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