fs/btrfs/inode.c

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