode might
3873          * have been evicted, and therefore the exact value of last_unlink_trans
3874          * lost, and set it to last_trans to avoid metadata inconsistencies
3875          * between the inode and its parent if the inode is fsync'ed and the log
3876          * replayed. For example, in the scenario:
3877          *
3878          * touch mydir/foo
3879          * ln mydir/foo mydir/bar
3880          * sync
3881          * unlink mydir/bar
3882          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3883          * xfs_io -c fsync mydir/foo
3884          * <power failure>
3885          * mount fs, triggers fsync log replay
3886          *
3887          * We must make sure that when we fsync our inode foo we also log its
3888          * parent inode, otherwise after log replay the parent still has the
3889          * dentry with the "bar" name but our inode foo has a link count of 1
3890          * and doesn't have an inode ref with the name "bar" anymore.
3891          *
3892          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3893          * but it guarantees correctness at the expense of occasional full
3894          * transaction commits on fsync if our inode is a directory, or if our
3895          * inode is not a directory, logging its parent unnecessarily.
3896          */
3897         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3898
3899         path->slots[0]++;
3900         if (inode->i_nlink != 1 ||
3901             path->slots[0] >= btrfs_header_nritems(leaf))
3902                 goto cache_acl;
3903
3904         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3905         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3906                 goto cache_acl;
3907
3908         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3909         if (location.type == BTRFS_INODE_REF_KEY) {
3910                 struct btrfs_inode_ref *ref;
3911
3912                 ref = (struct btrfs_inode_ref *)ptr;
3913                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3914         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3915                 struct btrfs_inode_extref *extref;
3916
3917                 extref = (struct btrfs_inode_extref *)ptr;
3918                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3919                                                                      extref);
3920         }
3921 cache_acl:
3922         /*
3923          * try to precache a NULL acl entry for files that don't have
3924          * any xattrs or acls
3925          */
3926         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3927                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3928         if (first_xattr_slot != -1) {
3929                 path->slots[0] = first_xattr_slot;
3930                 ret = btrfs_load_inode_props(inode, path);
3931                 if (ret)
3932                         btrfs_err(fs_info,
3933                                   "error loading props for ino %llu (root %llu): %d",
3934                                   btrfs_ino(BTRFS_I(inode)),
3935                                   root->root_key.objectid, ret);
3936         }
3937         btrfs_free_path(path);
3938
3939         if (!maybe_acls)
3940                 cache_no_acl(inode);
3941
3942         switch (inode->i_mode & S_IFMT) {
3943         case S_IFREG:
3944                 inode->i_mapping->a_ops = &btrfs_aops;
3945                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3946                 inode->i_fop = &btrfs_file_operations;
3947                 inode->i_op = &btrfs_file_inode_operations;
3948                 break;
3949         case S_IFDIR:
3950                 inode->i_fop = &btrfs_dir_file_operations;
3951                 inode->i_op = &btrfs_dir_inode_operations;
3952                 break;
3953         case S_IFLNK:
3954                 inode->i_op = &btrfs_symlink_inode_operations;
3955                 inode_nohighmem(inode);
3956                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3957                 break;
3958         default:
3959                 inode->i_op = &btrfs_special_inode_operations;
3960                 init_special_inode(inode, inode->i_mode, rdev);
3961                 break;
3962         }
3963
3964         btrfs_update_iflags(inode);
3965         return 0;
3966
3967 make_bad:
3968         btrfs_free_path(path);
3969         make_bad_inode(inode);
3970         return ret;
3971 }
3972
3973 /*
3974  * given a leaf and an inode, copy the inode fields into the leaf
3975  */
3976 static void fill_inode_item(struct btrfs_trans_handle *trans,
3977                             struct extent_buffer *leaf,
3978                             struct btrfs_inode_item *item,
3979                             struct inode *inode)
3980 {
3981         struct btrfs_map_token token;
3982
3983         btrfs_init_map_token(&token);
3984
3985         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3986         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3987         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3988                                    &token);
3989         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3990         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3991
3992         btrfs_set_token_timespec_sec(leaf, &item->atime,
3993                                      inode->i_atime.tv_sec, &token);
3994         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3995                                       inode->i_atime.tv_nsec, &token);
3996
3997         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3998                                      inode->i_mtime.tv_sec, &token);
3999         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
4000                                       inode->i_mtime.tv_nsec, &token);
4001
4002         btrfs_set_token_timespec_sec(leaf, &item->ctime,
4003                                      inode->i_ctime.tv_sec, &token);
4004         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
4005                                       inode->i_ctime.tv_nsec, &token);
4006
4007         btrfs_set_token_timespec_sec(leaf, &item->otime,
4008                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
4009         btrfs_set_token_timespec_nsec(leaf, &item->otime,
4010                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
4011
4012         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
4013                                      &token);
4014         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
4015                                          &token);
4016         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
4017                                        &token);
4018         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
4019         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
4020         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
4021         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
4022 }
4023
4024 /*
4025  * copy everything in the in-memory inode into the btree.
4026  */
4027 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
4028                                 struct btrfs_root *root, struct inode *inode)
4029 {
4030         struct btrfs_inode_item *inode_item;
4031         struct btrfs_path *path;
4032         struct extent_buffer *leaf;
4033         int ret;
4034
4035         path = btrfs_alloc_path();
4036         if (!path)
4037                 return -ENOMEM;
4038
4039         path->leave_spinning = 1;
4040         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
4041                                  1);
4042         if (ret) {
4043                 if (ret > 0)
4044                         ret = -ENOENT;
4045                 goto failed;
4046         }
4047
4048         leaf = path->nodes[0];
4049         inode_item = btrfs_item_ptr(leaf, path->slots[0],
4050                                     struct btrfs_inode_item);
4051
4052         fill_inode_item(trans, leaf, inode_item, inode);
4053         btrfs_mark_buffer_dirty(leaf);
4054         btrfs_set_inode_last_trans(trans, inode);
4055         ret = 0;
4056 failed:
4057         btrfs_free_path(path);
4058         return ret;
4059 }
4060
4061 /*
4062  * copy everything in the in-memory inode into the btree.
4063  */
4064 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
4065                                 struct btrfs_root *root, struct inode *inode)
4066 {
4067         struct btrfs_fs_info *fs_info = root->fs_info;
4068         int ret;
4069
4070         /*
4071          * If the inode is a free space inode, we can deadlock during commit
4072          * if we put it into the delayed code.
4073          *
4074          * The data relocation inode should also be directly updated
4075          * without delay
4076          */
4077         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
4078             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
4079             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
4080                 btrfs_update_root_times(trans, root);
4081
4082                 ret = btrfs_delayed_update_inode(trans, root, inode);
4083                 if (!ret)
4084                         btrfs_set_inode_last_trans(trans, inode);
4085                 return ret;
4086         }
4087
4088         return btrfs_update_inode_item(trans, root, inode);
4089 }
4090
4091 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4092                                          struct btrfs_root *root,
4093                                          struct inode *inode)
4094 {
4095         int ret;
4096
4097         ret = btrfs_update_inode(trans, root, inode);
4098         if (ret == -ENOSPC)
4099                 return btrfs_update_inode_item(trans, root, inode);
4100         return ret;
4101 }
4102
4103 /*
4104  * unlink helper that gets used here in inode.c and in the tree logging
4105  * recovery code.  It remove a link in a directory with a given name, and
4106  * also drops the back refs in the inode to the directory
4107  */
4108 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4109                                 struct btrfs_root *root,
4110                                 struct btrfs_inode *dir,
4111                                 struct btrfs_inode *inode,
4112                                 const char *name, int name_len)
4113 {
4114         struct btrfs_fs_info *fs_info = root->fs_info;
4115         struct btrfs_path *path;
4116         int ret = 0;
4117         struct extent_buffer *leaf;
4118         struct btrfs_dir_item *di;
4119         struct btrfs_key key;
4120         u64 index;
4121         u64 ino = btrfs_ino(inode);
4122         u64 dir_ino = btrfs_ino(dir);
4123
4124         path = btrfs_alloc_path();
4125         if (!path) {
4126                 ret = -ENOMEM;
4127                 goto out;
4128         }
4129
4130         path->leave_spinning = 1;
4131         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4132                                     name, name_len, -1);
4133         if (IS_ERR(di)) {
4134                 ret = PTR_ERR(di);
4135                 goto err;
4136         }
4137         if (!di) {
4138                 ret = -ENOENT;
4139                 goto err;
4140         }
4141         leaf = path->nodes[0];
4142         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4143         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4144         if (ret)
4145                 goto err;
4146         btrfs_release_path(path);
4147
4148         /*
4149          * If we don't have dir index, we have to get it by looking up
4150          * the inode ref, since we get the inode ref, remove it directly,
4151          * it is unnecessary to do delayed deletion.
4152          *
4153          * But if we have dir index, needn't search inode ref to get it.
4154          * Since the inode ref is close to the inode item, it is better
4155          * that we delay to delete it, and just do this deletion when
4156          * we update the inode item.
4157          */
4158         if (inode->dir_index) {
4159                 ret = btrfs_delayed_delete_inode_ref(inode);
4160                 if (!ret) {
4161                         index = inode->dir_index;
4162                         goto skip_backref;
4163                 }
4164         }
4165
4166         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4167                                   dir_ino, &index);
4168         if (ret) {
4169                 btrfs_info(fs_info,
4170                         "failed to delete reference to %.*s, inode %llu parent %llu",
4171                         name_len, name, ino, dir_ino);
4172                 btrfs_abort_transaction(trans, ret);
4173                 goto err;
4174         }
4175 skip_backref:
4176         ret = btrfs_delete_delayed_dir_index(trans, fs_info, dir, index);
4177         if (ret) {
4178                 btrfs_abort_transaction(trans, ret);
4179                 goto err;
4180         }
4181
4182         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4183                         dir_ino);
4184         if (ret != 0 && ret != -ENOENT) {
4185                 btrfs_abort_transaction(trans, ret);
4186                 goto err;
4187         }
4188
4189         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4190                         index);
4191         if (ret == -ENOENT)
4192                 ret = 0;
4193         else if (ret)
4194                 btrfs_abort_transaction(trans, ret);
4195 err:
4196         btrfs_free_path(path);
4197         if (ret)
4198                 goto out;
4199
4200         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4201         inode_inc_iversion(&inode->vfs_inode);
4202         inode_inc_iversion(&dir->vfs_inode);
4203         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4204                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4205         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4206 out:
4207         return ret;
4208 }
4209
4210 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4211                        struct btrfs_root *root,
4212                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4213                        const char *name, int name_len)
4214 {
4215         int ret;
4216         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4217         if (!ret) {
4218                 drop_nlink(&inode->vfs_inode);
4219                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4220         }
4221         return ret;
4222 }
4223
4224 /*
4225  * helper to start transaction for unlink and rmdir.
4226  *
4227  * unlink and rmdir are special in btrfs, they do not always free space, so
4228  * if we cannot make our reservations the normal way try and see if there is
4229  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4230  * allow the unlink to occur.
4231  */
4232 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4233 {
4234         struct btrfs_root *root = BTRFS_I(dir)->root;
4235
4236         /*
4237          * 1 for the possible orphan item
4238          * 1 for the dir item
4239          * 1 for the dir index
4240          * 1 for the inode ref
4241          * 1 for the inode
4242          */
4243         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4244 }
4245
4246 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4247 {
4248         struct btrfs_root *root = BTRFS_I(dir)->root;
4249         struct btrfs_trans_handle *trans;
4250         struct inode *inode = d_inode(dentry);
4251         int ret;
4252
4253         trans = __unlink_start_trans(dir);
4254         if (IS_ERR(trans))
4255                 return PTR_ERR(trans);
4256
4257         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4258                         0);
4259
4260         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4261                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4262                         dentry->d_name.len);
4263         if (ret)
4264                 goto out;
4265
4266         if (inode->i_nlink == 0) {
4267                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4268                 if (ret)
4269                         goto out;
4270         }
4271
4272 out:
4273         btrfs_end_transaction(trans);
4274         btrfs_btree_balance_dirty(root->fs_info);
4275         return ret;
4276 }
4277
4278 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4279                         struct btrfs_root *root,
4280                         struct inode *dir, u64 objectid,
4281                         const char *name, int name_len)
4282 {
4283         struct btrfs_fs_info *fs_info = root->fs_info;
4284         struct btrfs_path *path;
4285         struct extent_buffer *leaf;
4286         struct btrfs_dir_item *di;
4287         struct btrfs_key key;
4288         u64 index;
4289         int ret;
4290         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4291
4292         path = btrfs_alloc_path();
4293         if (!path)
4294                 return -ENOMEM;
4295
4296         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4297                                    name, name_len, -1);
4298         if (IS_ERR_OR_NULL(di)) {
4299                 if (!di)
4300                         ret = -ENOENT;
4301                 else
4302                         ret = PTR_ERR(di);
4303                 goto out;
4304         }
4305
4306         leaf = path->nodes[0];
4307         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4308         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4309         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4310         if (ret) {
4311                 btrfs_abort_transaction(trans, ret);
4312                 goto out;
4313         }
4314         btrfs_release_path(path);
4315
4316         ret = btrfs_del_root_ref(trans, fs_info, objectid,
4317                                  root->root_key.objectid, dir_ino,
4318                                  &index, name, name_len);
4319         if (ret < 0) {
4320                 if (ret != -ENOENT) {
4321                         btrfs_abort_transaction(trans, ret);
4322                         goto out;
4323                 }
4324                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4325                                                  name, name_len);
4326                 if (IS_ERR_OR_NULL(di)) {
4327                         if (!di)
4328                                 ret = -ENOENT;
4329                         else
4330                                 ret = PTR_ERR(di);
4331                         btrfs_abort_transaction(trans, ret);
4332                         goto out;
4333                 }
4334
4335                 leaf = path->nodes[0];
4336                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4337                 btrfs_release_path(path);
4338                 index = key.offset;
4339         }
4340         btrfs_release_path(path);
4341
4342         ret = btrfs_delete_delayed_dir_index(trans, fs_info, BTRFS_I(dir), index);
4343         if (ret) {
4344                 btrfs_abort_transaction(trans, ret);
4345                 goto out;
4346         }
4347
4348         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4349         inode_inc_iversion(dir);
4350         dir->i_mtime = dir->i_ctime = current_time(dir);
4351         ret = btrfs_update_inode_fallback(trans, root, dir);
4352         if (ret)
4353                 btrfs_abort_transaction(trans, ret);
4354 out:
4355         btrfs_free_path(path);
4356         return ret;
4357 }
4358
4359 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4360 {
4361         struct inode *inode = d_inode(dentry);
4362         int err = 0;
4363         struct btrfs_root *root = BTRFS_I(dir)->root;
4364         struct btrfs_trans_handle *trans;
4365         u64 last_unlink_trans;
4366
4367         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4368                 return -ENOTEMPTY;
4369         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4370                 return -EPERM;
4371
4372         trans = __unlink_start_trans(dir);
4373         if (IS_ERR(trans))
4374                 return PTR_ERR(trans);
4375
4376         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4377                 err = btrfs_unlink_subvol(trans, root, dir,
4378                                           BTRFS_I(inode)->location.objectid,
4379                                           dentry->d_name.name,
4380                                           dentry->d_name.len);
4381                 goto out;
4382         }
4383
4384         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4385         if (err)
4386                 goto out;
4387
4388         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4389
4390         /* now the directory is empty */
4391         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4392                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4393                         dentry->d_name.len);
4394         if (!err) {
4395                 btrfs_i_size_write(BTRFS_I(inode), 0);
4396                 /*
4397                  * Propagate the last_unlink_trans value of the deleted dir to
4398                  * its parent directory. This is to prevent an unrecoverable
4399                  * log tree in the case we do something like this:
4400                  * 1) create dir foo
4401                  * 2) create snapshot under dir foo
4402                  * 3) delete the snapshot
4403                  * 4) rmdir foo
4404                  * 5) mkdir foo
4405                  * 6) fsync foo or some file inside foo
4406                  */
4407                 if (last_unlink_trans >= trans->transid)
4408                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4409         }
4410 out:
4411         btrfs_end_transaction(trans);
4412         btrfs_btree_balance_dirty(root->fs_info);
4413
4414         return err;
4415 }
4416
4417 static int truncate_space_check(struct btrfs_trans_handle *trans,
4418                                 struct btrfs_root *root,
4419                                 u64 bytes_deleted)
4420 {
4421         struct btrfs_fs_info *fs_info = root->fs_info;
4422         int ret;
4423
4424         /*
4425          * This is only used to apply pressure to the enospc system, we don't
4426          * intend to use this reservation at all.
4427          */
4428         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4429         bytes_deleted *= fs_info->nodesize;
4430         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4431                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4432         if (!ret) {
4433                 trace_btrfs_space_reservation(fs_info, "transaction",
4434                                               trans->transid,
4435                                               bytes_deleted, 1);
4436                 trans->bytes_reserved += bytes_deleted;
4437         }
4438         return ret;
4439
4440 }
4441
4442 /*
4443  * Return this if we need to call truncate_block for the last bit of the
4444  * truncate.
4445  */
4446 #define NEED_TRUNCATE_BLOCK 1
4447
4448 /*
4449  * this can truncate away extent items, csum items and directory items.
4450  * It starts at a high offset and removes keys until it can't find
4451  * any higher than new_size
4452  *
4453  * csum items that cross the new i_size are truncated to the new size
4454  * as well.
4455  *
4456  * min_type is the minimum key type to truncate down to.  If set to 0, this
4457  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4458  */
4459 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4460                                struct btrfs_root *root,
4461                                struct inode *inode,
4462                                u64 new_size, u32 min_type)
4463 {
4464         struct btrfs_fs_info *fs_info = root->fs_info;
4465         struct btrfs_path *path;
4466         struct extent_buffer *leaf;
4467         struct btrfs_file_extent_item *fi;
4468         struct btrfs_key key;
4469         struct btrfs_key found_key;
4470         u64 extent_start = 0;
4471         u64 extent_num_bytes = 0;
4472         u64 extent_offset = 0;
4473         u64 item_end = 0;
4474         u64 last_size = new_size;
4475         u32 found_type = (u8)-1;
4476         int found_extent;
4477         int del_item;
4478         int pending_del_nr = 0;
4479         int pending_del_slot = 0;
4480         int extent_type = -1;
4481         int ret;
4482         int err = 0;
4483         u64 ino = btrfs_ino(BTRFS_I(inode));
4484         u64 bytes_deleted = 0;
4485         bool be_nice = false;
4486         bool should_throttle = false;
4487         bool should_end = false;
4488
4489         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4490
4491         /*
4492          * for non-free space inodes and ref cows, we want to back off from
4493          * time to time
4494          */
4495         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4496             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4497                 be_nice = true;
4498
4499         path = btrfs_alloc_path();
4500         if (!path)
4501                 return -ENOMEM;
4502         path->reada = READA_BACK;
4503
4504         /*
4505          * We want to drop from the next block forward in case this new size is
4506          * not block aligned since we will be keeping the last block of the
4507          * extent just the way it is.
4508          */
4509         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4510             root == fs_info->tree_root)
4511                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4512                                         fs_info->sectorsize),
4513                                         (u64)-1, 0);
4514
4515         /*
4516          * This function is also used to drop the items in the log tree before
4517          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4518          * it is used to drop the loged items. So we shouldn't kill the delayed
4519          * items.
4520          */
4521         if (min_type == 0 && root == BTRFS_I(inode)->root)
4522                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4523
4524         key.objectid = ino;
4525         key.offset = (u64)-1;
4526         key.type = (u8)-1;
4527
4528 search_again:
4529         /*
4530          * with a 16K leaf size and 128MB extents, you can actually queue
4531          * up a huge file in a single leaf.  Most of the time that
4532          * bytes_deleted is > 0, it will be huge by the time we get here
4533          */
4534         if (be_nice && bytes_deleted > SZ_32M) {
4535                 if (btrfs_should_end_transaction(trans)) {
4536                         err = -EAGAIN;
4537                         goto error;
4538                 }
4539         }
4540
4541
4542         path->leave_spinning = 1;
4543         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4544         if (ret < 0) {
4545                 err = ret;
4546                 goto out;
4547         }
4548
4549         if (ret > 0) {
4550                 /* there are no items in the tree for us to truncate, we're
4551                  * done
4552                  */
4553                 if (path->slots[0] == 0)
4554                         goto out;
4555                 path->slots[0]--;
4556         }
4557
4558         while (1) {
4559                 fi = NULL;
4560                 leaf = path->nodes[0];
4561                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4562                 found_type = found_key.type;
4563
4564                 if (found_key.objectid != ino)
4565                         break;
4566
4567                 if (found_type < min_type)
4568                         break;
4569
4570                 item_end = found_key.offset;
4571                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4572                         fi = btrfs_item_ptr(leaf, path->slots[0],
4573                                             struct btrfs_file_extent_item);
4574                         extent_type = btrfs_file_extent_type(leaf, fi);
4575                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4576                                 item_end +=
4577                                     btrfs_file_extent_num_bytes(leaf, fi);
4578
4579                                 trace_btrfs_truncate_show_fi_regular(
4580                                         BTRFS_I(inode), leaf, fi,
4581                                         found_key.offset);
4582                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4583                                 item_end += btrfs_file_extent_inline_len(leaf,
4584                                                          path->slots[0], fi);
4585
4586                                 trace_btrfs_truncate_show_fi_inline(
4587                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4588                                         found_key.offset);
4589                         }
4590                         item_end--;
4591                 }
4592                 if (found_type > min_type) {
4593                         del_item = 1;
4594                 } else {
4595                         if (item_end < new_size)
4596                                 break;
4597                         if (found_key.offset >= new_size)
4598                                 del_item = 1;
4599                         else
4600                                 del_item = 0;
4601                 }
4602                 found_extent = 0;
4603                 /* FIXME, shrink the extent if the ref count is only 1 */
4604                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4605                         goto delete;
4606
4607                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4608                         u64 num_dec;
4609                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4610                         if (!del_item) {
4611                                 u64 orig_num_bytes =
4612                                         btrfs_file_extent_num_bytes(leaf, fi);
4613                                 extent_num_bytes = ALIGN(new_size -
4614                                                 found_key.offset,
4615                                                 fs_info->sectorsize);
4616                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4617                                                          extent_num_bytes);
4618                                 num_dec = (orig_num_bytes -
4619                                            extent_num_bytes);
4620                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4621                                              &root->state) &&
4622                                     extent_start != 0)
4623                                         inode_sub_bytes(inode, num_dec);
4624                                 btrfs_mark_buffer_dirty(leaf);
4625                         } else {
4626                                 extent_num_bytes =
4627                                         btrfs_file_extent_disk_num_bytes(leaf,
4628                                                                          fi);
4629                                 extent_offset = found_key.offset -
4630                                         btrfs_file_extent_offset(leaf, fi);
4631
4632                                 /* FIXME blocksize != 4096 */
4633                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4634                                 if (extent_start != 0) {
4635                                         found_extent = 1;
4636                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4637                                                      &root->state))
4638                                                 inode_sub_bytes(inode, num_dec);
4639                                 }
4640                         }
4641                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4642                         /*
4643                          * we can't truncate inline items that have had
4644                          * special encodings
4645                          */
4646                         if (!del_item &&
4647                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4648                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4649                             btrfs_file_extent_compression(leaf, fi) == 0) {
4650                                 u32 size = (u32)(new_size - found_key.offset);
4651
4652                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4653                                 size = btrfs_file_extent_calc_inline_size(size);
4654                                 btrfs_truncate_item(root->fs_info, path, size, 1);
4655                         } else if (!del_item) {
4656                                 /*
4657                                  * We have to bail so the last_size is set to
4658                                  * just before this extent.
4659                                  */
4660                                 err = NEED_TRUNCATE_BLOCK;
4661                                 break;
4662                         }
4663
4664                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4665                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4666                 }
4667 delete:
4668                 if (del_item)
4669                         last_size = found_key.offset;
4670                 else
4671                         last_size = new_size;
4672                 if (del_item) {
4673                         if (!pending_del_nr) {
4674                                 /* no pending yet, add ourselves */
4675                                 pending_del_slot = path->slots[0];
4676                                 pending_del_nr = 1;
4677                         } else if (pending_del_nr &&
4678                                    path->slots[0] + 1 == pending_del_slot) {
4679                                 /* hop on the pending chunk */
4680                                 pending_del_nr++;
4681                                 pending_del_slot = path->slots[0];
4682                         } else {
4683                                 BUG();
4684                         }
4685                 } else {
4686                         break;
4687                 }
4688                 should_throttle = false;
4689
4690                 if (found_extent &&
4691                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4692                      root == fs_info->tree_root)) {
4693                         btrfs_set_path_blocking(path);
4694                         bytes_deleted += extent_num_bytes;
4695                         ret = btrfs_free_extent(trans, root, extent_start,
4696                                                 extent_num_bytes, 0,
4697                                                 btrfs_header_owner(leaf),
4698                                                 ino, extent_offset);
4699                         BUG_ON(ret);
4700                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4701                                 btrfs_async_run_delayed_refs(fs_info,
4702                                         trans->delayed_ref_updates * 2,
4703                                         trans->transid, 0);
4704                         if (be_nice) {
4705                                 if (truncate_space_check(trans, root,
4706                                                          extent_num_bytes)) {
4707                                         should_end = true;
4708                                 }
4709                                 if (btrfs_should_throttle_delayed_refs(trans,
4710                                                                        fs_info))
4711                                         should_throttle = true;
4712                         }
4713                 }
4714
4715                 if (found_type == BTRFS_INODE_ITEM_KEY)
4716                         break;
4717
4718                 if (path->slots[0] == 0 ||
4719                     path->slots[0] != pending_del_slot ||
4720                     should_throttle || should_end) {
4721                         if (pending_del_nr) {
4722                                 ret = btrfs_del_items(trans, root, path,
4723                                                 pending_del_slot,
4724                                                 pending_del_nr);
4725                                 if (ret) {
4726                                         btrfs_abort_transaction(trans, ret);
4727                                         goto error;
4728                                 }
4729                                 pending_del_nr = 0;
4730                         }
4731                         btrfs_release_path(path);
4732                         if (should_throttle) {
4733                                 unsigned long updates = trans->delayed_ref_updates;
4734                                 if (updates) {
4735                                         trans->delayed_ref_updates = 0;
4736                                         ret = btrfs_run_delayed_refs(trans,
4737                                                                    fs_info,
4738                                                                    updates * 2);
4739                                         if (ret && !err)
4740                                                 err = ret;
4741                                 }
4742                         }
4743                         /*
4744                          * if we failed to refill our space rsv, bail out
4745                          * and let the transaction restart
4746                          */
4747                         if (should_end) {
4748                                 err = -EAGAIN;
4749                                 goto error;
4750                         }
4751                         goto search_again;
4752                 } else {
4753                         path->slots[0]--;
4754                 }
4755         }
4756 out:
4757         if (pending_del_nr) {
4758                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4759                                       pending_del_nr);
4760                 if (ret)
4761                         btrfs_abort_transaction(trans, ret);
4762         }
4763 error:
4764         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4765                 ASSERT(last_size >= new_size);
4766                 if (!err && last_size > new_size)
4767                         last_size = new_size;
4768                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4769         }
4770
4771         btrfs_free_path(path);
4772
4773         if (be_nice && bytes_deleted > SZ_32M) {
4774                 unsigned long updates = trans->delayed_ref_updates;
4775                 if (updates) {
4776                         trans->delayed_ref_updates = 0;
4777                         ret = btrfs_run_delayed_refs(trans, fs_info,
4778                                                      updates * 2);
4779                         if (ret && !err)
4780                                 err = ret;
4781                 }
4782         }
4783         return err;
4784 }
4785
4786 /*
4787  * btrfs_truncate_block - read, zero a chunk and write a block
4788  * @inode - inode that we're zeroing
4789  * @from - the offset to start zeroing
4790  * @len - the length to zero, 0 to zero the entire range respective to the
4791  *      offset
4792  * @front - zero up to the offset instead of from the offset on
4793  *
4794  * This will find the block for the "from" offset and cow the block and zero the
4795  * part we want to zero.  This is used with truncate and hole punching.
