fs/btrfs/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/bio.h>
   8 #include <linux/file.h>
   9 #include <linux/fs.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/highmem.h>
  12 #include <linux/time.h>
  13 #include <linux/init.h>
  14 #include <linux/string.h>
  15 #include <linux/backing-dev.h>
  16 #include <linux/writeback.h>
  17 #include <linux/compat.h>
  18 #include <linux/xattr.h>
  19 #include <linux/posix_acl.h>
  20 #include <linux/falloc.h>
  21 #include <linux/slab.h>
  22 #include <linux/ratelimit.h>
  23 #include <linux/btrfs.h>
  24 #include <linux/blkdev.h>
  25 #include <linux/posix_acl_xattr.h>
  26 #include <linux/uio.h>
  27 #include <linux/magic.h>
  28 #include <linux/iversion.h>
  29 #include <linux/swap.h>
  30 #include <linux/migrate.h>
  31 #include <linux/sched/mm.h>
  32 #include <asm/unaligned.h>
  33 #include "misc.h"
  34 #include "ctree.h"
  35 #include "disk-io.h"
  36 #include "transaction.h"
  37 #include "btrfs_inode.h"
  38 #include "print-tree.h"
  39 #include "ordered-data.h"
  40 #include "xattr.h"
  41 #include "tree-log.h"
  42 #include "volumes.h"
  43 #include "compression.h"
  44 #include "locking.h"
  45 #include "free-space-cache.h"
  46 #include "inode-map.h"
  47 #include "props.h"
  48 #include "qgroup.h"
  49 #include "delalloc-space.h"
  50 #include "block-group.h"
  51 #include "space-info.h"
  52
  53 struct btrfs_iget_args {
  54         u64 ino;
  55         struct btrfs_root *root;
  56 };
  57
  58 struct btrfs_dio_data {
  59         u64 reserve;
  60         loff_t length;
  61         ssize_t submitted;
  62         struct extent_changeset *data_reserved;
  63 };
  64
  65 static const struct inode_operations btrfs_dir_inode_operations;
  66 static const struct inode_operations btrfs_symlink_inode_operations;
  67 static const struct inode_operations btrfs_special_inode_operations;
  68 static const struct inode_operations btrfs_file_inode_operations;
  69 static const struct address_space_operations btrfs_aops;
  70 static const struct file_operations btrfs_dir_file_operations;
  71 static const struct extent_io_ops btrfs_extent_io_ops;
  72
  73 static struct kmem_cache *btrfs_inode_cachep;
  74 struct kmem_cache *btrfs_trans_handle_cachep;
  75 struct kmem_cache *btrfs_path_cachep;
  76 struct kmem_cache *btrfs_free_space_cachep;
  77 struct kmem_cache *btrfs_free_space_bitmap_cachep;
  78
  79 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  80 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
  81 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  82 static noinline int cow_file_range(struct inode *inode,
  83                                    struct page *locked_page,
  84                                    u64 start, u64 end, int *page_started,
  85                                    unsigned long *nr_written, int unlock);
  86 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
  87                                        u64 orig_start, u64 block_start,
  88                                        u64 block_len, u64 orig_block_len,
  89                                        u64 ram_bytes, int compress_type,
  90                                        int type);
  91
  92 static void __endio_write_update_ordered(struct inode *inode,
  93                                          const u64 offset, const u64 bytes,
  94                                          const bool uptodate);
  95
  96 /*
  97  * Cleanup all submitted ordered extents in specified range to handle errors
  98  * from the btrfs_run_delalloc_range() callback.
  99  *
 100  * NOTE: caller must ensure that when an error happens, it can not call
 101  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 102  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 103  * to be released, which we want to happen only when finishing the ordered
 104  * extent (btrfs_finish_ordered_io()).
 105  */
 106 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 107                                                  struct page *locked_page,
 108                                                  u64 offset, u64 bytes)
 109 {
 110         unsigned long index = offset >> PAGE_SHIFT;
 111         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 112         u64 page_start = page_offset(locked_page);
 113         u64 page_end = page_start + PAGE_SIZE - 1;
 114
 115         struct page *page;
 116
 117         while (index <= end_index) {
 118                 page = find_get_page(inode->i_mapping, index);
 119                 index++;
 120                 if (!page)
 121                         continue;
 122                 ClearPagePrivate2(page);
 123                 put_page(page);
 124         }
 125
 126         /*
 127          * In case this page belongs to the delalloc range being instantiated
 128          * then skip it, since the first page of a range is going to be
 129          * properly cleaned up by the caller of run_delalloc_range
 130          */
 131         if (page_start >= offset && page_end <= (offset + bytes - 1)) {
 132                 offset += PAGE_SIZE;
 133                 bytes -= PAGE_SIZE;
 134         }
 135
 136         return __endio_write_update_ordered(inode, offset, bytes, false);
 137 }
 138
 139 static int btrfs_dirty_inode(struct inode *inode);
 140
 141 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 142 void btrfs_test_inode_set_ops(struct inode *inode)
 143 {
 144         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 145 }
 146 #endif
 147
 148 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 149                                      struct inode *inode,  struct inode *dir,
 150                                      const struct qstr *qstr)
 151 {
 152         int err;
 153
 154         err = btrfs_init_acl(trans, inode, dir);
 155         if (!err)
 156                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 157         return err;
 158 }
 159
 160 /*
 161  * this does all the hard work for inserting an inline extent into
 162  * the btree.  The caller should have done a btrfs_drop_extents so that
 163  * no overlapping inline items exist in the btree
 164  */
 165 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 166                                 struct btrfs_path *path, int extent_inserted,
 167                                 struct btrfs_root *root, struct inode *inode,
 168                                 u64 start, size_t size, size_t compressed_size,
 169                                 int compress_type,
 170                                 struct page **compressed_pages)
 171 {
 172         struct extent_buffer *leaf;
 173         struct page *page = NULL;
 174         char *kaddr;
 175         unsigned long ptr;
 176         struct btrfs_file_extent_item *ei;
 177         int ret;
 178         size_t cur_size = size;
 179         unsigned long offset;
 180
 181         ASSERT((compressed_size > 0 && compressed_pages) ||
 182                (compressed_size == 0 && !compressed_pages));
 183
 184         if (compressed_size && compressed_pages)
 185                 cur_size = compressed_size;
 186
 187         inode_add_bytes(inode, size);
 188
 189         if (!extent_inserted) {
 190                 struct btrfs_key key;
 191                 size_t datasize;
 192
 193                 key.objectid = btrfs_ino(BTRFS_I(inode));
 194                 key.offset = start;
 195                 key.type = BTRFS_EXTENT_DATA_KEY;
 196
 197                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 198                 path->leave_spinning = 1;
 199                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 200                                               datasize);
 201                 if (ret)
 202                         goto fail;
 203         }
 204         leaf = path->nodes[0];
 205         ei = btrfs_item_ptr(leaf, path->slots[0],
 206                             struct btrfs_file_extent_item);
 207         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 208         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 209         btrfs_set_file_extent_encryption(leaf, ei, 0);
 210         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 211         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 212         ptr = btrfs_file_extent_inline_start(ei);
 213
 214         if (compress_type != BTRFS_COMPRESS_NONE) {
 215                 struct page *cpage;
 216                 int i = 0;
 217                 while (compressed_size > 0) {
 218                         cpage = compressed_pages[i];
 219                         cur_size = min_t(unsigned long, compressed_size,
 220                                        PAGE_SIZE);
 221
 222                         kaddr = kmap_atomic(cpage);
 223                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 224                         kunmap_atomic(kaddr);
 225
 226                         i++;
 227                         ptr += cur_size;
 228                         compressed_size -= cur_size;
 229                 }
 230                 btrfs_set_file_extent_compression(leaf, ei,
 231                                                   compress_type);
 232         } else {
 233                 page = find_get_page(inode->i_mapping,
 234                                      start >> PAGE_SHIFT);
 235                 btrfs_set_file_extent_compression(leaf, ei, 0);
 236                 kaddr = kmap_atomic(page);
 237                 offset = offset_in_page(start);
 238                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 239                 kunmap_atomic(kaddr);
 240                 put_page(page);
 241         }
 242         btrfs_mark_buffer_dirty(leaf);
 243         btrfs_release_path(path);
 244
 245         /*
 246          * We align size to sectorsize for inline extents just for simplicity
 247          * sake.
 248          */
 249         size = ALIGN(size, root->fs_info->sectorsize);
 250         ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode), start, size);
 251         if (ret)
 252                 goto fail;
 253
 254         /*
 255          * we're an inline extent, so nobody can
 256          * extend the file past i_size without locking
 257          * a page we already have locked.
 258          *
 259          * We must do any isize and inode updates
 260          * before we unlock the pages.  Otherwise we
 261          * could end up racing with unlink.
 262          */
 263         BTRFS_I(inode)->disk_i_size = inode->i_size;
 264         ret = btrfs_update_inode(trans, root, inode);
 265
 266 fail:
 267         return ret;
 268 }
 269
 270
 271 /*
 272  * conditionally insert an inline extent into the file.  This
 273  * does the checks required to make sure the data is small enough
 274  * to fit as an inline extent.
 275  */
 276 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 277                                           u64 end, size_t compressed_size,
 278                                           int compress_type,
 279                                           struct page **compressed_pages)
 280 {
 281         struct btrfs_root *root = BTRFS_I(inode)->root;
 282         struct btrfs_fs_info *fs_info = root->fs_info;
 283         struct btrfs_trans_handle *trans;
 284         u64 isize = i_size_read(inode);
 285         u64 actual_end = min(end + 1, isize);
 286         u64 inline_len = actual_end - start;
 287         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 288         u64 data_len = inline_len;
 289         int ret;
 290         struct btrfs_path *path;
 291         int extent_inserted = 0;
 292         u32 extent_item_size;
 293
 294         if (compressed_size)
 295                 data_len = compressed_size;
 296
 297         if (start > 0 ||
 298             actual_end > fs_info->sectorsize ||
 299             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 300             (!compressed_size &&
 301             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 302             end + 1 < isize ||
 303             data_len > fs_info->max_inline) {
 304                 return 1;
 305         }
 306
 307         path = btrfs_alloc_path();
 308         if (!path)
 309                 return -ENOMEM;
 310
 311         trans = btrfs_join_transaction(root);
 312         if (IS_ERR(trans)) {
 313                 btrfs_free_path(path);
 314                 return PTR_ERR(trans);
 315         }
 316         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 317
 318         if (compressed_size && compressed_pages)
 319                 extent_item_size = btrfs_file_extent_calc_inline_size(
 320                    compressed_size);
 321         else
 322                 extent_item_size = btrfs_file_extent_calc_inline_size(
 323                     inline_len);
 324
 325         ret = __btrfs_drop_extents(trans, root, inode, path,
 326                                    start, aligned_end, NULL,
 327                                    1, 1, extent_item_size, &extent_inserted);
 328         if (ret) {
 329                 btrfs_abort_transaction(trans, ret);
 330                 goto out;
 331         }
 332
 333         if (isize > actual_end)
 334                 inline_len = min_t(u64, isize, actual_end);
 335         ret = insert_inline_extent(trans, path, extent_inserted,
 336                                    root, inode, start,
 337                                    inline_len, compressed_size,
 338                                    compress_type, compressed_pages);
 339         if (ret && ret != -ENOSPC) {
 340                 btrfs_abort_transaction(trans, ret);
 341                 goto out;
 342         } else if (ret == -ENOSPC) {
 343                 ret = 1;
 344                 goto out;
 345         }
 346
 347         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 348         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 349 out:
 350         /*
 351          * Don't forget to free the reserved space, as for inlined extent
 352          * it won't count as data extent, free them directly here.
 353          * And at reserve time, it's always aligned to page size, so
 354          * just free one page here.
 355          */
 356         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 357         btrfs_free_path(path);
 358         btrfs_end_transaction(trans);
 359         return ret;
 360 }
 361
 362 struct async_extent {
 363         u64 start;
 364         u64 ram_size;
 365         u64 compressed_size;
 366         struct page **pages;
 367         unsigned long nr_pages;
 368         int compress_type;
 369         struct list_head list;
 370 };
 371
 372 struct async_chunk {
 373         struct inode *inode;
 374         struct page *locked_page;
 375         u64 start;
 376         u64 end;
 377         unsigned int write_flags;
 378         struct list_head extents;
 379         struct cgroup_subsys_state *blkcg_css;
 380         struct btrfs_work work;
 381         atomic_t *pending;
 382 };
 383
 384 struct async_cow {
 385         /* Number of chunks in flight; must be first in the structure */
 386         atomic_t num_chunks;
 387         struct async_chunk chunks[];
 388 };
 389
 390 static noinline int add_async_extent(struct async_chunk *cow,
 391                                      u64 start, u64 ram_size,
 392                                      u64 compressed_size,
 393                                      struct page **pages,
 394                                      unsigned long nr_pages,
 395                                      int compress_type)
 396 {
 397         struct async_extent *async_extent;
 398
 399         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 400         BUG_ON(!async_extent); /* -ENOMEM */
 401         async_extent->start = start;
 402         async_extent->ram_size = ram_size;
 403         async_extent->compressed_size = compressed_size;
 404         async_extent->pages = pages;
 405         async_extent->nr_pages = nr_pages;
 406         async_extent->compress_type = compress_type;
 407         list_add_tail(&async_extent->list, &cow->extents);
 408         return 0;
 409 }
 410
 411 /*
 412  * Check if the inode has flags compatible with compression
 413  */
 414 static inline bool inode_can_compress(struct inode *inode)
 415 {
 416         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW ||
 417             BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
 418                 return false;
 419         return true;
 420 }
 421
 422 /*
 423  * Check if the inode needs to be submitted to compression, based on mount
 424  * options, defragmentation, properties or heuristics.
 425  */
 426 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 427 {
 428         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 429
 430         if (!inode_can_compress(inode)) {
 431                 WARN(IS_ENABLED(CONFIG_BTRFS_DEBUG),
 432                         KERN_ERR "BTRFS: unexpected compression for ino %llu\n",
 433                         btrfs_ino(BTRFS_I(inode)));
 434                 return 0;
 435         }
 436         /* force compress */
 437         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 438                 return 1;
 439         /* defrag ioctl */
 440         if (BTRFS_I(inode)->defrag_compress)
 441                 return 1;
 442         /* bad compression ratios */
 443         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 444                 return 0;
 445         if (btrfs_test_opt(fs_info, COMPRESS) ||
 446             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 447             BTRFS_I(inode)->prop_compress)
 448                 return btrfs_compress_heuristic(inode, start, end);
 449         return 0;
 450 }
 451
 452 static inline void inode_should_defrag(struct btrfs_inode *inode,
 453                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 454 {
 455         /* If this is a small write inside eof, kick off a defrag */
 456         if (num_bytes < small_write &&
 457             (start > 0 || end + 1 < inode->disk_i_size))
 458                 btrfs_add_inode_defrag(NULL, inode);
 459 }
 460
 461 /*
 462  * we create compressed extents in two phases.  The first
 463  * phase compresses a range of pages that have already been
 464  * locked (both pages and state bits are locked).
 465  *
 466  * This is done inside an ordered work queue, and the compression
 467  * is spread across many cpus.  The actual IO submission is step
 468  * two, and the ordered work queue takes care of making sure that
 469  * happens in the same order things were put onto the queue by
 470  * writepages and friends.
 471  *
 472  * If this code finds it can't get good compression, it puts an
 473  * entry onto the work queue to write the uncompressed bytes.  This
 474  * makes sure that both compressed inodes and uncompressed inodes
 475  * are written in the same order that the flusher thread sent them
 476  * down.
 477  */
 478 static noinline int compress_file_range(struct async_chunk *async_chunk)
 479 {
 480         struct inode *inode = async_chunk->inode;
 481         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 482         u64 blocksize = fs_info->sectorsize;
 483         u64 start = async_chunk->start;
 484         u64 end = async_chunk->end;
 485         u64 actual_end;
 486         u64 i_size;
 487         int ret = 0;
 488         struct page **pages = NULL;
 489         unsigned long nr_pages;
 490         unsigned long total_compressed = 0;
 491         unsigned long total_in = 0;
 492         int i;
 493         int will_compress;
 494         int compress_type = fs_info->compress_type;
 495         int compressed_extents = 0;
 496         int redirty = 0;
 497
 498         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 499                         SZ_16K);
 500
 501         /*
 502          * We need to save i_size before now because it could change in between
 503          * us evaluating the size and assigning it.  This is because we lock and
 504          * unlock the page in truncate and fallocate, and then modify the i_size
 505          * later on.
 506          *
 507          * The barriers are to emulate READ_ONCE, remove that once i_size_read
 508          * does that for us.
 509          */
 510         barrier();
 511         i_size = i_size_read(inode);
 512         barrier();
 513         actual_end = min_t(u64, i_size, end + 1);
 514 again:
 515         will_compress = 0;
 516         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 517         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 518         nr_pages = min_t(unsigned long, nr_pages,
 519                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 520
 521         /*
 522          * we don't want to send crud past the end of i_size through
 523          * compression, that's just a waste of CPU time.  So, if the
 524          * end of the file is before the start of our current
 525          * requested range of bytes, we bail out to the uncompressed
 526          * cleanup code that can deal with all of this.
 527          *
 528          * It isn't really the fastest way to fix things, but this is a
 529          * very uncommon corner.
 530          */
 531         if (actual_end <= start)
 532                 goto cleanup_and_bail_uncompressed;
 533
 534         total_compressed = actual_end - start;
 535
 536         /*
 537          * skip compression for a small file range(<=blocksize) that
 538          * isn't an inline extent, since it doesn't save disk space at all.
 539          */
 540         if (total_compressed <= blocksize &&
 541            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 542                 goto cleanup_and_bail_uncompressed;
 543
 544         total_compressed = min_t(unsigned long, total_compressed,
 545                         BTRFS_MAX_UNCOMPRESSED);
 546         total_in = 0;
 547         ret = 0;
 548
 549         /*
 550          * we do compression for mount -o compress and when the
 551          * inode has not been flagged as nocompress.  This flag can
 552          * change at any time if we discover bad compression ratios.
 553          */
 554         if (inode_need_compress(inode, start, end)) {
 555                 WARN_ON(pages);
 556                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 557                 if (!pages) {
 558                         /* just bail out to the uncompressed code */
 559                         nr_pages = 0;
 560                         goto cont;
 561                 }
 562
 563                 if (BTRFS_I(inode)->defrag_compress)
 564                         compress_type = BTRFS_I(inode)->defrag_compress;
 565                 else if (BTRFS_I(inode)->prop_compress)
 566                         compress_type = BTRFS_I(inode)->prop_compress;
 567
 568                 /*
 569                  * we need to call clear_page_dirty_for_io on each
 570                  * page in the range.  Otherwise applications with the file
 571                  * mmap'd can wander in and change the page contents while
 572                  * we are compressing them.
 573                  *
 574                  * If the compression fails for any reason, we set the pages
 575                  * dirty again later on.
 576                  *
 577                  * Note that the remaining part is redirtied, the start pointer
 578                  * has moved, the end is the original one.
 579                  */
 580                 if (!redirty) {
 581                         extent_range_clear_dirty_for_io(inode, start, end);
 582                         redirty = 1;
 583                 }
 584
 585                 /* Compression level is applied here and only here */
 586                 ret = btrfs_compress_pages(
 587                         compress_type | (fs_info->compress_level << 4),
 588                                            inode->i_mapping, start,
 589                                            pages,
 590                                            &nr_pages,
 591                                            &total_in,
 592                                            &total_compressed);
 593
 594                 if (!ret) {
 595                         unsigned long offset = offset_in_page(total_compressed);
 596                         struct page *page = pages[nr_pages - 1];
 597                         char *kaddr;
 598
 599                         /* zero the tail end of the last page, we might be
 600                          * sending it down to disk
 601                          */
 602                         if (offset) {
 603                                 kaddr = kmap_atomic(page);
 604                                 memset(kaddr + offset, 0,
 605                                        PAGE_SIZE - offset);
 606                                 kunmap_atomic(kaddr);
 607                         }
 608                         will_compress = 1;
 609                 }
 610         }
 611 cont:
 612         if (start == 0) {
 613                 /* lets try to make an inline extent */
 614                 if (ret || total_in < actual_end) {
 615                         /* we didn't compress the entire range, try
 616                          * to make an uncompressed inline extent.
 617                          */
 618                         ret = cow_file_range_inline(inode, start, end, 0,
 619                                                     BTRFS_COMPRESS_NONE, NULL);
 620                 } else {
 621                         /* try making a compressed inline extent */
 622                         ret = cow_file_range_inline(inode, start, end,
 623                                                     total_compressed,
 624                                                     compress_type, pages);
 625                 }
 626                 if (ret <= 0) {
 627                         unsigned long clear_flags = EXTENT_DELALLOC |
 628                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 629                                 EXTENT_DO_ACCOUNTING;
 630                         unsigned long page_error_op;
 631
 632                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 633
 634                         /*
 635                          * inline extent creation worked or returned error,
 636                          * we don't need to create any more async work items.
 637                          * Unlock and free up our temp pages.
 638                          *
 639                          * We use DO_ACCOUNTING here because we need the
 640                          * delalloc_release_metadata to be done _after_ we drop
 641                          * our outstanding extent for clearing delalloc for this
 642                          * range.
 643                          */
 644                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 645                                                      clear_flags,
 646                                                      PAGE_UNLOCK |
 647                                                      PAGE_CLEAR_DIRTY |
 648                                                      PAGE_SET_WRITEBACK |
 649                                                      page_error_op |
 650                                                      PAGE_END_WRITEBACK);
 651
 652                         for (i = 0; i < nr_pages; i++) {
 653                                 WARN_ON(pages[i]->mapping);
 654                                 put_page(pages[i]);
 655                         }
 656                         kfree(pages);
 657
 658                         return 0;
 659                 }
 660         }
 661
 662         if (will_compress) {
 663                 /*
 664                  * we aren't doing an inline extent round the compressed size
 665                  * up to a block size boundary so the allocator does sane
 666                  * things
 667                  */
 668                 total_compressed = ALIGN(total_compressed, blocksize);
 669
 670                 /*
 671                  * one last check to make sure the compression is really a
 672                  * win, compare the page count read with the blocks on disk,
 673                  * compression must free at least one sector size
 674                  */
 675                 total_in = ALIGN(total_in, PAGE_SIZE);
 676                 if (total_compressed + blocksize <= total_in) {
 677                         compressed_extents++;
 678
 679                         /*
 680                          * The async work queues will take care of doing actual
 681                          * allocation on disk for these compressed pages, and
 682                          * will submit them to the elevator.
 683                          */
 684                         add_async_extent(async_chunk, start, total_in,
 685                                         total_compressed, pages, nr_pages,
 686                                         compress_type);
 687
 688                         if (start + total_in < end) {
 689                                 start += total_in;
 690                                 pages = NULL;
 691                                 cond_resched();
 692                                 goto again;
 693                         }
 694                         return compressed_extents;
 695                 }
 696         }
 697         if (pages) {
 698                 /*
 699                  * the compression code ran but failed to make things smaller,
 700                  * free any pages it allocated and our page pointer array
 701                  */
 702                 for (i = 0; i < nr_pages; i++) {
 703                         WARN_ON(pages[i]->mapping);
 704                         put_page(pages[i]);
 705                 }
 706                 kfree(pages);
 707                 pages = NULL;
 708                 total_compressed = 0;
 709                 nr_pages = 0;
 710
 711                 /* flag the file so we don't compress in the future */
 712                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 713                     !(BTRFS_I(inode)->prop_compress)) {
 714                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 715                 }
 716         }
 717 cleanup_and_bail_uncompressed:
 718         /*
 719          * No compression, but we still need to write the pages in the file
 720          * we've been given so far.  redirty the locked page if it corresponds
 721          * to our extent and set things up for the async work queue to run
 722          * cow_file_range to do the normal delalloc dance.
 723          */
 724         if (async_chunk->locked_page &&
 725             (page_offset(async_chunk->locked_page) >= start &&
 726              page_offset(async_chunk->locked_page)) <= end) {
 727                 __set_page_dirty_nobuffers(async_chunk->locked_page);
 728                 /* unlocked later on in the async handlers */
 729         }
 730
 731         if (redirty)
 732                 extent_range_redirty_for_io(inode, start, end);
 733         add_async_extent(async_chunk, start, end - start + 1, 0, NULL, 0,
 734                          BTRFS_COMPRESS_NONE);
 735         compressed_extents++;
 736
 737         return compressed_extents;
 738 }
 739
 740 static void free_async_extent_pages(struct async_extent *async_extent)
 741 {
 742         int i;
 743
 744         if (!async_extent->pages)
 745                 return;
 746
 747         for (i = 0; i < async_extent->nr_pages; i++) {
 748                 WARN_ON(async_extent->pages[i]->mapping);
 749                 put_page(async_extent->pages[i]);
 750         }
 751         kfree(async_extent->pages);
 752         async_extent->nr_pages = 0;
 753         async_extent->pages = NULL;
 754 }
 755
 756 /*
 757  * phase two of compressed writeback.  This is the ordered portion
 758  * of the code, which only gets called in the order the work was
 759  * queued.  We walk all the async extents created by compress_file_range
 760  * and send them down to the disk.
 761  */
 762 static noinline void submit_compressed_extents(struct async_chunk *async_chunk)
 763 {
 764         struct inode *inode = async_chunk->inode;
 765         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 766         struct async_extent *async_extent;
 767         u64 alloc_hint = 0;
 768         struct btrfs_key ins;
 769         struct extent_map *em;
 770         struct btrfs_root *root = BTRFS_I(inode)->root;
 771         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
 772         int ret = 0;
 773
 774 again:
 775         while (!list_empty(&async_chunk->extents)) {
 776                 async_extent = list_entry(async_chunk->extents.next,
 777                                           struct async_extent, list);
 778                 list_del(&async_extent->list);
 779
 780 retry:
 781                 lock_extent(io_tree, async_extent->start,
 782                             async_extent->start + async_extent->ram_size - 1);
 783                 /* did the compression code fall back to uncompressed IO? */
 784                 if (!async_extent->pages) {
 785                         int page_started = 0;
 786                         unsigned long nr_written = 0;
 787
 788                         /* allocate blocks */
 789                         ret = cow_file_range(inode, async_chunk->locked_page,
 790                                              async_extent->start,
 791                                              async_extent->start +
 792                                              async_extent->ram_size - 1,
 793                                              &page_started, &nr_written, 0);
 794
 795                         /* JDM XXX */
 796
 797                         /*
 798                          * if page_started, cow_file_range inserted an
 799                          * inline extent and took care of all the unlocking
 800                          * and IO for us.  Otherwise, we need to submit
 801                          * all those pages down to the drive.
 802                          */
 803                         if (!page_started && !ret)
 804                                 extent_write_locked_range(inode,
 805                                                   async_extent->start,
 806                                                   async_extent->start +
 807                                                   async_extent->ram_size - 1,
 808                                                   WB_SYNC_ALL);
 809                         else if (ret && async_chunk->locked_page)
 810                                 unlock_page(async_chunk->locked_page);
 811                         kfree(async_extent);
 812                         cond_resched();
 813                         continue;
 814                 }
 815
 816                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 817                                            async_extent->compressed_size,
 818                                            async_extent->compressed_size,
 819                                            0, alloc_hint, &ins, 1, 1);
 820                 if (ret) {
 821                         free_async_extent_pages(async_extent);
 822
 823                         if (ret == -ENOSPC) {
 824                                 unlock_extent(io_tree, async_extent->start,
 825                                               async_extent->start +
 826                                               async_extent->ram_size - 1);
 827
 828                                 /*
 829                                  * we need to redirty the pages if we decide to
 830                                  * fallback to uncompressed IO, otherwise we
 831                                  * will not submit these pages down to lower
 832                                  * layers.
 833                                  */
 834                                 extent_range_redirty_for_io(inode,
 835                                                 async_extent->start,
 836                                                 async_extent->start +
 837                                                 async_extent->ram_size - 1);
 838
 839                                 goto retry;
 840                         }
 841                         goto out_free;
 842                 }
 843                 /*
 844                  * here we're doing allocation and writeback of the
 845                  * compressed pages
 846                  */
 847                 em = create_io_em(inode, async_extent->start,
 848                                   async_extent->ram_size, /* len */
 849                                   async_extent->start, /* orig_start */
 850                                   ins.objectid, /* block_start */
 851                                   ins.offset, /* block_len */
 852                                   ins.offset, /* orig_block_len */
 853                                   async_extent->ram_size, /* ram_bytes */
 854                                   async_extent->compress_type,
 855                                   BTRFS_ORDERED_COMPRESSED);
 856                 if (IS_ERR(em))
 857                         /* ret value is not necessary due to void function */
 858                         goto out_free_reserve;
 859                 free_extent_map(em);
 860
 861                 ret = btrfs_add_ordered_extent_compress(inode,
 862                                                 async_extent->start,
 863                                                 ins.objectid,
 864                                                 async_extent->ram_size,
 865                                                 ins.offset,
 866                                                 BTRFS_ORDERED_COMPRESSED,
 867                                                 async_extent->compress_type);
 868                 if (ret) {
 869                         btrfs_drop_extent_cache(BTRFS_I(inode),
 870                                                 async_extent->start,
 871                                                 async_extent->start +
 872                                                 async_extent->ram_size - 1, 0);
 873                         goto out_free_reserve;
 874                 }
 875                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 876
 877                 /*
 878                  * clear dirty, set writeback and unlock the pages.
 879                  */
 880                 extent_clear_unlock_delalloc(inode, async_extent->start,
 881                                 async_extent->start +
 882                                 async_extent->ram_size - 1,
 883                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 884                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 885                                 PAGE_SET_WRITEBACK);
 886                 if (btrfs_submit_compressed_write(inode,
 887                                     async_extent->start,
 888                                     async_extent->ram_size,
 889                                     ins.objectid,
 890                                     ins.offset, async_extent->pages,
 891                                     async_extent->nr_pages,
 892                                     async_chunk->write_flags,
 893                                     async_chunk->blkcg_css)) {
 894                         struct page *p = async_extent->pages[0];
 895                         const u64 start = async_extent->start;
 896                         const u64 end = start + async_extent->ram_size - 1;
 897
 898                         p->mapping = inode->i_mapping;
 899                         btrfs_writepage_endio_finish_ordered(p, start, end, 0);
 900
 901                         p->mapping = NULL;
 902                         extent_clear_unlock_delalloc(inode, start, end,
 903                                                      NULL, 0,
 904                                                      PAGE_END_WRITEBACK |
 905                                                      PAGE_SET_ERROR);
 906                         free_async_extent_pages(async_extent);
 907                 }
 908                 alloc_hint = ins.objectid + ins.offset;
 909                 kfree(async_extent);
 910                 cond_resched();
 911         }
 912         return;
 913 out_free_reserve:
 914         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 915         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 916 out_free:
 917         extent_clear_unlock_delalloc(inode, async_extent->start,
 918                                      async_extent->start +
 919                                      async_extent->ram_size - 1,
 920                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 921                                      EXTENT_DELALLOC_NEW |
 922                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 923                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 924                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 925                                      PAGE_SET_ERROR);
 926         free_async_extent_pages(async_extent);
 927         kfree(async_extent);
 928         goto again;
 929 }
 930
 931 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 932                                       u64 num_bytes)
 933 {
 934         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 935         struct extent_map *em;
 936         u64 alloc_hint = 0;
 937
 938         read_lock(&em_tree->lock);
 939         em = search_extent_mapping(em_tree, start, num_bytes);
 940         if (em) {
 941                 /*
 942                  * if block start isn't an actual block number then find the
 943                  * first block in this inode and use that as a hint.  If that
 944                  * block is also bogus then just don't worry about it.
 945                  */
 946                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 947                         free_extent_map(em);
 948                         em = search_extent_mapping(em_tree, 0, 0);
 949                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 950                                 alloc_hint = em->block_start;
 951                         if (em)
 952                                 free_extent_map(em);
 953                 } else {
 954                         alloc_hint = em->block_start;
 955                         free_extent_map(em);
 956                 }
 957         }
 958         read_unlock(&em_tree->lock);
 959
 960         return alloc_hint;
 961 }
 962
 963 /*
 964  * when extent_io.c finds a delayed allocation range in the file,
 965  * the call backs end up in this code.  The basic idea is to
 966  * allocate extents on disk for the range, and create ordered data structs
 967  * in ram to track those extents.
 968  *
 969  * locked_page is the page that writepage had locked already.  We use
 970  * it to make sure we don't do extra locks or unlocks.
 971  *
 972  * *page_started is set to one if we unlock locked_page and do everything
 973  * required to start IO on it.  It may be clean and already done with
 974  * IO when we return.