4796  */
4797 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4798                         int front)
4799 {
4800         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4801         struct address_space *mapping = inode->i_mapping;
4802         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4803         struct btrfs_ordered_extent *ordered;
4804         struct extent_state *cached_state = NULL;
4805         struct extent_changeset *data_reserved = NULL;
4806         char *kaddr;
4807         u32 blocksize = fs_info->sectorsize;
4808         pgoff_t index = from >> PAGE_SHIFT;
4809         unsigned offset = from & (blocksize - 1);
4810         struct page *page;
4811         gfp_t mask = btrfs_alloc_write_mask(mapping);
4812         int ret = 0;
4813         u64 block_start;
4814         u64 block_end;
4815
4816         if (IS_ALIGNED(offset, blocksize) &&
4817             (!len || IS_ALIGNED(len, blocksize)))
4818                 goto out;
4819
4820         block_start = round_down(from, blocksize);
4821         block_end = block_start + blocksize - 1;
4822
4823         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4824                                            block_start, blocksize);
4825         if (ret)
4826                 goto out;
4827
4828 again:
4829         page = find_or_create_page(mapping, index, mask);
4830         if (!page) {
4831                 btrfs_delalloc_release_space(inode, data_reserved,
4832                                              block_start, blocksize);
4833                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4834                 ret = -ENOMEM;
4835                 goto out;
4836         }
4837
4838         if (!PageUptodate(page)) {
4839                 ret = btrfs_readpage(NULL, page);
4840                 lock_page(page);
4841                 if (page->mapping != mapping) {
4842                         unlock_page(page);
4843                         put_page(page);
4844                         goto again;
4845                 }
4846                 if (!PageUptodate(page)) {
4847                         ret = -EIO;
4848                         goto out_unlock;
4849                 }
4850         }
4851         wait_on_page_writeback(page);
4852
4853         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4854         set_page_extent_mapped(page);
4855
4856         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4857         if (ordered) {
4858                 unlock_extent_cached(io_tree, block_start, block_end,
4859                                      &cached_state);
4860                 unlock_page(page);
4861                 put_page(page);
4862                 btrfs_start_ordered_extent(inode, ordered, 1);
4863                 btrfs_put_ordered_extent(ordered);
4864                 goto again;
4865         }
4866
4867         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4868                           EXTENT_DIRTY | EXTENT_DELALLOC |
4869                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4870                           0, 0, &cached_state);
4871
4872         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4873                                         &cached_state, 0);
4874         if (ret) {
4875                 unlock_extent_cached(io_tree, block_start, block_end,
4876                                      &cached_state);
4877                 goto out_unlock;
4878         }
4879
4880         if (offset != blocksize) {
4881                 if (!len)
4882                         len = blocksize - offset;
4883                 kaddr = kmap(page);
4884                 if (front)
4885                         memset(kaddr + (block_start - page_offset(page)),
4886                                 0, offset);
4887                 else
4888                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4889                                 0, len);
4890                 flush_dcache_page(page);
4891                 kunmap(page);
4892         }
4893         ClearPageChecked(page);
4894         set_page_dirty(page);
4895         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
4896
4897 out_unlock:
4898         if (ret)
4899                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4900                                              blocksize);
4901         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4902         unlock_page(page);
4903         put_page(page);
4904 out:
4905         extent_changeset_free(data_reserved);
4906         return ret;
4907 }
4908
4909 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4910                              u64 offset, u64 len)
4911 {
4912         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4913         struct btrfs_trans_handle *trans;
4914         int ret;
4915
4916         /*
4917          * Still need to make sure the inode looks like it's been updated so
4918          * that any holes get logged if we fsync.
4919          */
4920         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4921                 BTRFS_I(inode)->last_trans = fs_info->generation;
4922                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4923                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4924                 return 0;
4925         }
4926
4927         /*
4928          * 1 - for the one we're dropping
4929          * 1 - for the one we're adding
4930          * 1 - for updating the inode.
4931          */
4932         trans = btrfs_start_transaction(root, 3);
4933         if (IS_ERR(trans))
4934                 return PTR_ERR(trans);
4935
4936         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4937         if (ret) {
4938                 btrfs_abort_transaction(trans, ret);
4939                 btrfs_end_transaction(trans);
4940                 return ret;
4941         }
4942
4943         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4944                         offset, 0, 0, len, 0, len, 0, 0, 0);
4945         if (ret)
4946                 btrfs_abort_transaction(trans, ret);
4947         else
4948                 btrfs_update_inode(trans, root, inode);
4949         btrfs_end_transaction(trans);
4950         return ret;
4951 }
4952
4953 /*
4954  * This function puts in dummy file extents for the area we're creating a hole
4955  * for.  So if we are truncating this file to a larger size we need to insert
4956  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4957  * the range between oldsize and size
4958  */
4959 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4960 {
4961         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4962         struct btrfs_root *root = BTRFS_I(inode)->root;
4963         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4964         struct extent_map *em = NULL;
4965         struct extent_state *cached_state = NULL;
4966         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4967         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4968         u64 block_end = ALIGN(size, fs_info->sectorsize);
4969         u64 last_byte;
4970         u64 cur_offset;
4971         u64 hole_size;
4972         int err = 0;
4973
4974         /*
4975          * If our size started in the middle of a block we need to zero out the
4976          * rest of the block before we expand the i_size, otherwise we could
4977          * expose stale data.
4978          */
4979         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4980         if (err)
4981                 return err;
4982
4983         if (size <= hole_start)
4984                 return 0;
4985
4986         while (1) {
4987                 struct btrfs_ordered_extent *ordered;
4988
4989                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4990                                  &cached_state);
4991                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
4992                                                      block_end - hole_start);
4993                 if (!ordered)
4994                         break;
4995                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4996                                      &cached_state);
4997                 btrfs_start_ordered_extent(inode, ordered, 1);
4998                 btrfs_put_ordered_extent(ordered);
4999         }
5000
5001         cur_offset = hole_start;
5002         while (1) {
5003                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
5004                                 block_end - cur_offset, 0);
5005                 if (IS_ERR(em)) {
5006                         err = PTR_ERR(em);
5007                         em = NULL;
5008                         break;
5009                 }
5010                 last_byte = min(extent_map_end(em), block_end);
5011                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
5012                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
5013                         struct extent_map *hole_em;
5014                         hole_size = last_byte - cur_offset;
5015
5016                         err = maybe_insert_hole(root, inode, cur_offset,
5017                                                 hole_size);
5018                         if (err)
5019                                 break;
5020                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
5021                                                 cur_offset + hole_size - 1, 0);
5022                         hole_em = alloc_extent_map();
5023                         if (!hole_em) {
5024                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5025                                         &BTRFS_I(inode)->runtime_flags);
5026                                 goto next;
5027                         }
5028                         hole_em->start = cur_offset;
5029                         hole_em->len = hole_size;
5030                         hole_em->orig_start = cur_offset;
5031
5032                         hole_em->block_start = EXTENT_MAP_HOLE;
5033                         hole_em->block_len = 0;
5034                         hole_em->orig_block_len = 0;
5035                         hole_em->ram_bytes = hole_size;
5036                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
5037                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5038                         hole_em->generation = fs_info->generation;
5039
5040                         while (1) {
5041                                 write_lock(&em_tree->lock);
5042                                 err = add_extent_mapping(em_tree, hole_em, 1);
5043                                 write_unlock(&em_tree->lock);
5044                                 if (err != -EEXIST)
5045                                         break;
5046                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5047                                                         cur_offset,
5048                                                         cur_offset +
5049                                                         hole_size - 1, 0);
5050                         }
5051                         free_extent_map(hole_em);
5052                 }
5053 next:
5054                 free_extent_map(em);
5055                 em = NULL;
5056                 cur_offset = last_byte;
5057                 if (cur_offset >= block_end)
5058                         break;
5059         }
5060         free_extent_map(em);
5061         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5062         return err;
5063 }
5064
5065 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5066 {
5067         struct btrfs_root *root = BTRFS_I(inode)->root;
5068         struct btrfs_trans_handle *trans;
5069         loff_t oldsize = i_size_read(inode);
5070         loff_t newsize = attr->ia_size;
5071         int mask = attr->ia_valid;
5072         int ret;
5073
5074         /*
5075          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5076          * special case where we need to update the times despite not having
5077          * these flags set.  For all other operations the VFS set these flags
5078          * explicitly if it wants a timestamp update.
5079          */
5080         if (newsize != oldsize) {
5081                 inode_inc_iversion(inode);
5082                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5083                         inode->i_ctime = inode->i_mtime =
5084                                 current_time(inode);
5085         }
5086
5087         if (newsize > oldsize) {
5088                 /*
5089                  * Don't do an expanding truncate while snapshotting is ongoing.
5090                  * This is to ensure the snapshot captures a fully consistent
5091                  * state of this file - if the snapshot captures this expanding
5092                  * truncation, it must capture all writes that happened before
5093                  * this truncation.
5094                  */
5095                 btrfs_wait_for_snapshot_creation(root);
5096                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5097                 if (ret) {
5098                         btrfs_end_write_no_snapshotting(root);
5099                         return ret;
5100                 }
5101
5102                 trans = btrfs_start_transaction(root, 1);
5103                 if (IS_ERR(trans)) {
5104                         btrfs_end_write_no_snapshotting(root);
5105                         return PTR_ERR(trans);
5106                 }
5107
5108                 i_size_write(inode, newsize);
5109                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5110                 pagecache_isize_extended(inode, oldsize, newsize);
5111                 ret = btrfs_update_inode(trans, root, inode);
5112                 btrfs_end_write_no_snapshotting(root);
5113                 btrfs_end_transaction(trans);
5114         } else {
5115
5116                 /*
5117                  * We're truncating a file that used to have good data down to
5118                  * zero. Make sure it gets into the ordered flush list so that
5119                  * any new writes get down to disk quickly.
5120                  */
5121                 if (newsize == 0)
5122                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5123                                 &BTRFS_I(inode)->runtime_flags);
5124
5125                 /*
5126                  * 1 for the orphan item we're going to add
5127                  * 1 for the orphan item deletion.
5128                  */
5129                 trans = btrfs_start_transaction(root, 2);
5130                 if (IS_ERR(trans))
5131                         return PTR_ERR(trans);
5132
5133                 /*
5134                  * We need to do this in case we fail at _any_ point during the
5135                  * actual truncate.  Once we do the truncate_setsize we could
5136                  * invalidate pages which forces any outstanding ordered io to
5137                  * be instantly completed which will give us extents that need
5138                  * to be truncated.  If we fail to get an orphan inode down we
5139                  * could have left over extents that were never meant to live,
5140                  * so we need to guarantee from this point on that everything
5141                  * will be consistent.
5142                  */
5143                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
5144                 btrfs_end_transaction(trans);
5145                 if (ret)
5146                         return ret;
5147
5148                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5149                 truncate_setsize(inode, newsize);
5150
5151                 /* Disable nonlocked read DIO to avoid the end less truncate */
5152                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5153                 inode_dio_wait(inode);
5154                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5155
5156                 ret = btrfs_truncate(inode, newsize == oldsize);
5157                 if (ret && inode->i_nlink) {
5158                         int err;
5159
5160                         /* To get a stable disk_i_size */
5161                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5162                         if (err) {
5163                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5164                                 return err;
5165                         }
5166
5167                         /*
5168                          * failed to truncate, disk_i_size is only adjusted down
5169                          * as we remove extents, so it should represent the true
5170                          * size of the inode, so reset the in memory size and
5171                          * delete our orphan entry.
5172                          */
5173                         trans = btrfs_join_transaction(root);
5174                         if (IS_ERR(trans)) {
5175                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5176                                 return ret;
5177                         }
5178                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5179                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
5180                         if (err)
5181                                 btrfs_abort_transaction(trans, err);
5182                         btrfs_end_transaction(trans);
5183                 }
5184         }
5185
5186         return ret;
5187 }
5188
5189 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5190 {
5191         struct inode *inode = d_inode(dentry);
5192         struct btrfs_root *root = BTRFS_I(inode)->root;
5193         int err;
5194
5195         if (btrfs_root_readonly(root))
5196                 return -EROFS;
5197
5198         err = setattr_prepare(dentry, attr);
5199         if (err)
5200                 return err;
5201
5202         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5203                 err = btrfs_setsize(inode, attr);
5204                 if (err)
5205                         return err;
5206         }
5207
5208         if (attr->ia_valid) {
5209                 setattr_copy(inode, attr);
5210                 inode_inc_iversion(inode);
5211                 err = btrfs_dirty_inode(inode);
5212
5213                 if (!err && attr->ia_valid & ATTR_MODE)
5214                         err = posix_acl_chmod(inode, inode->i_mode);
5215         }
5216
5217         return err;
5218 }
5219
5220 /*
5221  * While truncating the inode pages during eviction, we get the VFS calling
5222  * btrfs_invalidatepage() against each page of the inode. This is slow because
5223  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5224  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5225  * extent_state structures over and over, wasting lots of time.
5226  *
5227  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5228  * those expensive operations on a per page basis and do only the ordered io
5229  * finishing, while we release here the extent_map and extent_state structures,
5230  * without the excessive merging and splitting.
5231  */
5232 static void evict_inode_truncate_pages(struct inode *inode)
5233 {
5234         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5235         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5236         struct rb_node *node;
5237
5238         ASSERT(inode->i_state & I_FREEING);
5239         truncate_inode_pages_final(&inode->i_data);
5240
5241         write_lock(&map_tree->lock);
5242         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5243                 struct extent_map *em;
5244
5245                 node = rb_first(&map_tree->map);
5246                 em = rb_entry(node, struct extent_map, rb_node);
5247                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5248                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5249                 remove_extent_mapping(map_tree, em);
5250                 free_extent_map(em);
5251                 if (need_resched()) {
5252                         write_unlock(&map_tree->lock);
5253                         cond_resched();
5254                         write_lock(&map_tree->lock);
5255                 }
5256         }
5257         write_unlock(&map_tree->lock);
5258
5259         /*
5260          * Keep looping until we have no more ranges in the io tree.
5261          * We can have ongoing bios started by readpages (called from readahead)
5262          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5263          * still in progress (unlocked the pages in the bio but did not yet
5264          * unlocked the ranges in the io tree). Therefore this means some
5265          * ranges can still be locked and eviction started because before
5266          * submitting those bios, which are executed by a separate task (work
5267          * queue kthread), inode references (inode->i_count) were not taken
5268          * (which would be dropped in the end io callback of each bio).
5269          * Therefore here we effectively end up waiting for those bios and
5270          * anyone else holding locked ranges without having bumped the inode's
5271          * reference count - if we don't do it, when they access the inode's
5272          * io_tree to unlock a range it may be too late, leading to an
5273          * use-after-free issue.
5274          */
5275         spin_lock(&io_tree->lock);
5276         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5277                 struct extent_state *state;
5278                 struct extent_state *cached_state = NULL;
5279                 u64 start;
5280                 u64 end;
5281
5282                 node = rb_first(&io_tree->state);
5283                 state = rb_entry(node, struct extent_state, rb_node);
5284                 start = state->start;
5285                 end = state->end;
5286                 spin_unlock(&io_tree->lock);
5287
5288                 lock_extent_bits(io_tree, start, end, &cached_state);
5289
5290                 /*
5291                  * If still has DELALLOC flag, the extent didn't reach disk,
5292                  * and its reserved space won't be freed by delayed_ref.
5293                  * So we need to free its reserved space here.
5294                  * (Refer to comment in btrfs_invalidatepage, case 2)
5295                  *
5296                  * Note, end is the bytenr of last byte, so we need + 1 here.
5297                  */
5298                 if (state->state & EXTENT_DELALLOC)
5299                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5300
5301                 clear_extent_bit(io_tree, start, end,
5302                                  EXTENT_LOCKED | EXTENT_DIRTY |
5303                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5304                                  EXTENT_DEFRAG, 1, 1, &cached_state);
5305
5306                 cond_resched();
5307                 spin_lock(&io_tree->lock);
5308         }
5309         spin_unlock(&io_tree->lock);
5310 }
5311
5312 void btrfs_evict_inode(struct inode *inode)
5313 {
5314         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5315         struct btrfs_trans_handle *trans;
5316         struct btrfs_root *root = BTRFS_I(inode)->root;
5317         struct btrfs_block_rsv *rsv, *global_rsv;
5318         int steal_from_global = 0;
5319         u64 min_size;
5320         int ret;
5321
5322         trace_btrfs_inode_evict(inode);
5323
5324         if (!root) {
5325                 clear_inode(inode);
5326                 return;
5327         }
5328
5329         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5330
5331         evict_inode_truncate_pages(inode);
5332
5333         if (inode->i_nlink &&
5334             ((btrfs_root_refs(&root->root_item) != 0 &&
5335               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5336              btrfs_is_free_space_inode(BTRFS_I(inode))))
5337                 goto no_delete;
5338
5339         if (is_bad_inode(inode)) {
5340                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5341                 goto no_delete;
5342         }
5343         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5344         if (!special_file(inode->i_mode))
5345                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5346
5347         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5348
5349         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
5350                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5351                                  &BTRFS_I(inode)->runtime_flags));
5352                 goto no_delete;
5353         }
5354
5355         if (inode->i_nlink > 0) {
5356                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5357                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5358                 goto no_delete;
5359         }
5360
5361         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5362         if (ret) {
5363                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5364                 goto no_delete;
5365         }
5366
5367         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5368         if (!rsv) {
5369                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5370                 goto no_delete;
5371         }
5372         rsv->size = min_size;
5373         rsv->failfast = 1;
5374         global_rsv = &fs_info->global_block_rsv;
5375
5376         btrfs_i_size_write(BTRFS_I(inode), 0);
5377
5378         /*
5379          * This is a bit simpler than btrfs_truncate since we've already
5380          * reserved our space for our orphan item in the unlink, so we just
5381          * need to reserve some slack space in case we add bytes and update
5382          * inode item when doing the truncate.
5383          */
5384         while (1) {
5385                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5386                                              BTRFS_RESERVE_FLUSH_LIMIT);
5387
5388                 /*
5389                  * Try and steal from the global reserve since we will
5390                  * likely not use this space anyway, we want to try as
5391                  * hard as possible to get this to work.
5392                  */
5393                 if (ret)
5394                         steal_from_global++;
5395                 else
5396                         steal_from_global = 0;
5397                 ret = 0;
5398
5399                 /*
5400                  * steal_from_global == 0: we reserved stuff, hooray!
5401                  * steal_from_global == 1: we didn't reserve stuff, boo!
5402                  * steal_from_global == 2: we've committed, still not a lot of
5403                  * room but maybe we'll have room in the global reserve this
5404                  * time.
5405                  * steal_from_global == 3: abandon all hope!
5406                  */
5407                 if (steal_from_global > 2) {
5408                         btrfs_warn(fs_info,
5409                                    "Could not get space for a delete, will truncate on mount %d",
5410                                    ret);
5411                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5412                         btrfs_free_block_rsv(fs_info, rsv);
5413                         goto no_delete;
5414                 }
5415
5416                 trans = btrfs_join_transaction(root);
5417                 if (IS_ERR(trans)) {
5418                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5419                         btrfs_free_block_rsv(fs_info, rsv);
5420                         goto no_delete;
5421                 }
5422
5423                 /*
5424                  * We can't just steal from the global reserve, we need to make
5425                  * sure there is room to do it, if not we need to commit and try
5426                  * again.
5427                  */
5428                 if (steal_from_global) {
5429                         if (!btrfs_check_space_for_delayed_refs(trans, fs_info))
5430                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5431                                                               min_size, 0);
5432                         else
5433                                 ret = -ENOSPC;
5434                 }
5435
5436                 /*
5437                  * Couldn't steal from the global reserve, we have too much
5438                  * pending stuff built up, commit the transaction and try it
5439                  * again.
5440                  */
5441                 if (ret) {
5442                         ret = btrfs_commit_transaction(trans);
5443                         if (ret) {
5444                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5445                                 btrfs_free_block_rsv(fs_info, rsv);
5446                                 goto no_delete;
5447                         }
5448                         continue;
5449                 } else {
5450                         steal_from_global = 0;
5451                 }
5452
5453                 trans->block_rsv = rsv;
5454
5455                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5456                 if (ret != -ENOSPC && ret != -EAGAIN)
5457                         break;
5458
5459                 trans->block_rsv = &fs_info->trans_block_rsv;
5460                 btrfs_end_transaction(trans);
5461                 trans = NULL;
5462                 btrfs_btree_balance_dirty(fs_info);
5463         }
5464
5465         btrfs_free_block_rsv(fs_info, rsv);
5466
5467         /*
5468          * Errors here aren't a big deal, it just means we leave orphan items
5469          * in the tree.  They will be cleaned up on the next mount.
5470          */
5471         if (ret == 0) {
5472                 trans->block_rsv = root->orphan_block_rsv;
5473                 btrfs_orphan_del(trans, BTRFS_I(inode));
5474         } else {
5475                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5476         }
5477
5478         trans->block_rsv = &fs_info->trans_block_rsv;
5479         if (!(root == fs_info->tree_root ||
5480               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5481                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5482
5483         btrfs_end_transaction(trans);
5484         btrfs_btree_balance_dirty(fs_info);
5485 no_delete:
5486         btrfs_remove_delayed_node(BTRFS_I(inode));
5487         clear_inode(inode);
5488 }
5489
5490 /*
5491  * this returns the key found in the dir entry in the location pointer.
5492  * If no dir entries were found, returns -ENOENT.
5493  * If found a corrupted location in dir entry, returns -EUCLEAN.
5494  */
5495 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5496                                struct btrfs_key *location)
5497 {
5498         const char *name = dentry->d_name.name;
5499         int namelen = dentry->d_name.len;
5500         struct btrfs_dir_item *di;
5501         struct btrfs_path *path;
5502         struct btrfs_root *root = BTRFS_I(dir)->root;
5503         int ret = 0;
5504
5505         path = btrfs_alloc_path();
5506         if (!path)
5507                 return -ENOMEM;
5508
5509         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5510                         name, namelen, 0);
5511         if (!di) {
5512                 ret = -ENOENT;
5513                 goto out;
5514         }
5515         if (IS_ERR(di)) {
5516                 ret = PTR_ERR(di);
5517                 goto out;
5518         }
5519
5520         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5521         if (location->type != BTRFS_INODE_ITEM_KEY &&
5522             location->type != BTRFS_ROOT_ITEM_KEY) {
5523                 ret = -EUCLEAN;
5524                 btrfs_warn(root->fs_info,
5525 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5526                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5527                            location->objectid, location->type, location->offset);
5528         }
5529 out:
5530         btrfs_free_path(path);
5531         return ret;
5532 }
5533
5534 /*
5535  * when we hit a tree root in a directory, the btrfs part of the inode
5536  * needs to be changed to reflect the root directory of the tree root.  This
5537  * is kind of like crossing a mount point.
5538  */
5539 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5540                                     struct inode *dir,
5541                                     struct dentry *dentry,
5542                                     struct btrfs_key *location,
5543                                     struct btrfs_root **sub_root)
5544 {
5545         struct btrfs_path *path;
5546         struct btrfs_root *new_root;
5547         struct btrfs_root_ref *ref;
5548         struct extent_buffer *leaf;
5549         struct btrfs_key key;
5550         int ret;
5551         int err = 0;
5552
5553         path = btrfs_alloc_path();
5554         if (!path) {
5555                 err = -ENOMEM;
5556                 goto out;
5557         }
5558
5559         err = -ENOENT;
5560         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5561         key.type = BTRFS_ROOT_REF_KEY;
5562         key.offset = location->objectid;
5563
5564         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5565         if (ret) {
5566                 if (ret < 0)
5567                         err = ret;
5568                 goto out;
5569         }
5570
5571         leaf = path->nodes[0];
5572         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5573         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5574             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5575                 goto out;
5576
5577         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5578                                    (unsigned long)(ref + 1),
5579                                    dentry->d_name.len);
5580         if (ret)
5581                 goto out;
5582
5583         btrfs_release_path(path);
5584
5585         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5586         if (IS_ERR(new_root)) {
5587                 err = PTR_ERR(new_root);
5588                 goto out;
5589         }
5590
5591         *sub_root = new_root;
5592         location->objectid = btrfs_root_dirid(&new_root->root_item);
5593         location->type = BTRFS_INODE_ITEM_KEY;
5594         location->offset = 0;
5595         err = 0;
5596 out:
5597         btrfs_free_path(path);
5598         return err;
5599 }
5600
5601 static void inode_tree_add(struct inode *inode)
5602 {
5603         struct btrfs_root *root = BTRFS_I(inode)->root;
5604         struct btrfs_inode *entry;
5605         struct rb_node **p;
5606         struct rb_node *parent;
5607         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5608         u64 ino = btrfs_ino(BTRFS_I(inode));
5609
5610         if (inode_unhashed(inode))
5611                 return;
5612         parent = NULL;
5613         spin_lock(&root->inode_lock);
5614         p = &root->inode_tree.rb_node;
5615         while (*p) {
5616                 parent = *p;
5617                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5618
5619                 if (ino < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5620                         p = &parent->rb_left;
5621                 else if (ino > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5622                         p = &parent->rb_right;
5623                 else {
5624                         WARN_ON(!(entry->vfs_inode.i_state &
5625                                   (I_WILL_FREE | I_FREEING)));
5626                         rb_replace_node(parent, new, &root->inode_tree);
5627                         RB_CLEAR_NODE(parent);
5628                         spin_unlock(&root->inode_lock);
5629                         return;
5630                 }
5631         }
5632         rb_link_node(new, parent, p);
5633         rb_insert_color(new, &root->inode_tree);
5634         spin_unlock(&root->inode_lock);
5635 }
5636
5637 static void inode_tree_del(struct inode *inode)
5638 {
5639         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5640         struct btrfs_root *root = BTRFS_I(inode)->root;
5641         int empty = 0;
5642
5643         spin_lock(&root->inode_lock);
5644         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5645                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5646                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5647                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5648         }
5649         spin_unlock(&root->inode_lock);
5650
5651         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5652                 synchronize_srcu(&fs_info->subvol_srcu);
5653                 spin_lock(&root->inode_lock);
5654                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5655                 spin_unlock(&root->inode_lock);
5656                 if (empty)
5657                         btrfs_add_dead_root(root);
5658         }
5659 }
5660
5661 void btrfs_invalidate_inodes(struct btrfs_root *root)
5662 {
5663         struct btrfs_fs_info *fs_info = root->fs_info;
5664         struct rb_node *node;
5665         struct rb_node *prev;
5666         struct btrfs_inode *entry;
5667         struct inode *inode;
5668         u64 objectid = 0;
5669
5670         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
5671                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5672
5673         spin_lock(&root->inode_lock);
5674 again:
5675         node = root->inode_tree.rb_node;
5676         prev = NULL;
5677         while (node) {
5678                 prev = node;
5679                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5680
5681                 if (objectid < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5682                         node = node->rb_left;
5683                 else if (objectid > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5684                         node = node->rb_right;
5685                 else
5686                         break;
5687         }
5688         if (!node) {
5689                 while (prev) {
5690                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5691                         if (objectid <= btrfs_ino(BTRFS_I(&entry->vfs_inode))) {
5692                                 node = prev;
5693                                 break;
5694                         }
5695                         prev = rb_next(prev);
5696                 }
5697         }
5698         while (node) {
5699                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5700                 objectid = btrfs_ino(BTRFS_I(&entry->vfs_inode)) + 1;
5701                 inode = igrab(&entry->vfs_inode);
5702                 if (inode) {
5703                         spin_unlock(&root->inode_lock);
5704                         if (atomic_read(&inode->i_count) > 1)
5705                                 d_prune_aliases(inode);
5706                         /*
5707                          * btrfs_drop_inode will have it removed from
5708                          * the inode cache when its usage count
5709                          * hits zero.
5710                          */
5711                         iput(inode);
5712                         cond_resched();
5713                         spin_lock(&root->inode_lock);
5714                         goto again;
5715                 }
5716
5717                 if (cond_resched_lock(&root->inode_lock))
5718                         goto again;
5719
5720                 node = rb_next(node);
5721         }
5722         spin_unlock(&root->inode_lock);
5723 }
5724
5725 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5726 {
5727         struct btrfs_iget_args *args = p;
5728         inode->i_ino = args->location->objectid;
5729         memcpy(&BTRFS_I(inode)->location, args->location,
5730                sizeof(*args->location));
5731         BTRFS_I(inode)->root = args->root;
5732         return 0;
5733 }
5734
5735 static int btrfs_find_actor(struct inode *inode, void *opaque)
5736 {
5737         struct btrfs_iget_args *args = opaque;
5738         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5739                 args->root == BTRFS_I(inode)->root;
5740 }
5741
5742 static struct inode *btrfs_iget_locked(struct super_block *s,
5743                                        struct btrfs_key *location,
5744                                        struct btrfs_root *root)
5745 {
5746         struct inode *inode;
5747         struct btrfs_iget_args args;
5748         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5749
5750         args.location = location;
5751         args.root = root;
5752
5753         inode = iget5_locked(s, hashval, btrfs_find_actor,
5754                              btrfs_init_locked_inode,
5755                              (void *)&args);
5756         return inode;
5757 }
5758
5759 /* Get an inode object given its location and corresponding root.
5760  * Returns in *is_new if the inode was read from disk
5761  */
5762 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5763                          struct btrfs_root *root, int *new)
5764 {
5765         struct inode *inode;
5766
5767         inode = btrfs_iget_locked(s, location, root);
5768         if (!inode)
5769                 return ERR_PTR(-ENOMEM);
5770
5771         if (inode->i_state & I_NEW) {
5772                 int ret;
5773
5774                 ret = btrfs_read_locked_inode(inode);
5775                 if (!is_bad_inode(inode)) {
5776                         inode_tree_add(inode);
5777                         unlock_new_inode(inode);
5778                         if (new)
5779                                 *new = 1;
5780                 } else {
5781                         unlock_new_inode(inode);
5782                         iput(inode);
5783                         ASSERT(ret < 0);
5784                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5785                 }
5786         }
5787
5788         return inode;
5789 }
5790
5791 static struct inode *new_simple_dir(struct super_block *s,
5792                                     struct btrfs_key *key,
5793                                     struct btrfs_root *root)
5794 {
5795         struct inode *inode = new_inode(s);
5796
5797         if (!inode)
5798                 return ERR_PTR(-ENOMEM);
5799
5800         BTRFS_I(inode)->root = root;
5801         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5802         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5803
5804         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5805         inode->i_op = &btrfs_dir_ro_inode_operations;
5806         inode->i_opflags &= ~IOP_XATTR;
5807         inode->i_fop = &simple_dir_operations;
5808         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5809         inode->i_mtime = current_time(inode);
5810         inode->i_atime = inode->i_mtime;
5811         inode->i_ctime = inode->i_mtime;
5812         BTRFS_I(inode)->i_otime = inode->i_mtime;
5813
5814         return inode;
5815 }
5816
5817 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5818 {
5819         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5820         struct inode *inode;
5821         struct btrfs_root *root = BTRFS_I(dir)->root;
5822         struct btrfs_root *sub_root = root;
5823         struct btrfs_key location;
5824         int index;
5825         int ret = 0;
5826
5827         if (dentry->d_name.len > BTRFS_NAME_LEN)
5828                 return ERR_PTR(-ENAMETOOLONG);
5829
5830         ret = btrfs_inode_by_name(dir, dentry, &location);
5831         if (ret < 0)
5832                 return ERR_PTR(ret);
5833
5834         if (location.type == BTRFS_INODE_ITEM_KEY) {
5835                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5836                 return inode;
5837         }
5838
5839         index = srcu_read_lock(&fs_info->subvol_srcu);
5840         ret = fixup_tree_root_location(fs_info, dir, dentry,
5841                                        &location, &sub_root);
5842         if (ret < 0) {
5843                 if (ret != -ENOENT)
5844                         inode = ERR_PTR(ret);
5845                 else
5846                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5847         } else {
5848                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5849         }
5850         srcu_read_unlock(&fs_info->subvol_srcu, index);
5851
5852         if (!IS_ERR(inode) && root != sub_root) {
5853                 down_read(&fs_info->cleanup_work_sem);
5854                 if (!sb_rdonly(inode->i_sb))
5855                         ret = btrfs_orphan_cleanup(sub_root);
5856                 up_read(&fs_info->cleanup_work_sem);
5857                 if (ret) {
5858                         iput(inode);
5859                         inode = ERR_PTR(ret);
5860                 }
5861         }
5862
5863         return inode;
5864 }
5865
5866 static int btrfs_dentry_delete(const struct dentry *dentry)
5867 {
5868         struct btrfs_root *root;
5869         struct inode *inode = d_inode(dentry);
5870
5871         if (!inode && !IS_ROOT(dentry))
5872                 inode = d_inode(dentry->d_parent);
5873
5874         if (inode) {
5875                 root = BTRFS_I(inode)->root;
5876                 if (btrfs_root_refs(&root->root_item) == 0)
5877                         return 1;
5878
5879                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5880                         return 1;
5881         }
5882         return 0;
5883 }
5884
5885 static void btrfs_dentry_release(struct dentry *dentry)
5886 {
5887         kfree(dentry->d_fsdata);
5888 }
5889
5890 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5891                                    unsigned int flags)
5892 {
5893         struct inode *inode;
5894
5895         inode = btrfs_lookup_dentry(dir, dentry);
5896         if (IS_ERR(inode)) {
5897                 if (PTR_ERR(inode) == -ENOENT)
5898                         inode = NULL;
5899                 else
5900                         return ERR_CAST(inode);
5901         }
5902
5903         return d_splice_alias(inode, dentry);
5904 }
5905
5906 unsigned char btrfs_filetype_table[] = {
5907         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5908 };
5909
5910 /*
5911  * All this infrastructure exists because dir_emit can fault, and we are holding
5912  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5913  * our information into that, and then dir_emit from the buffer.  This is
5914  * similar to what NFS does, only we don't keep the buffer around in pagecache
5915  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5916  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5917  * tree lock.
5918  */
5919 static int btrfs_opendir(struct inode *inode, struct file *file)
5920 {
5921         struct btrfs_file_private *private;
5922
5923         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5924         if (!private)
5925                 return -ENOMEM;
5926         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5927         if (!private->filldir_buf) {
5928                 kfree(private);
5929                 return -ENOMEM;
5930         }
5931         file->private_data = private;
5932         return 0;
5933 }
5934
5935 struct dir_entry {
5936         u64 ino;
5937         u64 offset;
5938         unsigned type;
5939         int name_len;
5940 };
5941
5942 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5943 {
5944         while (entries--) {
5945                 struct dir_entry *entry = addr;
5946                 char *name = (char *)(entry + 1);
5947
5948                 ctx->pos = entry->offset;
5949                 if (!dir_emit(ctx, name, entry->name_len, entry->ino,
5950                               entry->type))
5951                         return 1;
5952                 addr += sizeof(struct dir_entry) + entry->name_len;
5953                 ctx->pos++;
5954         }
5955         return 0;
5956 }
5957
5958 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5959 {
5960         struct inode *inode = file_inode(file);
5961         struct btrfs_root *root = BTRFS_I(inode)->root;
5962         struct btrfs_file_private *private = file->private_data;
5963         struct btrfs_dir_item *di;
5964         struct btrfs_key key;
5965         struct btrfs_key found_key;
5966         struct btrfs_path *path;
5967         void *addr;
5968         struct list_head ins_list;
5969         struct list_head del_list;
5970         int ret;
5971         struct extent_buffer *leaf;
5972         int slot;
5973         char *name_ptr;
5974         int name_len;
5975         int entries = 0;
5976         int total_len = 0;
5977         bool put = false;
5978         struct btrfs_key location;
5979
5980         if (!dir_emit_dots(file, ctx))
5981                 return 0;
5982
5983         path = btrfs_alloc_path();
5984         if (!path)
5985                 return -ENOMEM;
5986
5987         addr = private->filldir_buf;
5988         path->reada = READA_FORWARD;
5989
5990         INIT_LIST_HEAD(&ins_list);
5991         INIT_LIST_HEAD(&del_list);
5992         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5993
5994 again:
5995         key.type = BTRFS_DIR_INDEX_KEY;
5996         key.offset = ctx->pos;
5997         key.objectid = btrfs_ino(BTRFS_I(inode));
5998
5999         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6000         if (ret < 0)
6001                 goto err;
6002
6003         while (1) {
6004                 struct dir_entry *entry;
6005
6006                 leaf = path->nodes[0];
6007                 slot = path->slots[0];
6008                 if (slot >= btrfs_header_nritems(leaf)) {
6009                         ret = btrfs_next_leaf(root, path);
6010                         if (ret < 0)
6011                                 goto err;
6012                         else if (ret > 0)
6013                                 break;
6014                         continue;
6015                 }
6016
6017                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
6018
6019                 if (found_key.objectid != key.objectid)
6020                         break;
6021                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
6022                         break;
6023                 if (found_key.offset < ctx->pos)
6024                         goto next;
6025                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
6026                         goto next;
6027                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
6028                 name_len = btrfs_dir_name_len(leaf, di);
6029                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
6030                     PAGE_SIZE) {
6031                         btrfs_release_path(path);
6032                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6033                         if (ret)
6034                                 goto nopos;
6035                         addr = private->filldir_buf;
6036                         entries = 0;
6037                         total_len = 0;
6038                         goto again;
6039                 }
6040
6041                 entry = addr;
6042                 entry->name_len = name_len;
6043                 name_ptr = (char *)(entry + 1);
6044                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
6045                                    name_len);
6046                 entry->type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
6047                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
6048                 entry->ino = location.objectid;
6049                 entry->offset = found_key.offset;
6050                 entries++;
6051                 addr += sizeof(struct dir_entry) + name_len;
6052                 total_len += sizeof(struct dir_entry) + name_len;
6053 next:
6054                 path->slots[0]++;
6055         }
6056         btrfs_release_path(path);
6057
6058         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6059         if (ret)
6060                 goto nopos;
6061
6062         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6063         if (ret)
6064                 goto nopos;
6065
6066         /*
6067          * Stop new entries from being returned after we return the last
6068          * entry.
6069          *
6070          * New directory entries are assigned a strictly increasing
6071          * offset.  This means that new entries created during readdir
6072          * are *guaranteed* to be seen in the future by that readdir.
6073          * This has broken buggy programs which operate on names as
6074          * they're returned by readdir.  Until we re-use freed offsets
6075          * we have this hack to stop new entries from being returned
6076          * under the assumption that they'll never reach this huge
6077          * offset.
6078          *
6079          * This is being careful not to overflow 32bit loff_t unless the
6080          * last entry requires it because doing so has broken 32bit apps
6081          * in the past.
6082          */
6083         if (ctx->pos >= INT_MAX)
6084                 ctx->pos = LLONG_MAX;
6085         else
6086                 ctx->pos = INT_MAX;
6087 nopos:
6088         ret = 0;
6089 err:
6090         if (put)
6091                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6092         btrfs_free_path(path);
6093         return ret;
6094 }
6095
6096 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
6097 {
6098         struct btrfs_root *root = BTRFS_I(inode)->root;
6099         struct btrfs_trans_handle *trans;
6100         int ret = 0;
6101         bool nolock = false;
6102
6103         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6104                 return 0;
6105
6106         if (btrfs_fs_closing(root->fs_info) &&
6107                         btrfs_is_free_space_inode(BTRFS_I(inode)))
6108                 nolock = true;
6109
6110         if (wbc->sync_mode == WB_SYNC_ALL) {
6111                 if (nolock)
6112                         trans = btrfs_join_transaction_nolock(root);
6113                 else
6114                         trans = btrfs_join_transaction(root);
6115                 if (IS_ERR(trans))
6116                         return PTR_ERR(trans);
6117                 ret = btrfs_commit_transaction(trans);
6118         }
6119         return ret;
6120 }
6121
6122 /*
6123  * This is somewhat expensive, updating the tree every time the
6124  * inode changes.  But, it is most likely to find the inode in cache.
6125  * FIXME, needs more benchmarking...there are no reasons other than performance
6126  * to keep or drop this code.
6127  */
6128 static int btrfs_dirty_inode(struct inode *inode)
6129 {
6130         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6131         struct btrfs_root *root = BTRFS_I(inode)->root;
6132         struct btrfs_trans_handle *trans;
6133         int ret;
6134
6135         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6136                 return 0;
6137
6138         trans = btrfs_join_transaction(root);
6139         if (IS_ERR(trans))
6140                 return PTR_ERR(trans);
6141
6142         ret = btrfs_update_inode(trans, root, inode);
6143         if (ret && ret == -ENOSPC) {
6144                 /* whoops, lets try again with the full transaction */
6145                 btrfs_end_transaction(trans);
6146                 trans = btrfs_start_transaction(root, 1);
6147                 if (IS_ERR(trans))
6148                         return PTR_ERR(trans);
6149
6150                 ret = btrfs_update_inode(trans, root, inode);
6151         }
6152         btrfs_end_transaction(trans);
6153         if (BTRFS_I(inode)->delayed_node)
6154                 btrfs_balance_delayed_items(fs_info);
6155
6156         return ret;
6157 }
6158
6159 /*
6160  * This is a copy of file_update_time.  We need this so we can return error on
6161  * ENOSPC for updating the inode in the case of file write and mmap writes.
6162  */
6163 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6164                              int flags)
6165 {
6166         struct btrfs_root *root = BTRFS_I(inode)->root;
6167         bool dirty = flags & ~S_VERSION;
6168
6169         if (btrfs_root_readonly(root))
6170                 return -EROFS;
6171
6172         if (flags & S_VERSION)
6173                 dirty |= inode_maybe_inc_iversion(inode, dirty);
6174         if (flags & S_CTIME)
6175                 inode->i_ctime = *now;
6176         if (flags & S_MTIME)
6177                 inode->i_mtime = *now;
6178         if (flags & S_ATIME)
6179                 inode->i_atime = *now;
6180         return dirty ? btrfs_dirty_inode(inode) : 0;
6181 }
6182
6183 /*
6184  * find the highest existing sequence number in a directory
6185  * and then set the in-memory index_cnt variable to reflect
6186  * free sequence numbers
6187  */
6188 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6189 {
6190         struct btrfs_root *root = inode->root;
6191         struct btrfs_key key, found_key;
6192         struct btrfs_path *path;
6193         struct extent_buffer *leaf;
6194         int ret;
6195
6196         key.objectid = btrfs_ino(inode);
6197         key.type = BTRFS_DIR_INDEX_KEY;
6198         key.offset = (u64)-1;
6199
6200         path = btrfs_alloc_path();
6201         if (!path)
6202                 return -ENOMEM;
6203
6204         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6205         if (ret < 0)
6206                 goto out;
6207         /* FIXME: we should be able to handle this */
6208         if (ret == 0)
6209                 goto out;
6210         ret = 0;
6211
6212         /*
6213          * MAGIC NUMBER EXPLANATION:
6214          * since we search a directory based on f_pos we have to start at 2
6215          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6216          * else has to start at 2
6217          */
6218         if (path->slots[0] == 0) {
6219                 inode->index_cnt = 2;
6220                 goto out;
6221         }
6222
6223         path->slots[0]--;
6224
6225         leaf = path->nodes[0];
6226         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6227
6228         if (found_key.objectid != btrfs_ino(inode) ||
6229             found_key.type != BTRFS_DIR_INDEX_KEY) {
6230                 inode->index_cnt = 2;
6231                 goto out;
6232         }
6233
6234         inode->index_cnt = found_key.offset + 1;
6235 out:
6236         btrfs_free_path(path);
6237         return ret;
6238 }
6239
6240 /*
6241  * helper to find a free sequence number in a given directory.  This current
6242  * code is very simple, later versions will do smarter things in the btree
6243  */
6244 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6245 {
6246         int ret = 0;
6247
6248         if (dir->index_cnt == (u64)-1) {
6249                 ret = btrfs_inode_delayed_dir_index_count(dir);
6250                 if (ret) {
6251                         ret = btrfs_set_inode_index_count(dir);
6252                         if (ret)
6253                                 return ret;
6254                 }
6255         }
6256
6257         *index = dir->index_cnt;
6258         dir->index_cnt++;
6259
6260         return ret;
6261 }
6262
6263 static int btrfs_insert_inode_locked(struct inode *inode)
6264 {
6265         struct btrfs_iget_args args;
6266         args.location = &BTRFS_I(inode)->location;
6267         args.root = BTRFS_I(inode)->root;
6268
6269         return insert_inode_locked4(inode,
6270                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6271                    btrfs_find_actor, &args);
6272 }
6273
6274 /*
6275  * Inherit flags from the parent inode.
6276  *
6277  * Currently only the compression flags and the cow flags are inherited.
6278  */
6279 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6280 {
6281         unsigned int flags;
6282
6283         if (!dir)
6284                 return;
6285
6286         flags = BTRFS_I(dir)->flags;
6287
6288         if (flags & BTRFS_INODE_NOCOMPRESS) {
6289                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6290                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6291         } else if (flags & BTRFS_INODE_COMPRESS) {
6292                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6293                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6294         }
6295
6296         if (flags & BTRFS_INODE_NODATACOW) {
6297                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6298                 if (S_ISREG(inode->i_mode))
6299                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6300         }
6301
6302         btrfs_update_iflags(inode);
6303 }
6304
6305 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6306                                      struct btrfs_root *root,
6307                                      struct inode *dir,
6308                                      const char *name, int name_len,
6309                                      u64 ref_objectid, u64 objectid,
6310                                      umode_t mode, u64 *index)
6311 {
6312         struct btrfs_fs_info *fs_info = root->fs_info;
6313         struct inode *inode;
6314         struct btrfs_inode_item *inode_item;
6315         struct btrfs_key *location;
6316         struct btrfs_path *path;
6317         struct btrfs_inode_ref *ref;
6318         struct btrfs_key key[2];
6319         u32 sizes[2];
6320         int nitems = name ? 2 : 1;
6321         unsigned long ptr;
6322         int ret;
6323
6324         path = btrfs_alloc_path();
6325         if (!path)
6326                 return ERR_PTR(-ENOMEM);
6327
6328         inode = new_inode(fs_info->sb);
6329         if (!inode) {
6330                 btrfs_free_path(path);
6331                 return ERR_PTR(-ENOMEM);
6332         }
6333
6334         /*
6335          * O_TMPFILE, set link count to 0, so that after this point,
6336          * we fill in an inode item with the correct link count.
6337          */
6338         if (!name)
6339                 set_nlink(inode, 0);
6340
6341         /*
6342          * we have to initialize this early, so we can reclaim the inode
6343          * number if we fail afterwards in this function.
6344          */
6345         inode->i_ino = objectid;
6346
6347         if (dir && name) {
6348                 trace_btrfs_inode_request(dir);
6349
6350                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6351                 if (ret) {
6352                         btrfs_free_path(path);
6353                         iput(inode);
6354                         return ERR_PTR(ret);
6355                 }
6356         } else if (dir) {
6357                 *index = 0;
6358         }
6359         /*
6360          * index_cnt is ignored for everything but a dir,
6361          * btrfs_set_inode_index_count has an explanation for the magic
6362          * number
6363          */
6364         BTRFS_I(inode)->index_cnt = 2;
6365         BTRFS_I(inode)->dir_index = *index;
6366         BTRFS_I(inode)->root = root;
6367         BTRFS_I(inode)->generation = trans->transid;
6368         inode->i_generation = BTRFS_I(inode)->generation;
6369
6370         /*
6371          * We could have gotten an inode number from somebody who was fsynced
6372          * and then removed in this same transaction, so let's just set full
6373          * sync since it will be a full sync anyway and this will blow away the
6374          * old info in the log.
6375          */
6376         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6377
6378         key[0].objectid = objectid;
6379         key[0].type = BTRFS_INODE_ITEM_KEY;
6380         key[0].offset = 0;
6381
6382         sizes[0] = sizeof(struct btrfs_inode_item);
6383
6384         if (name) {
6385                 /*
6386                  * Start new inodes with an inode_ref. This is slightly more
6387                  * efficient for small numbers of hard links since they will
6388                  * be packed into one item. Extended refs will kick in if we
6389                  * add more hard links than can fit in the ref item.
6390                  */
6391                 key[1].objectid = objectid;
6392                 key[1].type = BTRFS_INODE_REF_KEY;
6393                 key[1].offset = ref_objectid;
6394
6395                 sizes[1] = name_len + sizeof(*ref);
6396         }
6397
6398         location = &BTRFS_I(inode)->location;
6399         location->objectid = objectid;
6400         location->offset = 0;
6401         location->type = BTRFS_INODE_ITEM_KEY;
6402
6403         ret = btrfs_insert_inode_locked(inode);
6404         if (ret < 0)
6405                 goto fail;
6406
6407         path->leave_spinning = 1;
6408         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6409         if (ret != 0)
6410                 goto fail_unlock;
6411
6412         inode_init_owner(inode, dir, mode);
6413         inode_set_bytes(inode, 0);
6414
6415         inode->i_mtime = current_time(inode);
6416         inode->i_atime = inode->i_mtime;
6417         inode->i_ctime = inode->i_mtime;
6418         BTRFS_I(inode)->i_otime = inode->i_mtime;
6419
6420         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6421                                   struct btrfs_inode_item);
6422         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6423                              sizeof(*inode_item));
6424         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6425
6426         if (name) {
6427                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6428                                      struct btrfs_inode_ref);
6429                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6430                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6431                 ptr = (unsigned long)(ref + 1);
6432                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6433         }
6434
6435         btrfs_mark_buffer_dirty(path->nodes[0]);
6436         btrfs_free_path(path);
6437
6438         btrfs_inherit_iflags(inode, dir);
6439
6440         if (S_ISREG(mode)) {
6441                 if (btrfs_test_opt(fs_info, NODATASUM))
6442                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6443                 if (btrfs_test_opt(fs_info, NODATACOW))
6444                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6445                                 BTRFS_INODE_NODATASUM;
6446         }
6447
6448         inode_tree_add(inode);
6449
6450         trace_btrfs_inode_new(inode);
6451         btrfs_set_inode_last_trans(trans, inode);
6452
6453         btrfs_update_root_times(trans, root);
6454
6455         ret = btrfs_inode_inherit_props(trans, inode, dir);
6456         if (ret)
6457                 btrfs_err(fs_info,
6458                           "error inheriting props for ino %llu (root %llu): %d",
6459                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6460
6461         return inode;
6462
6463 fail_unlock:
6464         unlock_new_inode(inode);
6465 fail:
6466         if (dir && name)
6467                 BTRFS_I(dir)->index_cnt--;
6468         btrfs_free_path(path);
6469         iput(inode);
6470         return ERR_PTR(ret);
6471 }
6472
6473 static inline u8 btrfs_inode_type(struct inode *inode)
6474 {
6475         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6476 }
6477
6478 /*
6479  * utility function to add 'inode' into 'parent_inode' with
6480  * a give name and a given sequence number.
6481  * if 'add_backref' is true, also insert a backref from the
6482  * inode to the parent directory.