 975  */
 976 static noinline int cow_file_range(struct inode *inode,
 977                                    struct page *locked_page,
 978                                    u64 start, u64 end, int *page_started,
 979                                    unsigned long *nr_written, int unlock)
 980 {
 981         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 982         struct btrfs_root *root = BTRFS_I(inode)->root;
 983         u64 alloc_hint = 0;
 984         u64 num_bytes;
 985         unsigned long ram_size;
 986         u64 cur_alloc_size = 0;
 987         u64 blocksize = fs_info->sectorsize;
 988         struct btrfs_key ins;
 989         struct extent_map *em;
 990         unsigned clear_bits;
 991         unsigned long page_ops;
 992         bool extent_reserved = false;
 993         int ret = 0;
 994
 995         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 996                 WARN_ON_ONCE(1);
 997                 ret = -EINVAL;
 998                 goto out_unlock;
 999         }
1000
1001         num_bytes = ALIGN(end - start + 1, blocksize);
1002         num_bytes = max(blocksize,  num_bytes);
1003         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
1004
1005         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
1006
1007         if (start == 0) {
1008                 /* lets try to make an inline extent */
1009                 ret = cow_file_range_inline(inode, start, end, 0,
1010                                             BTRFS_COMPRESS_NONE, NULL);
1011                 if (ret == 0) {
1012                         /*
1013                          * We use DO_ACCOUNTING here because we need the
1014                          * delalloc_release_metadata to be run _after_ we drop
1015                          * our outstanding extent for clearing delalloc for this
1016                          * range.
1017                          */
1018                         extent_clear_unlock_delalloc(inode, start, end, NULL,
1019                                      EXTENT_LOCKED | EXTENT_DELALLOC |
1020                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1021                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1022                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1023                                      PAGE_END_WRITEBACK);
1024                         *nr_written = *nr_written +
1025                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1026                         *page_started = 1;
1027                         goto out;
1028                 } else if (ret < 0) {
1029                         goto out_unlock;
1030                 }
1031         }
1032
1033         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1034         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1035                         start + num_bytes - 1, 0);
1036
1037         while (num_bytes > 0) {
1038                 cur_alloc_size = num_bytes;
1039                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1040                                            fs_info->sectorsize, 0, alloc_hint,
1041                                            &ins, 1, 1);
1042                 if (ret < 0)
1043                         goto out_unlock;
1044                 cur_alloc_size = ins.offset;
1045                 extent_reserved = true;
1046
1047                 ram_size = ins.offset;
1048                 em = create_io_em(inode, start, ins.offset, /* len */
1049                                   start, /* orig_start */
1050                                   ins.objectid, /* block_start */
1051                                   ins.offset, /* block_len */
1052                                   ins.offset, /* orig_block_len */
1053                                   ram_size, /* ram_bytes */
1054                                   BTRFS_COMPRESS_NONE, /* compress_type */
1055                                   BTRFS_ORDERED_REGULAR /* type */);
1056                 if (IS_ERR(em)) {
1057                         ret = PTR_ERR(em);
1058                         goto out_reserve;
1059                 }
1060                 free_extent_map(em);
1061
1062                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1063                                                ram_size, cur_alloc_size, 0);
1064                 if (ret)
1065                         goto out_drop_extent_cache;
1066
1067                 if (root->root_key.objectid ==
1068                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1069                         ret = btrfs_reloc_clone_csums(inode, start,
1070                                                       cur_alloc_size);
1071                         /*
1072                          * Only drop cache here, and process as normal.
1073                          *
1074                          * We must not allow extent_clear_unlock_delalloc()
1075                          * at out_unlock label to free meta of this ordered
1076                          * extent, as its meta should be freed by
1077                          * btrfs_finish_ordered_io().
1078                          *
1079                          * So we must continue until @start is increased to
1080                          * skip current ordered extent.
1081                          */
1082                         if (ret)
1083                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1084                                                 start + ram_size - 1, 0);
1085                 }
1086
1087                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1088
1089                 /* we're not doing compressed IO, don't unlock the first
1090                  * page (which the caller expects to stay locked), don't
1091                  * clear any dirty bits and don't set any writeback bits
1092                  *
1093                  * Do set the Private2 bit so we know this page was properly
1094                  * setup for writepage
1095                  */
1096                 page_ops = unlock ? PAGE_UNLOCK : 0;
1097                 page_ops |= PAGE_SET_PRIVATE2;
1098
1099                 extent_clear_unlock_delalloc(inode, start,
1100                                              start + ram_size - 1,
1101                                              locked_page,
1102                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1103                                              page_ops);
1104                 if (num_bytes < cur_alloc_size)
1105                         num_bytes = 0;
1106                 else
1107                         num_bytes -= cur_alloc_size;
1108                 alloc_hint = ins.objectid + ins.offset;
1109                 start += cur_alloc_size;
1110                 extent_reserved = false;
1111
1112                 /*
1113                  * btrfs_reloc_clone_csums() error, since start is increased
1114                  * extent_clear_unlock_delalloc() at out_unlock label won't
1115                  * free metadata of current ordered extent, we're OK to exit.
1116                  */
1117                 if (ret)
1118                         goto out_unlock;
1119         }
1120 out:
1121         return ret;
1122
1123 out_drop_extent_cache:
1124         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1125 out_reserve:
1126         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1127         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1128 out_unlock:
1129         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1130                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1131         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1132                 PAGE_END_WRITEBACK;
1133         /*
1134          * If we reserved an extent for our delalloc range (or a subrange) and
1135          * failed to create the respective ordered extent, then it means that
1136          * when we reserved the extent we decremented the extent's size from
1137          * the data space_info's bytes_may_use counter and incremented the
1138          * space_info's bytes_reserved counter by the same amount. We must make
1139          * sure extent_clear_unlock_delalloc() does not try to decrement again
1140          * the data space_info's bytes_may_use counter, therefore we do not pass
1141          * it the flag EXTENT_CLEAR_DATA_RESV.
1142          */
1143         if (extent_reserved) {
1144                 extent_clear_unlock_delalloc(inode, start,
1145                                              start + cur_alloc_size - 1,
1146                                              locked_page,
1147                                              clear_bits,
1148                                              page_ops);
1149                 start += cur_alloc_size;
1150                 if (start >= end)
1151                         goto out;
1152         }
1153         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1154                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1155                                      page_ops);
1156         goto out;
1157 }
1158
1159 /*
1160  * work queue call back to started compression on a file and pages
1161  */
1162 static noinline void async_cow_start(struct btrfs_work *work)
1163 {
1164         struct async_chunk *async_chunk;
1165         int compressed_extents;
1166
1167         async_chunk = container_of(work, struct async_chunk, work);
1168
1169         compressed_extents = compress_file_range(async_chunk);
1170         if (compressed_extents == 0) {
1171                 btrfs_add_delayed_iput(async_chunk->inode);
1172                 async_chunk->inode = NULL;
1173         }
1174 }
1175
1176 /*
1177  * work queue call back to submit previously compressed pages
1178  */
1179 static noinline void async_cow_submit(struct btrfs_work *work)
1180 {
1181         struct async_chunk *async_chunk = container_of(work, struct async_chunk,
1182                                                      work);
1183         struct btrfs_fs_info *fs_info = btrfs_work_owner(work);
1184         unsigned long nr_pages;
1185
1186         nr_pages = (async_chunk->end - async_chunk->start + PAGE_SIZE) >>
1187                 PAGE_SHIFT;
1188
1189         /* atomic_sub_return implies a barrier */
1190         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1191             5 * SZ_1M)
1192                 cond_wake_up_nomb(&fs_info->async_submit_wait);
1193
1194         /*
1195          * ->inode could be NULL if async_chunk_start has failed to compress,
1196          * in which case we don't have anything to submit, yet we need to
1197          * always adjust ->async_delalloc_pages as its paired with the init
1198          * happening in cow_file_range_async
1199          */
1200         if (async_chunk->inode)
1201                 submit_compressed_extents(async_chunk);
1202 }
1203
1204 static noinline void async_cow_free(struct btrfs_work *work)
1205 {
1206         struct async_chunk *async_chunk;
1207
1208         async_chunk = container_of(work, struct async_chunk, work);
1209         if (async_chunk->inode)
1210                 btrfs_add_delayed_iput(async_chunk->inode);
1211         if (async_chunk->blkcg_css)
1212                 css_put(async_chunk->blkcg_css);
1213         /*
1214          * Since the pointer to 'pending' is at the beginning of the array of
1215          * async_chunk's, freeing it ensures the whole array has been freed.
1216          */
1217         if (atomic_dec_and_test(async_chunk->pending))
1218                 kvfree(async_chunk->pending);
1219 }
1220
1221 static int cow_file_range_async(struct inode *inode,
1222                                 struct writeback_control *wbc,
1223                                 struct page *locked_page,
1224                                 u64 start, u64 end, int *page_started,
1225                                 unsigned long *nr_written)
1226 {
1227         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1228         struct cgroup_subsys_state *blkcg_css = wbc_blkcg_css(wbc);
1229         struct async_cow *ctx;
1230         struct async_chunk *async_chunk;
1231         unsigned long nr_pages;
1232         u64 cur_end;
1233         u64 num_chunks = DIV_ROUND_UP(end - start, SZ_512K);
1234         int i;
1235         bool should_compress;
1236         unsigned nofs_flag;
1237         const unsigned int write_flags = wbc_to_write_flags(wbc);
1238
1239         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1240
1241         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1242             !btrfs_test_opt(fs_info, FORCE_COMPRESS)) {
1243                 num_chunks = 1;
1244                 should_compress = false;
1245         } else {
1246                 should_compress = true;
1247         }
1248
1249         nofs_flag = memalloc_nofs_save();
1250         ctx = kvmalloc(struct_size(ctx, chunks, num_chunks), GFP_KERNEL);
1251         memalloc_nofs_restore(nofs_flag);
1252
1253         if (!ctx) {
1254                 unsigned clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC |
1255                         EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1256                         EXTENT_DO_ACCOUNTING;
1257                 unsigned long page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1258                         PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
1259                         PAGE_SET_ERROR;
1260
1261                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1262                                              clear_bits, page_ops);
1263                 return -ENOMEM;
1264         }
1265
1266         async_chunk = ctx->chunks;
1267         atomic_set(&ctx->num_chunks, num_chunks);
1268
1269         for (i = 0; i < num_chunks; i++) {
1270                 if (should_compress)
1271                         cur_end = min(end, start + SZ_512K - 1);
1272                 else
1273                         cur_end = end;
1274
1275                 /*
1276                  * igrab is called higher up in the call chain, take only the
1277                  * lightweight reference for the callback lifetime
1278                  */
1279                 ihold(inode);
1280                 async_chunk[i].pending = &ctx->num_chunks;
1281                 async_chunk[i].inode = inode;
1282                 async_chunk[i].start = start;
1283                 async_chunk[i].end = cur_end;
1284                 async_chunk[i].write_flags = write_flags;
1285                 INIT_LIST_HEAD(&async_chunk[i].extents);
1286
1287                 /*
1288                  * The locked_page comes all the way from writepage and its
1289                  * the original page we were actually given.  As we spread
1290                  * this large delalloc region across multiple async_chunk
1291                  * structs, only the first struct needs a pointer to locked_page
1292                  *
1293                  * This way we don't need racey decisions about who is supposed
1294                  * to unlock it.
1295                  */
1296                 if (locked_page) {
1297                         /*
1298                          * Depending on the compressibility, the pages might or
1299                          * might not go through async.  We want all of them to
1300                          * be accounted against wbc once.  Let's do it here
1301                          * before the paths diverge.  wbc accounting is used
1302                          * only for foreign writeback detection and doesn't
1303                          * need full accuracy.  Just account the whole thing
1304                          * against the first page.
1305                          */
1306                         wbc_account_cgroup_owner(wbc, locked_page,
1307                                                  cur_end - start);
1308                         async_chunk[i].locked_page = locked_page;
1309                         locked_page = NULL;
1310                 } else {
1311                         async_chunk[i].locked_page = NULL;
1312                 }
1313
1314                 if (blkcg_css != blkcg_root_css) {
1315                         css_get(blkcg_css);
1316                         async_chunk[i].blkcg_css = blkcg_css;
1317                 } else {
1318                         async_chunk[i].blkcg_css = NULL;
1319                 }
1320
1321                 btrfs_init_work(&async_chunk[i].work, async_cow_start,
1322                                 async_cow_submit, async_cow_free);
1323
1324                 nr_pages = DIV_ROUND_UP(cur_end - start, PAGE_SIZE);
1325                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1326
1327                 btrfs_queue_work(fs_info->delalloc_workers, &async_chunk[i].work);
1328
1329                 *nr_written += nr_pages;
1330                 start = cur_end + 1;
1331         }
1332         *page_started = 1;
1333         return 0;
1334 }
1335
1336 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1337                                         u64 bytenr, u64 num_bytes)
1338 {
1339         int ret;
1340         struct btrfs_ordered_sum *sums;
1341         LIST_HEAD(list);
1342
1343         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1344                                        bytenr + num_bytes - 1, &list, 0);
1345         if (ret == 0 && list_empty(&list))
1346                 return 0;
1347
1348         while (!list_empty(&list)) {
1349                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1350                 list_del(&sums->list);
1351                 kfree(sums);
1352         }
1353         if (ret < 0)
1354                 return ret;
1355         return 1;
1356 }
1357
1358 static int fallback_to_cow(struct inode *inode, struct page *locked_page,
1359                            const u64 start, const u64 end,
1360                            int *page_started, unsigned long *nr_written)
1361 {
1362         const bool is_space_ino = btrfs_is_free_space_inode(BTRFS_I(inode));
1363         const u64 range_bytes = end + 1 - start;
1364         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1365         u64 range_start = start;
1366         u64 count;
1367
1368         /*
1369          * If EXTENT_NORESERVE is set it means that when the buffered write was
1370          * made we had not enough available data space and therefore we did not
1371          * reserve data space for it, since we though we could do NOCOW for the
1372          * respective file range (either there is prealloc extent or the inode
1373          * has the NOCOW bit set).
1374          *
1375          * However when we need to fallback to COW mode (because for example the
1376          * block group for the corresponding extent was turned to RO mode by a
1377          * scrub or relocation) we need to do the following:
1378          *
1379          * 1) We increment the bytes_may_use counter of the data space info.
1380          *    If COW succeeds, it allocates a new data extent and after doing
1381          *    that it decrements the space info's bytes_may_use counter and
1382          *    increments its bytes_reserved counter by the same amount (we do
1383          *    this at btrfs_add_reserved_bytes()). So we need to increment the
1384          *    bytes_may_use counter to compensate (when space is reserved at
1385          *    buffered write time, the bytes_may_use counter is incremented);
1386          *
1387          * 2) We clear the EXTENT_NORESERVE bit from the range. We do this so
1388          *    that if the COW path fails for any reason, it decrements (through
1389          *    extent_clear_unlock_delalloc()) the bytes_may_use counter of the
1390          *    data space info, which we incremented in the step above.
1391          *
1392          * If we need to fallback to cow and the inode corresponds to a free
1393          * space cache inode, we must also increment bytes_may_use of the data
1394          * space_info for the same reason. Space caches always get a prealloc
1395          * extent for them, however scrub or balance may have set the block
1396          * group that contains that extent to RO mode.
1397          */
1398         count = count_range_bits(io_tree, &range_start, end, range_bytes,
1399                                  EXTENT_NORESERVE, 0);
1400         if (count > 0 || is_space_ino) {
1401                 const u64 bytes = is_space_ino ? range_bytes : count;
1402                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1403                 struct btrfs_space_info *sinfo = fs_info->data_sinfo;
1404
1405                 spin_lock(&sinfo->lock);
1406                 btrfs_space_info_update_bytes_may_use(fs_info, sinfo, bytes);
1407                 spin_unlock(&sinfo->lock);
1408
1409                 if (count > 0)
1410                         clear_extent_bit(io_tree, start, end, EXTENT_NORESERVE,
1411                                          0, 0, NULL);
1412         }
1413
1414         return cow_file_range(inode, locked_page, start, end, page_started,
1415                               nr_written, 1);
1416 }
1417
1418 /*
1419  * when nowcow writeback call back.  This checks for snapshots or COW copies
1420  * of the extents that exist in the file, and COWs the file as required.
1421  *
1422  * If no cow copies or snapshots exist, we write directly to the existing
1423  * blocks on disk
1424  */
1425 static noinline int run_delalloc_nocow(struct inode *inode,
1426                                        struct page *locked_page,
1427                                        const u64 start, const u64 end,
1428                                        int *page_started, int force,
1429                                        unsigned long *nr_written)
1430 {
1431         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1432         struct btrfs_root *root = BTRFS_I(inode)->root;
1433         struct btrfs_path *path;
1434         u64 cow_start = (u64)-1;
1435         u64 cur_offset = start;
1436         int ret;
1437         bool check_prev = true;
1438         const bool freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
1439         u64 ino = btrfs_ino(BTRFS_I(inode));
1440         bool nocow = false;
1441         u64 disk_bytenr = 0;
1442
1443         path = btrfs_alloc_path();
1444         if (!path) {
1445                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1446                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1447                                              EXTENT_DO_ACCOUNTING |
1448                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1449                                              PAGE_CLEAR_DIRTY |
1450                                              PAGE_SET_WRITEBACK |
1451                                              PAGE_END_WRITEBACK);
1452                 return -ENOMEM;
1453         }
1454
1455         while (1) {
1456                 struct btrfs_key found_key;
1457                 struct btrfs_file_extent_item *fi;
1458                 struct extent_buffer *leaf;
1459                 u64 extent_end;
1460                 u64 extent_offset;
1461                 u64 num_bytes = 0;
1462                 u64 disk_num_bytes;
1463                 u64 ram_bytes;
1464                 int extent_type;
1465
1466                 nocow = false;
1467
1468                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1469                                                cur_offset, 0);
1470                 if (ret < 0)
1471                         goto error;
1472
1473                 /*
1474                  * If there is no extent for our range when doing the initial
1475                  * search, then go back to the previous slot as it will be the
1476                  * one containing the search offset
1477                  */
1478                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1479                         leaf = path->nodes[0];
1480                         btrfs_item_key_to_cpu(leaf, &found_key,
1481                                               path->slots[0] - 1);
1482                         if (found_key.objectid == ino &&
1483                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1484                                 path->slots[0]--;
1485                 }
1486                 check_prev = false;
1487 next_slot:
1488                 /* Go to next leaf if we have exhausted the current one */
1489                 leaf = path->nodes[0];
1490                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1491                         ret = btrfs_next_leaf(root, path);
1492                         if (ret < 0) {
1493                                 if (cow_start != (u64)-1)
1494                                         cur_offset = cow_start;
1495                                 goto error;
1496                         }
1497                         if (ret > 0)
1498                                 break;
1499                         leaf = path->nodes[0];
1500                 }
1501
1502                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1503
1504                 /* Didn't find anything for our INO */
1505                 if (found_key.objectid > ino)
1506                         break;
1507                 /*
1508                  * Keep searching until we find an EXTENT_ITEM or there are no
1509                  * more extents for this inode
1510                  */
1511                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1512                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1513                         path->slots[0]++;
1514                         goto next_slot;
1515                 }
1516
1517                 /* Found key is not EXTENT_DATA_KEY or starts after req range */
1518                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1519                     found_key.offset > end)
1520                         break;
1521
1522                 /*
1523                  * If the found extent starts after requested offset, then
1524                  * adjust extent_end to be right before this extent begins
1525                  */
1526                 if (found_key.offset > cur_offset) {
1527                         extent_end = found_key.offset;
1528                         extent_type = 0;
1529                         goto out_check;
1530                 }
1531
1532                 /*
1533                  * Found extent which begins before our range and potentially
1534                  * intersect it
1535                  */
1536                 fi = btrfs_item_ptr(leaf, path->slots[0],
1537                                     struct btrfs_file_extent_item);
1538                 extent_type = btrfs_file_extent_type(leaf, fi);
1539
1540                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1541                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1542                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1543                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1544                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1545                         extent_end = found_key.offset +
1546                                 btrfs_file_extent_num_bytes(leaf, fi);
1547                         disk_num_bytes =
1548                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1549                         /*
1550                          * If the extent we got ends before our current offset,
1551                          * skip to the next extent.
1552                          */
1553                         if (extent_end <= cur_offset) {
1554                                 path->slots[0]++;
1555                                 goto next_slot;
1556                         }
1557                         /* Skip holes */
1558                         if (disk_bytenr == 0)
1559                                 goto out_check;
1560                         /* Skip compressed/encrypted/encoded extents */
1561                         if (btrfs_file_extent_compression(leaf, fi) ||
1562                             btrfs_file_extent_encryption(leaf, fi) ||
1563                             btrfs_file_extent_other_encoding(leaf, fi))
1564                                 goto out_check;
1565                         /*
1566                          * If extent is created before the last volume's snapshot
1567                          * this implies the extent is shared, hence we can't do
1568                          * nocow. This is the same check as in
1569                          * btrfs_cross_ref_exist but without calling
1570                          * btrfs_search_slot.
1571                          */
1572                         if (!freespace_inode &&
1573                             btrfs_file_extent_generation(leaf, fi) <=
1574                             btrfs_root_last_snapshot(&root->root_item))
1575                                 goto out_check;
1576                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1577                                 goto out_check;
1578                         /* If extent is RO, we must COW it */
1579                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1580                                 goto out_check;
1581                         ret = btrfs_cross_ref_exist(root, ino,
1582                                                     found_key.offset -
1583                                                     extent_offset, disk_bytenr);
1584                         if (ret) {
1585                                 /*
1586                                  * ret could be -EIO if the above fails to read
1587                                  * metadata.
1588                                  */
1589                                 if (ret < 0) {
1590                                         if (cow_start != (u64)-1)
1591                                                 cur_offset = cow_start;
1592                                         goto error;
1593                                 }
1594
1595                                 WARN_ON_ONCE(freespace_inode);
1596                                 goto out_check;
1597                         }
1598                         disk_bytenr += extent_offset;
1599                         disk_bytenr += cur_offset - found_key.offset;
1600                         num_bytes = min(end + 1, extent_end) - cur_offset;
1601                         /*
1602                          * If there are pending snapshots for this root, we
1603                          * fall into common COW way
1604                          */
1605                         if (!freespace_inode && atomic_read(&root->snapshot_force_cow))
1606                                 goto out_check;
1607                         /*
1608                          * force cow if csum exists in the range.
1609                          * this ensure that csum for a given extent are
1610                          * either valid or do not exist.
1611                          */
1612                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1613                                                   num_bytes);
1614                         if (ret) {
1615                                 /*
1616                                  * ret could be -EIO if the above fails to read
1617                                  * metadata.
1618                                  */
1619                                 if (ret < 0) {
1620                                         if (cow_start != (u64)-1)
1621                                                 cur_offset = cow_start;
1622                                         goto error;
1623                                 }
1624                                 WARN_ON_ONCE(freespace_inode);
1625                                 goto out_check;
1626                         }
1627                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr))
1628                                 goto out_check;
1629                         nocow = true;
1630                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1631                         extent_end = found_key.offset + ram_bytes;
1632                         extent_end = ALIGN(extent_end, fs_info->sectorsize);
1633                         /* Skip extents outside of our requested range */
1634                         if (extent_end <= start) {
1635                                 path->slots[0]++;
1636                                 goto next_slot;
1637                         }
1638                 } else {
1639                         /* If this triggers then we have a memory corruption */
1640                         BUG();
1641                 }
1642 out_check:
1643                 /*
1644                  * If nocow is false then record the beginning of the range
1645                  * that needs to be COWed
1646                  */
1647                 if (!nocow) {
1648                         if (cow_start == (u64)-1)
1649                                 cow_start = cur_offset;
1650                         cur_offset = extent_end;
1651                         if (cur_offset > end)
1652                                 break;
1653                         path->slots[0]++;
1654                         goto next_slot;
1655                 }
1656
1657                 btrfs_release_path(path);
1658
1659                 /*
1660                  * COW range from cow_start to found_key.offset - 1. As the key
1661                  * will contain the beginning of the first extent that can be
1662                  * NOCOW, following one which needs to be COW'ed
1663                  */
1664                 if (cow_start != (u64)-1) {
1665                         ret = fallback_to_cow(inode, locked_page, cow_start,
1666                                               found_key.offset - 1,
1667                                               page_started, nr_written);
1668                         if (ret) {
1669                                 if (nocow)
1670                                         btrfs_dec_nocow_writers(fs_info,
1671                                                                 disk_bytenr);
1672                                 goto error;
1673                         }
1674                         cow_start = (u64)-1;
1675                 }
1676
1677                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1678                         u64 orig_start = found_key.offset - extent_offset;
1679                         struct extent_map *em;
1680
1681                         em = create_io_em(inode, cur_offset, num_bytes,
1682                                           orig_start,
1683                                           disk_bytenr, /* block_start */
1684                                           num_bytes, /* block_len */
1685                                           disk_num_bytes, /* orig_block_len */
1686                                           ram_bytes, BTRFS_COMPRESS_NONE,
1687                                           BTRFS_ORDERED_PREALLOC);
1688                         if (IS_ERR(em)) {
1689                                 if (nocow)
1690                                         btrfs_dec_nocow_writers(fs_info,
1691                                                                 disk_bytenr);
1692                                 ret = PTR_ERR(em);
1693                                 goto error;
1694                         }
1695                         free_extent_map(em);
1696                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1697                                                        disk_bytenr, num_bytes,
1698                                                        num_bytes,
1699                                                        BTRFS_ORDERED_PREALLOC);
1700                         if (ret) {
1701                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1702                                                         cur_offset,
1703                                                         cur_offset + num_bytes - 1,
1704                                                         0);
1705                                 goto error;
1706                         }
1707                 } else {
1708                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1709                                                        disk_bytenr, num_bytes,
1710                                                        num_bytes,
1711                                                        BTRFS_ORDERED_NOCOW);
1712                         if (ret)
1713                                 goto error;
1714                 }
1715
1716                 if (nocow)
1717                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1718                 nocow = false;
1719
1720                 if (root->root_key.objectid ==
1721                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1722                         /*
1723                          * Error handled later, as we must prevent
1724                          * extent_clear_unlock_delalloc() in error handler
1725                          * from freeing metadata of created ordered extent.
1726                          */
1727                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1728                                                       num_bytes);
1729
1730                 extent_clear_unlock_delalloc(inode, cur_offset,
1731                                              cur_offset + num_bytes - 1,
1732                                              locked_page, EXTENT_LOCKED |
1733                                              EXTENT_DELALLOC |
1734                                              EXTENT_CLEAR_DATA_RESV,
1735                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1736
1737                 cur_offset = extent_end;
1738
1739                 /*
1740                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1741                  * handler, as metadata for created ordered extent will only
1742                  * be freed by btrfs_finish_ordered_io().
1743                  */
1744                 if (ret)
1745                         goto error;
1746                 if (cur_offset > end)
1747                         break;
1748         }
1749         btrfs_release_path(path);
1750
1751         if (cur_offset <= end && cow_start == (u64)-1)
1752                 cow_start = cur_offset;
1753
1754         if (cow_start != (u64)-1) {
1755                 cur_offset = end;
1756                 ret = fallback_to_cow(inode, locked_page, cow_start, end,
1757                                       page_started, nr_written);
1758                 if (ret)
1759                         goto error;
1760         }
1761
1762 error:
1763         if (nocow)
1764                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1765
1766         if (ret && cur_offset < end)
1767                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1768                                              locked_page, EXTENT_LOCKED |
1769                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1770                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1771                                              PAGE_CLEAR_DIRTY |
1772                                              PAGE_SET_WRITEBACK |
1773                                              PAGE_END_WRITEBACK);
1774         btrfs_free_path(path);
1775         return ret;
1776 }
1777
1778 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1779 {
1780
1781         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1782             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1783                 return 0;
1784
1785         /*
1786          * @defrag_bytes is a hint value, no spinlock held here,
1787          * if is not zero, it means the file is defragging.
1788          * Force cow if given extent needs to be defragged.
1789          */
1790         if (BTRFS_I(inode)->defrag_bytes &&
1791             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1792                            EXTENT_DEFRAG, 0, NULL))
1793                 return 1;
1794
1795         return 0;
1796 }
1797
1798 /*
1799  * Function to process delayed allocation (create CoW) for ranges which are
1800  * being touched for the first time.
1801  */
1802 int btrfs_run_delalloc_range(struct inode *inode, struct page *locked_page,
1803                 u64 start, u64 end, int *page_started, unsigned long *nr_written,
1804                 struct writeback_control *wbc)
1805 {
1806         int ret;
1807         int force_cow = need_force_cow(inode, start, end);
1808
1809         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1810                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1811                                          page_started, 1, nr_written);
1812         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1813                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1814                                          page_started, 0, nr_written);
1815         } else if (!inode_can_compress(inode) ||
1816                    !inode_need_compress(inode, start, end)) {
1817                 ret = cow_file_range(inode, locked_page, start, end,
1818                                       page_started, nr_written, 1);
1819         } else {
1820                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1821                         &BTRFS_I(inode)->runtime_flags);
1822                 ret = cow_file_range_async(inode, wbc, locked_page, start, end,
1823                                            page_started, nr_written);
1824         }
1825         if (ret)
1826                 btrfs_cleanup_ordered_extents(inode, locked_page, start,
1827                                               end - start + 1);
1828         return ret;
1829 }
1830
1831 void btrfs_split_delalloc_extent(struct inode *inode,
1832                                  struct extent_state *orig, u64 split)
1833 {
1834         u64 size;
1835
1836         /* not delalloc, ignore it */
1837         if (!(orig->state & EXTENT_DELALLOC))
1838                 return;
1839
1840         size = orig->end - orig->start + 1;
1841         if (size > BTRFS_MAX_EXTENT_SIZE) {
1842                 u32 num_extents;
1843                 u64 new_size;
1844
1845                 /*
1846                  * See the explanation in btrfs_merge_delalloc_extent, the same
1847                  * applies here, just in reverse.
1848                  */
1849                 new_size = orig->end - split + 1;
1850                 num_extents = count_max_extents(new_size);
1851                 new_size = split - orig->start;
1852                 num_extents += count_max_extents(new_size);
1853                 if (count_max_extents(size) >= num_extents)
1854                         return;
1855         }
1856
1857         spin_lock(&BTRFS_I(inode)->lock);
1858         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1859         spin_unlock(&BTRFS_I(inode)->lock);
1860 }
1861
1862 /*
1863  * Handle merged delayed allocation extents so we can keep track of new extents
1864  * that are just merged onto old extents, such as when we are doing sequential
1865  * writes, so we can properly account for the metadata space we'll need.
1866  */
1867 void btrfs_merge_delalloc_extent(struct inode *inode, struct extent_state *new,
1868                                  struct extent_state *other)
1869 {
1870         u64 new_size, old_size;
1871         u32 num_extents;
1872
1873         /* not delalloc, ignore it */
1874         if (!(other->state & EXTENT_DELALLOC))
1875                 return;
1876
1877         if (new->start > other->start)
1878                 new_size = new->end - other->start + 1;
1879         else
1880                 new_size = other->end - new->start + 1;
1881
1882         /* we're not bigger than the max, unreserve the space and go */
1883         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1884                 spin_lock(&BTRFS_I(inode)->lock);
1885                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1886                 spin_unlock(&BTRFS_I(inode)->lock);
1887                 return;
1888         }
1889
1890         /*
1891          * We have to add up either side to figure out how many extents were
1892          * accounted for before we merged into one big extent.  If the number of
1893          * extents we accounted for is <= the amount we need for the new range
1894          * then we can return, otherwise drop.  Think of it like this
1895          *
1896          * [ 4k][MAX_SIZE]
1897          *
1898          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1899          * need 2 outstanding extents, on one side we have 1 and the other side
1900          * we have 1 so they are == and we can return.  But in this case
1901          *
1902          * [MAX_SIZE+4k][MAX_SIZE+4k]
1903          *
1904          * Each range on their own accounts for 2 extents, but merged together
1905          * they are only 3 extents worth of accounting, so we need to drop in
1906          * this case.