6483  */
6484 int btrfs_add_link(struct btrfs_trans_handle *trans,
6485                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6486                    const char *name, int name_len, int add_backref, u64 index)
6487 {
6488         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6489         int ret = 0;
6490         struct btrfs_key key;
6491         struct btrfs_root *root = parent_inode->root;
6492         u64 ino = btrfs_ino(inode);
6493         u64 parent_ino = btrfs_ino(parent_inode);
6494
6495         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6496                 memcpy(&key, &inode->root->root_key, sizeof(key));
6497         } else {
6498                 key.objectid = ino;
6499                 key.type = BTRFS_INODE_ITEM_KEY;
6500                 key.offset = 0;
6501         }
6502
6503         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6504                 ret = btrfs_add_root_ref(trans, fs_info, key.objectid,
6505                                          root->root_key.objectid, parent_ino,
6506                                          index, name, name_len);
6507         } else if (add_backref) {
6508                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6509                                              parent_ino, index);
6510         }
6511
6512         /* Nothing to clean up yet */
6513         if (ret)
6514                 return ret;
6515
6516         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6517                                     parent_inode, &key,
6518                                     btrfs_inode_type(&inode->vfs_inode), index);
6519         if (ret == -EEXIST || ret == -EOVERFLOW)
6520                 goto fail_dir_item;
6521         else if (ret) {
6522                 btrfs_abort_transaction(trans, ret);
6523                 return ret;
6524         }
6525
6526         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6527                            name_len * 2);
6528         inode_inc_iversion(&parent_inode->vfs_inode);
6529         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6530                 current_time(&parent_inode->vfs_inode);
6531         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6532         if (ret)
6533                 btrfs_abort_transaction(trans, ret);
6534         return ret;
6535
6536 fail_dir_item:
6537         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6538                 u64 local_index;
6539                 int err;
6540                 err = btrfs_del_root_ref(trans, fs_info, key.objectid,
6541                                          root->root_key.objectid, parent_ino,
6542                                          &local_index, name, name_len);
6543
6544         } else if (add_backref) {
6545                 u64 local_index;
6546                 int err;
6547
6548                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6549                                           ino, parent_ino, &local_index);
6550         }
6551         return ret;
6552 }
6553
6554 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6555                             struct btrfs_inode *dir, struct dentry *dentry,
6556                             struct btrfs_inode *inode, int backref, u64 index)
6557 {
6558         int err = btrfs_add_link(trans, dir, inode,
6559                                  dentry->d_name.name, dentry->d_name.len,
6560                                  backref, index);
6561         if (err > 0)
6562                 err = -EEXIST;
6563         return err;
6564 }
6565
6566 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6567                         umode_t mode, dev_t rdev)
6568 {
6569         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6570         struct btrfs_trans_handle *trans;
6571         struct btrfs_root *root = BTRFS_I(dir)->root;
6572         struct inode *inode = NULL;
6573         int err;
6574         int drop_inode = 0;
6575         u64 objectid;
6576         u64 index = 0;
6577
6578         /*
6579          * 2 for inode item and ref
6580          * 2 for dir items
6581          * 1 for xattr if selinux is on
6582          */
6583         trans = btrfs_start_transaction(root, 5);
6584         if (IS_ERR(trans))
6585                 return PTR_ERR(trans);
6586
6587         err = btrfs_find_free_ino(root, &objectid);
6588         if (err)
6589                 goto out_unlock;
6590
6591         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6592                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6593                         mode, &index);
6594         if (IS_ERR(inode)) {
6595                 err = PTR_ERR(inode);
6596                 goto out_unlock;
6597         }
6598
6599         /*
6600         * If the active LSM wants to access the inode during
6601         * d_instantiate it needs these. Smack checks to see
6602         * if the filesystem supports xattrs by looking at the
6603         * ops vector.
6604         */
6605         inode->i_op = &btrfs_special_inode_operations;
6606         init_special_inode(inode, inode->i_mode, rdev);
6607
6608         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6609         if (err)
6610                 goto out_unlock_inode;
6611
6612         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6613                         0, index);
6614         if (err) {
6615                 goto out_unlock_inode;
6616         } else {
6617                 btrfs_update_inode(trans, root, inode);
6618                 unlock_new_inode(inode);
6619                 d_instantiate(dentry, inode);
6620         }
6621
6622 out_unlock:
6623         btrfs_end_transaction(trans);
6624         btrfs_btree_balance_dirty(fs_info);
6625         if (drop_inode) {
6626                 inode_dec_link_count(inode);
6627                 iput(inode);
6628         }
6629         return err;
6630
6631 out_unlock_inode:
6632         drop_inode = 1;
6633         unlock_new_inode(inode);
6634         goto out_unlock;
6635
6636 }
6637
6638 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6639                         umode_t mode, bool excl)
6640 {
6641         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6642         struct btrfs_trans_handle *trans;
6643         struct btrfs_root *root = BTRFS_I(dir)->root;
6644         struct inode *inode = NULL;
6645         int drop_inode_on_err = 0;
6646         int err;
6647         u64 objectid;
6648         u64 index = 0;
6649
6650         /*
6651          * 2 for inode item and ref
6652          * 2 for dir items
6653          * 1 for xattr if selinux is on
6654          */
6655         trans = btrfs_start_transaction(root, 5);
6656         if (IS_ERR(trans))
6657                 return PTR_ERR(trans);
6658
6659         err = btrfs_find_free_ino(root, &objectid);
6660         if (err)
6661                 goto out_unlock;
6662
6663         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6664                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6665                         mode, &index);
6666         if (IS_ERR(inode)) {
6667                 err = PTR_ERR(inode);
6668                 goto out_unlock;
6669         }
6670         drop_inode_on_err = 1;
6671         /*
6672         * If the active LSM wants to access the inode during
6673         * d_instantiate it needs these. Smack checks to see
6674         * if the filesystem supports xattrs by looking at the
6675         * ops vector.
6676         */
6677         inode->i_fop = &btrfs_file_operations;
6678         inode->i_op = &btrfs_file_inode_operations;
6679         inode->i_mapping->a_ops = &btrfs_aops;
6680
6681         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6682         if (err)
6683                 goto out_unlock_inode;
6684
6685         err = btrfs_update_inode(trans, root, inode);
6686         if (err)
6687                 goto out_unlock_inode;
6688
6689         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6690                         0, index);
6691         if (err)
6692                 goto out_unlock_inode;
6693
6694         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6695         unlock_new_inode(inode);
6696         d_instantiate(dentry, inode);
6697
6698 out_unlock:
6699         btrfs_end_transaction(trans);
6700         if (err && drop_inode_on_err) {
6701                 inode_dec_link_count(inode);
6702                 iput(inode);
6703         }
6704         btrfs_btree_balance_dirty(fs_info);
6705         return err;
6706
6707 out_unlock_inode:
6708         unlock_new_inode(inode);
6709         goto out_unlock;
6710
6711 }
6712
6713 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6714                       struct dentry *dentry)
6715 {
6716         struct btrfs_trans_handle *trans = NULL;
6717         struct btrfs_root *root = BTRFS_I(dir)->root;
6718         struct inode *inode = d_inode(old_dentry);
6719         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6720         u64 index;
6721         int err;
6722         int drop_inode = 0;
6723
6724         /* do not allow sys_link's with other subvols of the same device */
6725         if (root->objectid != BTRFS_I(inode)->root->objectid)
6726                 return -EXDEV;
6727
6728         if (inode->i_nlink >= BTRFS_LINK_MAX)
6729                 return -EMLINK;
6730
6731         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6732         if (err)
6733                 goto fail;
6734
6735         /*
6736          * 2 items for inode and inode ref
6737          * 2 items for dir items
6738          * 1 item for parent inode
6739          */
6740         trans = btrfs_start_transaction(root, 5);
6741         if (IS_ERR(trans)) {
6742                 err = PTR_ERR(trans);
6743                 trans = NULL;
6744                 goto fail;
6745         }
6746
6747         /* There are several dir indexes for this inode, clear the cache. */
6748         BTRFS_I(inode)->dir_index = 0ULL;
6749         inc_nlink(inode);
6750         inode_inc_iversion(inode);
6751         inode->i_ctime = current_time(inode);
6752         ihold(inode);
6753         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6754
6755         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6756                         1, index);
6757
6758         if (err) {
6759                 drop_inode = 1;
6760         } else {
6761                 struct dentry *parent = dentry->d_parent;
6762                 err = btrfs_update_inode(trans, root, inode);
6763                 if (err)
6764                         goto fail;
6765                 if (inode->i_nlink == 1) {
6766                         /*
6767                          * If new hard link count is 1, it's a file created
6768                          * with open(2) O_TMPFILE flag.
6769                          */
6770                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6771                         if (err)
6772                                 goto fail;
6773                 }
6774                 d_instantiate(dentry, inode);
6775                 btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent);
6776         }
6777
6778 fail:
6779         if (trans)
6780                 btrfs_end_transaction(trans);
6781         if (drop_inode) {
6782                 inode_dec_link_count(inode);
6783                 iput(inode);
6784         }
6785         btrfs_btree_balance_dirty(fs_info);
6786         return err;
6787 }
6788
6789 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6790 {
6791         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6792         struct inode *inode = NULL;
6793         struct btrfs_trans_handle *trans;
6794         struct btrfs_root *root = BTRFS_I(dir)->root;
6795         int err = 0;
6796         int drop_on_err = 0;
6797         u64 objectid = 0;
6798         u64 index = 0;
6799
6800         /*
6801          * 2 items for inode and ref
6802          * 2 items for dir items
6803          * 1 for xattr if selinux is on
6804          */
6805         trans = btrfs_start_transaction(root, 5);
6806         if (IS_ERR(trans))
6807                 return PTR_ERR(trans);
6808
6809         err = btrfs_find_free_ino(root, &objectid);
6810         if (err)
6811                 goto out_fail;
6812
6813         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6814                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6815                         S_IFDIR | mode, &index);
6816         if (IS_ERR(inode)) {
6817                 err = PTR_ERR(inode);
6818                 goto out_fail;
6819         }
6820
6821         drop_on_err = 1;
6822         /* these must be set before we unlock the inode */
6823         inode->i_op = &btrfs_dir_inode_operations;
6824         inode->i_fop = &btrfs_dir_file_operations;
6825
6826         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6827         if (err)
6828                 goto out_fail_inode;
6829
6830         btrfs_i_size_write(BTRFS_I(inode), 0);
6831         err = btrfs_update_inode(trans, root, inode);
6832         if (err)
6833                 goto out_fail_inode;
6834
6835         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6836                         dentry->d_name.name,
6837                         dentry->d_name.len, 0, index);
6838         if (err)
6839                 goto out_fail_inode;
6840
6841         d_instantiate(dentry, inode);
6842         /*
6843          * mkdir is special.  We're unlocking after we call d_instantiate
6844          * to avoid a race with nfsd calling d_instantiate.
6845          */
6846         unlock_new_inode(inode);
6847         drop_on_err = 0;
6848
6849 out_fail:
6850         btrfs_end_transaction(trans);
6851         if (drop_on_err) {
6852                 inode_dec_link_count(inode);
6853                 iput(inode);
6854         }
6855         btrfs_btree_balance_dirty(fs_info);
6856         return err;
6857
6858 out_fail_inode:
6859         unlock_new_inode(inode);
6860         goto out_fail;
6861 }
6862
6863 static noinline int uncompress_inline(struct btrfs_path *path,
6864                                       struct page *page,
6865                                       size_t pg_offset, u64 extent_offset,
6866                                       struct btrfs_file_extent_item *item)
6867 {
6868         int ret;
6869         struct extent_buffer *leaf = path->nodes[0];
6870         char *tmp;
6871         size_t max_size;
6872         unsigned long inline_size;
6873         unsigned long ptr;
6874         int compress_type;
6875
6876         WARN_ON(pg_offset != 0);
6877         compress_type = btrfs_file_extent_compression(leaf, item);
6878         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6879         inline_size = btrfs_file_extent_inline_item_len(leaf,
6880                                         btrfs_item_nr(path->slots[0]));
6881         tmp = kmalloc(inline_size, GFP_NOFS);
6882         if (!tmp)
6883                 return -ENOMEM;
6884         ptr = btrfs_file_extent_inline_start(item);
6885
6886         read_extent_buffer(leaf, tmp, ptr, inline_size);
6887
6888         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6889         ret = btrfs_decompress(compress_type, tmp, page,
6890                                extent_offset, inline_size, max_size);
6891
6892         /*
6893          * decompression code contains a memset to fill in any space between the end
6894          * of the uncompressed data and the end of max_size in case the decompressed
6895          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6896          * the end of an inline extent and the beginning of the next block, so we
6897          * cover that region here.
6898          */
6899
6900         if (max_size + pg_offset < PAGE_SIZE) {
6901                 char *map = kmap(page);
6902                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6903                 kunmap(page);
6904         }
6905         kfree(tmp);
6906         return ret;
6907 }
6908
6909 /*
6910  * a bit scary, this does extent mapping from logical file offset to the disk.
6911  * the ugly parts come from merging extents from the disk with the in-ram
6912  * representation.  This gets more complex because of the data=ordered code,
6913  * where the in-ram extents might be locked pending data=ordered completion.
6914  *
6915  * This also copies inline extents directly into the page.
6916  */
6917 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6918                 struct page *page,
6919             size_t pg_offset, u64 start, u64 len,
6920                 int create)
6921 {
6922         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6923         int ret;
6924         int err = 0;
6925         u64 extent_start = 0;
6926         u64 extent_end = 0;
6927         u64 objectid = btrfs_ino(inode);
6928         u32 found_type;
6929         struct btrfs_path *path = NULL;
6930         struct btrfs_root *root = inode->root;
6931         struct btrfs_file_extent_item *item;
6932         struct extent_buffer *leaf;
6933         struct btrfs_key found_key;
6934         struct extent_map *em = NULL;
6935         struct extent_map_tree *em_tree = &inode->extent_tree;
6936         struct extent_io_tree *io_tree = &inode->io_tree;
6937         const bool new_inline = !page || create;
6938
6939         read_lock(&em_tree->lock);
6940         em = lookup_extent_mapping(em_tree, start, len);
6941         if (em)
6942                 em->bdev = fs_info->fs_devices->latest_bdev;
6943         read_unlock(&em_tree->lock);
6944
6945         if (em) {
6946                 if (em->start > start || em->start + em->len <= start)
6947                         free_extent_map(em);
6948                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6949                         free_extent_map(em);
6950                 else
6951                         goto out;
6952         }
6953         em = alloc_extent_map();
6954         if (!em) {
6955                 err = -ENOMEM;
6956                 goto out;
6957         }
6958         em->bdev = fs_info->fs_devices->latest_bdev;
6959         em->start = EXTENT_MAP_HOLE;
6960         em->orig_start = EXTENT_MAP_HOLE;
6961         em->len = (u64)-1;
6962         em->block_len = (u64)-1;
6963
6964         if (!path) {
6965                 path = btrfs_alloc_path();
6966                 if (!path) {
6967                         err = -ENOMEM;
6968                         goto out;
6969                 }
6970                 /*
6971                  * Chances are we'll be called again, so go ahead and do
6972                  * readahead
6973                  */
6974                 path->reada = READA_FORWARD;
6975         }
6976
6977         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6978         if (ret < 0) {
6979                 err = ret;
6980                 goto out;
6981         }
6982
6983         if (ret != 0) {
6984                 if (path->slots[0] == 0)
6985                         goto not_found;
6986                 path->slots[0]--;
6987         }
6988
6989         leaf = path->nodes[0];
6990         item = btrfs_item_ptr(leaf, path->slots[0],
6991                               struct btrfs_file_extent_item);
6992         /* are we inside the extent that was found? */
6993         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6994         found_type = found_key.type;
6995         if (found_key.objectid != objectid ||
6996             found_type != BTRFS_EXTENT_DATA_KEY) {
6997                 /*
6998                  * If we backup past the first extent we want to move forward
6999                  * and see if there is an extent in front of us, otherwise we'll
7000                  * say there is a hole for our whole search range which can
7001                  * cause problems.
7002                  */
7003                 extent_end = start;
7004                 goto next;
7005         }
7006
7007         found_type = btrfs_file_extent_type(leaf, item);
7008         extent_start = found_key.offset;
7009         if (found_type == BTRFS_FILE_EXTENT_REG ||
7010             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7011                 extent_end = extent_start +
7012                        btrfs_file_extent_num_bytes(leaf, item);
7013
7014                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
7015                                                        extent_start);
7016         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7017                 size_t size;
7018                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7019                 extent_end = ALIGN(extent_start + size,
7020                                    fs_info->sectorsize);
7021
7022                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
7023                                                       path->slots[0],
7024                                                       extent_start);
7025         }
7026 next:
7027         if (start >= extent_end) {
7028                 path->slots[0]++;
7029                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
7030                         ret = btrfs_next_leaf(root, path);
7031                         if (ret < 0) {
7032                                 err = ret;
7033                                 goto out;
7034                         }
7035                         if (ret > 0)
7036                                 goto not_found;
7037                         leaf = path->nodes[0];
7038                 }
7039                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7040                 if (found_key.objectid != objectid ||
7041                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7042                         goto not_found;
7043                 if (start + len <= found_key.offset)
7044                         goto not_found;
7045                 if (start > found_key.offset)
7046                         goto next;
7047                 em->start = start;
7048                 em->orig_start = start;
7049                 em->len = found_key.offset - start;
7050                 goto not_found_em;
7051         }
7052
7053         btrfs_extent_item_to_extent_map(inode, path, item,
7054                         new_inline, em);
7055
7056         if (found_type == BTRFS_FILE_EXTENT_REG ||
7057             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7058                 goto insert;
7059         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7060                 unsigned long ptr;
7061                 char *map;
7062                 size_t size;
7063                 size_t extent_offset;
7064                 size_t copy_size;
7065
7066                 if (new_inline)
7067                         goto out;
7068
7069                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7070                 extent_offset = page_offset(page) + pg_offset - extent_start;
7071                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7072                                   size - extent_offset);
7073                 em->start = extent_start + extent_offset;
7074                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7075                 em->orig_block_len = em->len;
7076                 em->orig_start = em->start;
7077                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7078                 if (!PageUptodate(page)) {
7079                         if (btrfs_file_extent_compression(leaf, item) !=
7080                             BTRFS_COMPRESS_NONE) {
7081                                 ret = uncompress_inline(path, page, pg_offset,
7082                                                         extent_offset, item);
7083                                 if (ret) {
7084                                         err = ret;
7085                                         goto out;
7086                                 }
7087                         } else {
7088                                 map = kmap(page);
7089                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7090                                                    copy_size);
7091                                 if (pg_offset + copy_size < PAGE_SIZE) {
7092                                         memset(map + pg_offset + copy_size, 0,
7093                                                PAGE_SIZE - pg_offset -
7094                                                copy_size);
7095                                 }
7096                                 kunmap(page);
7097                         }
7098                         flush_dcache_page(page);
7099                 }
7100                 set_extent_uptodate(io_tree, em->start,
7101                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7102                 goto insert;
7103         }
7104 not_found:
7105         em->start = start;
7106         em->orig_start = start;
7107         em->len = len;
7108 not_found_em:
7109         em->block_start = EXTENT_MAP_HOLE;
7110 insert:
7111         btrfs_release_path(path);
7112         if (em->start > start || extent_map_end(em) <= start) {
7113                 btrfs_err(fs_info,
7114                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7115                           em->start, em->len, start, len);
7116                 err = -EIO;
7117                 goto out;
7118         }
7119
7120         err = 0;
7121         write_lock(&em_tree->lock);
7122         err = btrfs_add_extent_mapping(em_tree, &em, start, len);
7123         write_unlock(&em_tree->lock);
7124 out:
7125
7126         trace_btrfs_get_extent(root, inode, em);
7127
7128         btrfs_free_path(path);
7129         if (err) {
7130                 free_extent_map(em);
7131                 return ERR_PTR(err);
7132         }
7133         BUG_ON(!em); /* Error is always set */
7134         return em;
7135 }
7136
7137 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7138                 struct page *page,
7139                 size_t pg_offset, u64 start, u64 len,
7140                 int create)
7141 {
7142         struct extent_map *em;
7143         struct extent_map *hole_em = NULL;
7144         u64 range_start = start;
7145         u64 end;
7146         u64 found;
7147         u64 found_end;
7148         int err = 0;
7149
7150         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7151         if (IS_ERR(em))
7152                 return em;
7153         /*
7154          * If our em maps to:
7155          * - a hole or
7156          * - a pre-alloc extent,
7157          * there might actually be delalloc bytes behind it.
7158          */
7159         if (em->block_start != EXTENT_MAP_HOLE &&
7160             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7161                 return em;
7162         else
7163                 hole_em = em;
7164
7165         /* check to see if we've wrapped (len == -1 or similar) */
7166         end = start + len;
7167         if (end < start)
7168                 end = (u64)-1;
7169         else
7170                 end -= 1;
7171
7172         em = NULL;
7173
7174         /* ok, we didn't find anything, lets look for delalloc */
7175         found = count_range_bits(&inode->io_tree, &range_start,
7176                                  end, len, EXTENT_DELALLOC, 1);
7177         found_end = range_start + found;
7178         if (found_end < range_start)
7179                 found_end = (u64)-1;
7180
7181         /*
7182          * we didn't find anything useful, return
7183          * the original results from get_extent()
7184          */
7185         if (range_start > end || found_end <= start) {
7186                 em = hole_em;
7187                 hole_em = NULL;
7188                 goto out;
7189         }
7190
7191         /* adjust the range_start to make sure it doesn't
7192          * go backwards from the start they passed in
7193          */
7194         range_start = max(start, range_start);
7195         found = found_end - range_start;
7196
7197         if (found > 0) {
7198                 u64 hole_start = start;
7199                 u64 hole_len = len;
7200
7201                 em = alloc_extent_map();
7202                 if (!em) {
7203                         err = -ENOMEM;
7204                         goto out;
7205                 }
7206                 /*
7207                  * when btrfs_get_extent can't find anything it
7208                  * returns one huge hole
7209                  *
7210                  * make sure what it found really fits our range, and
7211                  * adjust to make sure it is based on the start from
7212                  * the caller
7213                  */
7214                 if (hole_em) {
7215                         u64 calc_end = extent_map_end(hole_em);
7216
7217                         if (calc_end <= start || (hole_em->start > end)) {
7218                                 free_extent_map(hole_em);
7219                                 hole_em = NULL;
7220                         } else {
7221                                 hole_start = max(hole_em->start, start);
7222                                 hole_len = calc_end - hole_start;
7223                         }
7224                 }
7225                 em->bdev = NULL;
7226                 if (hole_em && range_start > hole_start) {
7227                         /* our hole starts before our delalloc, so we
7228                          * have to return just the parts of the hole
7229                          * that go until  the delalloc starts
7230                          */
7231                         em->len = min(hole_len,
7232                                       range_start - hole_start);
7233                         em->start = hole_start;
7234                         em->orig_start = hole_start;
7235                         /*
7236                          * don't adjust block start at all,
7237                          * it is fixed at EXTENT_MAP_HOLE
7238                          */
7239                         em->block_start = hole_em->block_start;
7240                         em->block_len = hole_len;
7241                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7242                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7243                 } else {
7244                         em->start = range_start;
7245                         em->len = found;
7246                         em->orig_start = range_start;
7247                         em->block_start = EXTENT_MAP_DELALLOC;
7248                         em->block_len = found;
7249                 }
7250         } else {
7251                 return hole_em;
7252         }
7253 out:
7254
7255         free_extent_map(hole_em);
7256         if (err) {
7257                 free_extent_map(em);
7258                 return ERR_PTR(err);
7259         }
7260         return em;
7261 }
7262
7263 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7264                                                   const u64 start,
7265                                                   const u64 len,
7266                                                   const u64 orig_start,
7267                                                   const u64 block_start,
7268                                                   const u64 block_len,
7269                                                   const u64 orig_block_len,
7270                                                   const u64 ram_bytes,
7271                                                   const int type)
7272 {
7273         struct extent_map *em = NULL;
7274         int ret;
7275
7276         if (type != BTRFS_ORDERED_NOCOW) {
7277                 em = create_io_em(inode, start, len, orig_start,
7278                                   block_start, block_len, orig_block_len,
7279                                   ram_bytes,
7280                                   BTRFS_COMPRESS_NONE, /* compress_type */
7281                                   type);
7282                 if (IS_ERR(em))
7283                         goto out;
7284         }
7285         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7286                                            len, block_len, type);
7287         if (ret) {
7288                 if (em) {
7289                         free_extent_map(em);
7290                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7291                                                 start + len - 1, 0);
7292                 }
7293                 em = ERR_PTR(ret);
7294         }
7295  out:
7296
7297         return em;
7298 }
7299
7300 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7301                                                   u64 start, u64 len)
7302 {
7303         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7304         struct btrfs_root *root = BTRFS_I(inode)->root;
7305         struct extent_map *em;
7306         struct btrfs_key ins;
7307         u64 alloc_hint;
7308         int ret;
7309
7310         alloc_hint = get_extent_allocation_hint(inode, start, len);
7311         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7312                                    0, alloc_hint, &ins, 1, 1);
7313         if (ret)
7314                 return ERR_PTR(ret);
7315
7316         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7317                                      ins.objectid, ins.offset, ins.offset,
7318                                      ins.offset, BTRFS_ORDERED_REGULAR);
7319         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7320         if (IS_ERR(em))
7321                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7322                                            ins.offset, 1);
7323
7324         return em;
7325 }
7326
7327 /*
7328  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7329  * block must be cow'd
7330  */
7331 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7332                               u64 *orig_start, u64 *orig_block_len,
7333                               u64 *ram_bytes)
7334 {
7335         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7336         struct btrfs_path *path;
7337         int ret;
7338         struct extent_buffer *leaf;
7339         struct btrfs_root *root = BTRFS_I(inode)->root;
7340         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7341         struct btrfs_file_extent_item *fi;
7342         struct btrfs_key key;
7343         u64 disk_bytenr;
7344         u64 backref_offset;
7345         u64 extent_end;
7346         u64 num_bytes;
7347         int slot;
7348         int found_type;
7349         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7350
7351         path = btrfs_alloc_path();
7352         if (!path)
7353                 return -ENOMEM;
7354
7355         ret = btrfs_lookup_file_extent(NULL, root, path,
7356                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7357         if (ret < 0)
7358                 goto out;
7359
7360         slot = path->slots[0];
7361         if (ret == 1) {
7362                 if (slot == 0) {
7363                         /* can't find the item, must cow */
7364                         ret = 0;
7365                         goto out;
7366                 }
7367                 slot--;
7368         }
7369         ret = 0;
7370         leaf = path->nodes[0];
7371         btrfs_item_key_to_cpu(leaf, &key, slot);
7372         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7373             key.type != BTRFS_EXTENT_DATA_KEY) {
7374                 /* not our file or wrong item type, must cow */
7375                 goto out;
7376         }
7377
7378         if (key.offset > offset) {
7379                 /* Wrong offset, must cow */
7380                 goto out;
7381         }
7382
7383         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7384         found_type = btrfs_file_extent_type(leaf, fi);
7385         if (found_type != BTRFS_FILE_EXTENT_REG &&
7386             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7387                 /* not a regular extent, must cow */
7388                 goto out;
7389         }
7390
7391         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7392                 goto out;
7393
7394         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7395         if (extent_end <= offset)
7396                 goto out;
7397
7398         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7399         if (disk_bytenr == 0)
7400                 goto out;
7401
7402         if (btrfs_file_extent_compression(leaf, fi) ||
7403             btrfs_file_extent_encryption(leaf, fi) ||
7404             btrfs_file_extent_other_encoding(leaf, fi))
7405                 goto out;
7406
7407         backref_offset = btrfs_file_extent_offset(leaf, fi);
7408
7409         if (orig_start) {
7410                 *orig_start = key.offset - backref_offset;
7411                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7412                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7413         }
7414
7415         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7416                 goto out;
7417
7418         num_bytes = min(offset + *len, extent_end) - offset;
7419         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7420                 u64 range_end;
7421
7422                 range_end = round_up(offset + num_bytes,
7423                                      root->fs_info->sectorsize) - 1;
7424                 ret = test_range_bit(io_tree, offset, range_end,
7425                                      EXTENT_DELALLOC, 0, NULL);
7426                 if (ret) {
7427                         ret = -EAGAIN;
7428                         goto out;
7429                 }
7430         }
7431
7432         btrfs_release_path(path);
7433
7434         /*
7435          * look for other files referencing this extent, if we
7436          * find any we must cow
7437          */
7438
7439         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7440                                     key.offset - backref_offset, disk_bytenr);
7441         if (ret) {
7442                 ret = 0;
7443                 goto out;
7444         }
7445
7446         /*
7447          * adjust disk_bytenr and num_bytes to cover just the bytes
7448          * in this extent we are about to write.  If there
7449          * are any csums in that range we have to cow in order
7450          * to keep the csums correct
7451          */
7452         disk_bytenr += backref_offset;
7453         disk_bytenr += offset - key.offset;
7454         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7455                 goto out;
7456         /*
7457          * all of the above have passed, it is safe to overwrite this extent
7458          * without cow
7459          */
7460         *len = num_bytes;
7461         ret = 1;
7462 out:
7463         btrfs_free_path(path);
7464         return ret;
7465 }
7466
7467 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7468 {
7469         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7470         bool found = false;
7471         void **pagep = NULL;
7472         struct page *page = NULL;
7473         unsigned long start_idx;
7474         unsigned long end_idx;
7475
7476         start_idx = start >> PAGE_SHIFT;
7477
7478         /*
7479          * end is the last byte in the last page.  end == start is legal
7480          */
7481         end_idx = end >> PAGE_SHIFT;
7482
7483         rcu_read_lock();
7484
7485         /* Most of the code in this while loop is lifted from
7486          * find_get_page.  It's been modified to begin searching from a
7487          * page and return just the first page found in that range.  If the
7488          * found idx is less than or equal to the end idx then we know that
7489          * a page exists.  If no pages are found or if those pages are
7490          * outside of the range then we're fine (yay!) */
7491         while (page == NULL &&
7492                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7493                 page = radix_tree_deref_slot(pagep);
7494                 if (unlikely(!page))
7495                         break;
7496
7497                 if (radix_tree_exception(page)) {
7498                         if (radix_tree_deref_retry(page)) {
7499                                 page = NULL;
7500                                 continue;
7501                         }
7502                         /*
7503                          * Otherwise, shmem/tmpfs must be storing a swap entry
7504                          * here as an exceptional entry: so return it without
7505                          * attempting to raise page count.
7506                          */
7507                         page = NULL;
7508                         break; /* TODO: Is this relevant for this use case? */
7509                 }
7510
7511                 if (!page_cache_get_speculative(page)) {
7512                         page = NULL;
7513                         continue;
7514                 }
7515
7516                 /*
7517                  * Has the page moved?
7518                  * This is part of the lockless pagecache protocol. See
7519                  * include/linux/pagemap.h for details.
7520                  */
7521                 if (unlikely(page != *pagep)) {
7522                         put_page(page);
7523                         page = NULL;
7524                 }
7525         }
7526
7527         if (page) {
7528                 if (page->index <= end_idx)
7529                         found = true;
7530                 put_page(page);
7531         }
7532
7533         rcu_read_unlock();
7534         return found;
7535 }
7536
7537 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7538                               struct extent_state **cached_state, int writing)
7539 {
7540         struct btrfs_ordered_extent *ordered;
7541         int ret = 0;
7542
7543         while (1) {
7544                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7545                                  cached_state);
7546                 /*
7547                  * We're concerned with the entire range that we're going to be
7548                  * doing DIO to, so we need to make sure there's no ordered
7549                  * extents in this range.
7550                  */
7551                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7552                                                      lockend - lockstart + 1);
7553
7554                 /*
7555                  * We need to make sure there are no buffered pages in this
7556                  * range either, we could have raced between the invalidate in
7557                  * generic_file_direct_write and locking the extent.  The
7558                  * invalidate needs to happen so that reads after a write do not
7559                  * get stale data.
7560                  */
7561                 if (!ordered &&
7562                     (!writing ||
7563                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7564                         break;
7565
7566                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7567                                      cached_state);
7568
7569                 if (ordered) {
7570                         /*
7571                          * If we are doing a DIO read and the ordered extent we
7572                          * found is for a buffered write, we can not wait for it
7573                          * to complete and retry, because if we do so we can
7574                          * deadlock with concurrent buffered writes on page
7575                          * locks. This happens only if our DIO read covers more
7576                          * than one extent map, if at this point has already
7577                          * created an ordered extent for a previous extent map
7578                          * and locked its range in the inode's io tree, and a
7579                          * concurrent write against that previous extent map's
7580                          * range and this range started (we unlock the ranges
7581                          * in the io tree only when the bios complete and
7582                          * buffered writes always lock pages before attempting
7583                          * to lock range in the io tree).
7584                          */
7585                         if (writing ||
7586                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7587                                 btrfs_start_ordered_extent(inode, ordered, 1);
7588                         else
7589                                 ret = -ENOTBLK;
7590                         btrfs_put_ordered_extent(ordered);
7591                 } else {
7592                         /*
7593                          * We could trigger writeback for this range (and wait
7594                          * for it to complete) and then invalidate the pages for
7595                          * this range (through invalidate_inode_pages2_range()),
7596                          * but that can lead us to a deadlock with a concurrent
7597                          * call to readpages() (a buffered read or a defrag call
7598                          * triggered a readahead) on a page lock due to an
7599                          * ordered dio extent we created before but did not have
7600                          * yet a corresponding bio submitted (whence it can not
7601                          * complete), which makes readpages() wait for that
7602                          * ordered extent to complete while holding a lock on
7603                          * that page.
7604                          */
7605                         ret = -ENOTBLK;
7606                 }
7607
7608                 if (ret)
7609                         break;
7610
7611                 cond_resched();
7612         }
7613
7614         return ret;
7615 }
7616
7617 /* The callers of this must take lock_extent() */
7618 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7619                                        u64 orig_start, u64 block_start,
7620                                        u64 block_len, u64 orig_block_len,
7621                                        u64 ram_bytes, int compress_type,
7622                                        int type)
7623 {
7624         struct extent_map_tree *em_tree;
7625         struct extent_map *em;
7626         struct btrfs_root *root = BTRFS_I(inode)->root;
7627         int ret;
7628
7629         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7630                type == BTRFS_ORDERED_COMPRESSED ||
7631                type == BTRFS_ORDERED_NOCOW ||
7632                type == BTRFS_ORDERED_REGULAR);
7633
7634         em_tree = &BTRFS_I(inode)->extent_tree;
7635         em = alloc_extent_map();
7636         if (!em)
7637                 return ERR_PTR(-ENOMEM);
7638
7639         em->start = start;
7640         em->orig_start = orig_start;
7641         em->len = len;
7642         em->block_len = block_len;
7643         em->block_start = block_start;
7644         em->bdev = root->fs_info->fs_devices->latest_bdev;
7645         em->orig_block_len = orig_block_len;
7646         em->ram_bytes = ram_bytes;
7647         em->generation = -1;
7648         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7649         if (type == BTRFS_ORDERED_PREALLOC) {
7650                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7651         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7652                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7653                 em->compress_type = compress_type;
7654         }
7655
7656         do {
7657                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7658                                 em->start + em->len - 1, 0);
7659                 write_lock(&em_tree->lock);
7660                 ret = add_extent_mapping(em_tree, em, 1);
7661                 write_unlock(&em_tree->lock);
7662                 /*
7663                  * The caller has taken lock_extent(), who could race with us
7664                  * to add em?
7665                  */
7666         } while (ret == -EEXIST);
7667
7668         if (ret) {
7669                 free_extent_map(em);
7670                 return ERR_PTR(ret);
7671         }
7672
7673         /* em got 2 refs now, callers needs to do free_extent_map once. */
7674         return em;
7675 }
7676
7677 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7678                                    struct buffer_head *bh_result, int create)
7679 {
7680         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7681         struct extent_map *em;
7682         struct extent_state *cached_state = NULL;
7683         struct btrfs_dio_data *dio_data = NULL;
7684         u64 start = iblock << inode->i_blkbits;
7685         u64 lockstart, lockend;
7686         u64 len = bh_result->b_size;
7687         int unlock_bits = EXTENT_LOCKED;
7688         int ret = 0;
7689
7690         if (create)
7691                 unlock_bits |= EXTENT_DIRTY;
7692         else
7693                 len = min_t(u64, len, fs_info->sectorsize);
7694
7695         lockstart = start;
7696         lockend = start + len - 1;
7697
7698         if (current->journal_info) {
7699                 /*
7700                  * Need to pull our outstanding extents and set journal_info to NULL so
7701                  * that anything that needs to check if there's a transaction doesn't get
7702                  * confused.
7703                  */
7704                 dio_data = current->journal_info;
7705                 current->journal_info = NULL;
7706         }
7707
7708         /*
7709          * If this errors out it's because we couldn't invalidate pagecache for
7710          * this range and we need to fallback to buffered.
7711          */
7712         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7713                                create)) {
7714                 ret = -ENOTBLK;
7715                 goto err;
7716         }
7717
7718         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7719         if (IS_ERR(em)) {
7720                 ret = PTR_ERR(em);
7721                 goto unlock_err;
7722         }
7723
7724         /*
7725          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7726          * io.  INLINE is special, and we could probably kludge it in here, but
7727          * it's still buffered so for safety lets just fall back to the generic
7728          * buffered path.
7729          *
7730          * For COMPRESSED we _have_ to read the entire extent in so we can
7731          * decompress it, so there will be buffering required no matter what we
7732          * do, so go ahead and fallback to buffered.
7733          *
7734          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7735          * to buffered IO.  Don't blame me, this is the price we pay for using
7736          * the generic code.
7737          */
7738         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7739             em->block_start == EXTENT_MAP_INLINE) {
7740                 free_extent_map(em);
7741                 ret = -ENOTBLK;
7742                 goto unlock_err;
7743         }
7744
7745         /* Just a good old fashioned hole, return */
7746         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7747                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7748                 free_extent_map(em);
7749                 goto unlock_err;
7750         }
7751
7752         /*
7753          * We don't allocate a new extent in the following cases
7754          *
7755          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7756          * existing extent.
7757          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7758          * just use the extent.
7759          *
7760          */
7761         if (!create) {
7762                 len = min(len, em->len - (start - em->start));
7763                 lockstart = start + len;
7764                 goto unlock;
7765         }
7766
7767         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7768             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7769              em->block_start != EXTENT_MAP_HOLE)) {
7770                 int type;
7771                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7772
7773                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7774                         type = BTRFS_ORDERED_PREALLOC;
7775                 else
7776                         type = BTRFS_ORDERED_NOCOW;
7777                 len = min(len, em->len - (start - em->start));
7778                 block_start = em->block_start + (start - em->start);
7779
7780                 if (can_nocow_extent(inode, start, &len, &orig_start,
7781                                      &orig_block_len, &ram_bytes) == 1 &&
7782                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7783                         struct extent_map *em2;
7784
7785                         em2 = btrfs_create_dio_extent(inode, start, len,
7786                                                       orig_start, block_start,
7787                                                       len, orig_block_len,
7788                                                       ram_bytes, type);
7789                         btrfs_dec_nocow_writers(fs_info, block_start);
7790                         if (type == BTRFS_ORDERED_PREALLOC) {
7791                                 free_extent_map(em);
7792                                 em = em2;
7793                         }
7794                         if (em2 && IS_ERR(em2)) {
7795                                 ret = PTR_ERR(em2);
7796                                 goto unlock_err;
7797                         }
7798                         /*
7799                          * For inode marked NODATACOW or extent marked PREALLOC,
7800                          * use the existing or preallocated extent, so does not
7801                          * need to adjust btrfs_space_info's bytes_may_use.
7802                          */
7803                         btrfs_free_reserved_data_space_noquota(inode,
7804                                         start, len);
7805                         goto unlock;
7806                 }
7807         }
7808
7809         /*
7810          * this will cow the extent, reset the len in case we changed
7811          * it above
7812          */
7813         len = bh_result->b_size;
7814         free_extent_map(em);
7815         em = btrfs_new_extent_direct(inode, start, len);
7816         if (IS_ERR(em)) {
7817                 ret = PTR_ERR(em);
7818                 goto unlock_err;
7819         }
7820         len = min(len, em->len - (start - em->start));
7821 unlock:
7822         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7823                 inode->i_blkbits;
7824         bh_result->b_size = len;
7825         bh_result->b_bdev = em->bdev;
7826         set_buffer_mapped(bh_result);
7827         if (create) {
7828                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7829                         set_buffer_new(bh_result);
7830
7831                 /*
7832                  * Need to update the i_size under the extent lock so buffered
7833                  * readers will get the updated i_size when we unlock.
7834                  */
7835                 if (!dio_data->overwrite && start + len > i_size_read(inode))
7836                         i_size_write(inode, start + len);
7837
7838                 WARN_ON(dio_data->reserve < len);
7839                 dio_data->reserve -= len;
7840                 dio_data->unsubmitted_oe_range_end = start + len;
7841                 current->journal_info = dio_data;
7842         }
7843
7844         /*
7845          * In the case of write we need to clear and unlock the entire range,
7846          * in the case of read we need to unlock only the end area that we
7847          * aren't using if there is any left over space.
7848          */
7849         if (lockstart < lockend) {
7850                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7851                                  lockend, unlock_bits, 1, 0,
7852                                  &cached_state);
7853         } else {
7854                 free_extent_state(cached_state);
7855         }
7856
7857         free_extent_map(em);
7858
7859         return 0;
7860
7861 unlock_err:
7862         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7863                          unlock_bits, 1, 0, &cached_state);
7864 err:
7865         if (dio_data)
7866                 current->journal_info = dio_data;
7867         return ret;
7868 }
7869
7870 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7871                                                  struct bio *bio,
7872                                                  int mirror_num)
7873 {
7874         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7875         blk_status_t ret;
7876
7877         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7878
7879         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7880         if (ret)
7881                 return ret;
7882
7883         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7884
7885         return ret;
7886 }
7887
7888 static int btrfs_check_dio_repairable(struct inode *inode,
7889                                       struct bio *failed_bio,
7890                                       struct io_failure_record *failrec,
7891                                       int failed_mirror)
7892 {
7893         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7894         int num_copies;
7895
7896         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7897         if (num_copies == 1) {
7898                 /*
7899                  * we only have a single copy of the data, so don't bother with
7900                  * all the retry and error correction code that follows. no
7901                  * matter what the error is, it is very likely to persist.
7902                  */
7903                 btrfs_debug(fs_info,
7904                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7905                         num_copies, failrec->this_mirror, failed_mirror);
7906                 return 0;
7907         }
7908
7909         failrec->failed_mirror = failed_mirror;
7910         failrec->this_mirror++;
7911         if (failrec->this_mirror == failed_mirror)
7912                 failrec->this_mirror++;
7913
7914         if (failrec->this_mirror > num_copies) {
7915                 btrfs_debug(fs_info,
7916                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7917                         num_copies, failrec->this_mirror, failed_mirror);
7918                 return 0;
7919         }
7920
7921         return 1;
7922 }
7923
7924 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7925                                    struct page *page, unsigned int pgoff,
7926                                    u64 start, u64 end, int failed_mirror,
7927                                    bio_end_io_t *repair_endio, void *repair_arg)
7928 {
7929         struct io_failure_record *failrec;
7930         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7931         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7932         struct bio *bio;
7933         int isector;
7934         unsigned int read_mode = 0;
7935         int segs;
7936         int ret;
7937         blk_status_t status;
7938         struct bio_vec bvec;
7939
7940         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7941
7942         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7943         if (ret)
7944                 return errno_to_blk_status(ret);
7945
7946         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7947                                          failed_mirror);
7948         if (!ret) {
7949                 free_io_failure(failure_tree, io_tree, failrec);
7950                 return BLK_STS_IOERR;
7951         }
7952
7953         segs = bio_segments(failed_bio);
7954         bio_get_first_bvec(failed_bio, &bvec);
7955         if (segs > 1 ||
7956             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7957                 read_mode |= REQ_FAILFAST_DEV;
7958
7959         isector = start - btrfs_io_bio(failed_bio)->logical;
7960         isector >>= inode->i_sb->s_blocksize_bits;
7961         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7962                                 pgoff, isector, repair_endio, repair_arg);
7963         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
7964
7965         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7966                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7967                     read_mode, failrec->this_mirror, failrec->in_validation);
7968
7969         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7970         if (status) {
7971                 free_io_failure(failure_tree, io_tree, failrec);
7972                 bio_put(bio);
7973         }
7974
7975         return status;
7976 }
7977
7978 struct btrfs_retry_complete {
7979         struct completion done;
7980         struct inode *inode;
7981         u64 start;
7982         int uptodate;
7983 };
7984
7985 static void btrfs_retry_endio_nocsum(struct bio *bio)
7986 {
7987         struct btrfs_retry_complete *done = bio->bi_private;
7988         struct inode *inode = done->inode;
7989         struct bio_vec *bvec;
7990         struct extent_io_tree *io_tree, *failure_tree;
7991         int i;
7992
7993         if (bio->bi_status)
7994                 goto end;
7995
7996         ASSERT(bio->bi_vcnt == 1);
7997         io_tree = &BTRFS_I(inode)->io_tree;
7998         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7999         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
8000
8001         done->uptodate = 1;
8002         ASSERT(!bio_flagged(bio, BIO_CLONED));
8003         bio_for_each_segment_all(bvec, bio, i)
8004                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
8005                                  io_tree, done->start, bvec->bv_page,
8006                                  btrfs_ino(BTRFS_I(inode)), 0);
8007 end:
8008         complete(&done->done);
8009         bio_put(bio);
8010 }
8011
8012 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
8013                                                 struct btrfs_io_bio *io_bio)
8014 {
8015         struct btrfs_fs_info *fs_info;
8016         struct bio_vec bvec;
8017         struct bvec_iter iter;
8018         struct btrfs_retry_complete done;
8019         u64 start;
8020         unsigned int pgoff;
8021         u32 sectorsize;
8022         int nr_sectors;
8023         blk_status_t ret;
8024         blk_status_t err = BLK_STS_OK;
8025
8026         fs_info = BTRFS_I(inode)->root->fs_info;
8027         sectorsize = fs_info->sectorsize;
8028
8029         start = io_bio->logical;
8030         done.inode = inode;
8031         io_bio->bio.bi_iter = io_bio->iter;
8032
8033         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8034                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8035                 pgoff = bvec.bv_offset;
8036
8037 next_block_or_try_again:
8038                 done.uptodate = 0;
8039                 done.start = start;
8040                 init_completion(&done.done);
8041
8042                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8043                                 pgoff, start, start + sectorsize - 1,
8044                                 io_bio->mirror_num,
8045                                 btrfs_retry_endio_nocsum, &done);
8046                 if (ret) {
8047                         err = ret;
8048                         goto next;
8049                 }
8050
8051                 wait_for_completion_io(&done.done);
8052
8053                 if (!done.uptodate) {
8054                         /* We might have another mirror, so try again */
8055                         goto next_block_or_try_again;
8056                 }
8057
8058 next:
8059                 start += sectorsize;
8060
8061                 nr_sectors--;
8062                 if (nr_sectors) {
8063                         pgoff += sectorsize;
8064                         ASSERT(pgoff < PAGE_SIZE);
8065                         goto next_block_or_try_again;
8066                 }
8067         }
8068
8069         return err;
8070 }
8071
8072 static void btrfs_retry_endio(struct bio *bio)
8073 {
8074         struct btrfs_retry_complete *done = bio->bi_private;
8075         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8076         struct extent_io_tree *io_tree, *failure_tree;
8077         struct inode *inode = done->inode;
8078         struct bio_vec *bvec;
8079         int uptodate;
8080         int ret;
8081         int i;
8082
8083         if (bio->bi_status)
8084                 goto end;
8085
8086         uptodate = 1;
8087
8088         ASSERT(bio->bi_vcnt == 1);
8089         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
8090
8091         io_tree = &BTRFS_I(inode)->io_tree;
8092         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8093
8094         ASSERT(!bio_flagged(bio, BIO_CLONED));
8095         bio_for_each_segment_all(bvec, bio, i) {
8096                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8097                                              bvec->bv_offset, done->start,
8098                                              bvec->bv_len);
8099                 if (!ret)
8100                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8101                                          failure_tree, io_tree, done->start,
8102                                          bvec->bv_page,
8103                                          btrfs_ino(BTRFS_I(inode)),
8104                                          bvec->bv_offset);
8105                 else
8106                         uptodate = 0;
8107         }
8108
8109         done->uptodate = uptodate;
8110 end:
8111         complete(&done->done);
8112         bio_put(bio);
8113 }
8114
8115 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8116                 struct btrfs_io_bio *io_bio, blk_status_t err)
8117 {
8118         struct btrfs_fs_info *fs_info;
8119         struct bio_vec bvec;
8120         struct bvec_iter iter;
8121         struct btrfs_retry_complete done;
8122         u64 start;
8123         u64 offset = 0;
8124         u32 sectorsize;
8125         int nr_sectors;
8126         unsigned int pgoff;
8127         int csum_pos;
8128         bool uptodate = (err == 0);
8129         int ret;
8130         blk_status_t status;
8131
8132         fs_info = BTRFS_I(inode)->root->fs_info;
8133         sectorsize = fs_info->sectorsize;
8134
8135         err = BLK_STS_OK;
8136         start = io_bio->logical;
8137         done.inode = inode;
8138         io_bio->bio.bi_iter = io_bio->iter;
8139
8140         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8141                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8142
8143                 pgoff = bvec.bv_offset;
8144 next_block:
8145                 if (uptodate) {
8146                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8147                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8148                                         bvec.bv_page, pgoff, start, sectorsize);
8149                         if (likely(!ret))
8150                                 goto next;
8151                 }
8152 try_again:
8153                 done.uptodate = 0;
8154                 done.start = start;
8155                 init_completion(&done.done);
8156
8157                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8158                                         pgoff, start, start + sectorsize - 1,
8159                                         io_bio->mirror_num, btrfs_retry_endio,
8160                                         &done);
8161                 if (status) {
8162                         err = status;
8163                         goto next;
8164                 }
8165
8166                 wait_for_completion_io(&done.done);
8167
8168                 if (!done.uptodate) {
8169                         /* We might have another mirror, so try again */
8170                         goto try_again;
8171                 }
8172 next:
8173                 offset += sectorsize;
8174                 start += sectorsize;
8175
8176                 ASSERT(nr_sectors);
8177
8178                 nr_sectors--;
8179                 if (nr_sectors) {
8180                         pgoff += sectorsize;
8181                         ASSERT(pgoff < PAGE_SIZE);
8182                         goto next_block;
8183                 }
8184         }
8185
8186         return err;
8187 }
8188
8189 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8190                 struct btrfs_io_bio *io_bio, blk_status_t err)
8191 {
8192         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8193
8194         if (skip_csum) {
8195                 if (unlikely(err))
8196                         return __btrfs_correct_data_nocsum(inode, io_bio);
8197                 else
8198                         return BLK_STS_OK;
8199         } else {
8200                 return __btrfs_subio_endio_read(inode, io_bio, err);
8201         }
8202 }
8203
8204 static void btrfs_endio_direct_read(struct bio *bio)
8205 {
8206         struct btrfs_dio_private *dip = bio->bi_private;
8207         struct inode *inode = dip->inode;
8208         struct bio *dio_bio;
8209         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8210         blk_status_t err = bio->bi_status;
8211
8212         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8213                 err = btrfs_subio_endio_read(inode, io_bio, err);
8214
8215         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8216                       dip->logical_offset + dip->bytes - 1);
8217         dio_bio = dip->dio_bio;
8218
8219         kfree(dip);
8220
8221         dio_bio->bi_status = err;
8222         dio_end_io(dio_bio);
8223
8224         if (io_bio->end_io)
8225                 io_bio->end_io(io_bio, blk_status_to_errno(err));
8226         bio_put(bio);
8227 }
8228
8229 static void __endio_write_update_ordered(struct inode *inode,
8230                                          const u64 offset, const u64 bytes,
8231                                          const bool uptodate)
8232 {
8233         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8234         struct btrfs_ordered_extent *ordered = NULL;
8235         struct btrfs_workqueue *wq;
8236         btrfs_work_func_t func;
8237         u64 ordered_offset = offset;
8238         u64 ordered_bytes = bytes;
8239         u64 last_offset;
8240         int ret;
8241
8242         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8243                 wq = fs_info->endio_freespace_worker;
8244                 func = btrfs_freespace_write_helper;
8245         } else {
8246                 wq = fs_info->endio_write_workers;
8247                 func = btrfs_endio_write_helper;
8248         }
8249
8250 again:
8251         last_offset = ordered_offset;
8252         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8253                                                    &ordered_offset,
8254                                                    ordered_bytes,
8255                                                    uptodate);
8256         if (!ret)
8257                 goto out_test;
8258
8259         btrfs_init_work(&ordered->work, func, finish_ordered_fn, NULL, NULL);
8260         btrfs_queue_work(wq, &ordered->work);
8261 out_test:
8262         /*
8263          * If btrfs_dec_test_ordered_pending does not find any ordered extent
8264          * in the range, we can exit.
8265          */
8266         if (ordered_offset == last_offset)
8267                 return;
8268         /*
8269          * our bio might span multiple ordered extents.  If we haven't
8270          * completed the accounting for the whole dio, go back and try again
8271          */
8272         if (ordered_offset < offset + bytes) {
8273                 ordered_bytes = offset + bytes - ordered_offset;
8274                 ordered = NULL;
8275                 goto again;
8276         }
8277 }
8278
8279 static void btrfs_endio_direct_write(struct bio *bio)
8280 {
8281         struct btrfs_dio_private *dip = bio->bi_private;
8282         struct bio *dio_bio = dip->dio_bio;
8283
8284         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8285                                      dip->bytes, !bio->bi_status);
8286
8287         kfree(dip);
8288
8289         dio_bio->bi_status = bio->bi_status;
8290         dio_end_io(dio_bio);
8291         bio_put(bio);
8292 }
8293
8294 static blk_status_t __btrfs_submit_bio_start_direct_io(void *private_data,
8295                                     struct bio *bio, int mirror_num,
8296                                     unsigned long bio_flags, u64 offset)
8297 {
8298         struct inode *inode = private_data;
8299         blk_status_t ret;
8300         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8301         BUG_ON(ret); /* -ENOMEM */
8302         return 0;
8303 }
8304
8305 static void btrfs_end_dio_bio(struct bio *bio)
8306 {
8307         struct btrfs_dio_private *dip = bio->bi_private;
8308         blk_status_t err = bio->bi_status;
8309
8310         if (err)
8311                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8312                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8313                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8314                            bio->bi_opf,
8315                            (unsigned long long)bio->bi_iter.bi_sector,
8316                            bio->bi_iter.bi_size, err);
8317
8318         if (dip->subio_endio)
8319                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8320
8321         if (err) {
8322                 /*
8323                  * We want to perceive the errors flag being set before
8324                  * decrementing the reference count. We don't need a barrier
8325                  * since atomic operations with a return value are fully
8326                  * ordered as per atomic_t.txt
8327                  */
8328                 dip->errors = 1;
8329         }
8330
8331         /* if there are more bios still pending for this dio, just exit */
8332         if (!atomic_dec_and_test(&dip->pending_bios))
8333                 goto out;
8334
8335         if (dip->errors) {
8336                 bio_io_error(dip->orig_bio);
8337         } else {
8338                 dip->dio_bio->bi_status = BLK_STS_OK;
8339                 bio_endio(dip->orig_bio);
8340         }
8341 out:
8342         bio_put(bio);
8343 }
8344
8345 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8346                                                  struct btrfs_dio_private *dip,
8347                                                  struct bio *bio,
8348                                                  u64 file_offset)
8349 {
8350         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8351         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8352         blk_status_t ret;
8353
8354         /*
8355          * We load all the csum data we need when we submit
8356          * the first bio to reduce the csum tree search and
8357          * contention.