1907          */
1908         old_size = other->end - other->start + 1;
1909         num_extents = count_max_extents(old_size);
1910         old_size = new->end - new->start + 1;
1911         num_extents += count_max_extents(old_size);
1912         if (count_max_extents(new_size) >= num_extents)
1913                 return;
1914
1915         spin_lock(&BTRFS_I(inode)->lock);
1916         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1917         spin_unlock(&BTRFS_I(inode)->lock);
1918 }
1919
1920 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1921                                       struct inode *inode)
1922 {
1923         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1924
1925         spin_lock(&root->delalloc_lock);
1926         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1927                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1928                               &root->delalloc_inodes);
1929                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1930                         &BTRFS_I(inode)->runtime_flags);
1931                 root->nr_delalloc_inodes++;
1932                 if (root->nr_delalloc_inodes == 1) {
1933                         spin_lock(&fs_info->delalloc_root_lock);
1934                         BUG_ON(!list_empty(&root->delalloc_root));
1935                         list_add_tail(&root->delalloc_root,
1936                                       &fs_info->delalloc_roots);
1937                         spin_unlock(&fs_info->delalloc_root_lock);
1938                 }
1939         }
1940         spin_unlock(&root->delalloc_lock);
1941 }
1942
1943
1944 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1945                                 struct btrfs_inode *inode)
1946 {
1947         struct btrfs_fs_info *fs_info = root->fs_info;
1948
1949         if (!list_empty(&inode->delalloc_inodes)) {
1950                 list_del_init(&inode->delalloc_inodes);
1951                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1952                           &inode->runtime_flags);
1953                 root->nr_delalloc_inodes--;
1954                 if (!root->nr_delalloc_inodes) {
1955                         ASSERT(list_empty(&root->delalloc_inodes));
1956                         spin_lock(&fs_info->delalloc_root_lock);
1957                         BUG_ON(list_empty(&root->delalloc_root));
1958                         list_del_init(&root->delalloc_root);
1959                         spin_unlock(&fs_info->delalloc_root_lock);
1960                 }
1961         }
1962 }
1963
1964 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1965                                      struct btrfs_inode *inode)
1966 {
1967         spin_lock(&root->delalloc_lock);
1968         __btrfs_del_delalloc_inode(root, inode);
1969         spin_unlock(&root->delalloc_lock);
1970 }
1971
1972 /*
1973  * Properly track delayed allocation bytes in the inode and to maintain the
1974  * list of inodes that have pending delalloc work to be done.
1975  */
1976 void btrfs_set_delalloc_extent(struct inode *inode, struct extent_state *state,
1977                                unsigned *bits)
1978 {
1979         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1980
1981         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1982                 WARN_ON(1);
1983         /*
1984          * set_bit and clear bit hooks normally require _irqsave/restore
1985          * but in this case, we are only testing for the DELALLOC
1986          * bit, which is only set or cleared with irqs on
1987          */
1988         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1989                 struct btrfs_root *root = BTRFS_I(inode)->root;
1990                 u64 len = state->end + 1 - state->start;
1991                 u32 num_extents = count_max_extents(len);
1992                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1993
1994                 spin_lock(&BTRFS_I(inode)->lock);
1995                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1996                 spin_unlock(&BTRFS_I(inode)->lock);
1997
1998                 /* For sanity tests */
1999                 if (btrfs_is_testing(fs_info))
2000                         return;
2001
2002                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
2003                                          fs_info->delalloc_batch);
2004                 spin_lock(&BTRFS_I(inode)->lock);
2005                 BTRFS_I(inode)->delalloc_bytes += len;
2006                 if (*bits & EXTENT_DEFRAG)
2007                         BTRFS_I(inode)->defrag_bytes += len;
2008                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
2009                                          &BTRFS_I(inode)->runtime_flags))
2010                         btrfs_add_delalloc_inodes(root, inode);
2011                 spin_unlock(&BTRFS_I(inode)->lock);
2012         }
2013
2014         if (!(state->state & EXTENT_DELALLOC_NEW) &&
2015             (*bits & EXTENT_DELALLOC_NEW)) {
2016                 spin_lock(&BTRFS_I(inode)->lock);
2017                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
2018                         state->start;
2019                 spin_unlock(&BTRFS_I(inode)->lock);
2020         }
2021 }
2022
2023 /*
2024  * Once a range is no longer delalloc this function ensures that proper
2025  * accounting happens.
2026  */
2027 void btrfs_clear_delalloc_extent(struct inode *vfs_inode,
2028                                  struct extent_state *state, unsigned *bits)
2029 {
2030         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
2031         struct btrfs_fs_info *fs_info = btrfs_sb(vfs_inode->i_sb);
2032         u64 len = state->end + 1 - state->start;
2033         u32 num_extents = count_max_extents(len);
2034
2035         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
2036                 spin_lock(&inode->lock);
2037                 inode->defrag_bytes -= len;
2038                 spin_unlock(&inode->lock);
2039         }
2040
2041         /*
2042          * set_bit and clear bit hooks normally require _irqsave/restore
2043          * but in this case, we are only testing for the DELALLOC
2044          * bit, which is only set or cleared with irqs on
2045          */
2046         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
2047                 struct btrfs_root *root = inode->root;
2048                 bool do_list = !btrfs_is_free_space_inode(inode);
2049
2050                 spin_lock(&inode->lock);
2051                 btrfs_mod_outstanding_extents(inode, -num_extents);
2052                 spin_unlock(&inode->lock);
2053
2054                 /*
2055                  * We don't reserve metadata space for space cache inodes so we
2056                  * don't need to call delalloc_release_metadata if there is an
2057                  * error.
2058                  */
2059                 if (*bits & EXTENT_CLEAR_META_RESV &&
2060                     root != fs_info->tree_root)
2061                         btrfs_delalloc_release_metadata(inode, len, false);
2062
2063                 /* For sanity tests. */
2064                 if (btrfs_is_testing(fs_info))
2065                         return;
2066
2067                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
2068                     do_list && !(state->state & EXTENT_NORESERVE) &&
2069                     (*bits & EXTENT_CLEAR_DATA_RESV))
2070                         btrfs_free_reserved_data_space_noquota(
2071                                         &inode->vfs_inode,
2072                                         state->start, len);
2073
2074                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
2075                                          fs_info->delalloc_batch);
2076                 spin_lock(&inode->lock);
2077                 inode->delalloc_bytes -= len;
2078                 if (do_list && inode->delalloc_bytes == 0 &&
2079                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
2080                                         &inode->runtime_flags))
2081                         btrfs_del_delalloc_inode(root, inode);
2082                 spin_unlock(&inode->lock);
2083         }
2084
2085         if ((state->state & EXTENT_DELALLOC_NEW) &&
2086             (*bits & EXTENT_DELALLOC_NEW)) {
2087                 spin_lock(&inode->lock);
2088                 ASSERT(inode->new_delalloc_bytes >= len);
2089                 inode->new_delalloc_bytes -= len;
2090                 spin_unlock(&inode->lock);
2091         }
2092 }
2093
2094 /*
2095  * btrfs_bio_fits_in_stripe - Checks whether the size of the given bio will fit
2096  * in a chunk's stripe. This function ensures that bios do not span a
2097  * stripe/chunk
2098  *
2099  * @page - The page we are about to add to the bio
2100  * @size - size we want to add to the bio
2101  * @bio - bio we want to ensure is smaller than a stripe
2102  * @bio_flags - flags of the bio
2103  *
2104  * return 1 if page cannot be added to the bio
2105  * return 0 if page can be added to the bio
2106  * return error otherwise
2107  */
2108 int btrfs_bio_fits_in_stripe(struct page *page, size_t size, struct bio *bio,
2109                              unsigned long bio_flags)
2110 {
2111         struct inode *inode = page->mapping->host;
2112         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2113         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
2114         u64 length = 0;
2115         u64 map_length;
2116         int ret;
2117         struct btrfs_io_geometry geom;
2118
2119         if (bio_flags & EXTENT_BIO_COMPRESSED)
2120                 return 0;
2121
2122         length = bio->bi_iter.bi_size;
2123         map_length = length;
2124         ret = btrfs_get_io_geometry(fs_info, btrfs_op(bio), logical, map_length,
2125                                     &geom);
2126         if (ret < 0)
2127                 return ret;
2128
2129         if (geom.len < length + size)
2130                 return 1;
2131         return 0;
2132 }
2133
2134 /*
2135  * in order to insert checksums into the metadata in large chunks,
2136  * we wait until bio submission time.   All the pages in the bio are
2137  * checksummed and sums are attached onto the ordered extent record.
2138  *
2139  * At IO completion time the cums attached on the ordered extent record
2140  * are inserted into the btree
2141  */
2142 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
2143                                     u64 bio_offset)
2144 {
2145         struct inode *inode = private_data;
2146         blk_status_t ret = 0;
2147
2148         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2149         BUG_ON(ret); /* -ENOMEM */
2150         return 0;
2151 }
2152
2153 /*
2154  * extent_io.c submission hook. This does the right thing for csum calculation
2155  * on write, or reading the csums from the tree before a read.
2156  *
2157  * Rules about async/sync submit,
2158  * a) read:                             sync submit
2159  *
2160  * b) write without checksum:           sync submit
2161  *
2162  * c) write with checksum:
2163  *    c-1) if bio is issued by fsync:   sync submit
2164  *         (sync_writers != 0)
2165  *
2166  *    c-2) if root is reloc root:       sync submit
2167  *         (only in case of buffered IO)
2168  *
2169  *    c-3) otherwise:                   async submit
2170  */
2171 static blk_status_t btrfs_submit_bio_hook(struct inode *inode, struct bio *bio,
2172                                           int mirror_num,
2173                                           unsigned long bio_flags)
2174
2175 {
2176         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2177         struct btrfs_root *root = BTRFS_I(inode)->root;
2178         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2179         blk_status_t ret = 0;
2180         int skip_sum;
2181         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2182
2183         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2184
2185         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2186                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2187
2188         if (bio_op(bio) != REQ_OP_WRITE) {
2189                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2190                 if (ret)
2191                         goto out;
2192
2193                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2194                         ret = btrfs_submit_compressed_read(inode, bio,
2195                                                            mirror_num,
2196                                                            bio_flags);
2197                         goto out;
2198                 } else if (!skip_sum) {
2199                         ret = btrfs_lookup_bio_sums(inode, bio, (u64)-1, NULL);
2200                         if (ret)
2201                                 goto out;
2202                 }
2203                 goto mapit;
2204         } else if (async && !skip_sum) {
2205                 /* csum items have already been cloned */
2206                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2207                         goto mapit;
2208                 /* we're doing a write, do the async checksumming */
2209                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2210                                           0, inode, btrfs_submit_bio_start);
2211                 goto out;
2212         } else if (!skip_sum) {
2213                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2214                 if (ret)
2215                         goto out;
2216         }
2217
2218 mapit:
2219         ret = btrfs_map_bio(fs_info, bio, mirror_num);
2220
2221 out:
2222         if (ret) {
2223                 bio->bi_status = ret;
2224                 bio_endio(bio);
2225         }
2226         return ret;
2227 }
2228
2229 /*
2230  * given a list of ordered sums record them in the inode.  This happens
2231  * at IO completion time based on sums calculated at bio submission time.
2232  */
2233 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2234                              struct inode *inode, struct list_head *list)
2235 {
2236         struct btrfs_ordered_sum *sum;
2237         int ret;
2238
2239         list_for_each_entry(sum, list, list) {
2240                 trans->adding_csums = true;
2241                 ret = btrfs_csum_file_blocks(trans,
2242                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2243                 trans->adding_csums = false;
2244                 if (ret)
2245                         return ret;
2246         }
2247         return 0;
2248 }
2249
2250 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2251                               unsigned int extra_bits,
2252                               struct extent_state **cached_state)
2253 {
2254         WARN_ON(PAGE_ALIGNED(end));
2255         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2256                                    extra_bits, cached_state);
2257 }
2258
2259 /* see btrfs_writepage_start_hook for details on why this is required */
2260 struct btrfs_writepage_fixup {
2261         struct page *page;
2262         struct inode *inode;
2263         struct btrfs_work work;
2264 };
2265
2266 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2267 {
2268         struct btrfs_writepage_fixup *fixup;
2269         struct btrfs_ordered_extent *ordered;
2270         struct extent_state *cached_state = NULL;
2271         struct extent_changeset *data_reserved = NULL;
2272         struct page *page;
2273         struct inode *inode;
2274         u64 page_start;
2275         u64 page_end;
2276         int ret = 0;
2277         bool free_delalloc_space = true;
2278
2279         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2280         page = fixup->page;
2281         inode = fixup->inode;
2282         page_start = page_offset(page);
2283         page_end = page_offset(page) + PAGE_SIZE - 1;
2284
2285         /*
2286          * This is similar to page_mkwrite, we need to reserve the space before
2287          * we take the page lock.
2288          */
2289         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2290                                            PAGE_SIZE);
2291 again:
2292         lock_page(page);
2293
2294         /*
2295          * Before we queued this fixup, we took a reference on the page.
2296          * page->mapping may go NULL, but it shouldn't be moved to a different
2297          * address space.
2298          */
2299         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2300                 /*
2301                  * Unfortunately this is a little tricky, either
2302                  *
2303                  * 1) We got here and our page had already been dealt with and
2304                  *    we reserved our space, thus ret == 0, so we need to just
2305                  *    drop our space reservation and bail.  This can happen the
2306                  *    first time we come into the fixup worker, or could happen
2307                  *    while waiting for the ordered extent.
2308                  * 2) Our page was already dealt with, but we happened to get an
2309                  *    ENOSPC above from the btrfs_delalloc_reserve_space.  In
2310                  *    this case we obviously don't have anything to release, but
2311                  *    because the page was already dealt with we don't want to
2312                  *    mark the page with an error, so make sure we're resetting
2313                  *    ret to 0.  This is why we have this check _before_ the ret
2314                  *    check, because we do not want to have a surprise ENOSPC
2315                  *    when the page was already properly dealt with.
2316                  */
2317                 if (!ret) {
2318                         btrfs_delalloc_release_extents(BTRFS_I(inode),
2319                                                        PAGE_SIZE);
2320                         btrfs_delalloc_release_space(inode, data_reserved,
2321                                                      page_start, PAGE_SIZE,
2322                                                      true);
2323                 }
2324                 ret = 0;
2325                 goto out_page;
2326         }
2327
2328         /*
2329          * We can't mess with the page state unless it is locked, so now that
2330          * it is locked bail if we failed to make our space reservation.
2331          */
2332         if (ret)
2333                 goto out_page;
2334
2335         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2336                          &cached_state);
2337
2338         /* already ordered? We're done */
2339         if (PagePrivate2(page))
2340                 goto out_reserved;
2341
2342         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2343                                         PAGE_SIZE);
2344         if (ordered) {
2345                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2346                                      page_end, &cached_state);
2347                 unlock_page(page);
2348                 btrfs_start_ordered_extent(inode, ordered, 1);
2349                 btrfs_put_ordered_extent(ordered);
2350                 goto again;
2351         }
2352
2353         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2354                                         &cached_state);
2355         if (ret)
2356                 goto out_reserved;
2357
2358         /*
2359          * Everything went as planned, we're now the owner of a dirty page with
2360          * delayed allocation bits set and space reserved for our COW
2361          * destination.
2362          *
2363          * The page was dirty when we started, nothing should have cleaned it.
2364          */
2365         BUG_ON(!PageDirty(page));
2366         free_delalloc_space = false;
2367 out_reserved:
2368         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2369         if (free_delalloc_space)
2370                 btrfs_delalloc_release_space(inode, data_reserved, page_start,
2371                                              PAGE_SIZE, true);
2372         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2373                              &cached_state);
2374 out_page:
2375         if (ret) {
2376                 /*
2377                  * We hit ENOSPC or other errors.  Update the mapping and page
2378                  * to reflect the errors and clean the page.
2379                  */
2380                 mapping_set_error(page->mapping, ret);
2381                 end_extent_writepage(page, ret, page_start, page_end);
2382                 clear_page_dirty_for_io(page);
2383                 SetPageError(page);
2384         }
2385         ClearPageChecked(page);
2386         unlock_page(page);
2387         put_page(page);
2388         kfree(fixup);
2389         extent_changeset_free(data_reserved);
2390         /*
2391          * As a precaution, do a delayed iput in case it would be the last iput
2392          * that could need flushing space. Recursing back to fixup worker would
2393          * deadlock.
2394          */
2395         btrfs_add_delayed_iput(inode);
2396 }
2397
2398 /*
2399  * There are a few paths in the higher layers of the kernel that directly
2400  * set the page dirty bit without asking the filesystem if it is a
2401  * good idea.  This causes problems because we want to make sure COW
2402  * properly happens and the data=ordered rules are followed.
2403  *
2404  * In our case any range that doesn't have the ORDERED bit set
2405  * hasn't been properly setup for IO.  We kick off an async process
2406  * to fix it up.  The async helper will wait for ordered extents, set
2407  * the delalloc bit and make it safe to write the page.
2408  */
2409 int btrfs_writepage_cow_fixup(struct page *page, u64 start, u64 end)
2410 {
2411         struct inode *inode = page->mapping->host;
2412         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2413         struct btrfs_writepage_fixup *fixup;
2414
2415         /* this page is properly in the ordered list */
2416         if (TestClearPagePrivate2(page))
2417                 return 0;
2418
2419         /*
2420          * PageChecked is set below when we create a fixup worker for this page,
2421          * don't try to create another one if we're already PageChecked()
2422          *
2423          * The extent_io writepage code will redirty the page if we send back
2424          * EAGAIN.
2425          */
2426         if (PageChecked(page))
2427                 return -EAGAIN;
2428
2429         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2430         if (!fixup)
2431                 return -EAGAIN;
2432
2433         /*
2434          * We are already holding a reference to this inode from
2435          * write_cache_pages.  We need to hold it because the space reservation
2436          * takes place outside of the page lock, and we can't trust
2437          * page->mapping outside of the page lock.
2438          */
2439         ihold(inode);
2440         SetPageChecked(page);
2441         get_page(page);
2442         btrfs_init_work(&fixup->work, btrfs_writepage_fixup_worker, NULL, NULL);
2443         fixup->page = page;
2444         fixup->inode = inode;
2445         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2446
2447         return -EAGAIN;
2448 }
2449
2450 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2451                                        struct inode *inode, u64 file_pos,
2452                                        u64 disk_bytenr, u64 disk_num_bytes,
2453                                        u64 num_bytes, u64 ram_bytes,
2454                                        u8 compression, u8 encryption,
2455                                        u16 other_encoding, int extent_type)
2456 {
2457         struct btrfs_root *root = BTRFS_I(inode)->root;
2458         struct btrfs_file_extent_item *fi;
2459         struct btrfs_path *path;
2460         struct extent_buffer *leaf;
2461         struct btrfs_key ins;
2462         u64 qg_released;
2463         int extent_inserted = 0;
2464         int ret;
2465
2466         path = btrfs_alloc_path();
2467         if (!path)
2468                 return -ENOMEM;
2469
2470         /*
2471          * we may be replacing one extent in the tree with another.
2472          * The new extent is pinned in the extent map, and we don't want
2473          * to drop it from the cache until it is completely in the btree.
2474          *
2475          * So, tell btrfs_drop_extents to leave this extent in the cache.
2476          * the caller is expected to unpin it and allow it to be merged
2477          * with the others.
2478          */
2479         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2480                                    file_pos + num_bytes, NULL, 0,
2481                                    1, sizeof(*fi), &extent_inserted);
2482         if (ret)
2483                 goto out;
2484
2485         if (!extent_inserted) {
2486                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2487                 ins.offset = file_pos;
2488                 ins.type = BTRFS_EXTENT_DATA_KEY;
2489
2490                 path->leave_spinning = 1;
2491                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2492                                               sizeof(*fi));
2493                 if (ret)
2494                         goto out;
2495         }
2496         leaf = path->nodes[0];
2497         fi = btrfs_item_ptr(leaf, path->slots[0],
2498                             struct btrfs_file_extent_item);
2499         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2500         btrfs_set_file_extent_type(leaf, fi, extent_type);
2501         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2502         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2503         btrfs_set_file_extent_offset(leaf, fi, 0);
2504         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2505         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2506         btrfs_set_file_extent_compression(leaf, fi, compression);
2507         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2508         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2509
2510         btrfs_mark_buffer_dirty(leaf);
2511         btrfs_release_path(path);
2512
2513         inode_add_bytes(inode, num_bytes);
2514
2515         ins.objectid = disk_bytenr;
2516         ins.offset = disk_num_bytes;
2517         ins.type = BTRFS_EXTENT_ITEM_KEY;
2518
2519         ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode), file_pos,
2520                                                 ram_bytes);
2521         if (ret)
2522                 goto out;
2523
2524         /*
2525          * Release the reserved range from inode dirty range map, as it is
2526          * already moved into delayed_ref_head
2527          */
2528         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2529         if (ret < 0)
2530                 goto out;
2531         qg_released = ret;
2532         ret = btrfs_alloc_reserved_file_extent(trans, root,
2533                                                btrfs_ino(BTRFS_I(inode)),
2534                                                file_pos, qg_released, &ins);
2535 out:
2536         btrfs_free_path(path);
2537
2538         return ret;
2539 }
2540
2541 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2542                                          u64 start, u64 len)
2543 {
2544         struct btrfs_block_group *cache;
2545
2546         cache = btrfs_lookup_block_group(fs_info, start);
2547         ASSERT(cache);
2548
2549         spin_lock(&cache->lock);
2550         cache->delalloc_bytes -= len;
2551         spin_unlock(&cache->lock);
2552
2553         btrfs_put_block_group(cache);
2554 }
2555
2556 /* as ordered data IO finishes, this gets called so we can finish
2557  * an ordered extent if the range of bytes in the file it covers are
2558  * fully written.
2559  */
2560 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2561 {
2562         struct inode *inode = ordered_extent->inode;
2563         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2564         struct btrfs_root *root = BTRFS_I(inode)->root;
2565         struct btrfs_trans_handle *trans = NULL;
2566         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2567         struct extent_state *cached_state = NULL;
2568         u64 start, end;
2569         int compress_type = 0;
2570         int ret = 0;
2571         u64 logical_len = ordered_extent->num_bytes;
2572         bool freespace_inode;
2573         bool truncated = false;
2574         bool range_locked = false;
2575         bool clear_new_delalloc_bytes = false;
2576         bool clear_reserved_extent = true;
2577         unsigned int clear_bits;
2578
2579         start = ordered_extent->file_offset;
2580         end = start + ordered_extent->num_bytes - 1;
2581
2582         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2583             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2584             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2585                 clear_new_delalloc_bytes = true;
2586
2587         freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
2588
2589         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2590                 ret = -EIO;
2591                 goto out;
2592         }
2593
2594         btrfs_free_io_failure_record(BTRFS_I(inode), start, end);
2595
2596         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2597                 truncated = true;
2598                 logical_len = ordered_extent->truncated_len;
2599                 /* Truncated the entire extent, don't bother adding */
2600                 if (!logical_len)
2601                         goto out;
2602         }
2603
2604         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2605                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2606
2607                 /*
2608                  * For mwrite(mmap + memset to write) case, we still reserve
2609                  * space for NOCOW range.
2610                  * As NOCOW won't cause a new delayed ref, just free the space
2611                  */
2612                 btrfs_qgroup_free_data(inode, NULL, start,
2613                                        ordered_extent->num_bytes);
2614                 btrfs_inode_safe_disk_i_size_write(inode, 0);
2615                 if (freespace_inode)
2616                         trans = btrfs_join_transaction_spacecache(root);
2617                 else
2618                         trans = btrfs_join_transaction(root);
2619                 if (IS_ERR(trans)) {
2620                         ret = PTR_ERR(trans);
2621                         trans = NULL;
2622                         goto out;
2623                 }
2624                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2625                 ret = btrfs_update_inode_fallback(trans, root, inode);
2626                 if (ret) /* -ENOMEM or corruption */
2627                         btrfs_abort_transaction(trans, ret);
2628                 goto out;
2629         }
2630
2631         range_locked = true;
2632         lock_extent_bits(io_tree, start, end, &cached_state);
2633
2634         if (freespace_inode)
2635                 trans = btrfs_join_transaction_spacecache(root);
2636         else
2637                 trans = btrfs_join_transaction(root);
2638         if (IS_ERR(trans)) {
2639                 ret = PTR_ERR(trans);
2640                 trans = NULL;
2641                 goto out;
2642         }
2643
2644         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2645
2646         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2647                 compress_type = ordered_extent->compress_type;
2648         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2649                 BUG_ON(compress_type);
2650                 btrfs_qgroup_free_data(inode, NULL, start,
2651                                        ordered_extent->num_bytes);
2652                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
2653                                                 ordered_extent->file_offset,
2654                                                 ordered_extent->file_offset +
2655                                                 logical_len);
2656         } else {
2657                 BUG_ON(root == fs_info->tree_root);
2658                 ret = insert_reserved_file_extent(trans, inode, start,
2659                                                 ordered_extent->disk_bytenr,
2660                                                 ordered_extent->disk_num_bytes,
2661                                                 logical_len, logical_len,
2662                                                 compress_type, 0, 0,
2663                                                 BTRFS_FILE_EXTENT_REG);
2664                 if (!ret) {
2665                         clear_reserved_extent = false;
2666                         btrfs_release_delalloc_bytes(fs_info,
2667                                                 ordered_extent->disk_bytenr,
2668                                                 ordered_extent->disk_num_bytes);
2669                 }
2670         }
2671         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2672                            ordered_extent->file_offset,
2673                            ordered_extent->num_bytes, trans->transid);
2674         if (ret < 0) {
2675                 btrfs_abort_transaction(trans, ret);
2676                 goto out;
2677         }
2678
2679         ret = add_pending_csums(trans, inode, &ordered_extent->list);
2680         if (ret) {
2681                 btrfs_abort_transaction(trans, ret);
2682                 goto out;
2683         }
2684
2685         btrfs_inode_safe_disk_i_size_write(inode, 0);
2686         ret = btrfs_update_inode_fallback(trans, root, inode);
2687         if (ret) { /* -ENOMEM or corruption */
2688                 btrfs_abort_transaction(trans, ret);
2689                 goto out;
2690         }
2691         ret = 0;
2692 out:
2693         clear_bits = EXTENT_DEFRAG;
2694         if (range_locked)
2695                 clear_bits |= EXTENT_LOCKED;
2696         if (clear_new_delalloc_bytes)
2697                 clear_bits |= EXTENT_DELALLOC_NEW;
2698         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits,
2699                          (clear_bits & EXTENT_LOCKED) ? 1 : 0, 0,
2700                          &cached_state);
2701
2702         if (trans)
2703                 btrfs_end_transaction(trans);
2704
2705         if (ret || truncated) {
2706                 u64 unwritten_start = start;
2707
2708                 if (truncated)
2709                         unwritten_start += logical_len;
2710                 clear_extent_uptodate(io_tree, unwritten_start, end, NULL);
2711
2712                 /* Drop the cache for the part of the extent we didn't write. */
2713                 btrfs_drop_extent_cache(BTRFS_I(inode), unwritten_start, end, 0);
2714
2715                 /*
2716                  * If the ordered extent had an IOERR or something else went
2717                  * wrong we need to return the space for this ordered extent
2718                  * back to the allocator.  We only free the extent in the
2719                  * truncated case if we didn't write out the extent at all.
2720                  *
2721                  * If we made it past insert_reserved_file_extent before we
2722                  * errored out then we don't need to do this as the accounting
2723                  * has already been done.
2724                  */
2725                 if ((ret || !logical_len) &&
2726                     clear_reserved_extent &&
2727                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2728                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2729                         /*
2730                          * Discard the range before returning it back to the
2731                          * free space pool
2732                          */
2733                         if (ret && btrfs_test_opt(fs_info, DISCARD_SYNC))
2734                                 btrfs_discard_extent(fs_info,
2735                                                 ordered_extent->disk_bytenr,
2736                                                 ordered_extent->disk_num_bytes,
2737                                                 NULL);
2738                         btrfs_free_reserved_extent(fs_info,
2739                                         ordered_extent->disk_bytenr,
2740                                         ordered_extent->disk_num_bytes, 1);
2741                 }
2742         }
2743
2744         /*
2745          * This needs to be done to make sure anybody waiting knows we are done
2746          * updating everything for this ordered extent.
2747          */
2748         btrfs_remove_ordered_extent(inode, ordered_extent);
2749
2750         /* once for us */
2751         btrfs_put_ordered_extent(ordered_extent);
2752         /* once for the tree */
2753         btrfs_put_ordered_extent(ordered_extent);
2754
2755         return ret;
2756 }
2757
2758 static void finish_ordered_fn(struct btrfs_work *work)
2759 {
2760         struct btrfs_ordered_extent *ordered_extent;
2761         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
2762         btrfs_finish_ordered_io(ordered_extent);
2763 }
2764
2765 void btrfs_writepage_endio_finish_ordered(struct page *page, u64 start,
2766                                           u64 end, int uptodate)
2767 {
2768         struct inode *inode = page->mapping->host;
2769         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2770         struct btrfs_ordered_extent *ordered_extent = NULL;
2771         struct btrfs_workqueue *wq;
2772
2773         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
2774
2775         ClearPagePrivate2(page);
2776         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
2777                                             end - start + 1, uptodate))
2778                 return;
2779
2780         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2781                 wq = fs_info->endio_freespace_worker;
2782         else
2783                 wq = fs_info->endio_write_workers;
2784
2785         btrfs_init_work(&ordered_extent->work, finish_ordered_fn, NULL, NULL);
2786         btrfs_queue_work(wq, &ordered_extent->work);
2787 }
2788
2789 static int check_data_csum(struct inode *inode, struct btrfs_io_bio *io_bio,
2790                            int icsum, struct page *page, int pgoff, u64 start,
2791                            size_t len)
2792 {
2793         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2794         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
2795         char *kaddr;
2796         u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2797         u8 *csum_expected;
2798         u8 csum[BTRFS_CSUM_SIZE];
2799
2800         csum_expected = ((u8 *)io_bio->csum) + icsum * csum_size;
2801
2802         kaddr = kmap_atomic(page);
2803         shash->tfm = fs_info->csum_shash;
2804
2805         crypto_shash_digest(shash, kaddr + pgoff, len, csum);
2806
2807         if (memcmp(csum, csum_expected, csum_size))
2808                 goto zeroit;
2809
2810         kunmap_atomic(kaddr);
2811         return 0;
2812 zeroit:
2813         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
2814                                     io_bio->mirror_num);
2815         memset(kaddr + pgoff, 1, len);
2816         flush_dcache_page(page);
2817         kunmap_atomic(kaddr);
2818         return -EIO;
2819 }
2820
2821 /*
2822  * when reads are done, we need to check csums to verify the data is correct
2823  * if there's a match, we allow the bio to finish.  If not, the code in
2824  * extent_io.c will try to find good copies for us.
2825  */
2826 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
2827                                       u64 phy_offset, struct page *page,
2828                                       u64 start, u64 end, int mirror)
2829 {
2830         size_t offset = start - page_offset(page);
2831         struct inode *inode = page->mapping->host;
2832         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2833         struct btrfs_root *root = BTRFS_I(inode)->root;
2834
2835         if (PageChecked(page)) {
2836                 ClearPageChecked(page);
2837                 return 0;
2838         }
2839
2840         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
2841                 return 0;
2842
2843         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
2844             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
2845                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
2846                 return 0;
2847         }
2848
2849         phy_offset >>= inode->i_sb->s_blocksize_bits;
2850         return check_data_csum(inode, io_bio, phy_offset, page, offset, start,
2851                                (size_t)(end - start + 1));
2852 }
2853
2854 /*
2855  * btrfs_add_delayed_iput - perform a delayed iput on @inode
2856  *
2857  * @inode: The inode we want to perform iput on
2858  *
2859  * This function uses the generic vfs_inode::i_count to track whether we should
2860  * just decrement it (in case it's > 1) or if this is the last iput then link
2861  * the inode to the delayed iput machinery. Delayed iputs are processed at
2862  * transaction commit time/superblock commit/cleaner kthread.
2863  */
2864 void btrfs_add_delayed_iput(struct inode *inode)
2865 {
2866         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2867         struct btrfs_inode *binode = BTRFS_I(inode);
2868
2869         if (atomic_add_unless(&inode->i_count, -1, 1))
2870                 return;
2871
2872         atomic_inc(&fs_info->nr_delayed_iputs);
2873         spin_lock(&fs_info->delayed_iput_lock);
2874         ASSERT(list_empty(&binode->delayed_iput));
2875         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
2876         spin_unlock(&fs_info->delayed_iput_lock);
2877         if (!test_bit(BTRFS_FS_CLEANER_RUNNING, &fs_info->flags))
2878                 wake_up_process(fs_info->cleaner_kthread);
2879 }
2880
2881 static void run_delayed_iput_locked(struct btrfs_fs_info *fs_info,
2882                                     struct btrfs_inode *inode)
2883 {
2884         list_del_init(&inode->delayed_iput);
2885         spin_unlock(&fs_info->delayed_iput_lock);
2886         iput(&inode->vfs_inode);
2887         if (atomic_dec_and_test(&fs_info->nr_delayed_iputs))
2888                 wake_up(&fs_info->delayed_iputs_wait);
2889         spin_lock(&fs_info->delayed_iput_lock);
2890 }
2891
2892 static void btrfs_run_delayed_iput(struct btrfs_fs_info *fs_info,
2893                                    struct btrfs_inode *inode)
2894 {
2895         if (!list_empty(&inode->delayed_iput)) {
2896                 spin_lock(&fs_info->delayed_iput_lock);
2897                 if (!list_empty(&inode->delayed_iput))
2898                         run_delayed_iput_locked(fs_info, inode);
2899                 spin_unlock(&fs_info->delayed_iput_lock);
2900         }
2901 }
2902
2903 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
2904 {
2905
2906         spin_lock(&fs_info->delayed_iput_lock);
2907         while (!list_empty(&fs_info->delayed_iputs)) {
2908                 struct btrfs_inode *inode;
2909
2910                 inode = list_first_entry(&fs_info->delayed_iputs,
2911                                 struct btrfs_inode, delayed_iput);
2912                 run_delayed_iput_locked(fs_info, inode);
2913         }
2914         spin_unlock(&fs_info->delayed_iput_lock);
2915 }
2916
2917 /**
2918  * btrfs_wait_on_delayed_iputs - wait on the delayed iputs to be done running
2919  * @fs_info - the fs_info for this fs
2920  * @return - EINTR if we were killed, 0 if nothing's pending
2921  *
2922  * This will wait on any delayed iputs that are currently running with KILLABLE
2923  * set.  Once they are all done running we will return, unless we are killed in
2924  * which case we return EINTR. This helps in user operations like fallocate etc
2925  * that might get blocked on the iputs.
2926  */
2927 int btrfs_wait_on_delayed_iputs(struct btrfs_fs_info *fs_info)
2928 {
2929         int ret = wait_event_killable(fs_info->delayed_iputs_wait,
2930                         atomic_read(&fs_info->nr_delayed_iputs) == 0);
2931         if (ret)
2932                 return -EINTR;
2933         return 0;
2934 }
2935
2936 /*
2937  * This creates an orphan entry for the given inode in case something goes wrong
2938  * in the middle of an unlink.
2939  */
2940 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
2941                      struct btrfs_inode *inode)
2942 {
2943         int ret;
2944
2945         ret = btrfs_insert_orphan_item(trans, inode->root, btrfs_ino(inode));
2946         if (ret && ret != -EEXIST) {
2947                 btrfs_abort_transaction(trans, ret);
2948                 return ret;
2949         }
2950
2951         return 0;
2952 }
2953
2954 /*
2955  * We have done the delete so we can go ahead and remove the orphan item for
2956  * this particular inode.
2957  */
2958 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
2959                             struct btrfs_inode *inode)
2960 {
2961         return btrfs_del_orphan_item(trans, inode->root, btrfs_ino(inode));
2962 }
2963
2964 /*
2965  * this cleans up any orphans that may be left on the list from the last use
2966  * of this root.
2967  */
2968 int btrfs_orphan_cleanup(struct btrfs_root *root)
2969 {
2970         struct btrfs_fs_info *fs_info = root->fs_info;
2971         struct btrfs_path *path;
2972         struct extent_buffer *leaf;
2973         struct btrfs_key key, found_key;
2974         struct btrfs_trans_handle *trans;
2975         struct inode *inode;
2976         u64 last_objectid = 0;
2977         int ret = 0, nr_unlink = 0;
2978
2979         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
2980                 return 0;
2981
2982         path = btrfs_alloc_path();
2983         if (!path) {
2984                 ret = -ENOMEM;
2985                 goto out;
2986         }
2987         path->reada = READA_BACK;
2988
2989         key.objectid = BTRFS_ORPHAN_OBJECTID;
2990         key.type = BTRFS_ORPHAN_ITEM_KEY;
2991         key.offset = (u64)-1;
2992
2993         while (1) {
2994                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2995                 if (ret < 0)
2996                         goto out;
2997
2998                 /*
2999                  * if ret == 0 means we found what we were searching for, which
3000                  * is weird, but possible, so only screw with path if we didn't
3001                  * find the key and see if we have stuff that matches
3002                  */
3003                 if (ret > 0) {
3004                         ret = 0;
3005                         if (path->slots[0] == 0)
3006                                 break;
3007                         path->slots[0]--;
3008                 }
3009
3010                 /* pull out the item */
3011                 leaf = path->nodes[0];
3012                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3013
3014                 /* make sure the item matches what we want */
3015                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3016                         break;
3017                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3018                         break;
3019
3020                 /* release the path since we're done with it */
3021                 btrfs_release_path(path);
3022
3023                 /*
3024                  * this is where we are basically btrfs_lookup, without the
3025                  * crossing root thing.  we store the inode number in the
3026                  * offset of the orphan item.
3027                  */
3028
3029                 if (found_key.offset == last_objectid) {
3030                         btrfs_err(fs_info,
3031                                   "Error removing orphan entry, stopping orphan cleanup");
3032                         ret = -EINVAL;
3033                         goto out;
3034                 }
3035
3036                 last_objectid = found_key.offset;
3037
3038                 found_key.objectid = found_key.offset;
3039                 found_key.type = BTRFS_INODE_ITEM_KEY;
3040                 found_key.offset = 0;
3041                 inode = btrfs_iget(fs_info->sb, last_objectid, root);
3042                 ret = PTR_ERR_OR_ZERO(inode);
3043                 if (ret && ret != -ENOENT)
3044                         goto out;
3045
3046                 if (ret == -ENOENT && root == fs_info->tree_root) {
3047                         struct btrfs_root *dead_root;
3048                         struct btrfs_fs_info *fs_info = root->fs_info;
3049                         int is_dead_root = 0;
3050
3051                         /*
3052                          * this is an orphan in the tree root. Currently these
3053                          * could come from 2 sources:
3054                          *  a) a snapshot deletion in progress
3055                          *  b) a free space cache inode
3056                          * We need to distinguish those two, as the snapshot
3057                          * orphan must not get deleted.
3058                          * find_dead_roots already ran before us, so if this
3059                          * is a snapshot deletion, we should find the root
3060                          * in the fs_roots radix tree.
3061                          */
3062
3063                         spin_lock(&fs_info->fs_roots_radix_lock);
3064                         dead_root = radix_tree_lookup(&fs_info->fs_roots_radix,
3065                                                          (unsigned long)found_key.objectid);
3066                         if (dead_root && btrfs_root_refs(&dead_root->root_item) == 0)
3067                                 is_dead_root = 1;
3068                         spin_unlock(&fs_info->fs_roots_radix_lock);
3069
3070                         if (is_dead_root) {
3071                                 /* prevent this orphan from being found again */
3072                                 key.offset = found_key.objectid - 1;
3073                                 continue;
3074                         }
3075
3076                 }
3077
3078                 /*
3079                  * If we have an inode with links, there are a couple of
3080                  * possibilities. Old kernels (before v3.12) used to create an
3081                  * orphan item for truncate indicating that there were possibly
3082                  * extent items past i_size that needed to be deleted. In v3.12,
3083                  * truncate was changed to update i_size in sync with the extent
3084                  * items, but the (useless) orphan item was still created. Since
3085                  * v4.18, we don't create the orphan item for truncate at all.
3086                  *
3087                  * So, this item could mean that we need to do a truncate, but
3088                  * only if this filesystem was last used on a pre-v3.12 kernel
3089                  * and was not cleanly unmounted. The odds of that are quite
3090                  * slim, and it's a pain to do the truncate now, so just delete
3091                  * the orphan item.
3092                  *
3093                  * It's also possible that this orphan item was supposed to be
3094                  * deleted but wasn't. The inode number may have been reused,
3095                  * but either way, we can delete the orphan item.
3096                  */
3097                 if (ret == -ENOENT || inode->i_nlink) {
3098                         if (!ret)
3099                                 iput(inode);
3100                         trans = btrfs_start_transaction(root, 1);
3101                         if (IS_ERR(trans)) {
3102                                 ret = PTR_ERR(trans);
3103                                 goto out;
3104                         }
3105                         btrfs_debug(fs_info, "auto deleting %Lu",
3106                                     found_key.objectid);
3107                         ret = btrfs_del_orphan_item(trans, root,
3108                                                     found_key.objectid);
3109                         btrfs_end_transaction(trans);
3110                         if (ret)
3111                                 goto out;
3112                         continue;
3113                 }
3114
3115                 nr_unlink++;
3116
3117                 /* this will do delete_inode and everything for us */
3118                 iput(inode);
3119         }
3120         /* release the path since we're done with it */
3121         btrfs_release_path(path);
3122
3123         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3124
3125         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3126                 trans = btrfs_join_transaction(root);
3127                 if (!IS_ERR(trans))
3128                         btrfs_end_transaction(trans);
3129         }
3130
3131         if (nr_unlink)
3132                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3133
3134 out:
3135         if (ret)
3136                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3137         btrfs_free_path(path);
3138         return ret;
3139 }
3140
3141 /*
3142  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3143  * don't find any xattrs, we know there can't be any acls.
3144  *
3145  * slot is the slot the inode is in, objectid is the objectid of the inode
3146  */
3147 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3148                                           int slot, u64 objectid,
3149                                           int *first_xattr_slot)
3150 {
3151         u32 nritems = btrfs_header_nritems(leaf);
3152         struct btrfs_key found_key;
3153         static u64 xattr_access = 0;
3154         static u64 xattr_default = 0;
3155         int scanned = 0;
3156
3157         if (!xattr_access) {
3158                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3159                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3160                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3161                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3162         }
3163
3164         slot++;
3165         *first_xattr_slot = -1;
3166         while (slot < nritems) {
3167                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3168
3169                 /* we found a different objectid, there must not be acls */
3170                 if (found_key.objectid != objectid)
3171                         return 0;
3172
3173                 /* we found an xattr, assume we've got an acl */
3174                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3175                         if (*first_xattr_slot == -1)
3176                                 *first_xattr_slot = slot;
3177                         if (found_key.offset == xattr_access ||
3178                             found_key.offset == xattr_default)
3179                                 return 1;
3180                 }
3181
3182                 /*
3183                  * we found a key greater than an xattr key, there can't
3184                  * be any acls later on
3185                  */
3186                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3187                         return 0;
3188
3189                 slot++;
3190                 scanned++;
3191
3192                 /*
3193                  * it goes inode, inode backrefs, xattrs, extents,
3194                  * so if there are a ton of hard links to an inode there can
3195                  * be a lot of backrefs.  Don't waste time searching too hard,
3196                  * this is just an optimization
3197                  */
3198                 if (scanned >= 8)
3199                         break;
3200         }
3201         /* we hit the end of the leaf before we found an xattr or
3202          * something larger than an xattr.  We have to assume the inode
3203          * has acls
3204          */
3205         if (*first_xattr_slot == -1)
3206                 *first_xattr_slot = slot;
3207         return 1;
3208 }
3209
3210 /*
3211  * read an inode from the btree into the in-memory inode
3212  */
3213 static int btrfs_read_locked_inode(struct inode *inode,
3214                                    struct btrfs_path *in_path)
3215 {
3216         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3217         struct btrfs_path *path = in_path;
3218         struct extent_buffer *leaf;
3219         struct btrfs_inode_item *inode_item;
3220         struct btrfs_root *root = BTRFS_I(inode)->root;
3221         struct btrfs_key location;
3222         unsigned long ptr;
3223         int maybe_acls;
3224         u32 rdev;
3225         int ret;
3226         bool filled = false;
3227         int first_xattr_slot;
3228
3229         ret = btrfs_fill_inode(inode, &rdev);
3230         if (!ret)
3231                 filled = true;
3232
3233         if (!path) {
3234                 path = btrfs_alloc_path();
3235                 if (!path)
3236                         return -ENOMEM;
3237         }
3238
3239         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3240
3241         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3242         if (ret) {
3243                 if (path != in_path)
3244                         btrfs_free_path(path);
3245                 return ret;
3246         }
3247
3248         leaf = path->nodes[0];
3249
3250         if (filled)
3251                 goto cache_index;
3252
3253         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3254                                     struct btrfs_inode_item);
3255         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3256         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3257         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3258         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3259         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3260         btrfs_inode_set_file_extent_range(BTRFS_I(inode), 0,
3261                         round_up(i_size_read(inode), fs_info->sectorsize));
3262
3263         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3264         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3265
3266         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3267         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3268
3269         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3270         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3271
3272         BTRFS_I(inode)->i_otime.tv_sec =
3273                 btrfs_timespec_sec(leaf, &inode_item->otime);
3274         BTRFS_I(inode)->i_otime.tv_nsec =
3275                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3276
3277         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3278         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3279         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3280
3281         inode_set_iversion_queried(inode,
3282                                    btrfs_inode_sequence(leaf, inode_item));
3283         inode->i_generation = BTRFS_I(inode)->generation;
3284         inode->i_rdev = 0;
3285         rdev = btrfs_inode_rdev(leaf, inode_item);
3286
3287         BTRFS_I(inode)->index_cnt = (u64)-1;
3288         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3289
3290 cache_index:
3291         /*
3292          * If we were modified in the current generation and evicted from memory
3293          * and then re-read we need to do a full sync since we don't have any
3294          * idea about which extents were modified before we were evicted from
3295          * cache.
3296          *
3297          * This is required for both inode re-read from disk and delayed inode
3298          * in delayed_nodes_tree.
3299          */
3300         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3301                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3302                         &BTRFS_I(inode)->runtime_flags);
3303
3304         /*
3305          * We don't persist the id of the transaction where an unlink operation
3306          * against the inode was last made. So here we assume the inode might
3307          * have been evicted, and therefore the exact value of last_unlink_trans
3308          * lost, and set it to last_trans to avoid metadata inconsistencies
3309          * between the inode and its parent if the inode is fsync'ed and the log
3310          * replayed. For example, in the scenario:
3311          *
3312          * touch mydir/foo
3313          * ln mydir/foo mydir/bar
3314          * sync
3315          * unlink mydir/bar
3316          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3317          * xfs_io -c fsync mydir/foo
3318          * <power failure>
3319          * mount fs, triggers fsync log replay
3320          *
3321          * We must make sure that when we fsync our inode foo we also log its
3322          * parent inode, otherwise after log replay the parent still has the
3323          * dentry with the "bar" name but our inode foo has a link count of 1
3324          * and doesn't have an inode ref with the name "bar" anymore.
3325          *
3326          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3327          * but it guarantees correctness at the expense of occasional full
3328          * transaction commits on fsync if our inode is a directory, or if our
3329          * inode is not a directory, logging its parent unnecessarily.
3330          */
3331         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3332
3333         path->slots[0]++;
3334         if (inode->i_nlink != 1 ||
3335             path->slots[0] >= btrfs_header_nritems(leaf))
3336                 goto cache_acl;
3337
3338         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3339         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3340                 goto cache_acl;
3341
3342         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3343         if (location.type == BTRFS_INODE_REF_KEY) {
3344                 struct btrfs_inode_ref *ref;
3345
3346                 ref = (struct btrfs_inode_ref *)ptr;
3347                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3348         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3349                 struct btrfs_inode_extref *extref;
3350
3351                 extref = (struct btrfs_inode_extref *)ptr;
3352                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3353                                                                      extref);
3354         }
3355 cache_acl:
3356         /*
3357          * try to precache a NULL acl entry for files that don't have
3358          * any xattrs or acls
3359          */
3360         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3361                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3362         if (first_xattr_slot != -1) {
3363                 path->slots[0] = first_xattr_slot;
3364                 ret = btrfs_load_inode_props(inode, path);
3365                 if (ret)
3366                         btrfs_err(fs_info,
3367                                   "error loading props for ino %llu (root %llu): %d",
3368                                   btrfs_ino(BTRFS_I(inode)),
3369                                   root->root_key.objectid, ret);
3370         }
3371         if (path != in_path)
3372                 btrfs_free_path(path);
3373
3374         if (!maybe_acls)
3375                 cache_no_acl(inode);
3376
3377         switch (inode->i_mode & S_IFMT) {
3378         case S_IFREG:
3379                 inode->i_mapping->a_ops = &btrfs_aops;
3380                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3381                 inode->i_fop = &btrfs_file_operations;
3382                 inode->i_op = &btrfs_file_inode_operations;
3383                 break;
3384         case S_IFDIR:
3385                 inode->i_fop = &btrfs_dir_file_operations;
3386                 inode->i_op = &btrfs_dir_inode_operations;
3387                 break;
3388         case S_IFLNK:
3389                 inode->i_op = &btrfs_symlink_inode_operations;
3390                 inode_nohighmem(inode);
3391                 inode->i_mapping->a_ops = &btrfs_aops;
3392                 break;
3393         default:
3394                 inode->i_op = &btrfs_special_inode_operations;
3395                 init_special_inode(inode, inode->i_mode, rdev);
3396                 break;
3397         }
3398
3399         btrfs_sync_inode_flags_to_i_flags(inode);
3400         return 0;
3401 }
3402
3403 /*
3404  * given a leaf and an inode, copy the inode fields into the leaf
3405  */
3406 static void fill_inode_item(struct btrfs_trans_handle *trans,
3407                             struct extent_buffer *leaf,
3408                             struct btrfs_inode_item *item,
3409                             struct inode *inode)
3410 {
3411         struct btrfs_map_token token;
3412
3413         btrfs_init_map_token(&token, leaf);
3414
3415         btrfs_set_token_inode_uid(&token, item, i_uid_read(inode));
3416         btrfs_set_token_inode_gid(&token, item, i_gid_read(inode));
3417         btrfs_set_token_inode_size(&token, item, BTRFS_I(inode)->disk_i_size);
3418         btrfs_set_token_inode_mode(&token, item, inode->i_mode);
3419         btrfs_set_token_inode_nlink(&token, item, inode->i_nlink);
3420
3421         btrfs_set_token_timespec_sec(&token, &item->atime,
3422                                      inode->i_atime.tv_sec);
3423         btrfs_set_token_timespec_nsec(&token, &item->atime,
3424                                       inode->i_atime.tv_nsec);
3425
3426         btrfs_set_token_timespec_sec(&token, &item->mtime,
3427                                      inode->i_mtime.tv_sec);
3428         btrfs_set_token_timespec_nsec(&token, &item->mtime,
3429                                       inode->i_mtime.tv_nsec);
3430
3431         btrfs_set_token_timespec_sec(&token, &item->ctime,
3432                                      inode->i_ctime.tv_sec);
3433         btrfs_set_token_timespec_nsec(&token, &item->ctime,
3434                                       inode->i_ctime.tv_nsec);
3435
3436         btrfs_set_token_timespec_sec(&token, &item->otime,
3437                                      BTRFS_I(inode)->i_otime.tv_sec);
3438         btrfs_set_token_timespec_nsec(&token, &item->otime,
3439                                       BTRFS_I(inode)->i_otime.tv_nsec);
3440
3441         btrfs_set_token_inode_nbytes(&token, item, inode_get_bytes(inode));
3442         btrfs_set_token_inode_generation(&token, item,
3443                                          BTRFS_I(inode)->generation);
3444         btrfs_set_token_inode_sequence(&token, item, inode_peek_iversion(inode));
3445         btrfs_set_token_inode_transid(&token, item, trans->transid);
3446         btrfs_set_token_inode_rdev(&token, item, inode->i_rdev);
3447         btrfs_set_token_inode_flags(&token, item, BTRFS_I(inode)->flags);
3448         btrfs_set_token_inode_block_group(&token, item, 0);
3449 }
3450
3451 /*
3452  * copy everything in the in-memory inode into the btree.
3453  */
3454 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3455                                 struct btrfs_root *root, struct inode *inode)
3456 {
3457         struct btrfs_inode_item *inode_item;
3458         struct btrfs_path *path;
3459         struct extent_buffer *leaf;
3460         int ret;
3461
3462         path = btrfs_alloc_path();
3463         if (!path)
3464                 return -ENOMEM;
3465
3466         path->leave_spinning = 1;
3467         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3468                                  1);
3469         if (ret) {
3470                 if (ret > 0)
3471                         ret = -ENOENT;
3472                 goto failed;
3473         }
3474
3475         leaf = path->nodes[0];
3476         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3477                                     struct btrfs_inode_item);
3478
3479         fill_inode_item(trans, leaf, inode_item, inode);
3480         btrfs_mark_buffer_dirty(leaf);
3481         btrfs_set_inode_last_trans(trans, inode);
3482         ret = 0;
3483 failed:
3484         btrfs_free_path(path);
3485         return ret;
3486 }
3487
3488 /*
3489  * copy everything in the in-memory inode into the btree.
3490  */
3491 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3492                                 struct btrfs_root *root, struct inode *inode)
3493 {
3494         struct btrfs_fs_info *fs_info = root->fs_info;
3495         int ret;
3496
3497         /*
3498          * If the inode is a free space inode, we can deadlock during commit
3499          * if we put it into the delayed code.
3500          *
3501          * The data relocation inode should also be directly updated
3502          * without delay
3503          */
3504         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
3505             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3506             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
3507                 btrfs_update_root_times(trans, root);
3508
3509                 ret = btrfs_delayed_update_inode(trans, root, inode);
3510                 if (!ret)
3511                         btrfs_set_inode_last_trans(trans, inode);
3512                 return ret;
3513         }
3514
3515         return btrfs_update_inode_item(trans, root, inode);
3516 }
3517
3518 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3519                                          struct btrfs_root *root,
3520                                          struct inode *inode)
3521 {
3522         int ret;
3523
3524         ret = btrfs_update_inode(trans, root, inode);
3525         if (ret == -ENOSPC)
3526                 return btrfs_update_inode_item(trans, root, inode);
3527         return ret;
3528 }
3529
3530 /*
3531  * unlink helper that gets used here in inode.c and in the tree logging
3532  * recovery code.  It remove a link in a directory with a given name, and
3533  * also drops the back refs in the inode to the directory
3534  */
3535 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3536                                 struct btrfs_root *root,
3537                                 struct btrfs_inode *dir,
3538                                 struct btrfs_inode *inode,
3539                                 const char *name, int name_len)
3540 {
3541         struct btrfs_fs_info *fs_info = root->fs_info;
3542         struct btrfs_path *path;
3543         int ret = 0;
3544         struct btrfs_dir_item *di;
3545         u64 index;
3546         u64 ino = btrfs_ino(inode);
3547         u64 dir_ino = btrfs_ino(dir);
3548
3549         path = btrfs_alloc_path();
3550         if (!path) {
3551                 ret = -ENOMEM;
3552                 goto out;
3553         }
3554
3555         path->leave_spinning = 1;
3556         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3557                                     name, name_len, -1);
3558         if (IS_ERR_OR_NULL(di)) {
3559                 ret = di ? PTR_ERR(di) : -ENOENT;
3560                 goto err;
3561         }
3562         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3563         if (ret)
3564                 goto err;
3565         btrfs_release_path(path);
3566
3567         /*
3568          * If we don't have dir index, we have to get it by looking up
3569          * the inode ref, since we get the inode ref, remove it directly,
3570          * it is unnecessary to do delayed deletion.
3571          *
3572          * But if we have dir index, needn't search inode ref to get it.
3573          * Since the inode ref is close to the inode item, it is better
3574          * that we delay to delete it, and just do this deletion when
3575          * we update the inode item.
3576          */
3577         if (inode->dir_index) {
3578                 ret = btrfs_delayed_delete_inode_ref(inode);
3579                 if (!ret) {
3580                         index = inode->dir_index;
3581                         goto skip_backref;
3582                 }
3583         }
3584
3585         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3586                                   dir_ino, &index);
3587         if (ret) {
3588                 btrfs_info(fs_info,
3589                         "failed to delete reference to %.*s, inode %llu parent %llu",
3590                         name_len, name, ino, dir_ino);
3591                 btrfs_abort_transaction(trans, ret);
3592                 goto err;
3593         }
3594 skip_backref:
3595         ret = btrfs_delete_delayed_dir_index(trans, dir, index);
3596         if (ret) {
3597                 btrfs_abort_transaction(trans, ret);
3598                 goto err;
3599         }
3600
3601         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
3602                         dir_ino);
3603         if (ret != 0 && ret != -ENOENT) {
3604                 btrfs_abort_transaction(trans, ret);
3605                 goto err;
3606         }
3607
3608         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
3609                         index);
3610         if (ret == -ENOENT)
3611                 ret = 0;
3612         else if (ret)
3613                 btrfs_abort_transaction(trans, ret);
3614
3615         /*
3616          * If we have a pending delayed iput we could end up with the final iput
3617          * being run in btrfs-cleaner context.  If we have enough of these built
3618          * up we can end up burning a lot of time in btrfs-cleaner without any
3619          * way to throttle the unlinks.  Since we're currently holding a ref on
3620          * the inode we can run the delayed iput here without any issues as the
3621          * final iput won't be done until after we drop the ref we're currently
3622          * holding.
3623          */
3624         btrfs_run_delayed_iput(fs_info, inode);
3625 err:
3626         btrfs_free_path(path);
3627         if (ret)
3628                 goto out;
3629
3630         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
3631         inode_inc_iversion(&inode->vfs_inode);
3632         inode_inc_iversion(&dir->vfs_inode);
3633         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
3634                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
3635         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
3636 out:
3637         return ret;
3638 }
3639
3640 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3641                        struct btrfs_root *root,
3642                        struct btrfs_inode *dir, struct btrfs_inode *inode,
3643                        const char *name, int name_len)
3644 {
3645         int ret;
3646         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
3647         if (!ret) {
3648                 drop_nlink(&inode->vfs_inode);
3649                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
3650         }
3651         return ret;
3652 }
3653
3654 /*
3655  * helper to start transaction for unlink and rmdir.
3656  *
3657  * unlink and rmdir are special in btrfs, they do not always free space, so
3658  * if we cannot make our reservations the normal way try and see if there is
3659  * plenty of slack room in the global reserve to migrate, otherwise we cannot
3660  * allow the unlink to occur.
3661  */
3662 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
3663 {
3664         struct btrfs_root *root = BTRFS_I(dir)->root;
3665
3666         /*
3667          * 1 for the possible orphan item
3668          * 1 for the dir item
3669          * 1 for the dir index
3670          * 1 for the inode ref
3671          * 1 for the inode
3672          */
3673         return btrfs_start_transaction_fallback_global_rsv(root, 5);
3674 }
3675
3676 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
3677 {
3678         struct btrfs_root *root = BTRFS_I(dir)->root;
3679         struct btrfs_trans_handle *trans;
3680         struct inode *inode = d_inode(dentry);
3681         int ret;
3682
3683         trans = __unlink_start_trans(dir);
3684         if (IS_ERR(trans))
3685                 return PTR_ERR(trans);
3686
3687         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
3688                         0);
3689
3690         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
3691                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
3692                         dentry->d_name.len);
3693         if (ret)
3694                 goto out;
3695
3696         if (inode->i_nlink == 0) {
3697                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3698                 if (ret)
3699                         goto out;
3700         }
3701
3702 out:
3703         btrfs_end_transaction(trans);
3704         btrfs_btree_balance_dirty(root->fs_info);
3705         return ret;
3706 }
3707
3708 static int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
3709                                struct inode *dir, struct dentry *dentry)
3710 {
3711         struct btrfs_root *root = BTRFS_I(dir)->root;
3712         struct btrfs_inode *inode = BTRFS_I(d_inode(dentry));
3713         struct btrfs_path *path;
3714         struct extent_buffer *leaf;
3715         struct btrfs_dir_item *di;
3716         struct btrfs_key key;
3717         const char *name = dentry->d_name.name;
3718         int name_len = dentry->d_name.len;
3719         u64 index;
3720         int ret;
3721         u64 objectid;
3722         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
3723
3724         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID) {
3725                 objectid = inode->root->root_key.objectid;
3726         } else if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID) {
3727                 objectid = inode->location.objectid;
3728         } else {
3729                 WARN_ON(1);
3730                 return -EINVAL;
3731         }
3732
3733         path = btrfs_alloc_path();
3734         if (!path)
3735                 return -ENOMEM;
3736
3737         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3738                                    name, name_len, -1);
3739         if (IS_ERR_OR_NULL(di)) {
3740                 ret = di ? PTR_ERR(di) : -ENOENT;
3741                 goto out;
3742         }
3743
3744         leaf = path->nodes[0];
3745         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3746         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
3747         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3748         if (ret) {
3749                 btrfs_abort_transaction(trans, ret);
3750                 goto out;
3751         }
3752         btrfs_release_path(path);
3753
3754         /*
3755          * This is a placeholder inode for a subvolume we didn't have a
3756          * reference to at the time of the snapshot creation.  In the meantime
3757          * we could have renamed the real subvol link into our snapshot, so
3758          * depending on btrfs_del_root_ref to return -ENOENT here is incorret.
3759          * Instead simply lookup the dir_index_item for this entry so we can
3760          * remove it.  Otherwise we know we have a ref to the root and we can
3761          * call btrfs_del_root_ref, and it _shouldn't_ fail.
3762          */
3763         if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID) {
3764                 di = btrfs_search_dir_index_item(root, path, dir_ino,
3765                                                  name, name_len);
3766                 if (IS_ERR_OR_NULL(di)) {
3767                         if (!di)
3768                                 ret = -ENOENT;
3769                         else
3770                                 ret = PTR_ERR(di);
3771                         btrfs_abort_transaction(trans, ret);
3772                         goto out;
3773                 }
3774
3775                 leaf = path->nodes[0];
3776                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3777                 index = key.offset;
3778                 btrfs_release_path(path);
3779         } else {
3780                 ret = btrfs_del_root_ref(trans, objectid,
3781                                          root->root_key.objectid, dir_ino,
3782                                          &index, name, name_len);
3783                 if (ret) {
3784                         btrfs_abort_transaction(trans, ret);
3785                         goto out;
3786                 }
3787         }
3788
3789         ret = btrfs_delete_delayed_dir_index(trans, BTRFS_I(dir), index);
3790         if (ret) {
3791                 btrfs_abort_transaction(trans, ret);
3792                 goto out;
3793         }
3794
3795         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
3796         inode_inc_iversion(dir);
3797         dir->i_mtime = dir->i_ctime = current_time(dir);
3798         ret = btrfs_update_inode_fallback(trans, root, dir);
3799         if (ret)
3800                 btrfs_abort_transaction(trans, ret);
3801 out:
3802         btrfs_free_path(path);
3803         return ret;
3804 }
3805
3806 /*
3807  * Helper to check if the subvolume references other subvolumes or if it's
3808  * default.
3809  */
3810 static noinline int may_destroy_subvol(struct btrfs_root *root)
3811 {
3812         struct btrfs_fs_info *fs_info = root->fs_info;
3813         struct btrfs_path *path;
3814         struct btrfs_dir_item *di;
3815         struct btrfs_key key;
3816         u64 dir_id;
3817         int ret;
3818
3819         path = btrfs_alloc_path();
3820         if (!path)
3821                 return -ENOMEM;
3822
3823         /* Make sure this root isn't set as the default subvol */
3824         dir_id = btrfs_super_root_dir(fs_info->super_copy);
3825         di = btrfs_lookup_dir_item(NULL, fs_info->tree_root, path,
3826                                    dir_id, "default", 7, 0);
3827         if (di && !IS_ERR(di)) {
3828                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
3829                 if (key.objectid == root->root_key.objectid) {
3830                         ret = -EPERM;
3831                         btrfs_err(fs_info,
3832                                   "deleting default subvolume %llu is not allowed",
3833                                   key.objectid);
3834                         goto out;
3835                 }
3836                 btrfs_release_path(path);
3837         }
3838
3839         key.objectid = root->root_key.objectid;
3840         key.type = BTRFS_ROOT_REF_KEY;
3841         key.offset = (u64)-1;
3842
3843         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
3844         if (ret < 0)
3845                 goto out;
3846         BUG_ON(ret == 0);
3847
3848         ret = 0;
3849         if (path->slots[0] > 0) {
3850                 path->slots[0]--;
3851                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3852                 if (key.objectid == root->root_key.objectid &&
3853                     key.type == BTRFS_ROOT_REF_KEY)
3854                         ret = -ENOTEMPTY;
3855         }
3856 out:
3857         btrfs_free_path(path);
3858         return ret;
3859 }
3860
3861 /* Delete all dentries for inodes belonging to the root */
3862 static void btrfs_prune_dentries(struct btrfs_root *root)
3863 {
3864         struct btrfs_fs_info *fs_info = root->fs_info;
3865         struct rb_node *node;
3866         struct rb_node *prev;
3867         struct btrfs_inode *entry;
3868         struct inode *inode;
3869         u64 objectid = 0;
3870
3871         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3872                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
3873
3874         spin_lock(&root->inode_lock);
3875 again:
3876         node = root->inode_tree.rb_node;
3877         prev = NULL;
3878         while (node) {
3879                 prev = node;
3880                 entry = rb_entry(node, struct btrfs_inode, rb_node);
3881
3882                 if (objectid < btrfs_ino(entry))
3883                         node = node->rb_left;
3884                 else if (objectid > btrfs_ino(entry))
3885                         node = node->rb_right;
3886                 else
3887                         break;
3888         }
3889         if (!node) {
3890                 while (prev) {
3891                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
3892                         if (objectid <= btrfs_ino(entry)) {
3893                                 node = prev;
3894                                 break;
3895                         }
3896                         prev = rb_next(prev);
3897                 }
3898         }
3899         while (node) {
3900                 entry = rb_entry(node, struct btrfs_inode, rb_node);
3901                 objectid = btrfs_ino(entry) + 1;
3902                 inode = igrab(&entry->vfs_inode);
3903                 if (inode) {
3904                         spin_unlock(&root->inode_lock);
3905                         if (atomic_read(&inode->i_count) > 1)
3906                                 d_prune_aliases(inode);
3907                         /*
3908                          * btrfs_drop_inode will have it removed from the inode
3909                          * cache when its usage count hits zero.