8358          */
8359         if (dip->logical_offset == file_offset) {
8360                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8361                                                 file_offset);
8362                 if (ret)
8363                         return ret;
8364         }
8365
8366         if (bio == dip->orig_bio)
8367                 return 0;
8368
8369         file_offset -= dip->logical_offset;
8370         file_offset >>= inode->i_sb->s_blocksize_bits;
8371         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8372
8373         return 0;
8374 }
8375
8376 static inline blk_status_t
8377 __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode, u64 file_offset,
8378                        int async_submit)
8379 {
8380         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8381         struct btrfs_dio_private *dip = bio->bi_private;
8382         bool write = bio_op(bio) == REQ_OP_WRITE;
8383         blk_status_t ret;
8384
8385         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8386         if (async_submit)
8387                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8388
8389         if (!write) {
8390                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8391                 if (ret)
8392                         goto err;
8393         }
8394
8395         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8396                 goto map;
8397
8398         if (write && async_submit) {
8399                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8400                                           file_offset, inode,
8401                                           __btrfs_submit_bio_start_direct_io,
8402                                           __btrfs_submit_bio_done);
8403                 goto err;
8404         } else if (write) {
8405                 /*
8406                  * If we aren't doing async submit, calculate the csum of the
8407                  * bio now.
8408                  */
8409                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8410                 if (ret)
8411                         goto err;
8412         } else {
8413                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8414                                                      file_offset);
8415                 if (ret)
8416                         goto err;
8417         }
8418 map:
8419         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8420 err:
8421         return ret;
8422 }
8423
8424 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8425 {
8426         struct inode *inode = dip->inode;
8427         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8428         struct bio *bio;
8429         struct bio *orig_bio = dip->orig_bio;
8430         u64 start_sector = orig_bio->bi_iter.bi_sector;
8431         u64 file_offset = dip->logical_offset;
8432         u64 map_length;
8433         int async_submit = 0;
8434         u64 submit_len;
8435         int clone_offset = 0;
8436         int clone_len;
8437         int ret;
8438         blk_status_t status;
8439
8440         map_length = orig_bio->bi_iter.bi_size;
8441         submit_len = map_length;
8442         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8443                               &map_length, NULL, 0);
8444         if (ret)
8445                 return -EIO;
8446
8447         if (map_length >= submit_len) {
8448                 bio = orig_bio;
8449                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8450                 goto submit;
8451         }
8452
8453         /* async crcs make it difficult to collect full stripe writes. */
8454         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8455                 async_submit = 0;
8456         else
8457                 async_submit = 1;
8458
8459         /* bio split */
8460         ASSERT(map_length <= INT_MAX);
8461         atomic_inc(&dip->pending_bios);
8462         do {
8463                 clone_len = min_t(int, submit_len, map_length);
8464
8465                 /*
8466                  * This will never fail as it's passing GPF_NOFS and
8467                  * the allocation is backed by btrfs_bioset.
8468                  */
8469                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8470                                               clone_len);
8471                 bio->bi_private = dip;
8472                 bio->bi_end_io = btrfs_end_dio_bio;
8473                 btrfs_io_bio(bio)->logical = file_offset;
8474
8475                 ASSERT(submit_len >= clone_len);
8476                 submit_len -= clone_len;
8477                 if (submit_len == 0)
8478                         break;
8479
8480                 /*
8481                  * Increase the count before we submit the bio so we know
8482                  * the end IO handler won't happen before we increase the
8483                  * count. Otherwise, the dip might get freed before we're
8484                  * done setting it up.
8485                  */
8486                 atomic_inc(&dip->pending_bios);
8487
8488                 status = __btrfs_submit_dio_bio(bio, inode, file_offset,
8489                                                 async_submit);
8490                 if (status) {
8491                         bio_put(bio);
8492                         atomic_dec(&dip->pending_bios);
8493                         goto out_err;
8494                 }
8495
8496                 clone_offset += clone_len;
8497                 start_sector += clone_len >> 9;
8498                 file_offset += clone_len;
8499
8500                 map_length = submit_len;
8501                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8502                                       start_sector << 9, &map_length, NULL, 0);
8503                 if (ret)
8504                         goto out_err;
8505         } while (submit_len > 0);
8506
8507 submit:
8508         status = __btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8509         if (!status)
8510                 return 0;
8511
8512         bio_put(bio);
8513 out_err:
8514         dip->errors = 1;
8515         /*
8516          * Before atomic variable goto zero, we must  make sure dip->errors is
8517          * perceived to be set. This ordering is ensured by the fact that an
8518          * atomic operations with a return value are fully ordered as per
8519          * atomic_t.txt
8520          */
8521         if (atomic_dec_and_test(&dip->pending_bios))
8522                 bio_io_error(dip->orig_bio);
8523
8524         /* bio_end_io() will handle error, so we needn't return it */
8525         return 0;
8526 }
8527
8528 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8529                                 loff_t file_offset)
8530 {
8531         struct btrfs_dio_private *dip = NULL;
8532         struct bio *bio = NULL;
8533         struct btrfs_io_bio *io_bio;
8534         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8535         int ret = 0;
8536
8537         bio = btrfs_bio_clone(dio_bio);
8538
8539         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8540         if (!dip) {
8541                 ret = -ENOMEM;
8542                 goto free_ordered;
8543         }
8544
8545         dip->private = dio_bio->bi_private;
8546         dip->inode = inode;
8547         dip->logical_offset = file_offset;
8548         dip->bytes = dio_bio->bi_iter.bi_size;
8549         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8550         bio->bi_private = dip;
8551         dip->orig_bio = bio;
8552         dip->dio_bio = dio_bio;
8553         atomic_set(&dip->pending_bios, 0);
8554         io_bio = btrfs_io_bio(bio);
8555         io_bio->logical = file_offset;
8556
8557         if (write) {
8558                 bio->bi_end_io = btrfs_endio_direct_write;
8559         } else {
8560                 bio->bi_end_io = btrfs_endio_direct_read;
8561                 dip->subio_endio = btrfs_subio_endio_read;
8562         }
8563
8564         /*
8565          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8566          * even if we fail to submit a bio, because in such case we do the
8567          * corresponding error handling below and it must not be done a second
8568          * time by btrfs_direct_IO().
8569          */
8570         if (write) {
8571                 struct btrfs_dio_data *dio_data = current->journal_info;
8572
8573                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8574                         dip->bytes;
8575                 dio_data->unsubmitted_oe_range_start =
8576                         dio_data->unsubmitted_oe_range_end;
8577         }
8578
8579         ret = btrfs_submit_direct_hook(dip);
8580         if (!ret)
8581                 return;
8582
8583         if (io_bio->end_io)
8584                 io_bio->end_io(io_bio, ret);
8585
8586 free_ordered:
8587         /*
8588          * If we arrived here it means either we failed to submit the dip
8589          * or we either failed to clone the dio_bio or failed to allocate the
8590          * dip. If we cloned the dio_bio and allocated the dip, we can just
8591          * call bio_endio against our io_bio so that we get proper resource
8592          * cleanup if we fail to submit the dip, otherwise, we must do the
8593          * same as btrfs_endio_direct_[write|read] because we can't call these
8594          * callbacks - they require an allocated dip and a clone of dio_bio.
8595          */
8596         if (bio && dip) {
8597                 bio_io_error(bio);
8598                 /*
8599                  * The end io callbacks free our dip, do the final put on bio
8600                  * and all the cleanup and final put for dio_bio (through
8601                  * dio_end_io()).
8602                  */
8603                 dip = NULL;
8604                 bio = NULL;
8605         } else {
8606                 if (write)
8607                         __endio_write_update_ordered(inode,
8608                                                 file_offset,
8609                                                 dio_bio->bi_iter.bi_size,
8610                                                 false);
8611                 else
8612                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8613                               file_offset + dio_bio->bi_iter.bi_size - 1);
8614
8615                 dio_bio->bi_status = BLK_STS_IOERR;
8616                 /*
8617                  * Releases and cleans up our dio_bio, no need to bio_put()
8618                  * nor bio_endio()/bio_io_error() against dio_bio.
8619                  */
8620                 dio_end_io(dio_bio);
8621         }
8622         if (bio)
8623                 bio_put(bio);
8624         kfree(dip);
8625 }
8626
8627 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8628                                const struct iov_iter *iter, loff_t offset)
8629 {
8630         int seg;
8631         int i;
8632         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8633         ssize_t retval = -EINVAL;
8634
8635         if (offset & blocksize_mask)
8636                 goto out;
8637
8638         if (iov_iter_alignment(iter) & blocksize_mask)
8639                 goto out;
8640
8641         /* If this is a write we don't need to check anymore */
8642         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8643                 return 0;
8644         /*
8645          * Check to make sure we don't have duplicate iov_base's in this
8646          * iovec, if so return EINVAL, otherwise we'll get csum errors
8647          * when reading back.
8648          */
8649         for (seg = 0; seg < iter->nr_segs; seg++) {
8650                 for (i = seg + 1; i < iter->nr_segs; i++) {
8651                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8652                                 goto out;
8653                 }
8654         }
8655         retval = 0;
8656 out:
8657         return retval;
8658 }
8659
8660 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8661 {
8662         struct file *file = iocb->ki_filp;
8663         struct inode *inode = file->f_mapping->host;
8664         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8665         struct btrfs_dio_data dio_data = { 0 };
8666         struct extent_changeset *data_reserved = NULL;
8667         loff_t offset = iocb->ki_pos;
8668         size_t count = 0;
8669         int flags = 0;
8670         bool wakeup = true;
8671         bool relock = false;
8672         ssize_t ret;
8673
8674         if (check_direct_IO(fs_info, iter, offset))
8675                 return 0;
8676
8677         inode_dio_begin(inode);
8678
8679         /*
8680          * The generic stuff only does filemap_write_and_wait_range, which
8681          * isn't enough if we've written compressed pages to this area, so
8682          * we need to flush the dirty pages again to make absolutely sure
8683          * that any outstanding dirty pages are on disk.
8684          */
8685         count = iov_iter_count(iter);
8686         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8687                      &BTRFS_I(inode)->runtime_flags))
8688                 filemap_fdatawrite_range(inode->i_mapping, offset,
8689                                          offset + count - 1);
8690
8691         if (iov_iter_rw(iter) == WRITE) {
8692                 /*
8693                  * If the write DIO is beyond the EOF, we need update
8694                  * the isize, but it is protected by i_mutex. So we can
8695                  * not unlock the i_mutex at this case.
8696                  */
8697                 if (offset + count <= inode->i_size) {
8698                         dio_data.overwrite = 1;
8699                         inode_unlock(inode);
8700                         relock = true;
8701                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8702                         ret = -EAGAIN;
8703                         goto out;
8704                 }
8705                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8706                                                    offset, count);
8707                 if (ret)
8708                         goto out;
8709
8710                 /*
8711                  * We need to know how many extents we reserved so that we can
8712                  * do the accounting properly if we go over the number we
8713                  * originally calculated.  Abuse current->journal_info for this.
8714                  */
8715                 dio_data.reserve = round_up(count,
8716                                             fs_info->sectorsize);
8717                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8718                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8719                 current->journal_info = &dio_data;
8720                 down_read(&BTRFS_I(inode)->dio_sem);
8721         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8722                                      &BTRFS_I(inode)->runtime_flags)) {
8723                 inode_dio_end(inode);
8724                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8725                 wakeup = false;
8726         }
8727
8728         ret = __blockdev_direct_IO(iocb, inode,
8729                                    fs_info->fs_devices->latest_bdev,
8730                                    iter, btrfs_get_blocks_direct, NULL,
8731                                    btrfs_submit_direct, flags);
8732         if (iov_iter_rw(iter) == WRITE) {
8733                 up_read(&BTRFS_I(inode)->dio_sem);
8734                 current->journal_info = NULL;
8735                 if (ret < 0 && ret != -EIOCBQUEUED) {
8736                         if (dio_data.reserve)
8737                                 btrfs_delalloc_release_space(inode, data_reserved,
8738                                         offset, dio_data.reserve);
8739                         /*
8740                          * On error we might have left some ordered extents
8741                          * without submitting corresponding bios for them, so
8742                          * cleanup them up to avoid other tasks getting them
8743                          * and waiting for them to complete forever.
8744                          */
8745                         if (dio_data.unsubmitted_oe_range_start <
8746                             dio_data.unsubmitted_oe_range_end)
8747                                 __endio_write_update_ordered(inode,
8748                                         dio_data.unsubmitted_oe_range_start,
8749                                         dio_data.unsubmitted_oe_range_end -
8750                                         dio_data.unsubmitted_oe_range_start,
8751                                         false);
8752                 } else if (ret >= 0 && (size_t)ret < count)
8753                         btrfs_delalloc_release_space(inode, data_reserved,
8754                                         offset, count - (size_t)ret);
8755                 btrfs_delalloc_release_extents(BTRFS_I(inode), count);
8756         }
8757 out:
8758         if (wakeup)
8759                 inode_dio_end(inode);
8760         if (relock)
8761                 inode_lock(inode);
8762
8763         extent_changeset_free(data_reserved);
8764         return ret;
8765 }
8766
8767 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8768
8769 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8770                 __u64 start, __u64 len)
8771 {
8772         int     ret;
8773
8774         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8775         if (ret)
8776                 return ret;
8777
8778         return extent_fiemap(inode, fieinfo, start, len);
8779 }
8780
8781 int btrfs_readpage(struct file *file, struct page *page)
8782 {
8783         struct extent_io_tree *tree;
8784         tree = &BTRFS_I(page->mapping->host)->io_tree;
8785         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8786 }
8787
8788 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8789 {
8790         struct inode *inode = page->mapping->host;
8791         int ret;
8792
8793         if (current->flags & PF_MEMALLOC) {
8794                 redirty_page_for_writepage(wbc, page);
8795                 unlock_page(page);
8796                 return 0;
8797         }
8798
8799         /*
8800          * If we are under memory pressure we will call this directly from the
8801          * VM, we need to make sure we have the inode referenced for the ordered
8802          * extent.  If not just return like we didn't do anything.
8803          */
8804         if (!igrab(inode)) {
8805                 redirty_page_for_writepage(wbc, page);
8806                 return AOP_WRITEPAGE_ACTIVATE;
8807         }
8808         ret = extent_write_full_page(page, wbc);
8809         btrfs_add_delayed_iput(inode);
8810         return ret;
8811 }
8812
8813 static int btrfs_writepages(struct address_space *mapping,
8814                             struct writeback_control *wbc)
8815 {
8816         struct extent_io_tree *tree;
8817
8818         tree = &BTRFS_I(mapping->host)->io_tree;
8819         return extent_writepages(tree, mapping, wbc);
8820 }
8821
8822 static int
8823 btrfs_readpages(struct file *file, struct address_space *mapping,
8824                 struct list_head *pages, unsigned nr_pages)
8825 {
8826         struct extent_io_tree *tree;
8827         tree = &BTRFS_I(mapping->host)->io_tree;
8828         return extent_readpages(tree, mapping, pages, nr_pages);
8829 }
8830 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8831 {
8832         struct extent_io_tree *tree;
8833         struct extent_map_tree *map;
8834         int ret;
8835
8836         tree = &BTRFS_I(page->mapping->host)->io_tree;
8837         map = &BTRFS_I(page->mapping->host)->extent_tree;
8838         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8839         if (ret == 1) {
8840                 ClearPagePrivate(page);
8841                 set_page_private(page, 0);
8842                 put_page(page);
8843         }
8844         return ret;
8845 }
8846
8847 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8848 {
8849         if (PageWriteback(page) || PageDirty(page))
8850                 return 0;
8851         return __btrfs_releasepage(page, gfp_flags);
8852 }
8853
8854 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8855                                  unsigned int length)
8856 {
8857         struct inode *inode = page->mapping->host;
8858         struct extent_io_tree *tree;
8859         struct btrfs_ordered_extent *ordered;
8860         struct extent_state *cached_state = NULL;
8861         u64 page_start = page_offset(page);
8862         u64 page_end = page_start + PAGE_SIZE - 1;
8863         u64 start;
8864         u64 end;
8865         int inode_evicting = inode->i_state & I_FREEING;
8866
8867         /*
8868          * we have the page locked, so new writeback can't start,
8869          * and the dirty bit won't be cleared while we are here.
8870          *
8871          * Wait for IO on this page so that we can safely clear
8872          * the PagePrivate2 bit and do ordered accounting
8873          */
8874         wait_on_page_writeback(page);
8875
8876         tree = &BTRFS_I(inode)->io_tree;
8877         if (offset) {
8878                 btrfs_releasepage(page, GFP_NOFS);
8879                 return;
8880         }
8881
8882         if (!inode_evicting)
8883                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8884 again:
8885         start = page_start;
8886         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8887                                         page_end - start + 1);
8888         if (ordered) {
8889                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8890                 /*
8891                  * IO on this page will never be started, so we need
8892                  * to account for any ordered extents now
8893                  */
8894                 if (!inode_evicting)
8895                         clear_extent_bit(tree, start, end,
8896                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8897                                          EXTENT_DELALLOC_NEW |
8898                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8899                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8900                 /*
8901                  * whoever cleared the private bit is responsible
8902                  * for the finish_ordered_io
8903                  */
8904                 if (TestClearPagePrivate2(page)) {
8905                         struct btrfs_ordered_inode_tree *tree;
8906                         u64 new_len;
8907
8908                         tree = &BTRFS_I(inode)->ordered_tree;
8909
8910                         spin_lock_irq(&tree->lock);
8911                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8912                         new_len = start - ordered->file_offset;
8913                         if (new_len < ordered->truncated_len)
8914                                 ordered->truncated_len = new_len;
8915                         spin_unlock_irq(&tree->lock);
8916
8917                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8918                                                            start,
8919                                                            end - start + 1, 1))
8920                                 btrfs_finish_ordered_io(ordered);
8921                 }
8922                 btrfs_put_ordered_extent(ordered);
8923                 if (!inode_evicting) {
8924                         cached_state = NULL;
8925                         lock_extent_bits(tree, start, end,
8926                                          &cached_state);
8927                 }
8928
8929                 start = end + 1;
8930                 if (start < page_end)
8931                         goto again;
8932         }
8933
8934         /*
8935          * Qgroup reserved space handler
8936          * Page here will be either
8937          * 1) Already written to disk
8938          *    In this case, its reserved space is released from data rsv map
8939          *    and will be freed by delayed_ref handler finally.
8940          *    So even we call qgroup_free_data(), it won't decrease reserved
8941          *    space.
8942          * 2) Not written to disk
8943          *    This means the reserved space should be freed here. However,
8944          *    if a truncate invalidates the page (by clearing PageDirty)
8945          *    and the page is accounted for while allocating extent
8946          *    in btrfs_check_data_free_space() we let delayed_ref to
8947          *    free the entire extent.
8948          */
8949         if (PageDirty(page))
8950                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8951         if (!inode_evicting) {
8952                 clear_extent_bit(tree, page_start, page_end,
8953                                  EXTENT_LOCKED | EXTENT_DIRTY |
8954                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8955                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8956                                  &cached_state);
8957
8958                 __btrfs_releasepage(page, GFP_NOFS);
8959         }
8960
8961         ClearPageChecked(page);
8962         if (PagePrivate(page)) {
8963                 ClearPagePrivate(page);
8964                 set_page_private(page, 0);
8965                 put_page(page);
8966         }
8967 }
8968
8969 /*
8970  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8971  * called from a page fault handler when a page is first dirtied. Hence we must
8972  * be careful to check for EOF conditions here. We set the page up correctly
8973  * for a written page which means we get ENOSPC checking when writing into
8974  * holes and correct delalloc and unwritten extent mapping on filesystems that
8975  * support these features.
8976  *
8977  * We are not allowed to take the i_mutex here so we have to play games to
8978  * protect against truncate races as the page could now be beyond EOF.  Because
8979  * vmtruncate() writes the inode size before removing pages, once we have the
8980  * page lock we can determine safely if the page is beyond EOF. If it is not
8981  * beyond EOF, then the page is guaranteed safe against truncation until we
8982  * unlock the page.
8983  */
8984 int btrfs_page_mkwrite(struct vm_fault *vmf)
8985 {
8986         struct page *page = vmf->page;
8987         struct inode *inode = file_inode(vmf->vma->vm_file);
8988         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8989         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8990         struct btrfs_ordered_extent *ordered;
8991         struct extent_state *cached_state = NULL;
8992         struct extent_changeset *data_reserved = NULL;
8993         char *kaddr;
8994         unsigned long zero_start;
8995         loff_t size;
8996         int ret;
8997         int reserved = 0;
8998         u64 reserved_space;
8999         u64 page_start;
9000         u64 page_end;
9001         u64 end;
9002
9003         reserved_space = PAGE_SIZE;
9004
9005         sb_start_pagefault(inode->i_sb);
9006         page_start = page_offset(page);
9007         page_end = page_start + PAGE_SIZE - 1;
9008         end = page_end;
9009
9010         /*
9011          * Reserving delalloc space after obtaining the page lock can lead to
9012          * deadlock. For example, if a dirty page is locked by this function
9013          * and the call to btrfs_delalloc_reserve_space() ends up triggering
9014          * dirty page write out, then the btrfs_writepage() function could
9015          * end up waiting indefinitely to get a lock on the page currently
9016          * being processed by btrfs_page_mkwrite() function.
9017          */
9018         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
9019                                            reserved_space);
9020         if (!ret) {
9021                 ret = file_update_time(vmf->vma->vm_file);
9022                 reserved = 1;
9023         }
9024         if (ret) {
9025                 if (ret == -ENOMEM)
9026                         ret = VM_FAULT_OOM;
9027                 else /* -ENOSPC, -EIO, etc */
9028                         ret = VM_FAULT_SIGBUS;
9029                 if (reserved)
9030                         goto out;
9031                 goto out_noreserve;
9032         }
9033
9034         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9035 again:
9036         lock_page(page);
9037         size = i_size_read(inode);
9038
9039         if ((page->mapping != inode->i_mapping) ||
9040             (page_start >= size)) {
9041                 /* page got truncated out from underneath us */
9042                 goto out_unlock;
9043         }
9044         wait_on_page_writeback(page);
9045
9046         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9047         set_page_extent_mapped(page);
9048
9049         /*
9050          * we can't set the delalloc bits if there are pending ordered
9051          * extents.  Drop our locks and wait for them to finish
9052          */
9053         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
9054                         PAGE_SIZE);
9055         if (ordered) {
9056                 unlock_extent_cached(io_tree, page_start, page_end,
9057                                      &cached_state);
9058                 unlock_page(page);
9059                 btrfs_start_ordered_extent(inode, ordered, 1);
9060                 btrfs_put_ordered_extent(ordered);
9061                 goto again;
9062         }
9063
9064         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9065                 reserved_space = round_up(size - page_start,
9066                                           fs_info->sectorsize);
9067                 if (reserved_space < PAGE_SIZE) {
9068                         end = page_start + reserved_space - 1;
9069                         btrfs_delalloc_release_space(inode, data_reserved,
9070                                         page_start, PAGE_SIZE - reserved_space);
9071                 }
9072         }
9073
9074         /*
9075          * page_mkwrite gets called when the page is firstly dirtied after it's
9076          * faulted in, but write(2) could also dirty a page and set delalloc
9077          * bits, thus in this case for space account reason, we still need to
9078          * clear any delalloc bits within this page range since we have to
9079          * reserve data&meta space before lock_page() (see above comments).
9080          */
9081         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9082                           EXTENT_DIRTY | EXTENT_DELALLOC |
9083                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9084                           0, 0, &cached_state);
9085
9086         ret = btrfs_set_extent_delalloc(inode, page_start, end, 0,
9087                                         &cached_state, 0);
9088         if (ret) {
9089                 unlock_extent_cached(io_tree, page_start, page_end,
9090                                      &cached_state);
9091                 ret = VM_FAULT_SIGBUS;
9092                 goto out_unlock;
9093         }
9094         ret = 0;
9095
9096         /* page is wholly or partially inside EOF */
9097         if (page_start + PAGE_SIZE > size)
9098                 zero_start = size & ~PAGE_MASK;
9099         else
9100                 zero_start = PAGE_SIZE;
9101
9102         if (zero_start != PAGE_SIZE) {
9103                 kaddr = kmap(page);
9104                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9105                 flush_dcache_page(page);
9106                 kunmap(page);
9107         }
9108         ClearPageChecked(page);
9109         set_page_dirty(page);
9110         SetPageUptodate(page);
9111
9112         BTRFS_I(inode)->last_trans = fs_info->generation;
9113         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9114         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9115
9116         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
9117
9118 out_unlock:
9119         if (!ret) {
9120                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9121                 sb_end_pagefault(inode->i_sb);
9122                 extent_changeset_free(data_reserved);
9123                 return VM_FAULT_LOCKED;
9124         }
9125         unlock_page(page);
9126 out:
9127         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9128         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9129                                      reserved_space);
9130 out_noreserve:
9131         sb_end_pagefault(inode->i_sb);
9132         extent_changeset_free(data_reserved);
9133         return ret;
9134 }
9135
9136 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
9137 {
9138         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9139         struct btrfs_root *root = BTRFS_I(inode)->root;
9140         struct btrfs_block_rsv *rsv;
9141         int ret = 0;
9142         int err = 0;
9143         struct btrfs_trans_handle *trans;
9144         u64 mask = fs_info->sectorsize - 1;
9145         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
9146
9147         if (!skip_writeback) {
9148                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9149                                                (u64)-1);
9150                 if (ret)
9151                         return ret;
9152         }
9153
9154         /*
9155          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9156          * 3 things going on here
9157          *
9158          * 1) We need to reserve space for our orphan item and the space to
9159          * delete our orphan item.  Lord knows we don't want to have a dangling
9160          * orphan item because we didn't reserve space to remove it.