3910                          */
3911                         iput(inode);
3912                         cond_resched();
3913                         spin_lock(&root->inode_lock);
3914                         goto again;
3915                 }
3916
3917                 if (cond_resched_lock(&root->inode_lock))
3918                         goto again;
3919
3920                 node = rb_next(node);
3921         }
3922         spin_unlock(&root->inode_lock);
3923 }
3924
3925 int btrfs_delete_subvolume(struct inode *dir, struct dentry *dentry)
3926 {
3927         struct btrfs_fs_info *fs_info = btrfs_sb(dentry->d_sb);
3928         struct btrfs_root *root = BTRFS_I(dir)->root;
3929         struct inode *inode = d_inode(dentry);
3930         struct btrfs_root *dest = BTRFS_I(inode)->root;
3931         struct btrfs_trans_handle *trans;
3932         struct btrfs_block_rsv block_rsv;
3933         u64 root_flags;
3934         int ret;
3935         int err;
3936
3937         /*
3938          * Don't allow to delete a subvolume with send in progress. This is
3939          * inside the inode lock so the error handling that has to drop the bit
3940          * again is not run concurrently.
3941          */
3942         spin_lock(&dest->root_item_lock);
3943         if (dest->send_in_progress) {
3944                 spin_unlock(&dest->root_item_lock);
3945                 btrfs_warn(fs_info,
3946                            "attempt to delete subvolume %llu during send",
3947                            dest->root_key.objectid);
3948                 return -EPERM;
3949         }
3950         root_flags = btrfs_root_flags(&dest->root_item);
3951         btrfs_set_root_flags(&dest->root_item,
3952                              root_flags | BTRFS_ROOT_SUBVOL_DEAD);
3953         spin_unlock(&dest->root_item_lock);
3954
3955         down_write(&fs_info->subvol_sem);
3956
3957         err = may_destroy_subvol(dest);
3958         if (err)
3959                 goto out_up_write;
3960
3961         btrfs_init_block_rsv(&block_rsv, BTRFS_BLOCK_RSV_TEMP);
3962         /*
3963          * One for dir inode,
3964          * two for dir entries,
3965          * two for root ref/backref.
3966          */
3967         err = btrfs_subvolume_reserve_metadata(root, &block_rsv, 5, true);
3968         if (err)
3969                 goto out_up_write;
3970
3971         trans = btrfs_start_transaction(root, 0);
3972         if (IS_ERR(trans)) {
3973                 err = PTR_ERR(trans);
3974                 goto out_release;
3975         }
3976         trans->block_rsv = &block_rsv;
3977         trans->bytes_reserved = block_rsv.size;
3978
3979         btrfs_record_snapshot_destroy(trans, BTRFS_I(dir));
3980
3981         ret = btrfs_unlink_subvol(trans, dir, dentry);
3982         if (ret) {
3983                 err = ret;
3984                 btrfs_abort_transaction(trans, ret);
3985                 goto out_end_trans;
3986         }
3987
3988         btrfs_record_root_in_trans(trans, dest);
3989
3990         memset(&dest->root_item.drop_progress, 0,
3991                 sizeof(dest->root_item.drop_progress));
3992         dest->root_item.drop_level = 0;
3993         btrfs_set_root_refs(&dest->root_item, 0);
3994
3995         if (!test_and_set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &dest->state)) {
3996                 ret = btrfs_insert_orphan_item(trans,
3997                                         fs_info->tree_root,
3998                                         dest->root_key.objectid);
3999                 if (ret) {
4000                         btrfs_abort_transaction(trans, ret);
4001                         err = ret;
4002                         goto out_end_trans;
4003                 }
4004         }
4005
4006         ret = btrfs_uuid_tree_remove(trans, dest->root_item.uuid,
4007                                   BTRFS_UUID_KEY_SUBVOL,
4008                                   dest->root_key.objectid);
4009         if (ret && ret != -ENOENT) {
4010                 btrfs_abort_transaction(trans, ret);
4011                 err = ret;
4012                 goto out_end_trans;
4013         }
4014         if (!btrfs_is_empty_uuid(dest->root_item.received_uuid)) {
4015                 ret = btrfs_uuid_tree_remove(trans,
4016                                           dest->root_item.received_uuid,
4017                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
4018                                           dest->root_key.objectid);
4019                 if (ret && ret != -ENOENT) {
4020                         btrfs_abort_transaction(trans, ret);
4021                         err = ret;
4022                         goto out_end_trans;
4023                 }
4024         }
4025
4026 out_end_trans:
4027         trans->block_rsv = NULL;
4028         trans->bytes_reserved = 0;
4029         ret = btrfs_end_transaction(trans);
4030         if (ret && !err)
4031                 err = ret;
4032         inode->i_flags |= S_DEAD;
4033 out_release:
4034         btrfs_subvolume_release_metadata(fs_info, &block_rsv);
4035 out_up_write:
4036         up_write(&fs_info->subvol_sem);
4037         if (err) {
4038                 spin_lock(&dest->root_item_lock);
4039                 root_flags = btrfs_root_flags(&dest->root_item);
4040                 btrfs_set_root_flags(&dest->root_item,
4041                                 root_flags & ~BTRFS_ROOT_SUBVOL_DEAD);
4042                 spin_unlock(&dest->root_item_lock);
4043         } else {
4044                 d_invalidate(dentry);
4045                 btrfs_prune_dentries(dest);
4046                 ASSERT(dest->send_in_progress == 0);
4047
4048                 /* the last ref */
4049                 if (dest->ino_cache_inode) {
4050                         iput(dest->ino_cache_inode);
4051                         dest->ino_cache_inode = NULL;
4052                 }
4053         }
4054
4055         return err;
4056 }
4057
4058 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4059 {
4060         struct inode *inode = d_inode(dentry);
4061         int err = 0;
4062         struct btrfs_root *root = BTRFS_I(dir)->root;
4063         struct btrfs_trans_handle *trans;
4064         u64 last_unlink_trans;
4065
4066         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4067                 return -ENOTEMPTY;
4068         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4069                 return btrfs_delete_subvolume(dir, dentry);
4070
4071         trans = __unlink_start_trans(dir);
4072         if (IS_ERR(trans))
4073                 return PTR_ERR(trans);
4074
4075         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4076                 err = btrfs_unlink_subvol(trans, dir, dentry);
4077                 goto out;
4078         }
4079
4080         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4081         if (err)
4082                 goto out;
4083
4084         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4085
4086         /* now the directory is empty */
4087         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4088                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4089                         dentry->d_name.len);
4090         if (!err) {
4091                 btrfs_i_size_write(BTRFS_I(inode), 0);
4092                 /*
4093                  * Propagate the last_unlink_trans value of the deleted dir to
4094                  * its parent directory. This is to prevent an unrecoverable
4095                  * log tree in the case we do something like this:
4096                  * 1) create dir foo
4097                  * 2) create snapshot under dir foo
4098                  * 3) delete the snapshot
4099                  * 4) rmdir foo
4100                  * 5) mkdir foo
4101                  * 6) fsync foo or some file inside foo
4102                  */
4103                 if (last_unlink_trans >= trans->transid)
4104                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4105         }
4106 out:
4107         btrfs_end_transaction(trans);
4108         btrfs_btree_balance_dirty(root->fs_info);
4109
4110         return err;
4111 }
4112
4113 /*
4114  * Return this if we need to call truncate_block for the last bit of the
4115  * truncate.
4116  */
4117 #define NEED_TRUNCATE_BLOCK 1
4118
4119 /*
4120  * this can truncate away extent items, csum items and directory items.
4121  * It starts at a high offset and removes keys until it can't find
4122  * any higher than new_size
4123  *
4124  * csum items that cross the new i_size are truncated to the new size
4125  * as well.
4126  *
4127  * min_type is the minimum key type to truncate down to.  If set to 0, this
4128  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4129  */
4130 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4131                                struct btrfs_root *root,
4132                                struct inode *inode,
4133                                u64 new_size, u32 min_type)
4134 {
4135         struct btrfs_fs_info *fs_info = root->fs_info;
4136         struct btrfs_path *path;
4137         struct extent_buffer *leaf;
4138         struct btrfs_file_extent_item *fi;
4139         struct btrfs_key key;
4140         struct btrfs_key found_key;
4141         u64 extent_start = 0;
4142         u64 extent_num_bytes = 0;
4143         u64 extent_offset = 0;
4144         u64 item_end = 0;
4145         u64 last_size = new_size;
4146         u32 found_type = (u8)-1;
4147         int found_extent;
4148         int del_item;
4149         int pending_del_nr = 0;
4150         int pending_del_slot = 0;
4151         int extent_type = -1;
4152         int ret;
4153         u64 ino = btrfs_ino(BTRFS_I(inode));
4154         u64 bytes_deleted = 0;
4155         bool be_nice = false;
4156         bool should_throttle = false;
4157         const u64 lock_start = ALIGN_DOWN(new_size, fs_info->sectorsize);
4158         struct extent_state *cached_state = NULL;
4159
4160         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4161
4162         /*
4163          * For non-free space inodes and non-shareable roots, we want to back
4164          * off from time to time.  This means all inodes in subvolume roots,
4165          * reloc roots, and data reloc roots.
4166          */
4167         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4168             test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
4169                 be_nice = true;
4170
4171         path = btrfs_alloc_path();
4172         if (!path)
4173                 return -ENOMEM;
4174         path->reada = READA_BACK;
4175
4176         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4177                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, (u64)-1,
4178                                  &cached_state);
4179
4180                 /*
4181                  * We want to drop from the next block forward in case this
4182                  * new size is not block aligned since we will be keeping the
4183                  * last block of the extent just the way it is.
4184                  */
4185                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4186                                         fs_info->sectorsize),
4187                                         (u64)-1, 0);
4188         }
4189
4190         /*
4191          * This function is also used to drop the items in the log tree before
4192          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4193          * it is used to drop the logged items. So we shouldn't kill the delayed
4194          * items.
4195          */
4196         if (min_type == 0 && root == BTRFS_I(inode)->root)
4197                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4198
4199         key.objectid = ino;
4200         key.offset = (u64)-1;
4201         key.type = (u8)-1;
4202
4203 search_again:
4204         /*
4205          * with a 16K leaf size and 128MB extents, you can actually queue
4206          * up a huge file in a single leaf.  Most of the time that
4207          * bytes_deleted is > 0, it will be huge by the time we get here
4208          */
4209         if (be_nice && bytes_deleted > SZ_32M &&
4210             btrfs_should_end_transaction(trans)) {
4211                 ret = -EAGAIN;
4212                 goto out;
4213         }
4214
4215         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4216         if (ret < 0)
4217                 goto out;
4218
4219         if (ret > 0) {
4220                 ret = 0;
4221                 /* there are no items in the tree for us to truncate, we're
4222                  * done
4223                  */
4224                 if (path->slots[0] == 0)
4225                         goto out;
4226                 path->slots[0]--;
4227         }
4228
4229         while (1) {
4230                 u64 clear_start = 0, clear_len = 0;
4231
4232                 fi = NULL;
4233                 leaf = path->nodes[0];
4234                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4235                 found_type = found_key.type;
4236
4237                 if (found_key.objectid != ino)
4238                         break;
4239
4240                 if (found_type < min_type)
4241                         break;
4242
4243                 item_end = found_key.offset;
4244                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4245                         fi = btrfs_item_ptr(leaf, path->slots[0],
4246                                             struct btrfs_file_extent_item);
4247                         extent_type = btrfs_file_extent_type(leaf, fi);
4248                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4249                                 item_end +=
4250                                     btrfs_file_extent_num_bytes(leaf, fi);
4251
4252                                 trace_btrfs_truncate_show_fi_regular(
4253                                         BTRFS_I(inode), leaf, fi,
4254                                         found_key.offset);
4255                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4256                                 item_end += btrfs_file_extent_ram_bytes(leaf,
4257                                                                         fi);
4258
4259                                 trace_btrfs_truncate_show_fi_inline(
4260                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4261                                         found_key.offset);
4262                         }
4263                         item_end--;
4264                 }
4265                 if (found_type > min_type) {
4266                         del_item = 1;
4267                 } else {
4268                         if (item_end < new_size)
4269                                 break;
4270                         if (found_key.offset >= new_size)
4271                                 del_item = 1;
4272                         else
4273                                 del_item = 0;
4274                 }
4275                 found_extent = 0;
4276                 /* FIXME, shrink the extent if the ref count is only 1 */
4277                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4278                         goto delete;
4279
4280                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4281                         u64 num_dec;
4282
4283                         clear_start = found_key.offset;
4284                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4285                         if (!del_item) {
4286                                 u64 orig_num_bytes =
4287                                         btrfs_file_extent_num_bytes(leaf, fi);
4288                                 extent_num_bytes = ALIGN(new_size -
4289                                                 found_key.offset,
4290                                                 fs_info->sectorsize);
4291                                 clear_start = ALIGN(new_size, fs_info->sectorsize);
4292                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4293                                                          extent_num_bytes);
4294                                 num_dec = (orig_num_bytes -
4295                                            extent_num_bytes);
4296                                 if (test_bit(BTRFS_ROOT_SHAREABLE,
4297                                              &root->state) &&
4298                                     extent_start != 0)
4299                                         inode_sub_bytes(inode, num_dec);
4300                                 btrfs_mark_buffer_dirty(leaf);
4301                         } else {
4302                                 extent_num_bytes =
4303                                         btrfs_file_extent_disk_num_bytes(leaf,
4304                                                                          fi);
4305                                 extent_offset = found_key.offset -
4306                                         btrfs_file_extent_offset(leaf, fi);
4307
4308                                 /* FIXME blocksize != 4096 */
4309                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4310                                 if (extent_start != 0) {
4311                                         found_extent = 1;
4312                                         if (test_bit(BTRFS_ROOT_SHAREABLE,
4313                                                      &root->state))
4314                                                 inode_sub_bytes(inode, num_dec);
4315                                 }
4316                         }
4317                         clear_len = num_dec;
4318                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4319                         /*
4320                          * we can't truncate inline items that have had
4321                          * special encodings
4322                          */
4323                         if (!del_item &&
4324                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4325                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4326                             btrfs_file_extent_compression(leaf, fi) == 0) {
4327                                 u32 size = (u32)(new_size - found_key.offset);
4328
4329                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4330                                 size = btrfs_file_extent_calc_inline_size(size);
4331                                 btrfs_truncate_item(path, size, 1);
4332                         } else if (!del_item) {
4333                                 /*
4334                                  * We have to bail so the last_size is set to
4335                                  * just before this extent.
4336                                  */
4337                                 ret = NEED_TRUNCATE_BLOCK;
4338                                 break;
4339                         } else {
4340                                 /*
4341                                  * Inline extents are special, we just treat
4342                                  * them as a full sector worth in the file
4343                                  * extent tree just for simplicity sake.
4344                                  */
4345                                 clear_len = fs_info->sectorsize;
4346                         }
4347
4348                         if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
4349                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4350                 }
4351 delete:
4352                 /*
4353                  * We use btrfs_truncate_inode_items() to clean up log trees for
4354                  * multiple fsyncs, and in this case we don't want to clear the
4355                  * file extent range because it's just the log.
4356                  */
4357                 if (root == BTRFS_I(inode)->root) {
4358                         ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
4359                                                   clear_start, clear_len);
4360                         if (ret) {
4361                                 btrfs_abort_transaction(trans, ret);
4362                                 break;
4363                         }
4364                 }
4365
4366                 if (del_item)
4367                         last_size = found_key.offset;
4368                 else
4369                         last_size = new_size;
4370                 if (del_item) {
4371                         if (!pending_del_nr) {
4372                                 /* no pending yet, add ourselves */
4373                                 pending_del_slot = path->slots[0];
4374                                 pending_del_nr = 1;
4375                         } else if (pending_del_nr &&
4376                                    path->slots[0] + 1 == pending_del_slot) {
4377                                 /* hop on the pending chunk */
4378                                 pending_del_nr++;
4379                                 pending_del_slot = path->slots[0];
4380                         } else {
4381                                 BUG();
4382                         }
4383                 } else {
4384                         break;
4385                 }
4386                 should_throttle = false;
4387
4388                 if (found_extent &&
4389                     root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4390                         struct btrfs_ref ref = { 0 };
4391
4392                         bytes_deleted += extent_num_bytes;
4393
4394                         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF,
4395                                         extent_start, extent_num_bytes, 0);
4396                         ref.real_root = root->root_key.objectid;
4397                         btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
4398                                         ino, extent_offset);
4399                         ret = btrfs_free_extent(trans, &ref);
4400                         if (ret) {
4401                                 btrfs_abort_transaction(trans, ret);
4402                                 break;
4403                         }
4404                         if (be_nice) {
4405                                 if (btrfs_should_throttle_delayed_refs(trans))
4406                                         should_throttle = true;
4407                         }
4408                 }
4409
4410                 if (found_type == BTRFS_INODE_ITEM_KEY)
4411                         break;
4412
4413                 if (path->slots[0] == 0 ||
4414                     path->slots[0] != pending_del_slot ||
4415                     should_throttle) {
4416                         if (pending_del_nr) {
4417                                 ret = btrfs_del_items(trans, root, path,
4418                                                 pending_del_slot,
4419                                                 pending_del_nr);
4420                                 if (ret) {
4421                                         btrfs_abort_transaction(trans, ret);
4422                                         break;
4423                                 }
4424                                 pending_del_nr = 0;
4425                         }
4426                         btrfs_release_path(path);
4427
4428                         /*
4429                          * We can generate a lot of delayed refs, so we need to
4430                          * throttle every once and a while and make sure we're
4431                          * adding enough space to keep up with the work we are
4432                          * generating.  Since we hold a transaction here we
4433                          * can't flush, and we don't want to FLUSH_LIMIT because
4434                          * we could have generated too many delayed refs to
4435                          * actually allocate, so just bail if we're short and
4436                          * let the normal reservation dance happen higher up.
4437                          */
4438                         if (should_throttle) {
4439                                 ret = btrfs_delayed_refs_rsv_refill(fs_info,
4440                                                         BTRFS_RESERVE_NO_FLUSH);
4441                                 if (ret) {
4442                                         ret = -EAGAIN;
4443                                         break;
4444                                 }
4445                         }
4446                         goto search_again;
4447                 } else {
4448                         path->slots[0]--;
4449                 }
4450         }
4451 out:
4452         if (ret >= 0 && pending_del_nr) {
4453                 int err;
4454
4455                 err = btrfs_del_items(trans, root, path, pending_del_slot,
4456                                       pending_del_nr);
4457                 if (err) {
4458                         btrfs_abort_transaction(trans, err);
4459                         ret = err;
4460                 }
4461         }
4462         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4463                 ASSERT(last_size >= new_size);
4464                 if (!ret && last_size > new_size)
4465                         last_size = new_size;
4466                 btrfs_inode_safe_disk_i_size_write(inode, last_size);
4467                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start,
4468                                      (u64)-1, &cached_state);
4469         }
4470
4471         btrfs_free_path(path);
4472         return ret;
4473 }
4474
4475 /*
4476  * btrfs_truncate_block - read, zero a chunk and write a block
4477  * @inode - inode that we're zeroing
4478  * @from - the offset to start zeroing
4479  * @len - the length to zero, 0 to zero the entire range respective to the
4480  *      offset
4481  * @front - zero up to the offset instead of from the offset on
4482  *
4483  * This will find the block for the "from" offset and cow the block and zero the
4484  * part we want to zero.  This is used with truncate and hole punching.
4485  */
4486 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4487                         int front)
4488 {
4489         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4490         struct address_space *mapping = inode->i_mapping;
4491         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4492         struct btrfs_ordered_extent *ordered;
4493         struct extent_state *cached_state = NULL;
4494         struct extent_changeset *data_reserved = NULL;
4495         char *kaddr;
4496         u32 blocksize = fs_info->sectorsize;
4497         pgoff_t index = from >> PAGE_SHIFT;
4498         unsigned offset = from & (blocksize - 1);
4499         struct page *page;
4500         gfp_t mask = btrfs_alloc_write_mask(mapping);
4501         int ret = 0;
4502         u64 block_start;
4503         u64 block_end;
4504
4505         if (IS_ALIGNED(offset, blocksize) &&
4506             (!len || IS_ALIGNED(len, blocksize)))
4507                 goto out;
4508
4509         block_start = round_down(from, blocksize);
4510         block_end = block_start + blocksize - 1;
4511
4512         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4513                                            block_start, blocksize);
4514         if (ret)
4515                 goto out;
4516
4517 again:
4518         page = find_or_create_page(mapping, index, mask);
4519         if (!page) {
4520                 btrfs_delalloc_release_space(inode, data_reserved,
4521                                              block_start, blocksize, true);
4522                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4523                 ret = -ENOMEM;
4524                 goto out;
4525         }
4526
4527         if (!PageUptodate(page)) {
4528                 ret = btrfs_readpage(NULL, page);
4529                 lock_page(page);
4530                 if (page->mapping != mapping) {
4531                         unlock_page(page);
4532                         put_page(page);
4533                         goto again;
4534                 }
4535                 if (!PageUptodate(page)) {
4536                         ret = -EIO;
4537                         goto out_unlock;
4538                 }
4539         }
4540         wait_on_page_writeback(page);
4541
4542         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4543         set_page_extent_mapped(page);
4544
4545         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4546         if (ordered) {
4547                 unlock_extent_cached(io_tree, block_start, block_end,
4548                                      &cached_state);
4549                 unlock_page(page);
4550                 put_page(page);
4551                 btrfs_start_ordered_extent(inode, ordered, 1);
4552                 btrfs_put_ordered_extent(ordered);
4553                 goto again;
4554         }
4555
4556         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4557                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4558                          0, 0, &cached_state);
4559
4560         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4561                                         &cached_state);
4562         if (ret) {
4563                 unlock_extent_cached(io_tree, block_start, block_end,
4564                                      &cached_state);
4565                 goto out_unlock;
4566         }
4567
4568         if (offset != blocksize) {
4569                 if (!len)
4570                         len = blocksize - offset;
4571                 kaddr = kmap(page);
4572                 if (front)
4573                         memset(kaddr + (block_start - page_offset(page)),
4574                                 0, offset);
4575                 else
4576                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4577                                 0, len);
4578                 flush_dcache_page(page);
4579                 kunmap(page);
4580         }
4581         ClearPageChecked(page);
4582         set_page_dirty(page);
4583         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
4584
4585 out_unlock:
4586         if (ret)
4587                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4588                                              blocksize, true);
4589         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4590         unlock_page(page);
4591         put_page(page);
4592 out:
4593         extent_changeset_free(data_reserved);
4594         return ret;
4595 }
4596
4597 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4598                              u64 offset, u64 len)
4599 {
4600         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4601         struct btrfs_trans_handle *trans;
4602         int ret;
4603
4604         /*
4605          * Still need to make sure the inode looks like it's been updated so
4606          * that any holes get logged if we fsync.
4607          */
4608         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4609                 BTRFS_I(inode)->last_trans = fs_info->generation;
4610                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4611                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4612                 return 0;
4613         }
4614
4615         /*
4616          * 1 - for the one we're dropping
4617          * 1 - for the one we're adding
4618          * 1 - for updating the inode.
4619          */
4620         trans = btrfs_start_transaction(root, 3);
4621         if (IS_ERR(trans))
4622                 return PTR_ERR(trans);
4623
4624         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4625         if (ret) {
4626                 btrfs_abort_transaction(trans, ret);
4627                 btrfs_end_transaction(trans);
4628                 return ret;
4629         }
4630
4631         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4632                         offset, 0, 0, len, 0, len, 0, 0, 0);
4633         if (ret)
4634                 btrfs_abort_transaction(trans, ret);
4635         else
4636                 btrfs_update_inode(trans, root, inode);
4637         btrfs_end_transaction(trans);
4638         return ret;
4639 }
4640
4641 /*
4642  * This function puts in dummy file extents for the area we're creating a hole
4643  * for.  So if we are truncating this file to a larger size we need to insert
4644  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4645  * the range between oldsize and size
4646  */
4647 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4648 {
4649         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4650         struct btrfs_root *root = BTRFS_I(inode)->root;
4651         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4652         struct extent_map *em = NULL;
4653         struct extent_state *cached_state = NULL;
4654         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4655         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4656         u64 block_end = ALIGN(size, fs_info->sectorsize);
4657         u64 last_byte;
4658         u64 cur_offset;
4659         u64 hole_size;
4660         int err = 0;
4661
4662         /*
4663          * If our size started in the middle of a block we need to zero out the
4664          * rest of the block before we expand the i_size, otherwise we could
4665          * expose stale data.
4666          */
4667         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4668         if (err)
4669                 return err;
4670
4671         if (size <= hole_start)
4672                 return 0;
4673
4674         btrfs_lock_and_flush_ordered_range(BTRFS_I(inode), hole_start,
4675                                            block_end - 1, &cached_state);
4676         cur_offset = hole_start;
4677         while (1) {
4678                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
4679                                       block_end - cur_offset);
4680                 if (IS_ERR(em)) {
4681                         err = PTR_ERR(em);
4682                         em = NULL;
4683                         break;
4684                 }
4685                 last_byte = min(extent_map_end(em), block_end);
4686                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
4687                 hole_size = last_byte - cur_offset;
4688
4689                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4690                         struct extent_map *hole_em;
4691
4692                         err = maybe_insert_hole(root, inode, cur_offset,
4693                                                 hole_size);
4694                         if (err)
4695                                 break;
4696
4697                         err = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
4698                                                         cur_offset, hole_size);
4699                         if (err)
4700                                 break;
4701
4702                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
4703                                                 cur_offset + hole_size - 1, 0);
4704                         hole_em = alloc_extent_map();
4705                         if (!hole_em) {
4706                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4707                                         &BTRFS_I(inode)->runtime_flags);
4708                                 goto next;
4709                         }
4710                         hole_em->start = cur_offset;
4711                         hole_em->len = hole_size;
4712                         hole_em->orig_start = cur_offset;
4713
4714                         hole_em->block_start = EXTENT_MAP_HOLE;
4715                         hole_em->block_len = 0;
4716                         hole_em->orig_block_len = 0;
4717                         hole_em->ram_bytes = hole_size;
4718                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4719                         hole_em->generation = fs_info->generation;
4720
4721                         while (1) {
4722                                 write_lock(&em_tree->lock);
4723                                 err = add_extent_mapping(em_tree, hole_em, 1);
4724                                 write_unlock(&em_tree->lock);
4725                                 if (err != -EEXIST)
4726                                         break;
4727                                 btrfs_drop_extent_cache(BTRFS_I(inode),
4728                                                         cur_offset,
4729                                                         cur_offset +
4730                                                         hole_size - 1, 0);
4731                         }
4732                         free_extent_map(hole_em);
4733                 } else {
4734                         err = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
4735                                                         cur_offset, hole_size);
4736                         if (err)
4737                                 break;
4738                 }
4739 next:
4740                 free_extent_map(em);
4741                 em = NULL;
4742                 cur_offset = last_byte;
4743                 if (cur_offset >= block_end)
4744                         break;
4745         }
4746         free_extent_map(em);
4747         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
4748         return err;
4749 }
4750
4751 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4752 {
4753         struct btrfs_root *root = BTRFS_I(inode)->root;
4754         struct btrfs_trans_handle *trans;
4755         loff_t oldsize = i_size_read(inode);
4756         loff_t newsize = attr->ia_size;
4757         int mask = attr->ia_valid;
4758         int ret;
4759
4760         /*
4761          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4762          * special case where we need to update the times despite not having
4763          * these flags set.  For all other operations the VFS set these flags
4764          * explicitly if it wants a timestamp update.
4765          */
4766         if (newsize != oldsize) {
4767                 inode_inc_iversion(inode);
4768                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4769                         inode->i_ctime = inode->i_mtime =
4770                                 current_time(inode);
4771         }
4772
4773         if (newsize > oldsize) {
4774                 /*
4775                  * Don't do an expanding truncate while snapshotting is ongoing.
4776                  * This is to ensure the snapshot captures a fully consistent
4777                  * state of this file - if the snapshot captures this expanding
4778                  * truncation, it must capture all writes that happened before
4779                  * this truncation.
4780                  */
4781                 btrfs_drew_write_lock(&root->snapshot_lock);
4782                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4783                 if (ret) {
4784                         btrfs_drew_write_unlock(&root->snapshot_lock);
4785                         return ret;
4786                 }
4787
4788                 trans = btrfs_start_transaction(root, 1);
4789                 if (IS_ERR(trans)) {
4790                         btrfs_drew_write_unlock(&root->snapshot_lock);
4791                         return PTR_ERR(trans);
4792                 }
4793
4794                 i_size_write(inode, newsize);
4795                 btrfs_inode_safe_disk_i_size_write(inode, 0);
4796                 pagecache_isize_extended(inode, oldsize, newsize);
4797                 ret = btrfs_update_inode(trans, root, inode);
4798                 btrfs_drew_write_unlock(&root->snapshot_lock);
4799                 btrfs_end_transaction(trans);
4800         } else {
4801
4802                 /*
4803                  * We're truncating a file that used to have good data down to
4804                  * zero. Make sure it gets into the ordered flush list so that
4805                  * any new writes get down to disk quickly.
4806                  */
4807                 if (newsize == 0)
4808                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4809                                 &BTRFS_I(inode)->runtime_flags);
4810
4811                 truncate_setsize(inode, newsize);
4812
4813                 inode_dio_wait(inode);
4814
4815                 ret = btrfs_truncate(inode, newsize == oldsize);
4816                 if (ret && inode->i_nlink) {
4817                         int err;
4818
4819                         /*
4820                          * Truncate failed, so fix up the in-memory size. We
4821                          * adjusted disk_i_size down as we removed extents, so
4822                          * wait for disk_i_size to be stable and then update the
4823                          * in-memory size to match.
4824                          */
4825                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
4826                         if (err)
4827                                 return err;
4828                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
4829                 }
4830         }
4831
4832         return ret;
4833 }
4834
4835 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
4836 {
4837         struct inode *inode = d_inode(dentry);
4838         struct btrfs_root *root = BTRFS_I(inode)->root;
4839         int err;
4840
4841         if (btrfs_root_readonly(root))
4842                 return -EROFS;
4843
4844         err = setattr_prepare(dentry, attr);
4845         if (err)
4846                 return err;
4847
4848         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
4849                 err = btrfs_setsize(inode, attr);
4850                 if (err)
4851                         return err;
4852         }
4853
4854         if (attr->ia_valid) {
4855                 setattr_copy(inode, attr);
4856                 inode_inc_iversion(inode);
4857                 err = btrfs_dirty_inode(inode);
4858
4859                 if (!err && attr->ia_valid & ATTR_MODE)
4860                         err = posix_acl_chmod(inode, inode->i_mode);
4861         }
4862
4863         return err;
4864 }
4865
4866 /*
4867  * While truncating the inode pages during eviction, we get the VFS calling
4868  * btrfs_invalidatepage() against each page of the inode. This is slow because
4869  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
4870  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
4871  * extent_state structures over and over, wasting lots of time.
4872  *
4873  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
4874  * those expensive operations on a per page basis and do only the ordered io
4875  * finishing, while we release here the extent_map and extent_state structures,
4876  * without the excessive merging and splitting.