9161          *
9162          * 2) We need to reserve space to update our inode.
9163          *
9164          * 3) We need to have something to cache all the space that is going to
9165          * be free'd up by the truncate operation, but also have some slack
9166          * space reserved in case it uses space during the truncate (thank you
9167          * very much snapshotting).
9168          *
9169          * And we need these to all be separate.  The fact is we can use a lot of
9170          * space doing the truncate, and we have no earthly idea how much space
9171          * we will use, so we need the truncate reservation to be separate so it
9172          * doesn't end up using space reserved for updating the inode or
9173          * removing the orphan item.  We also need to be able to stop the
9174          * transaction and start a new one, which means we need to be able to
9175          * update the inode several times, and we have no idea of knowing how
9176          * many times that will be, so we can't just reserve 1 item for the
9177          * entirety of the operation, so that has to be done separately as well.
9178          * Then there is the orphan item, which does indeed need to be held on
9179          * to for the whole operation, and we need nobody to touch this reserved
9180          * space except the orphan code.
9181          *
9182          * So that leaves us with
9183          *
9184          * 1) root->orphan_block_rsv - for the orphan deletion.
9185          * 2) rsv - for the truncate reservation, which we will steal from the
9186          * transaction reservation.
9187          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9188          * updating the inode.
9189          */
9190         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9191         if (!rsv)
9192                 return -ENOMEM;
9193         rsv->size = min_size;
9194         rsv->failfast = 1;
9195
9196         /*
9197          * 1 for the truncate slack space
9198          * 1 for updating the inode.
9199          */
9200         trans = btrfs_start_transaction(root, 2);
9201         if (IS_ERR(trans)) {
9202                 err = PTR_ERR(trans);
9203                 goto out;
9204         }
9205
9206         /* Migrate the slack space for the truncate to our reserve */
9207         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9208                                       min_size, 0);
9209         BUG_ON(ret);
9210
9211         /*
9212          * So if we truncate and then write and fsync we normally would just
9213          * write the extents that changed, which is a problem if we need to
9214          * first truncate that entire inode.  So set this flag so we write out
9215          * all of the extents in the inode to the sync log so we're completely
9216          * safe.
9217          */
9218         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9219         trans->block_rsv = rsv;
9220
9221         while (1) {
9222                 ret = btrfs_truncate_inode_items(trans, root, inode,
9223                                                  inode->i_size,
9224                                                  BTRFS_EXTENT_DATA_KEY);
9225                 trans->block_rsv = &fs_info->trans_block_rsv;
9226                 if (ret != -ENOSPC && ret != -EAGAIN) {
9227                         err = ret;
9228                         break;
9229                 }
9230
9231                 ret = btrfs_update_inode(trans, root, inode);
9232                 if (ret) {
9233                         err = ret;
9234                         break;
9235                 }
9236
9237                 btrfs_end_transaction(trans);
9238                 btrfs_btree_balance_dirty(fs_info);
9239
9240                 trans = btrfs_start_transaction(root, 2);
9241                 if (IS_ERR(trans)) {
9242                         ret = err = PTR_ERR(trans);
9243                         trans = NULL;
9244                         break;
9245                 }
9246
9247                 btrfs_block_rsv_release(fs_info, rsv, -1);
9248                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9249                                               rsv, min_size, 0);
9250                 BUG_ON(ret);    /* shouldn't happen */
9251                 trans->block_rsv = rsv;
9252         }
9253
9254         /*
9255          * We can't call btrfs_truncate_block inside a trans handle as we could
9256          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9257          * we've truncated everything except the last little bit, and can do
9258          * btrfs_truncate_block and then update the disk_i_size.
9259          */
9260         if (ret == NEED_TRUNCATE_BLOCK) {
9261                 btrfs_end_transaction(trans);
9262                 btrfs_btree_balance_dirty(fs_info);
9263
9264                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9265                 if (ret)
9266                         goto out;
9267                 trans = btrfs_start_transaction(root, 1);
9268                 if (IS_ERR(trans)) {
9269                         ret = PTR_ERR(trans);
9270                         goto out;
9271                 }
9272                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9273         }
9274
9275         if (ret == 0 && inode->i_nlink > 0) {
9276                 trans->block_rsv = root->orphan_block_rsv;
9277                 ret = btrfs_orphan_del(trans, BTRFS_I(inode));
9278                 if (ret)
9279                         err = ret;
9280         }
9281
9282         if (trans) {
9283                 trans->block_rsv = &fs_info->trans_block_rsv;
9284                 ret = btrfs_update_inode(trans, root, inode);
9285                 if (ret && !err)
9286                         err = ret;
9287
9288                 ret = btrfs_end_transaction(trans);
9289                 btrfs_btree_balance_dirty(fs_info);
9290         }
9291 out:
9292         btrfs_free_block_rsv(fs_info, rsv);
9293
9294         if (ret && !err)
9295                 err = ret;
9296
9297         return err;
9298 }
9299
9300 /*
9301  * create a new subvolume directory/inode (helper for the ioctl).
9302  */
9303 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9304                              struct btrfs_root *new_root,
9305                              struct btrfs_root *parent_root,
9306                              u64 new_dirid)
9307 {
9308         struct inode *inode;
9309         int err;
9310         u64 index = 0;
9311
9312         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9313                                 new_dirid, new_dirid,
9314                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9315                                 &index);
9316         if (IS_ERR(inode))
9317                 return PTR_ERR(inode);
9318         inode->i_op = &btrfs_dir_inode_operations;
9319         inode->i_fop = &btrfs_dir_file_operations;
9320
9321         set_nlink(inode, 1);
9322         btrfs_i_size_write(BTRFS_I(inode), 0);
9323         unlock_new_inode(inode);
9324
9325         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9326         if (err)
9327                 btrfs_err(new_root->fs_info,
9328                           "error inheriting subvolume %llu properties: %d",
9329                           new_root->root_key.objectid, err);
9330
9331         err = btrfs_update_inode(trans, new_root, inode);
9332
9333         iput(inode);
9334         return err;
9335 }
9336
9337 struct inode *btrfs_alloc_inode(struct super_block *sb)
9338 {
9339         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9340         struct btrfs_inode *ei;
9341         struct inode *inode;
9342
9343         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9344         if (!ei)
9345                 return NULL;
9346
9347         ei->root = NULL;
9348         ei->generation = 0;
9349         ei->last_trans = 0;
9350         ei->last_sub_trans = 0;
9351         ei->logged_trans = 0;
9352         ei->delalloc_bytes = 0;
9353         ei->new_delalloc_bytes = 0;
9354         ei->defrag_bytes = 0;
9355         ei->disk_i_size = 0;
9356         ei->flags = 0;
9357         ei->csum_bytes = 0;
9358         ei->index_cnt = (u64)-1;
9359         ei->dir_index = 0;
9360         ei->last_unlink_trans = 0;
9361         ei->last_log_commit = 0;
9362
9363         spin_lock_init(&ei->lock);
9364         ei->outstanding_extents = 0;
9365         if (sb->s_magic != BTRFS_TEST_MAGIC)
9366                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9367                                               BTRFS_BLOCK_RSV_DELALLOC);
9368         ei->runtime_flags = 0;
9369         ei->prop_compress = BTRFS_COMPRESS_NONE;
9370         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9371
9372         ei->delayed_node = NULL;
9373
9374         ei->i_otime.tv_sec = 0;
9375         ei->i_otime.tv_nsec = 0;
9376
9377         inode = &ei->vfs_inode;
9378         extent_map_tree_init(&ei->extent_tree);
9379         extent_io_tree_init(&ei->io_tree, inode);
9380         extent_io_tree_init(&ei->io_failure_tree, inode);
9381         ei->io_tree.track_uptodate = 1;
9382         ei->io_failure_tree.track_uptodate = 1;
9383         atomic_set(&ei->sync_writers, 0);
9384         mutex_init(&ei->log_mutex);
9385         mutex_init(&ei->delalloc_mutex);
9386         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9387         INIT_LIST_HEAD(&ei->delalloc_inodes);
9388         INIT_LIST_HEAD(&ei->delayed_iput);
9389         RB_CLEAR_NODE(&ei->rb_node);
9390         init_rwsem(&ei->dio_sem);
9391
9392         return inode;
9393 }
9394
9395 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9396 void btrfs_test_destroy_inode(struct inode *inode)
9397 {
9398         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9399         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9400 }
9401 #endif
9402
9403 static void btrfs_i_callback(struct rcu_head *head)
9404 {
9405         struct inode *inode = container_of(head, struct inode, i_rcu);
9406         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9407 }
9408
9409 void btrfs_destroy_inode(struct inode *inode)
9410 {
9411         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9412         struct btrfs_ordered_extent *ordered;
9413         struct btrfs_root *root = BTRFS_I(inode)->root;
9414
9415         WARN_ON(!hlist_empty(&inode->i_dentry));
9416         WARN_ON(inode->i_data.nrpages);
9417         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9418         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9419         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9420         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9421         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9422         WARN_ON(BTRFS_I(inode)->csum_bytes);
9423         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9424
9425         /*
9426          * This can happen where we create an inode, but somebody else also
9427          * created the same inode and we need to destroy the one we already
9428          * created.
9429          */
9430         if (!root)
9431                 goto free;
9432
9433         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9434                      &BTRFS_I(inode)->runtime_flags)) {
9435                 btrfs_info(fs_info, "inode %llu still on the orphan list",
9436                            btrfs_ino(BTRFS_I(inode)));
9437                 atomic_dec(&root->orphan_inodes);
9438         }
9439
9440         while (1) {
9441                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9442                 if (!ordered)
9443                         break;
9444                 else {
9445                         btrfs_err(fs_info,
9446                                   "found ordered extent %llu %llu on inode cleanup",
9447                                   ordered->file_offset, ordered->len);
9448                         btrfs_remove_ordered_extent(inode, ordered);
9449                         btrfs_put_ordered_extent(ordered);
9450                         btrfs_put_ordered_extent(ordered);
9451                 }
9452         }
9453         btrfs_qgroup_check_reserved_leak(inode);
9454         inode_tree_del(inode);
9455         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9456 free:
9457         call_rcu(&inode->i_rcu, btrfs_i_callback);
9458 }
9459
9460 int btrfs_drop_inode(struct inode *inode)
9461 {
9462         struct btrfs_root *root = BTRFS_I(inode)->root;
9463
9464         if (root == NULL)
9465                 return 1;
9466
9467         /* the snap/subvol tree is on deleting */
9468         if (btrfs_root_refs(&root->root_item) == 0)
9469                 return 1;
9470         else
9471                 return generic_drop_inode(inode);
9472 }
9473
9474 static void init_once(void *foo)
9475 {
9476         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9477
9478         inode_init_once(&ei->vfs_inode);
9479 }
9480
9481 void __cold btrfs_destroy_cachep(void)
9482 {
9483         /*
9484          * Make sure all delayed rcu free inodes are flushed before we
9485          * destroy cache.
9486          */
9487         rcu_barrier();
9488         kmem_cache_destroy(btrfs_inode_cachep);
9489         kmem_cache_destroy(btrfs_trans_handle_cachep);
9490         kmem_cache_destroy(btrfs_path_cachep);
9491         kmem_cache_destroy(btrfs_free_space_cachep);
9492 }
9493
9494 int __init btrfs_init_cachep(void)
9495 {
9496         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9497                         sizeof(struct btrfs_inode), 0,
9498                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9499                         init_once);
9500         if (!btrfs_inode_cachep)
9501                 goto fail;
9502
9503         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9504                         sizeof(struct btrfs_trans_handle), 0,
9505                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9506         if (!btrfs_trans_handle_cachep)
9507                 goto fail;
9508
9509         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9510                         sizeof(struct btrfs_path), 0,
9511                         SLAB_MEM_SPREAD, NULL);
9512         if (!btrfs_path_cachep)
9513                 goto fail;
9514
9515         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9516                         sizeof(struct btrfs_free_space), 0,
9517                         SLAB_MEM_SPREAD, NULL);
9518         if (!btrfs_free_space_cachep)
9519                 goto fail;
9520
9521         return 0;
9522 fail:
9523         btrfs_destroy_cachep();
9524         return -ENOMEM;
9525 }
9526
9527 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9528                          u32 request_mask, unsigned int flags)
9529 {
9530         u64 delalloc_bytes;
9531         struct inode *inode = d_inode(path->dentry);
9532         u32 blocksize = inode->i_sb->s_blocksize;
9533         u32 bi_flags = BTRFS_I(inode)->flags;
9534
9535         stat->result_mask |= STATX_BTIME;
9536         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9537         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9538         if (bi_flags & BTRFS_INODE_APPEND)
9539                 stat->attributes |= STATX_ATTR_APPEND;
9540         if (bi_flags & BTRFS_INODE_COMPRESS)
9541                 stat->attributes |= STATX_ATTR_COMPRESSED;
9542         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9543                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9544         if (bi_flags & BTRFS_INODE_NODUMP)
9545                 stat->attributes |= STATX_ATTR_NODUMP;
9546
9547         stat->attributes_mask |= (STATX_ATTR_APPEND |
9548                                   STATX_ATTR_COMPRESSED |
9549                                   STATX_ATTR_IMMUTABLE |
9550                                   STATX_ATTR_NODUMP);
9551
9552         generic_fillattr(inode, stat);
9553         stat->dev = BTRFS_I(inode)->root->anon_dev;
9554
9555         spin_lock(&BTRFS_I(inode)->lock);
9556         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9557         spin_unlock(&BTRFS_I(inode)->lock);
9558         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9559                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9560         return 0;
9561 }
9562
9563 static int btrfs_rename_exchange(struct inode *old_dir,
9564                               struct dentry *old_dentry,
9565                               struct inode *new_dir,
9566                               struct dentry *new_dentry)
9567 {
9568         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9569         struct btrfs_trans_handle *trans;
9570         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9571         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9572         struct inode *new_inode = new_dentry->d_inode;
9573         struct inode *old_inode = old_dentry->d_inode;
9574         struct timespec ctime = current_time(old_inode);
9575         struct dentry *parent;
9576         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9577         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9578         u64 old_idx = 0;
9579         u64 new_idx = 0;
9580         u64 root_objectid;
9581         int ret;
9582         bool root_log_pinned = false;
9583         bool dest_log_pinned = false;
9584
9585         /* we only allow rename subvolume link between subvolumes */
9586         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9587                 return -EXDEV;
9588
9589         /* close the race window with snapshot create/destroy ioctl */
9590         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9591                 down_read(&fs_info->subvol_sem);
9592         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9593                 down_read(&fs_info->subvol_sem);
9594
9595         /*
9596          * We want to reserve the absolute worst case amount of items.  So if
9597          * both inodes are subvols and we need to unlink them then that would
9598          * require 4 item modifications, but if they are both normal inodes it
9599          * would require 5 item modifications, so we'll assume their normal
9600          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9601          * should cover the worst case number of items we'll modify.
9602          */
9603         trans = btrfs_start_transaction(root, 12);
9604         if (IS_ERR(trans)) {
9605                 ret = PTR_ERR(trans);
9606                 goto out_notrans;
9607         }
9608
9609         /*
9610          * We need to find a free sequence number both in the source and
9611          * in the destination directory for the exchange.
9612          */
9613         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9614         if (ret)
9615                 goto out_fail;
9616         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9617         if (ret)
9618                 goto out_fail;
9619
9620         BTRFS_I(old_inode)->dir_index = 0ULL;
9621         BTRFS_I(new_inode)->dir_index = 0ULL;
9622
9623         /* Reference for the source. */
9624         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9625                 /* force full log commit if subvolume involved. */
9626                 btrfs_set_log_full_commit(fs_info, trans);
9627         } else {
9628                 btrfs_pin_log_trans(root);
9629                 root_log_pinned = true;
9630                 ret = btrfs_insert_inode_ref(trans, dest,
9631                                              new_dentry->d_name.name,
9632                                              new_dentry->d_name.len,
9633                                              old_ino,
9634                                              btrfs_ino(BTRFS_I(new_dir)),
9635                                              old_idx);
9636                 if (ret)
9637                         goto out_fail;
9638         }
9639
9640         /* And now for the dest. */
9641         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9642                 /* force full log commit if subvolume involved. */
9643                 btrfs_set_log_full_commit(fs_info, trans);
9644         } else {
9645                 btrfs_pin_log_trans(dest);
9646                 dest_log_pinned = true;
9647                 ret = btrfs_insert_inode_ref(trans, root,
9648                                              old_dentry->d_name.name,
9649                                              old_dentry->d_name.len,
9650                                              new_ino,
9651                                              btrfs_ino(BTRFS_I(old_dir)),
9652                                              new_idx);
9653                 if (ret)
9654                         goto out_fail;
9655         }
9656
9657         /* Update inode version and ctime/mtime. */
9658         inode_inc_iversion(old_dir);
9659         inode_inc_iversion(new_dir);
9660         inode_inc_iversion(old_inode);
9661         inode_inc_iversion(new_inode);
9662         old_dir->i_ctime = old_dir->i_mtime = ctime;
9663         new_dir->i_ctime = new_dir->i_mtime = ctime;
9664         old_inode->i_ctime = ctime;
9665         new_inode->i_ctime = ctime;
9666
9667         if (old_dentry->d_parent != new_dentry->d_parent) {
9668                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9669                                 BTRFS_I(old_inode), 1);
9670                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9671                                 BTRFS_I(new_inode), 1);
9672         }
9673
9674         /* src is a subvolume */
9675         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9676                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9677                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9678                                           root_objectid,
9679                                           old_dentry->d_name.name,
9680                                           old_dentry->d_name.len);
9681         } else { /* src is an inode */
9682                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9683                                            BTRFS_I(old_dentry->d_inode),
9684                                            old_dentry->d_name.name,
9685                                            old_dentry->d_name.len);
9686                 if (!ret)
9687                         ret = btrfs_update_inode(trans, root, old_inode);
9688         }
9689         if (ret) {
9690                 btrfs_abort_transaction(trans, ret);
9691                 goto out_fail;
9692         }
9693
9694         /* dest is a subvolume */
9695         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9696                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9697                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9698                                           root_objectid,
9699                                           new_dentry->d_name.name,
9700                                           new_dentry->d_name.len);
9701         } else { /* dest is an inode */
9702                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9703                                            BTRFS_I(new_dentry->d_inode),
9704                                            new_dentry->d_name.name,
9705                                            new_dentry->d_name.len);
9706                 if (!ret)
9707                         ret = btrfs_update_inode(trans, dest, new_inode);
9708         }
9709         if (ret) {
9710                 btrfs_abort_transaction(trans, ret);
9711                 goto out_fail;
9712         }
9713
9714         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9715                              new_dentry->d_name.name,
9716                              new_dentry->d_name.len, 0, old_idx);
9717         if (ret) {
9718                 btrfs_abort_transaction(trans, ret);
9719                 goto out_fail;
9720         }
9721
9722         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9723                              old_dentry->d_name.name,
9724                              old_dentry->d_name.len, 0, new_idx);
9725         if (ret) {
9726                 btrfs_abort_transaction(trans, ret);
9727                 goto out_fail;
9728         }
9729
9730         if (old_inode->i_nlink == 1)
9731                 BTRFS_I(old_inode)->dir_index = old_idx;
9732         if (new_inode->i_nlink == 1)
9733                 BTRFS_I(new_inode)->dir_index = new_idx;
9734
9735         if (root_log_pinned) {
9736                 parent = new_dentry->d_parent;
9737                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9738                                 parent);
9739                 btrfs_end_log_trans(root);
9740                 root_log_pinned = false;
9741         }
9742         if (dest_log_pinned) {
9743                 parent = old_dentry->d_parent;
9744                 btrfs_log_new_name(trans, BTRFS_I(new_inode), BTRFS_I(new_dir),
9745                                 parent);
9746                 btrfs_end_log_trans(dest);
9747                 dest_log_pinned = false;
9748         }
9749 out_fail:
9750         /*
9751          * If we have pinned a log and an error happened, we unpin tasks
9752          * trying to sync the log and force them to fallback to a transaction
9753          * commit if the log currently contains any of the inodes involved in
9754          * this rename operation (to ensure we do not persist a log with an
9755          * inconsistent state for any of these inodes or leading to any
9756          * inconsistencies when replayed). If the transaction was aborted, the
9757          * abortion reason is propagated to userspace when attempting to commit
9758          * the transaction. If the log does not contain any of these inodes, we
9759          * allow the tasks to sync it.
9760          */
9761         if (ret && (root_log_pinned || dest_log_pinned)) {
9762                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9763                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9764                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9765                     (new_inode &&
9766                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9767                         btrfs_set_log_full_commit(fs_info, trans);
9768
9769                 if (root_log_pinned) {
9770                         btrfs_end_log_trans(root);
9771                         root_log_pinned = false;
9772                 }
9773                 if (dest_log_pinned) {
9774                         btrfs_end_log_trans(dest);
9775                         dest_log_pinned = false;
9776                 }
9777         }
9778         ret = btrfs_end_transaction(trans);
9779 out_notrans:
9780         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9781                 up_read(&fs_info->subvol_sem);
9782         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9783                 up_read(&fs_info->subvol_sem);
9784
9785         return ret;
9786 }
9787
9788 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9789                                      struct btrfs_root *root,
9790                                      struct inode *dir,
9791                                      struct dentry *dentry)
9792 {
9793         int ret;
9794         struct inode *inode;
9795         u64 objectid;
9796         u64 index;
9797
9798         ret = btrfs_find_free_ino(root, &objectid);
9799         if (ret)
9800                 return ret;
9801
9802         inode = btrfs_new_inode(trans, root, dir,
9803                                 dentry->d_name.name,
9804                                 dentry->d_name.len,
9805                                 btrfs_ino(BTRFS_I(dir)),
9806                                 objectid,
9807                                 S_IFCHR | WHITEOUT_MODE,
9808                                 &index);
9809
9810         if (IS_ERR(inode)) {
9811                 ret = PTR_ERR(inode);
9812                 return ret;
9813         }
9814
9815         inode->i_op = &btrfs_special_inode_operations;
9816         init_special_inode(inode, inode->i_mode,
9817                 WHITEOUT_DEV);
9818
9819         ret = btrfs_init_inode_security(trans, inode, dir,
9820                                 &dentry->d_name);
9821         if (ret)
9822                 goto out;
9823
9824         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9825                                 BTRFS_I(inode), 0, index);
9826         if (ret)
9827                 goto out;
9828
9829         ret = btrfs_update_inode(trans, root, inode);
9830 out:
9831         unlock_new_inode(inode);
9832         if (ret)
9833                 inode_dec_link_count(inode);
9834         iput(inode);
9835
9836         return ret;
9837 }
9838
9839 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9840                            struct inode *new_dir, struct dentry *new_dentry,
9841                            unsigned int flags)
9842 {
9843         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9844         struct btrfs_trans_handle *trans;
9845         unsigned int trans_num_items;
9846         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9847         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9848         struct inode *new_inode = d_inode(new_dentry);
9849         struct inode *old_inode = d_inode(old_dentry);
9850         u64 index = 0;
9851         u64 root_objectid;
9852         int ret;
9853         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9854         bool log_pinned = false;
9855
9856         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9857                 return -EPERM;
9858
9859         /* we only allow rename subvolume link between subvolumes */
9860         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9861                 return -EXDEV;
9862
9863         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9864             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9865                 return -ENOTEMPTY;
9866
9867         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9868             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9869                 return -ENOTEMPTY;
9870
9871
9872         /* check for collisions, even if the  name isn't there */
9873         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9874                              new_dentry->d_name.name,
9875                              new_dentry->d_name.len);
9876
9877         if (ret) {
9878                 if (ret == -EEXIST) {
9879                         /* we shouldn't get
9880                          * eexist without a new_inode */
9881                         if (WARN_ON(!new_inode)) {
9882                                 return ret;
9883                         }
9884                 } else {
9885                         /* maybe -EOVERFLOW */
9886                         return ret;
9887                 }
9888         }
9889         ret = 0;
9890
9891         /*
9892          * we're using rename to replace one file with another.  Start IO on it
9893          * now so  we don't add too much work to the end of the transaction
9894          */
9895         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9896                 filemap_flush(old_inode->i_mapping);
9897
9898         /* close the racy window with snapshot create/destroy ioctl */
9899         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9900                 down_read(&fs_info->subvol_sem);
9901         /*
9902          * We want to reserve the absolute worst case amount of items.  So if
9903          * both inodes are subvols and we need to unlink them then that would
9904          * require 4 item modifications, but if they are both normal inodes it
9905          * would require 5 item modifications, so we'll assume they are normal
9906          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9907          * should cover the worst case number of items we'll modify.
9908          * If our rename has the whiteout flag, we need more 5 units for the
9909          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9910          * when selinux is enabled).