4877  */
4878 static void evict_inode_truncate_pages(struct inode *inode)
4879 {
4880         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4881         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
4882         struct rb_node *node;
4883
4884         ASSERT(inode->i_state & I_FREEING);
4885         truncate_inode_pages_final(&inode->i_data);
4886
4887         write_lock(&map_tree->lock);
4888         while (!RB_EMPTY_ROOT(&map_tree->map.rb_root)) {
4889                 struct extent_map *em;
4890
4891                 node = rb_first_cached(&map_tree->map);
4892                 em = rb_entry(node, struct extent_map, rb_node);
4893                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
4894                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
4895                 remove_extent_mapping(map_tree, em);
4896                 free_extent_map(em);
4897                 if (need_resched()) {
4898                         write_unlock(&map_tree->lock);
4899                         cond_resched();
4900                         write_lock(&map_tree->lock);
4901                 }
4902         }
4903         write_unlock(&map_tree->lock);
4904
4905         /*
4906          * Keep looping until we have no more ranges in the io tree.
4907          * We can have ongoing bios started by readahead that have
4908          * their endio callback (extent_io.c:end_bio_extent_readpage)
4909          * still in progress (unlocked the pages in the bio but did not yet
4910          * unlocked the ranges in the io tree). Therefore this means some
4911          * ranges can still be locked and eviction started because before
4912          * submitting those bios, which are executed by a separate task (work
4913          * queue kthread), inode references (inode->i_count) were not taken
4914          * (which would be dropped in the end io callback of each bio).
4915          * Therefore here we effectively end up waiting for those bios and
4916          * anyone else holding locked ranges without having bumped the inode's
4917          * reference count - if we don't do it, when they access the inode's
4918          * io_tree to unlock a range it may be too late, leading to an
4919          * use-after-free issue.
4920          */
4921         spin_lock(&io_tree->lock);
4922         while (!RB_EMPTY_ROOT(&io_tree->state)) {
4923                 struct extent_state *state;
4924                 struct extent_state *cached_state = NULL;
4925                 u64 start;
4926                 u64 end;
4927                 unsigned state_flags;
4928
4929                 node = rb_first(&io_tree->state);
4930                 state = rb_entry(node, struct extent_state, rb_node);
4931                 start = state->start;
4932                 end = state->end;
4933                 state_flags = state->state;
4934                 spin_unlock(&io_tree->lock);
4935
4936                 lock_extent_bits(io_tree, start, end, &cached_state);
4937
4938                 /*
4939                  * If still has DELALLOC flag, the extent didn't reach disk,
4940                  * and its reserved space won't be freed by delayed_ref.
4941                  * So we need to free its reserved space here.
4942                  * (Refer to comment in btrfs_invalidatepage, case 2)
4943                  *
4944                  * Note, end is the bytenr of last byte, so we need + 1 here.
4945                  */
4946                 if (state_flags & EXTENT_DELALLOC)
4947                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
4948
4949                 clear_extent_bit(io_tree, start, end,
4950                                  EXTENT_LOCKED | EXTENT_DELALLOC |
4951                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
4952                                  &cached_state);
4953
4954                 cond_resched();
4955                 spin_lock(&io_tree->lock);
4956         }
4957         spin_unlock(&io_tree->lock);
4958 }
4959
4960 static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
4961                                                         struct btrfs_block_rsv *rsv)
4962 {
4963         struct btrfs_fs_info *fs_info = root->fs_info;
4964         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4965         struct btrfs_trans_handle *trans;
4966         u64 delayed_refs_extra = btrfs_calc_insert_metadata_size(fs_info, 1);
4967         int ret;
4968
4969         /*
4970          * Eviction should be taking place at some place safe because of our
4971          * delayed iputs.  However the normal flushing code will run delayed
4972          * iputs, so we cannot use FLUSH_ALL otherwise we'll deadlock.
4973          *
4974          * We reserve the delayed_refs_extra here again because we can't use
4975          * btrfs_start_transaction(root, 0) for the same deadlocky reason as
4976          * above.  We reserve our extra bit here because we generate a ton of
4977          * delayed refs activity by truncating.
4978          *
4979          * If we cannot make our reservation we'll attempt to steal from the
4980          * global reserve, because we really want to be able to free up space.
4981          */
4982         ret = btrfs_block_rsv_refill(root, rsv, rsv->size + delayed_refs_extra,
4983                                      BTRFS_RESERVE_FLUSH_EVICT);
4984         if (ret) {
4985                 /*
4986                  * Try to steal from the global reserve if there is space for
4987                  * it.
4988                  */
4989                 if (btrfs_check_space_for_delayed_refs(fs_info) ||
4990                     btrfs_block_rsv_migrate(global_rsv, rsv, rsv->size, 0)) {
4991                         btrfs_warn(fs_info,
4992                                    "could not allocate space for delete; will truncate on mount");
4993                         return ERR_PTR(-ENOSPC);
4994                 }
4995                 delayed_refs_extra = 0;
4996         }
4997
4998         trans = btrfs_join_transaction(root);
4999         if (IS_ERR(trans))
5000                 return trans;
5001
5002         if (delayed_refs_extra) {
5003                 trans->block_rsv = &fs_info->trans_block_rsv;
5004                 trans->bytes_reserved = delayed_refs_extra;
5005                 btrfs_block_rsv_migrate(rsv, trans->block_rsv,
5006                                         delayed_refs_extra, 1);
5007         }
5008         return trans;
5009 }
5010
5011 void btrfs_evict_inode(struct inode *inode)
5012 {
5013         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5014         struct btrfs_trans_handle *trans;
5015         struct btrfs_root *root = BTRFS_I(inode)->root;
5016         struct btrfs_block_rsv *rsv;
5017         int ret;
5018
5019         trace_btrfs_inode_evict(inode);
5020
5021         if (!root) {
5022                 clear_inode(inode);
5023                 return;
5024         }
5025
5026         evict_inode_truncate_pages(inode);
5027
5028         if (inode->i_nlink &&
5029             ((btrfs_root_refs(&root->root_item) != 0 &&
5030               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5031              btrfs_is_free_space_inode(BTRFS_I(inode))))
5032                 goto no_delete;
5033
5034         if (is_bad_inode(inode))
5035                 goto no_delete;
5036
5037         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5038
5039         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
5040                 goto no_delete;
5041
5042         if (inode->i_nlink > 0) {
5043                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5044                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5045                 goto no_delete;
5046         }
5047
5048         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5049         if (ret)
5050                 goto no_delete;
5051
5052         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5053         if (!rsv)
5054                 goto no_delete;
5055         rsv->size = btrfs_calc_metadata_size(fs_info, 1);
5056         rsv->failfast = 1;
5057
5058         btrfs_i_size_write(BTRFS_I(inode), 0);
5059
5060         while (1) {
5061                 trans = evict_refill_and_join(root, rsv);
5062                 if (IS_ERR(trans))
5063                         goto free_rsv;
5064
5065                 trans->block_rsv = rsv;
5066
5067                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5068                 trans->block_rsv = &fs_info->trans_block_rsv;
5069                 btrfs_end_transaction(trans);
5070                 btrfs_btree_balance_dirty(fs_info);
5071                 if (ret && ret != -ENOSPC && ret != -EAGAIN)
5072                         goto free_rsv;
5073                 else if (!ret)
5074                         break;
5075         }
5076
5077         /*
5078          * Errors here aren't a big deal, it just means we leave orphan items in
5079          * the tree. They will be cleaned up on the next mount. If the inode
5080          * number gets reused, cleanup deletes the orphan item without doing
5081          * anything, and unlink reuses the existing orphan item.
5082          *
5083          * If it turns out that we are dropping too many of these, we might want
5084          * to add a mechanism for retrying these after a commit.
5085          */
5086         trans = evict_refill_and_join(root, rsv);
5087         if (!IS_ERR(trans)) {
5088                 trans->block_rsv = rsv;
5089                 btrfs_orphan_del(trans, BTRFS_I(inode));
5090                 trans->block_rsv = &fs_info->trans_block_rsv;
5091                 btrfs_end_transaction(trans);
5092         }
5093
5094         if (!(root == fs_info->tree_root ||
5095               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5096                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5097
5098 free_rsv:
5099         btrfs_free_block_rsv(fs_info, rsv);
5100 no_delete:
5101         /*
5102          * If we didn't successfully delete, the orphan item will still be in
5103          * the tree and we'll retry on the next mount. Again, we might also want
5104          * to retry these periodically in the future.
5105          */
5106         btrfs_remove_delayed_node(BTRFS_I(inode));
5107         clear_inode(inode);
5108 }
5109
5110 /*
5111  * Return the key found in the dir entry in the location pointer, fill @type
5112  * with BTRFS_FT_*, and return 0.
5113  *
5114  * If no dir entries were found, returns -ENOENT.
5115  * If found a corrupted location in dir entry, returns -EUCLEAN.
5116  */
5117 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5118                                struct btrfs_key *location, u8 *type)
5119 {
5120         const char *name = dentry->d_name.name;
5121         int namelen = dentry->d_name.len;
5122         struct btrfs_dir_item *di;
5123         struct btrfs_path *path;
5124         struct btrfs_root *root = BTRFS_I(dir)->root;
5125         int ret = 0;
5126
5127         path = btrfs_alloc_path();
5128         if (!path)
5129                 return -ENOMEM;
5130
5131         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5132                         name, namelen, 0);
5133         if (IS_ERR_OR_NULL(di)) {
5134                 ret = di ? PTR_ERR(di) : -ENOENT;
5135                 goto out;
5136         }
5137
5138         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5139         if (location->type != BTRFS_INODE_ITEM_KEY &&
5140             location->type != BTRFS_ROOT_ITEM_KEY) {
5141                 ret = -EUCLEAN;
5142                 btrfs_warn(root->fs_info,
5143 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5144                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5145                            location->objectid, location->type, location->offset);
5146         }
5147         if (!ret)
5148                 *type = btrfs_dir_type(path->nodes[0], di);
5149 out:
5150         btrfs_free_path(path);
5151         return ret;
5152 }
5153
5154 /*
5155  * when we hit a tree root in a directory, the btrfs part of the inode
5156  * needs to be changed to reflect the root directory of the tree root.  This
5157  * is kind of like crossing a mount point.
5158  */
5159 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5160                                     struct inode *dir,
5161                                     struct dentry *dentry,
5162                                     struct btrfs_key *location,
5163                                     struct btrfs_root **sub_root)
5164 {
5165         struct btrfs_path *path;
5166         struct btrfs_root *new_root;
5167         struct btrfs_root_ref *ref;
5168         struct extent_buffer *leaf;
5169         struct btrfs_key key;
5170         int ret;
5171         int err = 0;
5172
5173         path = btrfs_alloc_path();
5174         if (!path) {
5175                 err = -ENOMEM;
5176                 goto out;
5177         }
5178
5179         err = -ENOENT;
5180         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5181         key.type = BTRFS_ROOT_REF_KEY;
5182         key.offset = location->objectid;
5183
5184         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5185         if (ret) {
5186                 if (ret < 0)
5187                         err = ret;
5188                 goto out;
5189         }
5190
5191         leaf = path->nodes[0];
5192         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5193         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5194             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5195                 goto out;
5196
5197         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5198                                    (unsigned long)(ref + 1),
5199                                    dentry->d_name.len);
5200         if (ret)
5201                 goto out;
5202
5203         btrfs_release_path(path);
5204
5205         new_root = btrfs_get_fs_root(fs_info, location->objectid, true);
5206         if (IS_ERR(new_root)) {
5207                 err = PTR_ERR(new_root);
5208                 goto out;
5209         }
5210
5211         *sub_root = new_root;
5212         location->objectid = btrfs_root_dirid(&new_root->root_item);
5213         location->type = BTRFS_INODE_ITEM_KEY;
5214         location->offset = 0;
5215         err = 0;
5216 out:
5217         btrfs_free_path(path);
5218         return err;
5219 }
5220
5221 static void inode_tree_add(struct inode *inode)
5222 {
5223         struct btrfs_root *root = BTRFS_I(inode)->root;
5224         struct btrfs_inode *entry;
5225         struct rb_node **p;
5226         struct rb_node *parent;
5227         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5228         u64 ino = btrfs_ino(BTRFS_I(inode));
5229
5230         if (inode_unhashed(inode))
5231                 return;
5232         parent = NULL;
5233         spin_lock(&root->inode_lock);
5234         p = &root->inode_tree.rb_node;
5235         while (*p) {
5236                 parent = *p;
5237                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5238
5239                 if (ino < btrfs_ino(entry))
5240                         p = &parent->rb_left;
5241                 else if (ino > btrfs_ino(entry))
5242                         p = &parent->rb_right;
5243                 else {
5244                         WARN_ON(!(entry->vfs_inode.i_state &
5245                                   (I_WILL_FREE | I_FREEING)));
5246                         rb_replace_node(parent, new, &root->inode_tree);
5247                         RB_CLEAR_NODE(parent);
5248                         spin_unlock(&root->inode_lock);
5249                         return;
5250                 }
5251         }
5252         rb_link_node(new, parent, p);
5253         rb_insert_color(new, &root->inode_tree);
5254         spin_unlock(&root->inode_lock);
5255 }
5256
5257 static void inode_tree_del(struct inode *inode)
5258 {
5259         struct btrfs_root *root = BTRFS_I(inode)->root;
5260         int empty = 0;
5261
5262         spin_lock(&root->inode_lock);
5263         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5264                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5265                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5266                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5267         }
5268         spin_unlock(&root->inode_lock);
5269
5270         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5271                 spin_lock(&root->inode_lock);
5272                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5273                 spin_unlock(&root->inode_lock);
5274                 if (empty)
5275                         btrfs_add_dead_root(root);
5276         }
5277 }
5278
5279
5280 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5281 {
5282         struct btrfs_iget_args *args = p;
5283
5284         inode->i_ino = args->ino;
5285         BTRFS_I(inode)->location.objectid = args->ino;
5286         BTRFS_I(inode)->location.type = BTRFS_INODE_ITEM_KEY;
5287         BTRFS_I(inode)->location.offset = 0;
5288         BTRFS_I(inode)->root = btrfs_grab_root(args->root);
5289         BUG_ON(args->root && !BTRFS_I(inode)->root);
5290         return 0;
5291 }
5292
5293 static int btrfs_find_actor(struct inode *inode, void *opaque)
5294 {
5295         struct btrfs_iget_args *args = opaque;
5296
5297         return args->ino == BTRFS_I(inode)->location.objectid &&
5298                 args->root == BTRFS_I(inode)->root;
5299 }
5300
5301 static struct inode *btrfs_iget_locked(struct super_block *s, u64 ino,
5302                                        struct btrfs_root *root)
5303 {
5304         struct inode *inode;
5305         struct btrfs_iget_args args;
5306         unsigned long hashval = btrfs_inode_hash(ino, root);
5307
5308         args.ino = ino;
5309         args.root = root;
5310
5311         inode = iget5_locked(s, hashval, btrfs_find_actor,
5312                              btrfs_init_locked_inode,
5313                              (void *)&args);
5314         return inode;
5315 }
5316
5317 /*
5318  * Get an inode object given its inode number and corresponding root.
5319  * Path can be preallocated to prevent recursing back to iget through
5320  * allocator. NULL is also valid but may require an additional allocation
5321  * later.
5322  */
5323 struct inode *btrfs_iget_path(struct super_block *s, u64 ino,
5324                               struct btrfs_root *root, struct btrfs_path *path)
5325 {
5326         struct inode *inode;
5327
5328         inode = btrfs_iget_locked(s, ino, root);
5329         if (!inode)
5330                 return ERR_PTR(-ENOMEM);
5331
5332         if (inode->i_state & I_NEW) {
5333                 int ret;
5334
5335                 ret = btrfs_read_locked_inode(inode, path);
5336                 if (!ret) {
5337                         inode_tree_add(inode);
5338                         unlock_new_inode(inode);
5339                 } else {
5340                         iget_failed(inode);
5341                         /*
5342                          * ret > 0 can come from btrfs_search_slot called by
5343                          * btrfs_read_locked_inode, this means the inode item
5344                          * was not found.
5345                          */
5346                         if (ret > 0)
5347                                 ret = -ENOENT;
5348                         inode = ERR_PTR(ret);
5349                 }
5350         }
5351
5352         return inode;
5353 }
5354
5355 struct inode *btrfs_iget(struct super_block *s, u64 ino, struct btrfs_root *root)
5356 {
5357         return btrfs_iget_path(s, ino, root, NULL);
5358 }
5359
5360 static struct inode *new_simple_dir(struct super_block *s,
5361                                     struct btrfs_key *key,
5362                                     struct btrfs_root *root)
5363 {
5364         struct inode *inode = new_inode(s);
5365
5366         if (!inode)
5367                 return ERR_PTR(-ENOMEM);
5368
5369         BTRFS_I(inode)->root = btrfs_grab_root(root);
5370         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5371         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5372
5373         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5374         /*
5375          * We only need lookup, the rest is read-only and there's no inode
5376          * associated with the dentry
5377          */
5378         inode->i_op = &simple_dir_inode_operations;
5379         inode->i_opflags &= ~IOP_XATTR;
5380         inode->i_fop = &simple_dir_operations;
5381         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5382         inode->i_mtime = current_time(inode);
5383         inode->i_atime = inode->i_mtime;
5384         inode->i_ctime = inode->i_mtime;
5385         BTRFS_I(inode)->i_otime = inode->i_mtime;
5386
5387         return inode;
5388 }
5389
5390 static inline u8 btrfs_inode_type(struct inode *inode)
5391 {
5392         /*
5393          * Compile-time asserts that generic FT_* types still match
5394          * BTRFS_FT_* types
5395          */
5396         BUILD_BUG_ON(BTRFS_FT_UNKNOWN != FT_UNKNOWN);
5397         BUILD_BUG_ON(BTRFS_FT_REG_FILE != FT_REG_FILE);
5398         BUILD_BUG_ON(BTRFS_FT_DIR != FT_DIR);
5399         BUILD_BUG_ON(BTRFS_FT_CHRDEV != FT_CHRDEV);
5400         BUILD_BUG_ON(BTRFS_FT_BLKDEV != FT_BLKDEV);
5401         BUILD_BUG_ON(BTRFS_FT_FIFO != FT_FIFO);
5402         BUILD_BUG_ON(BTRFS_FT_SOCK != FT_SOCK);
5403         BUILD_BUG_ON(BTRFS_FT_SYMLINK != FT_SYMLINK);
5404
5405         return fs_umode_to_ftype(inode->i_mode);
5406 }
5407
5408 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5409 {
5410         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5411         struct inode *inode;
5412         struct btrfs_root *root = BTRFS_I(dir)->root;
5413         struct btrfs_root *sub_root = root;
5414         struct btrfs_key location;
5415         u8 di_type = 0;
5416         int ret = 0;
5417
5418         if (dentry->d_name.len > BTRFS_NAME_LEN)
5419                 return ERR_PTR(-ENAMETOOLONG);
5420
5421         ret = btrfs_inode_by_name(dir, dentry, &location, &di_type);
5422         if (ret < 0)
5423                 return ERR_PTR(ret);
5424
5425         if (location.type == BTRFS_INODE_ITEM_KEY) {
5426                 inode = btrfs_iget(dir->i_sb, location.objectid, root);
5427                 if (IS_ERR(inode))
5428                         return inode;
5429
5430                 /* Do extra check against inode mode with di_type */
5431                 if (btrfs_inode_type(inode) != di_type) {
5432                         btrfs_crit(fs_info,
5433 "inode mode mismatch with dir: inode mode=0%o btrfs type=%u dir type=%u",
5434                                   inode->i_mode, btrfs_inode_type(inode),
5435                                   di_type);
5436                         iput(inode);
5437                         return ERR_PTR(-EUCLEAN);
5438                 }
5439                 return inode;
5440         }
5441
5442         ret = fixup_tree_root_location(fs_info, dir, dentry,
5443                                        &location, &sub_root);
5444         if (ret < 0) {
5445                 if (ret != -ENOENT)
5446                         inode = ERR_PTR(ret);
5447                 else
5448                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5449         } else {
5450                 inode = btrfs_iget(dir->i_sb, location.objectid, sub_root);
5451         }
5452         if (root != sub_root)
5453                 btrfs_put_root(sub_root);
5454
5455         if (!IS_ERR(inode) && root != sub_root) {
5456                 down_read(&fs_info->cleanup_work_sem);
5457                 if (!sb_rdonly(inode->i_sb))
5458                         ret = btrfs_orphan_cleanup(sub_root);
5459                 up_read(&fs_info->cleanup_work_sem);
5460                 if (ret) {
5461                         iput(inode);
5462                         inode = ERR_PTR(ret);
5463                 }
5464         }
5465
5466         return inode;
5467 }
5468
5469 static int btrfs_dentry_delete(const struct dentry *dentry)
5470 {
5471         struct btrfs_root *root;
5472         struct inode *inode = d_inode(dentry);
5473
5474         if (!inode && !IS_ROOT(dentry))
5475                 inode = d_inode(dentry->d_parent);
5476
5477         if (inode) {
5478                 root = BTRFS_I(inode)->root;
5479                 if (btrfs_root_refs(&root->root_item) == 0)
5480                         return 1;
5481
5482                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5483                         return 1;
5484         }
5485         return 0;
5486 }
5487
5488 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5489                                    unsigned int flags)
5490 {
5491         struct inode *inode = btrfs_lookup_dentry(dir, dentry);
5492
5493         if (inode == ERR_PTR(-ENOENT))
5494                 inode = NULL;
5495         return d_splice_alias(inode, dentry);
5496 }
5497
5498 /*
5499  * All this infrastructure exists because dir_emit can fault, and we are holding
5500  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5501  * our information into that, and then dir_emit from the buffer.  This is
5502  * similar to what NFS does, only we don't keep the buffer around in pagecache
5503  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5504  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5505  * tree lock.
5506  */
5507 static int btrfs_opendir(struct inode *inode, struct file *file)
5508 {
5509         struct btrfs_file_private *private;
5510
5511         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5512         if (!private)
5513                 return -ENOMEM;
5514         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5515         if (!private->filldir_buf) {
5516                 kfree(private);
5517                 return -ENOMEM;
5518         }
5519         file->private_data = private;
5520         return 0;
5521 }
5522
5523 struct dir_entry {
5524         u64 ino;
5525         u64 offset;
5526         unsigned type;
5527         int name_len;
5528 };
5529
5530 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5531 {
5532         while (entries--) {
5533                 struct dir_entry *entry = addr;
5534                 char *name = (char *)(entry + 1);
5535
5536                 ctx->pos = get_unaligned(&entry->offset);
5537                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
5538                                          get_unaligned(&entry->ino),
5539                                          get_unaligned(&entry->type)))
5540                         return 1;
5541                 addr += sizeof(struct dir_entry) +
5542                         get_unaligned(&entry->name_len);
5543                 ctx->pos++;
5544         }
5545         return 0;
5546 }
5547
5548 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5549 {
5550         struct inode *inode = file_inode(file);
5551         struct btrfs_root *root = BTRFS_I(inode)->root;
5552         struct btrfs_file_private *private = file->private_data;
5553         struct btrfs_dir_item *di;
5554         struct btrfs_key key;
5555         struct btrfs_key found_key;
5556         struct btrfs_path *path;
5557         void *addr;
5558         struct list_head ins_list;
5559         struct list_head del_list;
5560         int ret;
5561         struct extent_buffer *leaf;
5562         int slot;
5563         char *name_ptr;
5564         int name_len;
5565         int entries = 0;
5566         int total_len = 0;
5567         bool put = false;
5568         struct btrfs_key location;
5569
5570         if (!dir_emit_dots(file, ctx))
5571                 return 0;
5572
5573         path = btrfs_alloc_path();
5574         if (!path)
5575                 return -ENOMEM;
5576
5577         addr = private->filldir_buf;
5578         path->reada = READA_FORWARD;
5579
5580         INIT_LIST_HEAD(&ins_list);
5581         INIT_LIST_HEAD(&del_list);
5582         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5583
5584 again:
5585         key.type = BTRFS_DIR_INDEX_KEY;
5586         key.offset = ctx->pos;
5587         key.objectid = btrfs_ino(BTRFS_I(inode));
5588
5589         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5590         if (ret < 0)
5591                 goto err;
5592
5593         while (1) {
5594                 struct dir_entry *entry;
5595
5596                 leaf = path->nodes[0];
5597                 slot = path->slots[0];
5598                 if (slot >= btrfs_header_nritems(leaf)) {
5599                         ret = btrfs_next_leaf(root, path);
5600                         if (ret < 0)
5601                                 goto err;
5602                         else if (ret > 0)
5603                                 break;
5604                         continue;
5605                 }
5606
5607                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5608
5609                 if (found_key.objectid != key.objectid)
5610                         break;
5611                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5612                         break;
5613                 if (found_key.offset < ctx->pos)
5614                         goto next;
5615                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5616                         goto next;
5617                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5618                 name_len = btrfs_dir_name_len(leaf, di);
5619                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
5620                     PAGE_SIZE) {
5621                         btrfs_release_path(path);
5622                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5623                         if (ret)
5624                                 goto nopos;
5625                         addr = private->filldir_buf;
5626                         entries = 0;
5627                         total_len = 0;
5628                         goto again;
5629                 }
5630
5631                 entry = addr;
5632                 put_unaligned(name_len, &entry->name_len);
5633                 name_ptr = (char *)(entry + 1);
5634                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5635                                    name_len);
5636                 put_unaligned(fs_ftype_to_dtype(btrfs_dir_type(leaf, di)),
5637                                 &entry->type);
5638                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5639                 put_unaligned(location.objectid, &entry->ino);
5640                 put_unaligned(found_key.offset, &entry->offset);
5641                 entries++;
5642                 addr += sizeof(struct dir_entry) + name_len;
5643                 total_len += sizeof(struct dir_entry) + name_len;
5644 next:
5645                 path->slots[0]++;
5646         }
5647         btrfs_release_path(path);
5648
5649         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5650         if (ret)
5651                 goto nopos;
5652
5653         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5654         if (ret)
5655                 goto nopos;
5656
5657         /*
5658          * Stop new entries from being returned after we return the last
5659          * entry.
5660          *
5661          * New directory entries are assigned a strictly increasing
5662          * offset.  This means that new entries created during readdir
5663          * are *guaranteed* to be seen in the future by that readdir.
5664          * This has broken buggy programs which operate on names as
5665          * they're returned by readdir.  Until we re-use freed offsets
5666          * we have this hack to stop new entries from being returned
5667          * under the assumption that they'll never reach this huge
5668          * offset.
5669          *
5670          * This is being careful not to overflow 32bit loff_t unless the
5671          * last entry requires it because doing so has broken 32bit apps
5672          * in the past.
5673          */
5674         if (ctx->pos >= INT_MAX)
5675                 ctx->pos = LLONG_MAX;
5676         else
5677                 ctx->pos = INT_MAX;
5678 nopos:
5679         ret = 0;
5680 err:
5681         if (put)
5682                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5683         btrfs_free_path(path);
5684         return ret;
5685 }
5686
5687 /*
5688  * This is somewhat expensive, updating the tree every time the
5689  * inode changes.  But, it is most likely to find the inode in cache.
5690  * FIXME, needs more benchmarking...there are no reasons other than performance
5691  * to keep or drop this code.
5692  */
5693 static int btrfs_dirty_inode(struct inode *inode)
5694 {
5695         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5696         struct btrfs_root *root = BTRFS_I(inode)->root;
5697         struct btrfs_trans_handle *trans;
5698         int ret;
5699
5700         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5701                 return 0;
5702
5703         trans = btrfs_join_transaction(root);
5704         if (IS_ERR(trans))
5705                 return PTR_ERR(trans);
5706
5707         ret = btrfs_update_inode(trans, root, inode);
5708         if (ret && ret == -ENOSPC) {
5709                 /* whoops, lets try again with the full transaction */
5710                 btrfs_end_transaction(trans);
5711                 trans = btrfs_start_transaction(root, 1);
5712                 if (IS_ERR(trans))
5713                         return PTR_ERR(trans);
5714
5715                 ret = btrfs_update_inode(trans, root, inode);
5716         }
5717         btrfs_end_transaction(trans);
5718         if (BTRFS_I(inode)->delayed_node)
5719                 btrfs_balance_delayed_items(fs_info);
5720
5721         return ret;
5722 }
5723
5724 /*
5725  * This is a copy of file_update_time.  We need this so we can return error on
5726  * ENOSPC for updating the inode in the case of file write and mmap writes.
5727  */
5728 static int btrfs_update_time(struct inode *inode, struct timespec64 *now,
5729                              int flags)
5730 {
5731         struct btrfs_root *root = BTRFS_I(inode)->root;
5732         bool dirty = flags & ~S_VERSION;
5733
5734         if (btrfs_root_readonly(root))
5735                 return -EROFS;
5736
5737         if (flags & S_VERSION)
5738                 dirty |= inode_maybe_inc_iversion(inode, dirty);
5739         if (flags & S_CTIME)
5740                 inode->i_ctime = *now;
5741         if (flags & S_MTIME)
5742                 inode->i_mtime = *now;
5743         if (flags & S_ATIME)
5744                 inode->i_atime = *now;
5745         return dirty ? btrfs_dirty_inode(inode) : 0;
5746 }
5747
5748 /*
5749  * find the highest existing sequence number in a directory
5750  * and then set the in-memory index_cnt variable to reflect
5751  * free sequence numbers
5752  */
5753 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
5754 {
5755         struct btrfs_root *root = inode->root;
5756         struct btrfs_key key, found_key;
5757         struct btrfs_path *path;
5758         struct extent_buffer *leaf;
5759         int ret;
5760
5761         key.objectid = btrfs_ino(inode);
5762         key.type = BTRFS_DIR_INDEX_KEY;
5763         key.offset = (u64)-1;
5764
5765         path = btrfs_alloc_path();
5766         if (!path)
5767                 return -ENOMEM;
5768
5769         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5770         if (ret < 0)
5771                 goto out;
5772         /* FIXME: we should be able to handle this */
5773         if (ret == 0)
5774                 goto out;
5775         ret = 0;
5776
5777         /*
5778          * MAGIC NUMBER EXPLANATION:
5779          * since we search a directory based on f_pos we have to start at 2
5780          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
5781          * else has to start at 2
5782          */
5783         if (path->slots[0] == 0) {
5784                 inode->index_cnt = 2;
5785                 goto out;
5786         }
5787
5788         path->slots[0]--;
5789
5790         leaf = path->nodes[0];
5791         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5792
5793         if (found_key.objectid != btrfs_ino(inode) ||
5794             found_key.type != BTRFS_DIR_INDEX_KEY) {
5795                 inode->index_cnt = 2;
5796                 goto out;
5797         }
5798
5799         inode->index_cnt = found_key.offset + 1;
5800 out:
5801         btrfs_free_path(path);
5802         return ret;
5803 }
5804
5805 /*
5806  * helper to find a free sequence number in a given directory.  This current
5807  * code is very simple, later versions will do smarter things in the btree
5808  */
5809 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
5810 {
5811         int ret = 0;
5812
5813         if (dir->index_cnt == (u64)-1) {
5814                 ret = btrfs_inode_delayed_dir_index_count(dir);
5815                 if (ret) {
5816                         ret = btrfs_set_inode_index_count(dir);
5817                         if (ret)
5818                                 return ret;
5819                 }
5820         }
5821
5822         *index = dir->index_cnt;
5823         dir->index_cnt++;
5824
5825         return ret;
5826 }
5827
5828 static int btrfs_insert_inode_locked(struct inode *inode)
5829 {
5830         struct btrfs_iget_args args;
5831
5832         args.ino = BTRFS_I(inode)->location.objectid;
5833         args.root = BTRFS_I(inode)->root;
5834
5835         return insert_inode_locked4(inode,
5836                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
5837                    btrfs_find_actor, &args);
5838 }
5839
5840 /*
5841  * Inherit flags from the parent inode.
5842  *
5843  * Currently only the compression flags and the cow flags are inherited.