9911          */
9912         trans_num_items = 11;
9913         if (flags & RENAME_WHITEOUT)
9914                 trans_num_items += 5;
9915         trans = btrfs_start_transaction(root, trans_num_items);
9916         if (IS_ERR(trans)) {
9917                 ret = PTR_ERR(trans);
9918                 goto out_notrans;
9919         }
9920
9921         if (dest != root)
9922                 btrfs_record_root_in_trans(trans, dest);
9923
9924         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9925         if (ret)
9926                 goto out_fail;
9927
9928         BTRFS_I(old_inode)->dir_index = 0ULL;
9929         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9930                 /* force full log commit if subvolume involved. */
9931                 btrfs_set_log_full_commit(fs_info, trans);
9932         } else {
9933                 btrfs_pin_log_trans(root);
9934                 log_pinned = true;
9935                 ret = btrfs_insert_inode_ref(trans, dest,
9936                                              new_dentry->d_name.name,
9937                                              new_dentry->d_name.len,
9938                                              old_ino,
9939                                              btrfs_ino(BTRFS_I(new_dir)), index);
9940                 if (ret)
9941                         goto out_fail;
9942         }
9943
9944         inode_inc_iversion(old_dir);
9945         inode_inc_iversion(new_dir);
9946         inode_inc_iversion(old_inode);
9947         old_dir->i_ctime = old_dir->i_mtime =
9948         new_dir->i_ctime = new_dir->i_mtime =
9949         old_inode->i_ctime = current_time(old_dir);
9950
9951         if (old_dentry->d_parent != new_dentry->d_parent)
9952                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9953                                 BTRFS_I(old_inode), 1);
9954
9955         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9956                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9957                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9958                                         old_dentry->d_name.name,
9959                                         old_dentry->d_name.len);
9960         } else {
9961                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9962                                         BTRFS_I(d_inode(old_dentry)),
9963                                         old_dentry->d_name.name,
9964                                         old_dentry->d_name.len);
9965                 if (!ret)
9966                         ret = btrfs_update_inode(trans, root, old_inode);
9967         }
9968         if (ret) {
9969                 btrfs_abort_transaction(trans, ret);
9970                 goto out_fail;
9971         }
9972
9973         if (new_inode) {
9974                 inode_inc_iversion(new_inode);
9975                 new_inode->i_ctime = current_time(new_inode);
9976                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9977                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9978                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9979                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9980                                                 root_objectid,
9981                                                 new_dentry->d_name.name,
9982                                                 new_dentry->d_name.len);
9983                         BUG_ON(new_inode->i_nlink == 0);
9984                 } else {
9985                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9986                                                  BTRFS_I(d_inode(new_dentry)),
9987                                                  new_dentry->d_name.name,
9988                                                  new_dentry->d_name.len);
9989                 }
9990                 if (!ret && new_inode->i_nlink == 0)
9991                         ret = btrfs_orphan_add(trans,
9992                                         BTRFS_I(d_inode(new_dentry)));
9993                 if (ret) {
9994                         btrfs_abort_transaction(trans, ret);
9995                         goto out_fail;
9996                 }
9997         }
9998
9999         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
10000                              new_dentry->d_name.name,
10001                              new_dentry->d_name.len, 0, index);
10002         if (ret) {
10003                 btrfs_abort_transaction(trans, ret);
10004                 goto out_fail;
10005         }
10006
10007         if (old_inode->i_nlink == 1)
10008                 BTRFS_I(old_inode)->dir_index = index;
10009
10010         if (log_pinned) {
10011                 struct dentry *parent = new_dentry->d_parent;
10012
10013                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
10014                                 parent);
10015                 btrfs_end_log_trans(root);
10016                 log_pinned = false;
10017         }
10018
10019         if (flags & RENAME_WHITEOUT) {
10020                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
10021                                                 old_dentry);
10022
10023                 if (ret) {
10024                         btrfs_abort_transaction(trans, ret);
10025                         goto out_fail;
10026                 }
10027         }
10028 out_fail:
10029         /*
10030          * If we have pinned the log and an error happened, we unpin tasks
10031          * trying to sync the log and force them to fallback to a transaction
10032          * commit if the log currently contains any of the inodes involved in
10033          * this rename operation (to ensure we do not persist a log with an
10034          * inconsistent state for any of these inodes or leading to any
10035          * inconsistencies when replayed). If the transaction was aborted, the
10036          * abortion reason is propagated to userspace when attempting to commit
10037          * the transaction. If the log does not contain any of these inodes, we
10038          * allow the tasks to sync it.
10039          */
10040         if (ret && log_pinned) {
10041                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
10042                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
10043                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10044                     (new_inode &&
10045                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10046                         btrfs_set_log_full_commit(fs_info, trans);
10047
10048                 btrfs_end_log_trans(root);
10049                 log_pinned = false;
10050         }
10051         btrfs_end_transaction(trans);
10052 out_notrans:
10053         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10054                 up_read(&fs_info->subvol_sem);
10055
10056         return ret;
10057 }
10058
10059 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10060                          struct inode *new_dir, struct dentry *new_dentry,
10061                          unsigned int flags)
10062 {
10063         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10064                 return -EINVAL;
10065
10066         if (flags & RENAME_EXCHANGE)
10067                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10068                                           new_dentry);
10069
10070         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10071 }
10072
10073 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10074 {
10075         struct btrfs_delalloc_work *delalloc_work;
10076         struct inode *inode;
10077
10078         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10079                                      work);
10080         inode = delalloc_work->inode;
10081         filemap_flush(inode->i_mapping);
10082         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10083                                 &BTRFS_I(inode)->runtime_flags))
10084                 filemap_flush(inode->i_mapping);
10085
10086         if (delalloc_work->delay_iput)
10087                 btrfs_add_delayed_iput(inode);
10088         else
10089                 iput(inode);
10090         complete(&delalloc_work->completion);
10091 }
10092
10093 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
10094                                                     int delay_iput)
10095 {
10096         struct btrfs_delalloc_work *work;
10097
10098         work = kmalloc(sizeof(*work), GFP_NOFS);
10099         if (!work)
10100                 return NULL;
10101
10102         init_completion(&work->completion);
10103         INIT_LIST_HEAD(&work->list);
10104         work->inode = inode;
10105         work->delay_iput = delay_iput;
10106         WARN_ON_ONCE(!inode);
10107         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10108                         btrfs_run_delalloc_work, NULL, NULL);
10109
10110         return work;
10111 }
10112
10113 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10114 {
10115         wait_for_completion(&work->completion);
10116         kfree(work);
10117 }
10118
10119 /*
10120  * some fairly slow code that needs optimization. This walks the list
10121  * of all the inodes with pending delalloc and forces them to disk.
10122  */
10123 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10124                                    int nr)
10125 {
10126         struct btrfs_inode *binode;
10127         struct inode *inode;
10128         struct btrfs_delalloc_work *work, *next;
10129         struct list_head works;
10130         struct list_head splice;
10131         int ret = 0;
10132
10133         INIT_LIST_HEAD(&works);
10134         INIT_LIST_HEAD(&splice);
10135
10136         mutex_lock(&root->delalloc_mutex);
10137         spin_lock(&root->delalloc_lock);
10138         list_splice_init(&root->delalloc_inodes, &splice);
10139         while (!list_empty(&splice)) {
10140                 binode = list_entry(splice.next, struct btrfs_inode,
10141                                     delalloc_inodes);
10142
10143                 list_move_tail(&binode->delalloc_inodes,
10144                                &root->delalloc_inodes);
10145                 inode = igrab(&binode->vfs_inode);
10146                 if (!inode) {
10147                         cond_resched_lock(&root->delalloc_lock);
10148                         continue;
10149                 }
10150                 spin_unlock(&root->delalloc_lock);
10151
10152                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10153                 if (!work) {
10154                         if (delay_iput)
10155                                 btrfs_add_delayed_iput(inode);
10156                         else
10157                                 iput(inode);
10158                         ret = -ENOMEM;
10159                         goto out;
10160                 }
10161                 list_add_tail(&work->list, &works);
10162                 btrfs_queue_work(root->fs_info->flush_workers,
10163                                  &work->work);
10164                 ret++;
10165                 if (nr != -1 && ret >= nr)
10166                         goto out;
10167                 cond_resched();
10168                 spin_lock(&root->delalloc_lock);
10169         }
10170         spin_unlock(&root->delalloc_lock);
10171
10172 out:
10173         list_for_each_entry_safe(work, next, &works, list) {
10174                 list_del_init(&work->list);
10175                 btrfs_wait_and_free_delalloc_work(work);
10176         }
10177
10178         if (!list_empty_careful(&splice)) {
10179                 spin_lock(&root->delalloc_lock);
10180                 list_splice_tail(&splice, &root->delalloc_inodes);
10181                 spin_unlock(&root->delalloc_lock);
10182         }
10183         mutex_unlock(&root->delalloc_mutex);
10184         return ret;
10185 }
10186
10187 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10188 {
10189         struct btrfs_fs_info *fs_info = root->fs_info;
10190         int ret;
10191
10192         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10193                 return -EROFS;
10194
10195         ret = __start_delalloc_inodes(root, delay_iput, -1);
10196         if (ret > 0)
10197                 ret = 0;
10198         return ret;
10199 }
10200
10201 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10202                                int nr)
10203 {
10204         struct btrfs_root *root;
10205         struct list_head splice;
10206         int ret;
10207
10208         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10209                 return -EROFS;
10210
10211         INIT_LIST_HEAD(&splice);
10212
10213         mutex_lock(&fs_info->delalloc_root_mutex);
10214         spin_lock(&fs_info->delalloc_root_lock);
10215         list_splice_init(&fs_info->delalloc_roots, &splice);
10216         while (!list_empty(&splice) && nr) {
10217                 root = list_first_entry(&splice, struct btrfs_root,
10218                                         delalloc_root);
10219                 root = btrfs_grab_fs_root(root);
10220                 BUG_ON(!root);
10221                 list_move_tail(&root->delalloc_root,
10222                                &fs_info->delalloc_roots);
10223                 spin_unlock(&fs_info->delalloc_root_lock);
10224
10225                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10226                 btrfs_put_fs_root(root);
10227                 if (ret < 0)
10228                         goto out;
10229
10230                 if (nr != -1) {
10231                         nr -= ret;
10232                         WARN_ON(nr < 0);
10233                 }
10234                 spin_lock(&fs_info->delalloc_root_lock);
10235         }
10236         spin_unlock(&fs_info->delalloc_root_lock);
10237
10238         ret = 0;
10239 out:
10240         if (!list_empty_careful(&splice)) {
10241                 spin_lock(&fs_info->delalloc_root_lock);
10242                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10243                 spin_unlock(&fs_info->delalloc_root_lock);
10244         }
10245         mutex_unlock(&fs_info->delalloc_root_mutex);
10246         return ret;
10247 }
10248
10249 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10250                          const char *symname)
10251 {
10252         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10253         struct btrfs_trans_handle *trans;
10254         struct btrfs_root *root = BTRFS_I(dir)->root;
10255         struct btrfs_path *path;
10256         struct btrfs_key key;
10257         struct inode *inode = NULL;
10258         int err;
10259         int drop_inode = 0;
10260         u64 objectid;
10261         u64 index = 0;
10262         int name_len;
10263         int datasize;
10264         unsigned long ptr;
10265         struct btrfs_file_extent_item *ei;
10266         struct extent_buffer *leaf;
10267
10268         name_len = strlen(symname);
10269         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10270                 return -ENAMETOOLONG;
10271
10272         /*
10273          * 2 items for inode item and ref
10274          * 2 items for dir items
10275          * 1 item for updating parent inode item
10276          * 1 item for the inline extent item
10277          * 1 item for xattr if selinux is on
10278          */
10279         trans = btrfs_start_transaction(root, 7);
10280         if (IS_ERR(trans))
10281                 return PTR_ERR(trans);
10282
10283         err = btrfs_find_free_ino(root, &objectid);
10284         if (err)
10285                 goto out_unlock;
10286
10287         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10288                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10289                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10290         if (IS_ERR(inode)) {
10291                 err = PTR_ERR(inode);
10292                 goto out_unlock;
10293         }
10294
10295         /*
10296         * If the active LSM wants to access the inode during
10297         * d_instantiate it needs these. Smack checks to see
10298         * if the filesystem supports xattrs by looking at the
10299         * ops vector.
10300         */
10301         inode->i_fop = &btrfs_file_operations;
10302         inode->i_op = &btrfs_file_inode_operations;
10303         inode->i_mapping->a_ops = &btrfs_aops;
10304         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10305
10306         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10307         if (err)
10308                 goto out_unlock_inode;
10309
10310         path = btrfs_alloc_path();
10311         if (!path) {
10312                 err = -ENOMEM;
10313                 goto out_unlock_inode;
10314         }
10315         key.objectid = btrfs_ino(BTRFS_I(inode));
10316         key.offset = 0;
10317         key.type = BTRFS_EXTENT_DATA_KEY;
10318         datasize = btrfs_file_extent_calc_inline_size(name_len);
10319         err = btrfs_insert_empty_item(trans, root, path, &key,
10320                                       datasize);
10321         if (err) {
10322                 btrfs_free_path(path);
10323                 goto out_unlock_inode;
10324         }
10325         leaf = path->nodes[0];
10326         ei = btrfs_item_ptr(leaf, path->slots[0],
10327                             struct btrfs_file_extent_item);
10328         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10329         btrfs_set_file_extent_type(leaf, ei,
10330                                    BTRFS_FILE_EXTENT_INLINE);
10331         btrfs_set_file_extent_encryption(leaf, ei, 0);
10332         btrfs_set_file_extent_compression(leaf, ei, 0);
10333         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10334         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10335
10336         ptr = btrfs_file_extent_inline_start(ei);
10337         write_extent_buffer(leaf, symname, ptr, name_len);
10338         btrfs_mark_buffer_dirty(leaf);
10339         btrfs_free_path(path);
10340
10341         inode->i_op = &btrfs_symlink_inode_operations;
10342         inode_nohighmem(inode);
10343         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10344         inode_set_bytes(inode, name_len);
10345         btrfs_i_size_write(BTRFS_I(inode), name_len);
10346         err = btrfs_update_inode(trans, root, inode);
10347         /*
10348          * Last step, add directory indexes for our symlink inode. This is the
10349          * last step to avoid extra cleanup of these indexes if an error happens
10350          * elsewhere above.
10351          */
10352         if (!err)
10353                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10354                                 BTRFS_I(inode), 0, index);
10355         if (err) {
10356                 drop_inode = 1;
10357                 goto out_unlock_inode;
10358         }
10359
10360         unlock_new_inode(inode);
10361         d_instantiate(dentry, inode);
10362
10363 out_unlock:
10364         btrfs_end_transaction(trans);
10365         if (drop_inode) {
10366                 inode_dec_link_count(inode);
10367                 iput(inode);
10368         }
10369         btrfs_btree_balance_dirty(fs_info);
10370         return err;
10371
10372 out_unlock_inode:
10373         drop_inode = 1;
10374         unlock_new_inode(inode);
10375         goto out_unlock;
10376 }
10377
10378 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10379                                        u64 start, u64 num_bytes, u64 min_size,
10380                                        loff_t actual_len, u64 *alloc_hint,
10381                                        struct btrfs_trans_handle *trans)
10382 {
10383         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10384         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10385         struct extent_map *em;
10386         struct btrfs_root *root = BTRFS_I(inode)->root;
10387         struct btrfs_key ins;
10388         u64 cur_offset = start;
10389         u64 i_size;
10390         u64 cur_bytes;
10391         u64 last_alloc = (u64)-1;
10392         int ret = 0;
10393         bool own_trans = true;
10394         u64 end = start + num_bytes - 1;
10395
10396         if (trans)
10397                 own_trans = false;
10398         while (num_bytes > 0) {
10399                 if (own_trans) {
10400                         trans = btrfs_start_transaction(root, 3);
10401                         if (IS_ERR(trans)) {
10402                                 ret = PTR_ERR(trans);
10403                                 break;
10404                         }
10405                 }
10406
10407                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10408                 cur_bytes = max(cur_bytes, min_size);
10409                 /*
10410                  * If we are severely fragmented we could end up with really
10411                  * small allocations, so if the allocator is returning small
10412                  * chunks lets make its job easier by only searching for those
10413                  * sized chunks.
10414                  */
10415                 cur_bytes = min(cur_bytes, last_alloc);
10416                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10417                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10418                 if (ret) {
10419                         if (own_trans)
10420                                 btrfs_end_transaction(trans);
10421                         break;
10422                 }
10423                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10424
10425                 last_alloc = ins.offset;
10426                 ret = insert_reserved_file_extent(trans, inode,
10427                                                   cur_offset, ins.objectid,
10428                                                   ins.offset, ins.offset,
10429                                                   ins.offset, 0, 0, 0,
10430                                                   BTRFS_FILE_EXTENT_PREALLOC);
10431                 if (ret) {
10432                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10433                                                    ins.offset, 0);
10434                         btrfs_abort_transaction(trans, ret);
10435                         if (own_trans)
10436                                 btrfs_end_transaction(trans);
10437                         break;
10438                 }
10439
10440                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10441                                         cur_offset + ins.offset -1, 0);
10442
10443                 em = alloc_extent_map();
10444                 if (!em) {
10445                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10446                                 &BTRFS_I(inode)->runtime_flags);
10447                         goto next;
10448                 }
10449
10450                 em->start = cur_offset;
10451                 em->orig_start = cur_offset;
10452                 em->len = ins.offset;
10453                 em->block_start = ins.objectid;
10454                 em->block_len = ins.offset;
10455                 em->orig_block_len = ins.offset;
10456                 em->ram_bytes = ins.offset;
10457                 em->bdev = fs_info->fs_devices->latest_bdev;
10458                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10459                 em->generation = trans->transid;
10460
10461                 while (1) {
10462                         write_lock(&em_tree->lock);
10463                         ret = add_extent_mapping(em_tree, em, 1);
10464                         write_unlock(&em_tree->lock);
10465                         if (ret != -EEXIST)
10466                                 break;
10467                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10468                                                 cur_offset + ins.offset - 1,
10469                                                 0);
10470                 }
10471                 free_extent_map(em);
10472 next:
10473                 num_bytes -= ins.offset;
10474                 cur_offset += ins.offset;
10475                 *alloc_hint = ins.objectid + ins.offset;
10476
10477                 inode_inc_iversion(inode);
10478                 inode->i_ctime = current_time(inode);
10479                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10480                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10481                     (actual_len > inode->i_size) &&
10482                     (cur_offset > inode->i_size)) {
10483                         if (cur_offset > actual_len)
10484                                 i_size = actual_len;
10485                         else
10486                                 i_size = cur_offset;
10487                         i_size_write(inode, i_size);
10488                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10489                 }
10490
10491                 ret = btrfs_update_inode(trans, root, inode);
10492
10493                 if (ret) {
10494                         btrfs_abort_transaction(trans, ret);
10495                         if (own_trans)
10496                                 btrfs_end_transaction(trans);
10497                         break;
10498                 }
10499
10500                 if (own_trans)
10501                         btrfs_end_transaction(trans);
10502         }
10503         if (cur_offset < end)
10504                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10505                         end - cur_offset + 1);
10506         return ret;
10507 }
10508
10509 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10510                               u64 start, u64 num_bytes, u64 min_size,
10511                               loff_t actual_len, u64 *alloc_hint)
10512 {
10513         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10514                                            min_size, actual_len, alloc_hint,
10515                                            NULL);
10516 }
10517
10518 int btrfs_prealloc_file_range_trans(struct inode *inode,
10519                                     struct btrfs_trans_handle *trans, int mode,
10520                                     u64 start, u64 num_bytes, u64 min_size,
10521                                     loff_t actual_len, u64 *alloc_hint)
10522 {
10523         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10524                                            min_size, actual_len, alloc_hint, trans);
10525 }
10526
10527 static int btrfs_set_page_dirty(struct page *page)
10528 {
10529         return __set_page_dirty_nobuffers(page);
10530 }
10531
10532 static int btrfs_permission(struct inode *inode, int mask)
10533 {
10534         struct btrfs_root *root = BTRFS_I(inode)->root;
10535         umode_t mode = inode->i_mode;
10536
10537         if (mask & MAY_WRITE &&
10538             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10539                 if (btrfs_root_readonly(root))
10540                         return -EROFS;
10541                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10542                         return -EACCES;
10543         }
10544         return generic_permission(inode, mask);
10545 }
10546
10547 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10548 {
10549         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10550         struct btrfs_trans_handle *trans;
10551         struct btrfs_root *root = BTRFS_I(dir)->root;
10552         struct inode *inode = NULL;
10553         u64 objectid;
10554         u64 index;
10555         int ret = 0;
10556
10557         /*
10558          * 5 units required for adding orphan entry
10559          */
10560         trans = btrfs_start_transaction(root, 5);
10561         if (IS_ERR(trans))
10562                 return PTR_ERR(trans);
10563
10564         ret = btrfs_find_free_ino(root, &objectid);
10565         if (ret)
10566                 goto out;
10567
10568         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10569                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10570         if (IS_ERR(inode)) {
10571                 ret = PTR_ERR(inode);
10572                 inode = NULL;
10573                 goto out;
10574         }
10575
10576         inode->i_fop = &btrfs_file_operations;
10577         inode->i_op = &btrfs_file_inode_operations;
10578
10579         inode->i_mapping->a_ops = &btrfs_aops;
10580         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10581
10582         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10583         if (ret)
10584                 goto out_inode;
10585
10586         ret = btrfs_update_inode(trans, root, inode);
10587         if (ret)
10588                 goto out_inode;
10589         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10590         if (ret)
10591                 goto out_inode;
10592
10593         /*
10594          * We set number of links to 0 in btrfs_new_inode(), and here we set
10595          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10596          * through:
10597          *
10598          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10599          */
10600         set_nlink(inode, 1);
10601         unlock_new_inode(inode);
10602         d_tmpfile(dentry, inode);
10603         mark_inode_dirty(inode);
10604
10605 out:
10606         btrfs_end_transaction(trans);
10607         if (ret)
10608                 iput(inode);
10609         btrfs_btree_balance_dirty(fs_info);
10610         return ret;
10611
10612 out_inode:
10613         unlock_new_inode(inode);
10614         goto out;
10615
10616 }
10617
10618 __attribute__((const))
10619 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10620 {
10621         return -EAGAIN;
10622 }
10623
10624 static struct btrfs_fs_info *iotree_fs_info(void *private_data)
10625 {
10626         struct inode *inode = private_data;
10627         return btrfs_sb(inode->i_sb);
10628 }
10629
10630 static void btrfs_check_extent_io_range(void *private_data, const char *caller,
10631                                         u64 start, u64 end)
10632 {
10633         struct inode *inode = private_data;
10634         u64 isize;
10635
10636         isize = i_size_read(inode);
10637         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
10638                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
10639                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
10640                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
10641         }
10642 }
10643
10644 void btrfs_set_range_writeback(void *private_data, u64 start, u64 end)
10645 {
10646         struct inode *inode = private_data;
10647         unsigned long index = start >> PAGE_SHIFT;
10648         unsigned long end_index = end >> PAGE_SHIFT;
10649         struct page *page;
10650
10651         while (index <= end_index) {
10652                 page = find_get_page(inode->i_mapping, index);
10653                 ASSERT(page); /* Pages should be in the extent_io_tree */
10654                 set_page_writeback(page);
10655                 put_page(page);
10656                 index++;
10657         }
10658 }
10659
10660 static const struct inode_operations btrfs_dir_inode_operations = {
10661         .getattr        = btrfs_getattr,
10662         .lookup         = btrfs_lookup,
10663         .create         = btrfs_create,
10664         .unlink         = btrfs_unlink,
10665         .link           = btrfs_link,
10666         .mkdir          = btrfs_mkdir,
10667         .rmdir          = btrfs_rmdir,
10668         .rename         = btrfs_rename2,
10669         .symlink        = btrfs_symlink,
10670         .setattr        = btrfs_setattr,
10671         .mknod          = btrfs_mknod,
10672         .listxattr      = btrfs_listxattr,
10673         .permission     = btrfs_permission,
10674         .get_acl        = btrfs_get_acl,
10675         .set_acl        = btrfs_set_acl,
10676         .update_time    = btrfs_update_time,
10677         .tmpfile        = btrfs_tmpfile,
10678 };
10679 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10680         .lookup         = btrfs_lookup,
10681         .permission     = btrfs_permission,
10682         .update_time    = btrfs_update_time,
10683 };
10684
10685 static const struct file_operations btrfs_dir_file_operations = {
10686         .llseek         = generic_file_llseek,
10687         .read           = generic_read_dir,
10688         .iterate_shared = btrfs_real_readdir,
10689         .open           = btrfs_opendir,
10690         .unlocked_ioctl = btrfs_ioctl,
10691 #ifdef CONFIG_COMPAT
10692         .compat_ioctl   = btrfs_compat_ioctl,
10693 #endif
10694         .release        = btrfs_release_file,
10695         .fsync          = btrfs_sync_file,
10696 };
10697
10698 static const struct extent_io_ops btrfs_extent_io_ops = {
10699         /* mandatory callbacks */
10700         .submit_bio_hook = btrfs_submit_bio_hook,
10701         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10702         .merge_bio_hook = btrfs_merge_bio_hook,
10703         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10704         .tree_fs_info = iotree_fs_info,
10705         .set_range_writeback = btrfs_set_range_writeback,
10706
10707         /* optional callbacks */
10708         .fill_delalloc = run_delalloc_range,
10709         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10710         .writepage_start_hook = btrfs_writepage_start_hook,
10711         .set_bit_hook = btrfs_set_bit_hook,
10712         .clear_bit_hook = btrfs_clear_bit_hook,
10713         .merge_extent_hook = btrfs_merge_extent_hook,
10714         .split_extent_hook = btrfs_split_extent_hook,
10715         .check_extent_io_range = btrfs_check_extent_io_range,
10716 };
10717
10718 /*
10719  * btrfs doesn't support the bmap operation because swapfiles
10720  * use bmap to make a mapping of extents in the file.  They assume
10721  * these extents won't change over the life of the file and they
10722  * use the bmap result to do IO directly to the drive.
10723  *
10724  * the btrfs bmap call would return logical addresses that aren't
10725  * suitable for IO and they also will change frequently as COW
10726  * operations happen.  So, swapfile + btrfs == corruption.
10727  *
10728  * For now we're avoiding this by dropping bmap.
10729  */
10730 static const struct address_space_operations btrfs_aops = {
10731         .readpage       = btrfs_readpage,
10732         .writepage      = btrfs_writepage,
10733         .writepages     = btrfs_writepages,
10734         .readpages      = btrfs_readpages,
10735         .direct_IO      = btrfs_direct_IO,
10736         .invalidatepage = btrfs_invalidatepage,
10737         .releasepage    = btrfs_releasepage,
10738         .set_page_dirty = btrfs_set_page_dirty,
10739         .error_remove_page = generic_error_remove_page,
10740 };
10741
10742 static const struct address_space_operations btrfs_symlink_aops = {
10743         .readpage       = btrfs_readpage,
10744         .writepage      = btrfs_writepage,
10745         .invalidatepage = btrfs_invalidatepage,
10746         .releasepage    = btrfs_releasepage,
10747 };
10748
10749 static const struct inode_operations btrfs_file_inode_operations = {
10750         .getattr        = btrfs_getattr,
10751         .setattr        = btrfs_setattr,
10752         .listxattr      = btrfs_listxattr,
10753         .permission     = btrfs_permission,
10754         .fiemap         = btrfs_fiemap,
10755         .get_acl        = btrfs_get_acl,
10756         .set_acl        = btrfs_set_acl,
10757         .update_time    = btrfs_update_time,
10758 };
10759 static const struct inode_operations btrfs_special_inode_operations = {
10760         .getattr        = btrfs_getattr,
10761         .setattr        = btrfs_setattr,
10762         .permission     = btrfs_permission,
10763         .listxattr      = btrfs_listxattr,
10764         .get_acl        = btrfs_get_acl,
10765         .set_acl        = btrfs_set_acl,
10766         .update_time    = btrfs_update_time,
10767 };
10768 static const struct inode_operations btrfs_symlink_inode_operations = {
10769         .get_link       = page_get_link,
10770         .getattr        = btrfs_getattr,
10771         .setattr        = btrfs_setattr,
10772         .permission     = btrfs_permission,
10773         .listxattr      = btrfs_listxattr,
10774         .update_time    = btrfs_update_time,
10775 };
10776
10777 const struct dentry_operations btrfs_dentry_operations = {
10778         .d_delete       = btrfs_dentry_delete,
10779         .d_release      = btrfs_dentry_release,
10780 };