5844  */
5845 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
5846 {
5847         unsigned int flags;
5848
5849         if (!dir)
5850                 return;
5851
5852         flags = BTRFS_I(dir)->flags;
5853
5854         if (flags & BTRFS_INODE_NOCOMPRESS) {
5855                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
5856                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
5857         } else if (flags & BTRFS_INODE_COMPRESS) {
5858                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
5859                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
5860         }
5861
5862         if (flags & BTRFS_INODE_NODATACOW) {
5863                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
5864                 if (S_ISREG(inode->i_mode))
5865                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
5866         }
5867
5868         btrfs_sync_inode_flags_to_i_flags(inode);
5869 }
5870
5871 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
5872                                      struct btrfs_root *root,
5873                                      struct inode *dir,
5874                                      const char *name, int name_len,
5875                                      u64 ref_objectid, u64 objectid,
5876                                      umode_t mode, u64 *index)
5877 {
5878         struct btrfs_fs_info *fs_info = root->fs_info;
5879         struct inode *inode;
5880         struct btrfs_inode_item *inode_item;
5881         struct btrfs_key *location;
5882         struct btrfs_path *path;
5883         struct btrfs_inode_ref *ref;
5884         struct btrfs_key key[2];
5885         u32 sizes[2];
5886         int nitems = name ? 2 : 1;
5887         unsigned long ptr;
5888         unsigned int nofs_flag;
5889         int ret;
5890
5891         path = btrfs_alloc_path();
5892         if (!path)
5893                 return ERR_PTR(-ENOMEM);
5894
5895         nofs_flag = memalloc_nofs_save();
5896         inode = new_inode(fs_info->sb);
5897         memalloc_nofs_restore(nofs_flag);
5898         if (!inode) {
5899                 btrfs_free_path(path);
5900                 return ERR_PTR(-ENOMEM);
5901         }
5902
5903         /*
5904          * O_TMPFILE, set link count to 0, so that after this point,
5905          * we fill in an inode item with the correct link count.
5906          */
5907         if (!name)
5908                 set_nlink(inode, 0);
5909
5910         /*
5911          * we have to initialize this early, so we can reclaim the inode
5912          * number if we fail afterwards in this function.
5913          */
5914         inode->i_ino = objectid;
5915
5916         if (dir && name) {
5917                 trace_btrfs_inode_request(dir);
5918
5919                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
5920                 if (ret) {
5921                         btrfs_free_path(path);
5922                         iput(inode);
5923                         return ERR_PTR(ret);
5924                 }
5925         } else if (dir) {
5926                 *index = 0;
5927         }
5928         /*
5929          * index_cnt is ignored for everything but a dir,
5930          * btrfs_set_inode_index_count has an explanation for the magic
5931          * number
5932          */
5933         BTRFS_I(inode)->index_cnt = 2;
5934         BTRFS_I(inode)->dir_index = *index;
5935         BTRFS_I(inode)->root = btrfs_grab_root(root);
5936         BTRFS_I(inode)->generation = trans->transid;
5937         inode->i_generation = BTRFS_I(inode)->generation;
5938
5939         /*
5940          * We could have gotten an inode number from somebody who was fsynced
5941          * and then removed in this same transaction, so let's just set full
5942          * sync since it will be a full sync anyway and this will blow away the
5943          * old info in the log.
5944          */
5945         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
5946
5947         key[0].objectid = objectid;
5948         key[0].type = BTRFS_INODE_ITEM_KEY;
5949         key[0].offset = 0;
5950
5951         sizes[0] = sizeof(struct btrfs_inode_item);
5952
5953         if (name) {
5954                 /*
5955                  * Start new inodes with an inode_ref. This is slightly more
5956                  * efficient for small numbers of hard links since they will
5957                  * be packed into one item. Extended refs will kick in if we
5958                  * add more hard links than can fit in the ref item.
5959                  */
5960                 key[1].objectid = objectid;
5961                 key[1].type = BTRFS_INODE_REF_KEY;
5962                 key[1].offset = ref_objectid;
5963
5964                 sizes[1] = name_len + sizeof(*ref);
5965         }
5966
5967         location = &BTRFS_I(inode)->location;
5968         location->objectid = objectid;
5969         location->offset = 0;
5970         location->type = BTRFS_INODE_ITEM_KEY;
5971
5972         ret = btrfs_insert_inode_locked(inode);
5973         if (ret < 0) {
5974                 iput(inode);
5975                 goto fail;
5976         }
5977
5978         path->leave_spinning = 1;
5979         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
5980         if (ret != 0)
5981                 goto fail_unlock;
5982
5983         inode_init_owner(inode, dir, mode);
5984         inode_set_bytes(inode, 0);
5985
5986         inode->i_mtime = current_time(inode);
5987         inode->i_atime = inode->i_mtime;
5988         inode->i_ctime = inode->i_mtime;
5989         BTRFS_I(inode)->i_otime = inode->i_mtime;
5990
5991         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
5992                                   struct btrfs_inode_item);
5993         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
5994                              sizeof(*inode_item));
5995         fill_inode_item(trans, path->nodes[0], inode_item, inode);
5996
5997         if (name) {
5998                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
5999                                      struct btrfs_inode_ref);
6000                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6001                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6002                 ptr = (unsigned long)(ref + 1);
6003                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6004         }
6005
6006         btrfs_mark_buffer_dirty(path->nodes[0]);
6007         btrfs_free_path(path);
6008
6009         btrfs_inherit_iflags(inode, dir);
6010
6011         if (S_ISREG(mode)) {
6012                 if (btrfs_test_opt(fs_info, NODATASUM))
6013                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6014                 if (btrfs_test_opt(fs_info, NODATACOW))
6015                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6016                                 BTRFS_INODE_NODATASUM;
6017         }
6018
6019         inode_tree_add(inode);
6020
6021         trace_btrfs_inode_new(inode);
6022         btrfs_set_inode_last_trans(trans, inode);
6023
6024         btrfs_update_root_times(trans, root);
6025
6026         ret = btrfs_inode_inherit_props(trans, inode, dir);
6027         if (ret)
6028                 btrfs_err(fs_info,
6029                           "error inheriting props for ino %llu (root %llu): %d",
6030                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6031
6032         return inode;
6033
6034 fail_unlock:
6035         discard_new_inode(inode);
6036 fail:
6037         if (dir && name)
6038                 BTRFS_I(dir)->index_cnt--;
6039         btrfs_free_path(path);
6040         return ERR_PTR(ret);
6041 }
6042
6043 /*
6044  * utility function to add 'inode' into 'parent_inode' with
6045  * a give name and a given sequence number.
6046  * if 'add_backref' is true, also insert a backref from the
6047  * inode to the parent directory.
6048  */
6049 int btrfs_add_link(struct btrfs_trans_handle *trans,
6050                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6051                    const char *name, int name_len, int add_backref, u64 index)
6052 {
6053         int ret = 0;
6054         struct btrfs_key key;
6055         struct btrfs_root *root = parent_inode->root;
6056         u64 ino = btrfs_ino(inode);
6057         u64 parent_ino = btrfs_ino(parent_inode);
6058
6059         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6060                 memcpy(&key, &inode->root->root_key, sizeof(key));
6061         } else {
6062                 key.objectid = ino;
6063                 key.type = BTRFS_INODE_ITEM_KEY;
6064                 key.offset = 0;
6065         }
6066
6067         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6068                 ret = btrfs_add_root_ref(trans, key.objectid,
6069                                          root->root_key.objectid, parent_ino,
6070                                          index, name, name_len);
6071         } else if (add_backref) {
6072                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6073                                              parent_ino, index);
6074         }
6075
6076         /* Nothing to clean up yet */
6077         if (ret)
6078                 return ret;
6079
6080         ret = btrfs_insert_dir_item(trans, name, name_len, parent_inode, &key,
6081                                     btrfs_inode_type(&inode->vfs_inode), index);
6082         if (ret == -EEXIST || ret == -EOVERFLOW)
6083                 goto fail_dir_item;
6084         else if (ret) {
6085                 btrfs_abort_transaction(trans, ret);
6086                 return ret;
6087         }
6088
6089         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6090                            name_len * 2);
6091         inode_inc_iversion(&parent_inode->vfs_inode);
6092         /*
6093          * If we are replaying a log tree, we do not want to update the mtime
6094          * and ctime of the parent directory with the current time, since the
6095          * log replay procedure is responsible for setting them to their correct
6096          * values (the ones it had when the fsync was done).
6097          */
6098         if (!test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags)) {
6099                 struct timespec64 now = current_time(&parent_inode->vfs_inode);
6100
6101                 parent_inode->vfs_inode.i_mtime = now;
6102                 parent_inode->vfs_inode.i_ctime = now;
6103         }
6104         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6105         if (ret)
6106                 btrfs_abort_transaction(trans, ret);
6107         return ret;
6108
6109 fail_dir_item:
6110         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6111                 u64 local_index;
6112                 int err;
6113                 err = btrfs_del_root_ref(trans, key.objectid,
6114                                          root->root_key.objectid, parent_ino,
6115                                          &local_index, name, name_len);
6116                 if (err)
6117                         btrfs_abort_transaction(trans, err);
6118         } else if (add_backref) {
6119                 u64 local_index;
6120                 int err;
6121
6122                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6123                                           ino, parent_ino, &local_index);
6124                 if (err)
6125                         btrfs_abort_transaction(trans, err);
6126         }
6127
6128         /* Return the original error code */
6129         return ret;
6130 }
6131
6132 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6133                             struct btrfs_inode *dir, struct dentry *dentry,
6134                             struct btrfs_inode *inode, int backref, u64 index)
6135 {
6136         int err = btrfs_add_link(trans, dir, inode,
6137                                  dentry->d_name.name, dentry->d_name.len,
6138                                  backref, index);
6139         if (err > 0)
6140                 err = -EEXIST;
6141         return err;
6142 }
6143
6144 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6145                         umode_t mode, dev_t rdev)
6146 {
6147         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6148         struct btrfs_trans_handle *trans;
6149         struct btrfs_root *root = BTRFS_I(dir)->root;
6150         struct inode *inode = NULL;
6151         int err;
6152         u64 objectid;
6153         u64 index = 0;
6154
6155         /*
6156          * 2 for inode item and ref
6157          * 2 for dir items
6158          * 1 for xattr if selinux is on
6159          */
6160         trans = btrfs_start_transaction(root, 5);
6161         if (IS_ERR(trans))
6162                 return PTR_ERR(trans);
6163
6164         err = btrfs_find_free_ino(root, &objectid);
6165         if (err)
6166                 goto out_unlock;
6167
6168         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6169                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6170                         mode, &index);
6171         if (IS_ERR(inode)) {
6172                 err = PTR_ERR(inode);
6173                 inode = NULL;
6174                 goto out_unlock;
6175         }
6176
6177         /*
6178         * If the active LSM wants to access the inode during
6179         * d_instantiate it needs these. Smack checks to see
6180         * if the filesystem supports xattrs by looking at the
6181         * ops vector.
6182         */
6183         inode->i_op = &btrfs_special_inode_operations;
6184         init_special_inode(inode, inode->i_mode, rdev);
6185
6186         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6187         if (err)
6188                 goto out_unlock;
6189
6190         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6191                         0, index);
6192         if (err)
6193                 goto out_unlock;
6194
6195         btrfs_update_inode(trans, root, inode);
6196         d_instantiate_new(dentry, inode);
6197
6198 out_unlock:
6199         btrfs_end_transaction(trans);
6200         btrfs_btree_balance_dirty(fs_info);
6201         if (err && inode) {
6202                 inode_dec_link_count(inode);
6203                 discard_new_inode(inode);
6204         }
6205         return err;
6206 }
6207
6208 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6209                         umode_t mode, bool excl)
6210 {
6211         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6212         struct btrfs_trans_handle *trans;
6213         struct btrfs_root *root = BTRFS_I(dir)->root;
6214         struct inode *inode = NULL;
6215         int err;
6216         u64 objectid;
6217         u64 index = 0;
6218
6219         /*
6220          * 2 for inode item and ref
6221          * 2 for dir items
6222          * 1 for xattr if selinux is on
6223          */
6224         trans = btrfs_start_transaction(root, 5);
6225         if (IS_ERR(trans))
6226                 return PTR_ERR(trans);
6227
6228         err = btrfs_find_free_ino(root, &objectid);
6229         if (err)
6230                 goto out_unlock;
6231
6232         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6233                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6234                         mode, &index);
6235         if (IS_ERR(inode)) {
6236                 err = PTR_ERR(inode);
6237                 inode = NULL;
6238                 goto out_unlock;
6239         }
6240         /*
6241         * If the active LSM wants to access the inode during
6242         * d_instantiate it needs these. Smack checks to see
6243         * if the filesystem supports xattrs by looking at the
6244         * ops vector.
6245         */
6246         inode->i_fop = &btrfs_file_operations;
6247         inode->i_op = &btrfs_file_inode_operations;
6248         inode->i_mapping->a_ops = &btrfs_aops;
6249
6250         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6251         if (err)
6252                 goto out_unlock;
6253
6254         err = btrfs_update_inode(trans, root, inode);
6255         if (err)
6256                 goto out_unlock;
6257
6258         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6259                         0, index);
6260         if (err)
6261                 goto out_unlock;
6262
6263         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6264         d_instantiate_new(dentry, inode);
6265
6266 out_unlock:
6267         btrfs_end_transaction(trans);
6268         if (err && inode) {
6269                 inode_dec_link_count(inode);
6270                 discard_new_inode(inode);
6271         }
6272         btrfs_btree_balance_dirty(fs_info);
6273         return err;
6274 }
6275
6276 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6277                       struct dentry *dentry)
6278 {
6279         struct btrfs_trans_handle *trans = NULL;
6280         struct btrfs_root *root = BTRFS_I(dir)->root;
6281         struct inode *inode = d_inode(old_dentry);
6282         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6283         u64 index;
6284         int err;
6285         int drop_inode = 0;
6286
6287         /* do not allow sys_link's with other subvols of the same device */
6288         if (root->root_key.objectid != BTRFS_I(inode)->root->root_key.objectid)
6289                 return -EXDEV;
6290
6291         if (inode->i_nlink >= BTRFS_LINK_MAX)
6292                 return -EMLINK;
6293
6294         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6295         if (err)
6296                 goto fail;
6297
6298         /*
6299          * 2 items for inode and inode ref
6300          * 2 items for dir items
6301          * 1 item for parent inode
6302          * 1 item for orphan item deletion if O_TMPFILE
6303          */
6304         trans = btrfs_start_transaction(root, inode->i_nlink ? 5 : 6);
6305         if (IS_ERR(trans)) {
6306                 err = PTR_ERR(trans);
6307                 trans = NULL;
6308                 goto fail;
6309         }
6310
6311         /* There are several dir indexes for this inode, clear the cache. */
6312         BTRFS_I(inode)->dir_index = 0ULL;
6313         inc_nlink(inode);
6314         inode_inc_iversion(inode);
6315         inode->i_ctime = current_time(inode);
6316         ihold(inode);
6317         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6318
6319         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6320                         1, index);
6321
6322         if (err) {
6323                 drop_inode = 1;
6324         } else {
6325                 struct dentry *parent = dentry->d_parent;
6326                 int ret;
6327
6328                 err = btrfs_update_inode(trans, root, inode);
6329                 if (err)
6330                         goto fail;
6331                 if (inode->i_nlink == 1) {
6332                         /*
6333                          * If new hard link count is 1, it's a file created
6334                          * with open(2) O_TMPFILE flag.
6335                          */
6336                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6337                         if (err)
6338                                 goto fail;
6339                 }
6340                 d_instantiate(dentry, inode);
6341                 ret = btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent,
6342                                          true, NULL);
6343                 if (ret == BTRFS_NEED_TRANS_COMMIT) {
6344                         err = btrfs_commit_transaction(trans);
6345                         trans = NULL;
6346                 }
6347         }
6348
6349 fail:
6350         if (trans)
6351                 btrfs_end_transaction(trans);
6352         if (drop_inode) {
6353                 inode_dec_link_count(inode);
6354                 iput(inode);
6355         }
6356         btrfs_btree_balance_dirty(fs_info);
6357         return err;
6358 }
6359
6360 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6361 {
6362         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6363         struct inode *inode = NULL;
6364         struct btrfs_trans_handle *trans;
6365         struct btrfs_root *root = BTRFS_I(dir)->root;
6366         int err = 0;
6367         u64 objectid = 0;
6368         u64 index = 0;
6369
6370         /*
6371          * 2 items for inode and ref
6372          * 2 items for dir items
6373          * 1 for xattr if selinux is on
6374          */
6375         trans = btrfs_start_transaction(root, 5);
6376         if (IS_ERR(trans))
6377                 return PTR_ERR(trans);
6378
6379         err = btrfs_find_free_ino(root, &objectid);
6380         if (err)
6381                 goto out_fail;
6382
6383         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6384                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6385                         S_IFDIR | mode, &index);
6386         if (IS_ERR(inode)) {
6387                 err = PTR_ERR(inode);
6388                 inode = NULL;
6389                 goto out_fail;
6390         }
6391
6392         /* these must be set before we unlock the inode */
6393         inode->i_op = &btrfs_dir_inode_operations;
6394         inode->i_fop = &btrfs_dir_file_operations;
6395
6396         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6397         if (err)
6398                 goto out_fail;
6399
6400         btrfs_i_size_write(BTRFS_I(inode), 0);
6401         err = btrfs_update_inode(trans, root, inode);
6402         if (err)
6403                 goto out_fail;
6404
6405         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6406                         dentry->d_name.name,
6407                         dentry->d_name.len, 0, index);
6408         if (err)
6409                 goto out_fail;
6410
6411         d_instantiate_new(dentry, inode);
6412
6413 out_fail:
6414         btrfs_end_transaction(trans);
6415         if (err && inode) {
6416                 inode_dec_link_count(inode);
6417                 discard_new_inode(inode);
6418         }
6419         btrfs_btree_balance_dirty(fs_info);
6420         return err;
6421 }
6422
6423 static noinline int uncompress_inline(struct btrfs_path *path,
6424                                       struct page *page,
6425                                       size_t pg_offset, u64 extent_offset,
6426                                       struct btrfs_file_extent_item *item)
6427 {
6428         int ret;
6429         struct extent_buffer *leaf = path->nodes[0];
6430         char *tmp;
6431         size_t max_size;
6432         unsigned long inline_size;
6433         unsigned long ptr;
6434         int compress_type;
6435
6436         WARN_ON(pg_offset != 0);
6437         compress_type = btrfs_file_extent_compression(leaf, item);
6438         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6439         inline_size = btrfs_file_extent_inline_item_len(leaf,
6440                                         btrfs_item_nr(path->slots[0]));
6441         tmp = kmalloc(inline_size, GFP_NOFS);
6442         if (!tmp)
6443                 return -ENOMEM;
6444         ptr = btrfs_file_extent_inline_start(item);
6445
6446         read_extent_buffer(leaf, tmp, ptr, inline_size);
6447
6448         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6449         ret = btrfs_decompress(compress_type, tmp, page,
6450                                extent_offset, inline_size, max_size);
6451
6452         /*
6453          * decompression code contains a memset to fill in any space between the end
6454          * of the uncompressed data and the end of max_size in case the decompressed
6455          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6456          * the end of an inline extent and the beginning of the next block, so we
6457          * cover that region here.
6458          */
6459
6460         if (max_size + pg_offset < PAGE_SIZE) {
6461                 char *map = kmap(page);
6462                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6463                 kunmap(page);
6464         }
6465         kfree(tmp);
6466         return ret;
6467 }
6468
6469 /**
6470  * btrfs_get_extent - Lookup the first extent overlapping a range in a file.
6471  * @inode:      file to search in
6472  * @page:       page to read extent data into if the extent is inline
6473  * @pg_offset:  offset into @page to copy to
6474  * @start:      file offset
6475  * @len:        length of range starting at @start
6476  *
6477  * This returns the first &struct extent_map which overlaps with the given
6478  * range, reading it from the B-tree and caching it if necessary. Note that
6479  * there may be more extents which overlap the given range after the returned
6480  * extent_map.
6481  *
6482  * If @page is not NULL and the extent is inline, this also reads the extent
6483  * data directly into the page and marks the extent up to date in the io_tree.
6484  *
6485  * Return: ERR_PTR on error, non-NULL extent_map on success.
6486  */
6487 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6488                                     struct page *page, size_t pg_offset,
6489                                     u64 start, u64 len)
6490 {
6491         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6492         int ret;
6493         int err = 0;
6494         u64 extent_start = 0;
6495         u64 extent_end = 0;
6496         u64 objectid = btrfs_ino(inode);
6497         int extent_type = -1;
6498         struct btrfs_path *path = NULL;
6499         struct btrfs_root *root = inode->root;
6500         struct btrfs_file_extent_item *item;
6501         struct extent_buffer *leaf;
6502         struct btrfs_key found_key;
6503         struct extent_map *em = NULL;
6504         struct extent_map_tree *em_tree = &inode->extent_tree;
6505         struct extent_io_tree *io_tree = &inode->io_tree;
6506
6507         read_lock(&em_tree->lock);
6508         em = lookup_extent_mapping(em_tree, start, len);
6509         read_unlock(&em_tree->lock);
6510
6511         if (em) {
6512                 if (em->start > start || em->start + em->len <= start)
6513                         free_extent_map(em);
6514                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6515                         free_extent_map(em);
6516                 else
6517                         goto out;
6518         }
6519         em = alloc_extent_map();
6520         if (!em) {
6521                 err = -ENOMEM;
6522                 goto out;
6523         }
6524         em->start = EXTENT_MAP_HOLE;
6525         em->orig_start = EXTENT_MAP_HOLE;
6526         em->len = (u64)-1;
6527         em->block_len = (u64)-1;
6528
6529         path = btrfs_alloc_path();
6530         if (!path) {
6531                 err = -ENOMEM;
6532                 goto out;
6533         }
6534
6535         /* Chances are we'll be called again, so go ahead and do readahead */
6536         path->reada = READA_FORWARD;
6537
6538         /*
6539          * Unless we're going to uncompress the inline extent, no sleep would
6540          * happen.
6541          */
6542         path->leave_spinning = 1;
6543
6544         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6545         if (ret < 0) {
6546                 err = ret;
6547                 goto out;
6548         } else if (ret > 0) {
6549                 if (path->slots[0] == 0)
6550                         goto not_found;
6551                 path->slots[0]--;
6552         }
6553
6554         leaf = path->nodes[0];
6555         item = btrfs_item_ptr(leaf, path->slots[0],
6556                               struct btrfs_file_extent_item);
6557         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6558         if (found_key.objectid != objectid ||
6559             found_key.type != BTRFS_EXTENT_DATA_KEY) {
6560                 /*
6561                  * If we backup past the first extent we want to move forward
6562                  * and see if there is an extent in front of us, otherwise we'll
6563                  * say there is a hole for our whole search range which can
6564                  * cause problems.
6565                  */
6566                 extent_end = start;
6567                 goto next;
6568         }
6569
6570         extent_type = btrfs_file_extent_type(leaf, item);
6571         extent_start = found_key.offset;
6572         extent_end = btrfs_file_extent_end(path);
6573         if (extent_type == BTRFS_FILE_EXTENT_REG ||
6574             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
6575                 /* Only regular file could have regular/prealloc extent */
6576                 if (!S_ISREG(inode->vfs_inode.i_mode)) {
6577                         ret = -EUCLEAN;
6578                         btrfs_crit(fs_info,
6579                 "regular/prealloc extent found for non-regular inode %llu",
6580                                    btrfs_ino(inode));
6581                         goto out;
6582                 }
6583                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
6584                                                        extent_start);
6585         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
6586                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
6587                                                       path->slots[0],
6588                                                       extent_start);
6589         }
6590 next:
6591         if (start >= extent_end) {
6592                 path->slots[0]++;
6593                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6594                         ret = btrfs_next_leaf(root, path);
6595                         if (ret < 0) {
6596                                 err = ret;
6597                                 goto out;
6598                         } else if (ret > 0) {
6599                                 goto not_found;
6600                         }
6601                         leaf = path->nodes[0];
6602                 }
6603                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6604                 if (found_key.objectid != objectid ||
6605                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6606                         goto not_found;
6607                 if (start + len <= found_key.offset)
6608                         goto not_found;
6609                 if (start > found_key.offset)
6610                         goto next;
6611
6612                 /* New extent overlaps with existing one */
6613                 em->start = start;
6614                 em->orig_start = start;
6615                 em->len = found_key.offset - start;
6616                 em->block_start = EXTENT_MAP_HOLE;
6617                 goto insert;
6618         }
6619
6620         btrfs_extent_item_to_extent_map(inode, path, item, !page, em);
6621
6622         if (extent_type == BTRFS_FILE_EXTENT_REG ||
6623             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
6624                 goto insert;
6625         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
6626                 unsigned long ptr;
6627                 char *map;
6628                 size_t size;
6629                 size_t extent_offset;
6630                 size_t copy_size;
6631
6632                 if (!page)
6633                         goto out;
6634
6635                 size = btrfs_file_extent_ram_bytes(leaf, item);
6636                 extent_offset = page_offset(page) + pg_offset - extent_start;
6637                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6638                                   size - extent_offset);
6639                 em->start = extent_start + extent_offset;
6640                 em->len = ALIGN(copy_size, fs_info->sectorsize);
6641                 em->orig_block_len = em->len;
6642                 em->orig_start = em->start;
6643                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6644
6645                 btrfs_set_path_blocking(path);
6646                 if (!PageUptodate(page)) {
6647                         if (btrfs_file_extent_compression(leaf, item) !=
6648                             BTRFS_COMPRESS_NONE) {
6649                                 ret = uncompress_inline(path, page, pg_offset,
6650                                                         extent_offset, item);
6651                                 if (ret) {
6652                                         err = ret;
6653                                         goto out;
6654                                 }
6655                         } else {
6656                                 map = kmap(page);
6657                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6658                                                    copy_size);
6659                                 if (pg_offset + copy_size < PAGE_SIZE) {
6660                                         memset(map + pg_offset + copy_size, 0,
6661                                                PAGE_SIZE - pg_offset -
6662                                                copy_size);
6663                                 }
6664                                 kunmap(page);
6665                         }
6666                         flush_dcache_page(page);
6667                 }
6668                 set_extent_uptodate(io_tree, em->start,
6669                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6670                 goto insert;
6671         }
6672 not_found:
6673         em->start = start;
6674         em->orig_start = start;
6675         em->len = len;
6676         em->block_start = EXTENT_MAP_HOLE;
6677 insert:
6678         btrfs_release_path(path);
6679         if (em->start > start || extent_map_end(em) <= start) {
6680                 btrfs_err(fs_info,
6681                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
6682                           em->start, em->len, start, len);
6683                 err = -EIO;
6684                 goto out;
6685         }
6686
6687         err = 0;
6688         write_lock(&em_tree->lock);
6689         err = btrfs_add_extent_mapping(fs_info, em_tree, &em, start, len);
6690         write_unlock(&em_tree->lock);
6691 out:
6692         btrfs_free_path(path);
6693
6694         trace_btrfs_get_extent(root, inode, em);
6695
6696         if (err) {
6697                 free_extent_map(em);
6698                 return ERR_PTR(err);
6699         }
6700         BUG_ON(!em); /* Error is always set */
6701         return em;
6702 }
6703
6704 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
6705                                            u64 start, u64 len)
6706 {
6707         struct extent_map *em;
6708         struct extent_map *hole_em = NULL;
6709         u64 delalloc_start = start;
6710         u64 end;
6711         u64 delalloc_len;
6712         u64 delalloc_end;
6713         int err = 0;
6714
6715         em = btrfs_get_extent(inode, NULL, 0, start, len);
6716         if (IS_ERR(em))
6717                 return em;
6718         /*
6719          * If our em maps to:
6720          * - a hole or
6721          * - a pre-alloc extent,
6722          * there might actually be delalloc bytes behind it.
6723          */
6724         if (em->block_start != EXTENT_MAP_HOLE &&
6725             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
6726                 return em;
6727         else
6728                 hole_em = em;
6729
6730         /* check to see if we've wrapped (len == -1 or similar) */
6731         end = start + len;
6732         if (end < start)
6733                 end = (u64)-1;
6734         else
6735                 end -= 1;
6736
6737         em = NULL;
6738
6739         /* ok, we didn't find anything, lets look for delalloc */
6740         delalloc_len = count_range_bits(&inode->io_tree, &delalloc_start,
6741                                  end, len, EXTENT_DELALLOC, 1);
6742         delalloc_end = delalloc_start + delalloc_len;
6743         if (delalloc_end < delalloc_start)
6744                 delalloc_end = (u64)-1;
6745
6746         /*
6747          * We didn't find anything useful, return the original results from
6748          * get_extent()
6749          */
6750         if (delalloc_start > end || delalloc_end <= start) {
6751                 em = hole_em;
6752                 hole_em = NULL;
6753                 goto out;
6754         }
6755
6756         /*
6757          * Adjust the delalloc_start to make sure it doesn't go backwards from
6758          * the start they passed in
6759          */
6760         delalloc_start = max(start, delalloc_start);
6761         delalloc_len = delalloc_end - delalloc_start;
6762
6763         if (delalloc_len > 0) {
6764                 u64 hole_start;
6765                 u64 hole_len;
6766                 const u64 hole_end = extent_map_end(hole_em);
6767
6768                 em = alloc_extent_map();
6769                 if (!em) {
6770                         err = -ENOMEM;
6771                         goto out;
6772                 }
6773
6774                 ASSERT(hole_em);
6775                 /*
6776                  * When btrfs_get_extent can't find anything it returns one
6777                  * huge hole
6778                  *
6779                  * Make sure what it found really fits our range, and adjust to
6780                  * make sure it is based on the start from the caller
6781                  */
6782                 if (hole_end <= start || hole_em->start > end) {
6783                        free_extent_map(hole_em);
6784                        hole_em = NULL;
6785                 } else {
6786                        hole_start = max(hole_em->start, start);
6787                        hole_len = hole_end - hole_start;
6788                 }
6789
6790                 if (hole_em && delalloc_start > hole_start) {
6791                         /*
6792                          * Our hole starts before our delalloc, so we have to
6793                          * return just the parts of the hole that go until the
6794                          * delalloc starts
6795                          */
6796                         em->len = min(hole_len, delalloc_start - hole_start);
6797                         em->start = hole_start;
6798                         em->orig_start = hole_start;
6799                         /*
6800                          * Don't adjust block start at all, it is fixed at
6801                          * EXTENT_MAP_HOLE
6802                          */
6803                         em->block_start = hole_em->block_start;
6804                         em->block_len = hole_len;
6805                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
6806                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
6807                 } else {
6808                         /*
6809                          * Hole is out of passed range or it starts after
6810                          * delalloc range
6811                          */
6812                         em->start = delalloc_start;
6813                         em->len = delalloc_len;
6814                         em->orig_start = delalloc_start;
6815                         em->block_start = EXTENT_MAP_DELALLOC;
6816                         em->block_len = delalloc_len;
6817                 }
6818         } else {
6819                 return hole_em;
6820         }
6821 out:
6822
6823         free_extent_map(hole_em);
6824         if (err) {
6825                 free_extent_map(em);
6826                 return ERR_PTR(err);
6827         }
6828         return em;
6829 }
6830
6831 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
6832                                                   const u64 start,
6833                                                   const u64 len,
6834                                                   const u64 orig_start,
6835                                                   const u64 block_start,
6836                                                   const u64 block_len,
6837                                                   const u64 orig_block_len,
6838                                                   const u64 ram_bytes,
6839                                                   const int type)
6840 {
6841         struct extent_map *em = NULL;
6842         int ret;
6843
6844         if (type != BTRFS_ORDERED_NOCOW) {
6845                 em = create_io_em(inode, start, len, orig_start,
6846                                   block_start, block_len, orig_block_len,
6847                                   ram_bytes,
6848                                   BTRFS_COMPRESS_NONE, /* compress_type */
6849                                   type);
6850                 if (IS_ERR(em))
6851                         goto out;
6852         }
6853         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
6854                                            len, block_len, type);
6855         if (ret) {
6856                 if (em) {
6857                         free_extent_map(em);
6858                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
6859                                                 start + len - 1, 0);
6860                 }
6861                 em = ERR_PTR(ret);
6862         }
6863  out:
6864
6865         return em;
6866 }
6867
6868 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
6869                                                   u64 start, u64 len)
6870 {
6871         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6872         struct btrfs_root *root = BTRFS_I(inode)->root;
6873         struct extent_map *em;
6874         struct btrfs_key ins;
6875         u64 alloc_hint;
6876         int ret;
6877
6878         alloc_hint = get_extent_allocation_hint(inode, start, len);
6879         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
6880                                    0, alloc_hint, &ins, 1, 1);
6881         if (ret)
6882                 return ERR_PTR(ret);
6883
6884         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
6885                                      ins.objectid, ins.offset, ins.offset,
6886                                      ins.offset, BTRFS_ORDERED_REGULAR);
6887         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
6888         if (IS_ERR(em))
6889                 btrfs_free_reserved_extent(fs_info, ins.objectid,
6890                                            ins.offset, 1);
6891
6892         return em;
6893 }
6894
6895 /*
6896  * returns 1 when the nocow is safe, < 1 on error, 0 if the
6897  * block must be cow'd
6898  */
6899 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
6900                               u64 *orig_start, u64 *orig_block_len,
6901                               u64 *ram_bytes)
6902 {
6903         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6904         struct btrfs_path *path;
6905         int ret;
6906         struct extent_buffer *leaf;
6907         struct btrfs_root *root = BTRFS_I(inode)->root;
6908         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6909         struct btrfs_file_extent_item *fi;
6910         struct btrfs_key key;
6911         u64 disk_bytenr;
6912         u64 backref_offset;
6913         u64 extent_end;
6914         u64 num_bytes;
6915         int slot;
6916         int found_type;
6917         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
6918
6919         path = btrfs_alloc_path();
6920         if (!path)
6921                 return -ENOMEM;
6922
6923         ret = btrfs_lookup_file_extent(NULL, root, path,
6924                         btrfs_ino(BTRFS_I(inode)), offset, 0);
6925         if (ret < 0)
6926                 goto out;
6927
6928         slot = path->slots[0];
6929         if (ret == 1) {
6930                 if (slot == 0) {
6931                         /* can't find the item, must cow */
6932                         ret = 0;
6933                         goto out;
6934                 }
6935                 slot--;
6936         }
6937         ret = 0;
6938         leaf = path->nodes[0];
6939         btrfs_item_key_to_cpu(leaf, &key, slot);
6940         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
6941             key.type != BTRFS_EXTENT_DATA_KEY) {
6942                 /* not our file or wrong item type, must cow */
6943                 goto out;
6944         }
6945
6946         if (key.offset > offset) {
6947                 /* Wrong offset, must cow */
6948                 goto out;
6949         }
6950
6951         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
6952         found_type = btrfs_file_extent_type(leaf, fi);
6953         if (found_type != BTRFS_FILE_EXTENT_REG &&
6954             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
6955                 /* not a regular extent, must cow */
6956                 goto out;
6957         }
6958
6959         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
6960                 goto out;
6961
6962         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
6963         if (extent_end <= offset)
6964                 goto out;
6965
6966         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
6967         if (disk_bytenr == 0)
6968                 goto out;
6969
6970         if (btrfs_file_extent_compression(leaf, fi) ||
6971             btrfs_file_extent_encryption(leaf, fi) ||
6972             btrfs_file_extent_other_encoding(leaf, fi))
6973                 goto out;
6974
6975         /*
6976          * Do the same check as in btrfs_cross_ref_exist but without the
6977          * unnecessary search.
6978          */
6979         if (btrfs_file_extent_generation(leaf, fi) <=
6980             btrfs_root_last_snapshot(&root->root_item))
6981                 goto out;
6982
6983         backref_offset = btrfs_file_extent_offset(leaf, fi);
6984
6985         if (orig_start) {
6986                 *orig_start = key.offset - backref_offset;
6987                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
6988                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
6989         }
6990
6991         if (btrfs_extent_readonly(fs_info, disk_bytenr))
6992                 goto out;
6993
6994         num_bytes = min(offset + *len, extent_end) - offset;
6995         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6996                 u64 range_end;
6997
6998                 range_end = round_up(offset + num_bytes,
6999                                      root->fs_info->sectorsize) - 1;
7000                 ret = test_range_bit(io_tree, offset, range_end,
7001                                      EXTENT_DELALLOC, 0, NULL);
7002                 if (ret) {
7003                         ret = -EAGAIN;
7004                         goto out;
7005                 }
7006         }
7007
7008         btrfs_release_path(path);
7009
7010         /*
7011          * look for other files referencing this extent, if we
7012          * find any we must cow
7013          */
7014
7015         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7016                                     key.offset - backref_offset, disk_bytenr);
7017         if (ret) {
7018                 ret = 0;
7019                 goto out;
7020         }
7021
7022         /*
7023          * adjust disk_bytenr and num_bytes to cover just the bytes
7024          * in this extent we are about to write.  If there
7025          * are any csums in that range we have to cow in order
7026          * to keep the csums correct
7027          */
7028         disk_bytenr += backref_offset;
7029         disk_bytenr += offset - key.offset;
7030         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7031                 goto out;
7032         /*
7033          * all of the above have passed, it is safe to overwrite this extent
7034          * without cow
7035          */
7036         *len = num_bytes;
7037         ret = 1;
7038 out:
7039         btrfs_free_path(path);
7040         return ret;
7041 }
7042
7043 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7044                               struct extent_state **cached_state, bool writing)
7045 {
7046         struct btrfs_ordered_extent *ordered;
7047         int ret = 0;
7048
7049         while (1) {
7050                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7051                                  cached_state);
7052                 /*
7053                  * We're concerned with the entire range that we're going to be
7054                  * doing DIO to, so we need to make sure there's no ordered
7055                  * extents in this range.
7056                  */
7057                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7058                                                      lockend - lockstart + 1);
7059
7060                 /*
7061                  * We need to make sure there are no buffered pages in this
7062                  * range either, we could have raced between the invalidate in
7063                  * generic_file_direct_write and locking the extent.  The
7064                  * invalidate needs to happen so that reads after a write do not
7065                  * get stale data.
7066                  */
7067                 if (!ordered &&
7068                     (!writing || !filemap_range_has_page(inode->i_mapping,
7069                                                          lockstart, lockend)))
7070                         break;
7071
7072                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7073                                      cached_state);
7074
7075                 if (ordered) {
7076                         /*
7077                          * If we are doing a DIO read and the ordered extent we
7078                          * found is for a buffered write, we can not wait for it
7079                          * to complete and retry, because if we do so we can
7080                          * deadlock with concurrent buffered writes on page
7081                          * locks. This happens only if our DIO read covers more
7082                          * than one extent map, if at this point has already
7083                          * created an ordered extent for a previous extent map
7084                          * and locked its range in the inode's io tree, and a
7085                          * concurrent write against that previous extent map's
7086                          * range and this range started (we unlock the ranges
7087                          * in the io tree only when the bios complete and
7088                          * buffered writes always lock pages before attempting
7089                          * to lock range in the io tree).
7090                          */
7091                         if (writing ||
7092                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7093                                 btrfs_start_ordered_extent(inode, ordered, 1);
7094                         else
7095                                 ret = -ENOTBLK;
7096                         btrfs_put_ordered_extent(ordered);
7097                 } else {
7098                         /*
7099                          * We could trigger writeback for this range (and wait
7100                          * for it to complete) and then invalidate the pages for
7101                          * this range (through invalidate_inode_pages2_range()),
7102                          * but that can lead us to a deadlock with a concurrent
7103                          * call to readahead (a buffered read or a defrag call
7104                          * triggered a readahead) on a page lock due to an
7105                          * ordered dio extent we created before but did not have
7106                          * yet a corresponding bio submitted (whence it can not
7107                          * complete), which makes readahead wait for that
7108                          * ordered extent to complete while holding a lock on
7109                          * that page.
7110                          */
7111                         ret = -ENOTBLK;
7112                 }
7113
7114                 if (ret)
7115                         break;
7116
7117                 cond_resched();
7118         }
7119
7120         return ret;
7121 }
7122
7123 /* The callers of this must take lock_extent() */
7124 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7125                                        u64 orig_start, u64 block_start,
7126                                        u64 block_len, u64 orig_block_len,
7127                                        u64 ram_bytes, int compress_type,
7128                                        int type)
7129 {
7130         struct extent_map_tree *em_tree;
7131         struct extent_map *em;
7132         int ret;
7133
7134         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7135                type == BTRFS_ORDERED_COMPRESSED ||
7136                type == BTRFS_ORDERED_NOCOW ||
7137                type == BTRFS_ORDERED_REGULAR);
7138
7139         em_tree = &BTRFS_I(inode)->extent_tree;
7140         em = alloc_extent_map();
7141         if (!em)
7142                 return ERR_PTR(-ENOMEM);
7143
7144         em->start = start;
7145         em->orig_start = orig_start;
7146         em->len = len;
7147         em->block_len = block_len;
7148         em->block_start = block_start;
7149         em->orig_block_len = orig_block_len;
7150         em->ram_bytes = ram_bytes;
7151         em->generation = -1;
7152         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7153         if (type == BTRFS_ORDERED_PREALLOC) {
7154                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7155         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7156                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7157                 em->compress_type = compress_type;
7158         }
7159
7160         do {
7161                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7162                                 em->start + em->len - 1, 0);
7163                 write_lock(&em_tree->lock);
7164                 ret = add_extent_mapping(em_tree, em, 1);
7165                 write_unlock(&em_tree->lock);
7166                 /*
7167                  * The caller has taken lock_extent(), who could race with us
7168                  * to add em?
7169                  */
7170         } while (ret == -EEXIST);
7171
7172         if (ret) {
7173                 free_extent_map(em);
7174                 return ERR_PTR(ret);
7175         }
7176
7177         /* em got 2 refs now, callers needs to do free_extent_map once. */
7178         return em;
7179 }
7180
7181
7182 static int btrfs_get_blocks_direct_write(struct extent_map **map,
7183                                          struct inode *inode,
7184                                          struct btrfs_dio_data *dio_data,
7185                                          u64 start, u64 len)
7186 {
7187         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7188         struct extent_map *em = *map;
7189         int ret = 0;
7190
7191         /*
7192          * We don't allocate a new extent in the following cases
7193          *
7194          * 1) The inode is marked as NODATACOW. In this case we'll just use the
7195          * existing extent.
7196          * 2) The extent is marked as PREALLOC. We're good to go here and can
7197          * just use the extent.
7198          *
7199          */
7200         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7201             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7202              em->block_start != EXTENT_MAP_HOLE)) {
7203                 int type;
7204                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7205
7206                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7207                         type = BTRFS_ORDERED_PREALLOC;
7208                 else
7209                         type = BTRFS_ORDERED_NOCOW;
7210                 len = min(len, em->len - (start - em->start));
7211                 block_start = em->block_start + (start - em->start);
7212
7213                 if (can_nocow_extent(inode, start, &len, &orig_start,
7214                                      &orig_block_len, &ram_bytes) == 1 &&
7215                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7216                         struct extent_map *em2;
7217
7218                         em2 = btrfs_create_dio_extent(inode, start, len,
7219                                                       orig_start, block_start,
7220                                                       len, orig_block_len,
7221                                                       ram_bytes, type);
7222                         btrfs_dec_nocow_writers(fs_info, block_start);
7223                         if (type == BTRFS_ORDERED_PREALLOC) {
7224                                 free_extent_map(em);
7225                                 *map = em = em2;
7226                         }
7227
7228                         if (em2 && IS_ERR(em2)) {
7229                                 ret = PTR_ERR(em2);
7230                                 goto out;
7231                         }
7232                         /*
7233                          * For inode marked NODATACOW or extent marked PREALLOC,
7234                          * use the existing or preallocated extent, so does not
7235                          * need to adjust btrfs_space_info's bytes_may_use.
7236                          */
7237                         btrfs_free_reserved_data_space_noquota(inode, start,
7238                                                                len);
7239                         goto skip_cow;
7240                 }
7241         }
7242
7243         /* this will cow the extent */
7244         free_extent_map(em);
7245         *map = em = btrfs_new_extent_direct(inode, start, len);
7246         if (IS_ERR(em)) {
7247                 ret = PTR_ERR(em);
7248                 goto out;
7249         }
7250
7251         len = min(len, em->len - (start - em->start));
7252
7253 skip_cow:
7254         /*
7255          * Need to update the i_size under the extent lock so buffered
7256          * readers will get the updated i_size when we unlock.
7257          */
7258         if (start + len > i_size_read(inode))
7259                 i_size_write(inode, start + len);
7260
7261         dio_data->reserve -= len;
7262 out:
7263         return ret;
7264 }
7265
7266 static int btrfs_dio_iomap_begin(struct inode *inode, loff_t start,
7267                 loff_t length, unsigned flags, struct iomap *iomap,
7268                 struct iomap *srcmap)
7269 {
7270         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7271         struct extent_map *em;
7272         struct extent_state *cached_state = NULL;
7273         struct btrfs_dio_data *dio_data = NULL;
7274         u64 lockstart, lockend;
7275         const bool write = !!(flags & IOMAP_WRITE);
7276         int ret = 0;
7277         u64 len = length;
7278         bool unlock_extents = false;
7279
7280         if (!write)
7281                 len = min_t(u64, len, fs_info->sectorsize);
7282
7283         lockstart = start;
7284         lockend = start + len - 1;
7285
7286         /*
7287          * The generic stuff only does filemap_write_and_wait_range, which
7288          * isn't enough if we've written compressed pages to this area, so we
7289          * need to flush the dirty pages again to make absolutely sure that any
7290          * outstanding dirty pages are on disk.
7291          */
7292         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
7293                      &BTRFS_I(inode)->runtime_flags))
7294                 ret = filemap_fdatawrite_range(inode->i_mapping, start,
7295                                                start + length - 1);
7296
7297         dio_data = kzalloc(sizeof(*dio_data), GFP_NOFS);
7298         if (!dio_data)
7299                 return -ENOMEM;
7300
7301         dio_data->length = length;
7302         if (write) {
7303                 dio_data->reserve = round_up(length, fs_info->sectorsize);
7304                 ret = btrfs_delalloc_reserve_space(inode,
7305                                 &dio_data->data_reserved,
7306                                 start, dio_data->reserve);
7307                 if (ret) {
7308                         extent_changeset_free(dio_data->data_reserved);
7309                         kfree(dio_data);
7310                         return ret;
7311                 }
7312         }
7313         iomap->private = dio_data;
7314
7315
7316         /*
7317          * If this errors out it's because we couldn't invalidate pagecache for
7318          * this range and we need to fallback to buffered.
7319          */
7320         if (lock_extent_direct(inode, lockstart, lockend, &cached_state, write)) {
7321                 ret = -ENOTBLK;
7322                 goto err;
7323         }
7324
7325         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len);
7326         if (IS_ERR(em)) {
7327                 ret = PTR_ERR(em);
7328                 goto unlock_err;
7329         }
7330
7331         /*
7332          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7333          * io.  INLINE is special, and we could probably kludge it in here, but
7334          * it's still buffered so for safety lets just fall back to the generic
7335          * buffered path.
7336          *
7337          * For COMPRESSED we _have_ to read the entire extent in so we can
7338          * decompress it, so there will be buffering required no matter what we
7339          * do, so go ahead and fallback to buffered.
7340          *
7341          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7342          * to buffered IO.  Don't blame me, this is the price we pay for using
7343          * the generic code.
7344          */
7345         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7346             em->block_start == EXTENT_MAP_INLINE) {
7347                 free_extent_map(em);
7348                 ret = -ENOTBLK;
7349                 goto unlock_err;
7350         }
7351
7352         len = min(len, em->len - (start - em->start));
7353         if (write) {
7354                 ret = btrfs_get_blocks_direct_write(&em, inode, dio_data,
7355                                                     start, len);
7356                 if (ret < 0)
7357                         goto unlock_err;
7358                 unlock_extents = true;
7359                 /* Recalc len in case the new em is smaller than requested */
7360                 len = min(len, em->len - (start - em->start));
7361         } else {
7362                 /*
7363                  * We need to unlock only the end area that we aren't using.
7364                  * The rest is going to be unlocked by the endio routine.
7365                  */
7366                 lockstart = start + len;
7367                 if (lockstart < lockend)
7368                         unlock_extents = true;
7369         }
7370
7371         if (unlock_extents)
7372                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
7373                                      lockstart, lockend, &cached_state);
7374         else
7375                 free_extent_state(cached_state);
7376
7377         /*
7378          * Translate extent map information to iomap.
7379          * We trim the extents (and move the addr) even though iomap code does
7380          * that, since we have locked only the parts we are performing I/O in.
7381          */
7382         if ((em->block_start == EXTENT_MAP_HOLE) ||
7383             (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) && !write)) {
7384                 iomap->addr = IOMAP_NULL_ADDR;
7385                 iomap->type = IOMAP_HOLE;
7386         } else {
7387                 iomap->addr = em->block_start + (start - em->start);
7388                 iomap->type = IOMAP_MAPPED;
7389         }
7390         iomap->offset = start;
7391         iomap->bdev = fs_info->fs_devices->latest_bdev;
7392         iomap->length = len;
7393
7394         free_extent_map(em);
7395
7396         return 0;
7397
7398 unlock_err:
7399         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7400                              &cached_state);
7401 err:
7402         if (dio_data) {
7403                 btrfs_delalloc_release_space(inode, dio_data->data_reserved,
7404                                 start, dio_data->reserve, true);
7405                 btrfs_delalloc_release_extents(BTRFS_I(inode), dio_data->reserve);
7406                 extent_changeset_free(dio_data->data_reserved);
7407                 kfree(dio_data);
7408         }
7409         return ret;
7410 }
7411
7412 static int btrfs_dio_iomap_end(struct inode *inode, loff_t pos, loff_t length,
7413                 ssize_t written, unsigned flags, struct iomap *iomap)
7414 {
7415         int ret = 0;
7416         struct btrfs_dio_data *dio_data = iomap->private;
7417         size_t submitted = dio_data->submitted;
7418         const bool write = !!(flags & IOMAP_WRITE);
7419
7420         if (!write && (iomap->type == IOMAP_HOLE)) {
7421                 /* If reading from a hole, unlock and return */
7422                 unlock_extent(&BTRFS_I(inode)->io_tree, pos, pos + length - 1);
7423                 goto out;
7424         }
7425
7426         if (submitted < length) {
7427                 pos += submitted;
7428                 length -= submitted;
7429                 if (write)
7430                         __endio_write_update_ordered(inode, pos, length, false);
7431                 else
7432                         unlock_extent(&BTRFS_I(inode)->io_tree, pos,
7433                                       pos + length - 1);
7434                 ret = -ENOTBLK;
7435         }
7436
7437         if (write) {
7438                 if (dio_data->reserve)
7439                         btrfs_delalloc_release_space(inode,
7440                                         dio_data->data_reserved, pos,
7441                                         dio_data->reserve, true);
7442                 btrfs_delalloc_release_extents(BTRFS_I(inode), dio_data->length);
7443                 extent_changeset_free(dio_data->data_reserved);
7444         }
7445 out:
7446         kfree(dio_data);
7447         iomap->private = NULL;
7448
7449         return ret;
7450 }
7451
7452 static void btrfs_dio_private_put(struct btrfs_dio_private *dip)
7453 {
7454         /*
7455          * This implies a barrier so that stores to dio_bio->bi_status before
7456          * this and loads of dio_bio->bi_status after this are fully ordered.
7457          */
7458         if (!refcount_dec_and_test(&dip->refs))
7459                 return;
7460
7461         if (bio_op(dip->dio_bio) == REQ_OP_WRITE) {
7462                 __endio_write_update_ordered(dip->inode, dip->logical_offset,
7463                                              dip->bytes,
7464                                              !dip->dio_bio->bi_status);
7465         } else {
7466                 unlock_extent(&BTRFS_I(dip->inode)->io_tree,
7467                               dip->logical_offset,
7468                               dip->logical_offset + dip->bytes - 1);
7469         }
7470
7471         bio_endio(dip->dio_bio);
7472         kfree(dip);
7473 }
7474
7475 static blk_status_t submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7476                                           int mirror_num,
7477                                           unsigned long bio_flags)
7478 {
7479         struct btrfs_dio_private *dip = bio->bi_private;
7480         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7481         blk_status_t ret;
7482
7483         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7484
7485         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
7486         if (ret)
7487                 return ret;
7488
7489         refcount_inc(&dip->refs);
7490         ret = btrfs_map_bio(fs_info, bio, mirror_num);
7491         if (ret)
7492                 refcount_dec(&dip->refs);
7493         return ret;
7494 }
7495
7496 static blk_status_t btrfs_check_read_dio_bio(struct inode *inode,
7497                                              struct btrfs_io_bio *io_bio,
7498                                              const bool uptodate)
7499 {
7500         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
7501         const u32 sectorsize = fs_info->sectorsize;
7502         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7503         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7504         const bool csum = !(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM);
7505         struct bio_vec bvec;
7506         struct bvec_iter iter;
7507         u64 start = io_bio->logical;
7508         int icsum = 0;
7509         blk_status_t err = BLK_STS_OK;
7510
7511         __bio_for_each_segment(bvec, &io_bio->bio, iter, io_bio->iter) {
7512                 unsigned int i, nr_sectors, pgoff;
7513
7514                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7515                 pgoff = bvec.bv_offset;
7516                 for (i = 0; i < nr_sectors; i++) {
7517                         ASSERT(pgoff < PAGE_SIZE);
7518                         if (uptodate &&
7519                             (!csum || !check_data_csum(inode, io_bio, icsum,
7520                                                        bvec.bv_page, pgoff,
7521                                                        start, sectorsize))) {
7522                                 clean_io_failure(fs_info, failure_tree, io_tree,
7523                                                  start, bvec.bv_page,
7524                                                  btrfs_ino(BTRFS_I(inode)),
7525                                                  pgoff);
7526                         } else {
7527                                 blk_status_t status;
7528
7529                                 status = btrfs_submit_read_repair(inode,
7530                                                         &io_bio->bio,
7531                                                         start - io_bio->logical,
7532                                                         bvec.bv_page, pgoff,
7533                                                         start,
7534                                                         start + sectorsize - 1,
7535                                                         io_bio->mirror_num,
7536                                                         submit_dio_repair_bio);
7537                                 if (status)
7538                                         err = status;
7539                         }
7540                         start += sectorsize;
7541                         icsum++;
7542                         pgoff += sectorsize;
7543                 }
7544         }
7545         return err;
7546 }
7547
7548 static void __endio_write_update_ordered(struct inode *inode,
7549                                          const u64 offset, const u64 bytes,
7550                                          const bool uptodate)
7551 {
7552         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7553         struct btrfs_ordered_extent *ordered = NULL;
7554         struct btrfs_workqueue *wq;
7555         u64 ordered_offset = offset;
7556         u64 ordered_bytes = bytes;
7557         u64 last_offset;
7558
7559         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
7560                 wq = fs_info->endio_freespace_worker;
7561         else
7562                 wq = fs_info->endio_write_workers;
7563
7564         while (ordered_offset < offset + bytes) {
7565                 last_offset = ordered_offset;
7566                 if (btrfs_dec_test_first_ordered_pending(inode, &ordered,
7567                                                            &ordered_offset,
7568                                                            ordered_bytes,
7569                                                            uptodate)) {
7570                         btrfs_init_work(&ordered->work, finish_ordered_fn, NULL,
7571                                         NULL);
7572                         btrfs_queue_work(wq, &ordered->work);
7573                 }
7574                 /*
7575                  * If btrfs_dec_test_ordered_pending does not find any ordered
7576                  * extent in the range, we can exit.
7577                  */
7578                 if (ordered_offset == last_offset)
7579                         return;
7580                 /*
7581                  * Our bio might span multiple ordered extents. In this case
7582                  * we keep going until we have accounted the whole dio.
7583                  */
7584                 if (ordered_offset < offset + bytes) {
7585                         ordered_bytes = offset + bytes - ordered_offset;
7586                         ordered = NULL;
7587                 }
7588         }
7589 }
7590
7591 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
7592                                     struct bio *bio, u64 offset)
7593 {
7594         struct inode *inode = private_data;
7595         blk_status_t ret;
7596         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
7597         BUG_ON(ret); /* -ENOMEM */
7598         return 0;
7599 }
7600
7601 static void btrfs_end_dio_bio(struct bio *bio)
7602 {
7603         struct btrfs_dio_private *dip = bio->bi_private;
7604         blk_status_t err = bio->bi_status;
7605
7606         if (err)
7607                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
7608                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
7609                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
7610                            bio->bi_opf,
7611                            (unsigned long long)bio->bi_iter.bi_sector,
7612                            bio->bi_iter.bi_size, err);
7613
7614         if (bio_op(bio) == REQ_OP_READ) {
7615                 err = btrfs_check_read_dio_bio(dip->inode, btrfs_io_bio(bio),
7616                                                !err);
7617         }
7618
7619         if (err)
7620                 dip->dio_bio->bi_status = err;
7621
7622         bio_put(bio);
7623         btrfs_dio_private_put(dip);
7624 }
7625
7626 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
7627                 struct inode *inode, u64 file_offset, int async_submit)
7628 {
7629         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7630         struct btrfs_dio_private *dip = bio->bi_private;
7631         bool write = bio_op(bio) == REQ_OP_WRITE;
7632         blk_status_t ret;
7633
7634         /* Check btrfs_submit_bio_hook() for rules about async submit. */
7635         if (async_submit)
7636                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
7637
7638         if (!write) {
7639                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
7640                 if (ret)
7641                         goto err;
7642         }
7643
7644         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
7645                 goto map;
7646
7647         if (write && async_submit) {
7648                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
7649                                           file_offset, inode,
7650                                           btrfs_submit_bio_start_direct_io);
7651                 goto err;
7652         } else if (write) {
7653                 /*
7654                  * If we aren't doing async submit, calculate the csum of the
7655                  * bio now.
7656                  */
7657                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
7658                 if (ret)
7659                         goto err;
7660         } else {
7661                 u64 csum_offset;
7662
7663                 csum_offset = file_offset - dip->logical_offset;
7664                 csum_offset >>= inode->i_sb->s_blocksize_bits;
7665                 csum_offset *= btrfs_super_csum_size(fs_info->super_copy);
7666                 btrfs_io_bio(bio)->csum = dip->csums + csum_offset;
7667         }
7668 map:
7669         ret = btrfs_map_bio(fs_info, bio, 0);
7670 err:
7671         return ret;
7672 }
7673
7674 /*
7675  * If this succeeds, the btrfs_dio_private is responsible for cleaning up locked
7676  * or ordered extents whether or not we submit any bios.
7677  */
7678 static struct btrfs_dio_private *btrfs_create_dio_private(struct bio *dio_bio,
7679                                                           struct inode *inode,
7680                                                           loff_t file_offset)
7681 {
7682         const bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
7683         const bool csum = !(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM);
7684         size_t dip_size;
7685         struct btrfs_dio_private *dip;
7686
7687         dip_size = sizeof(*dip);
7688         if (!write && csum) {
7689                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7690                 const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
7691                 size_t nblocks;
7692
7693                 nblocks = dio_bio->bi_iter.bi_size >> inode->i_sb->s_blocksize_bits;
7694                 dip_size += csum_size * nblocks;
7695         }
7696
7697         dip = kzalloc(dip_size, GFP_NOFS);
7698         if (!dip)
7699                 return NULL;
7700
7701         dip->inode = inode;
7702         dip->logical_offset = file_offset;
7703         dip->bytes = dio_bio->bi_iter.bi_size;
7704         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
7705         dip->dio_bio = dio_bio;
7706         refcount_set(&dip->refs, 1);
7707         return dip;
7708 }
7709
7710 static blk_qc_t btrfs_submit_direct(struct inode *inode, struct iomap *iomap,
7711                 struct bio *dio_bio, loff_t file_offset)
7712 {
7713         const bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
7714         const bool csum = !(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM);
7715         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7716         const bool raid56 = (btrfs_data_alloc_profile(fs_info) &
7717                              BTRFS_BLOCK_GROUP_RAID56_MASK);
7718         struct btrfs_dio_private *dip;
7719         struct bio *bio;
7720         u64 start_sector;
7721         int async_submit = 0;
7722         u64 submit_len;
7723         int clone_offset = 0;
7724         int clone_len;
7725         int ret;
7726         blk_status_t status;
7727         struct btrfs_io_geometry geom;
7728         struct btrfs_dio_data *dio_data = iomap->private;
7729
7730         dip = btrfs_create_dio_private(dio_bio, inode, file_offset);
7731         if (!dip) {
7732                 if (!write) {
7733                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
7734                                 file_offset + dio_bio->bi_iter.bi_size - 1);
7735                 }
7736                 dio_bio->bi_status = BLK_STS_RESOURCE;
7737                 bio_endio(dio_bio);
7738                 return BLK_QC_T_NONE;
7739         }
7740
7741         if (!write && csum) {
7742                 /*
7743                  * Load the csums up front to reduce csum tree searches and
7744                  * contention when submitting bios.
7745                  */
7746                 status = btrfs_lookup_bio_sums(inode, dio_bio, file_offset,
7747                                                dip->csums);
7748                 if (status != BLK_STS_OK)
7749                         goto out_err;
7750         }
7751
7752         start_sector = dio_bio->bi_iter.bi_sector;
7753         submit_len = dio_bio->bi_iter.bi_size;
7754
7755         do {
7756                 ret = btrfs_get_io_geometry(fs_info, btrfs_op(dio_bio),
7757                                             start_sector << 9, submit_len,
7758                                             &geom);
7759                 if (ret) {
7760                         status = errno_to_blk_status(ret);
7761                         goto out_err;
7762                 }
7763                 ASSERT(geom.len <= INT_MAX);
7764
7765                 clone_len = min_t(int, submit_len, geom.len);
7766
7767                 /*
7768                  * This will never fail as it's passing GPF_NOFS and
7769                  * the allocation is backed by btrfs_bioset.
7770                  */
7771                 bio = btrfs_bio_clone_partial(dio_bio, clone_offset, clone_len);
7772                 bio->bi_private = dip;
7773                 bio->bi_end_io = btrfs_end_dio_bio;
7774                 btrfs_io_bio(bio)->logical = file_offset;
7775
7776                 ASSERT(submit_len >= clone_len);
7777                 submit_len -= clone_len;
7778
7779                 /*
7780                  * Increase the count before we submit the bio so we know
7781                  * the end IO handler won't happen before we increase the
7782                  * count. Otherwise, the dip might get freed before we're
7783                  * done setting it up.
7784                  *
7785                  * We transfer the initial reference to the last bio, so we
7786                  * don't need to increment the reference count for the last one.
7787                  */
7788                 if (submit_len > 0) {
7789                         refcount_inc(&dip->refs);
7790                         /*
7791                          * If we are submitting more than one bio, submit them
7792                          * all asynchronously. The exception is RAID 5 or 6, as
7793                          * asynchronous checksums make it difficult to collect
7794                          * full stripe writes.
7795                          */
7796                         if (!raid56)
7797                                 async_submit = 1;
7798                 }
7799
7800                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
7801                                                 async_submit);
7802                 if (status) {
7803                         bio_put(bio);
7804                         if (submit_len > 0)
7805                                 refcount_dec(&dip->refs);
7806                         goto out_err;
7807                 }
7808
7809                 dio_data->submitted += clone_len;
7810                 clone_offset += clone_len;
7811                 start_sector += clone_len >> 9;
7812                 file_offset += clone_len;
7813         } while (submit_len > 0);
7814         return BLK_QC_T_NONE;
7815
7816 out_err:
7817         dip->dio_bio->bi_status = status;
7818         btrfs_dio_private_put(dip);
7819         return BLK_QC_T_NONE;
7820 }
7821
7822 const struct iomap_ops btrfs_dio_iomap_ops = {
7823         .iomap_begin            = btrfs_dio_iomap_begin,
7824         .iomap_end              = btrfs_dio_iomap_end,
7825 };
7826
7827 const struct iomap_dio_ops btrfs_dops = {
7828         .submit_io              = btrfs_submit_direct,
7829 };
7830
7831 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
7832
7833 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
7834                 __u64 start, __u64 len)
7835 {
7836         int     ret;
7837
7838         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
7839         if (ret)
7840                 return ret;
7841
7842         return extent_fiemap(inode, fieinfo, start, len);
7843 }
7844
7845 int btrfs_readpage(struct file *file, struct page *page)
7846 {
7847         return extent_read_full_page(page, btrfs_get_extent, 0);
7848 }
7849
7850 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
7851 {
7852         struct inode *inode = page->mapping->host;
7853         int ret;
7854
7855         if (current->flags & PF_MEMALLOC) {
7856                 redirty_page_for_writepage(wbc, page);
7857                 unlock_page(page);
7858                 return 0;
7859         }
7860
7861         /*
7862          * If we are under memory pressure we will call this directly from the
7863          * VM, we need to make sure we have the inode referenced for the ordered
7864          * extent.  If not just return like we didn't do anything.
7865          */
7866         if (!igrab(inode)) {
7867                 redirty_page_for_writepage(wbc, page);
7868                 return AOP_WRITEPAGE_ACTIVATE;
7869         }
7870         ret = extent_write_full_page(page, wbc);
7871         btrfs_add_delayed_iput(inode);
7872         return ret;
7873 }
7874
7875 static int btrfs_writepages(struct address_space *mapping,
7876                             struct writeback_control *wbc)
7877 {
7878         return extent_writepages(mapping, wbc);
7879 }
7880
7881 static void btrfs_readahead(struct readahead_control *rac)
7882 {
7883         extent_readahead(rac);
7884 }
7885
7886 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
7887 {
7888         int ret = try_release_extent_mapping(page, gfp_flags);
7889         if (ret == 1)
7890                 detach_page_private(page);
7891         return ret;
7892 }
7893
7894 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
7895 {
7896         if (PageWriteback(page) || PageDirty(page))
7897                 return 0;
7898         return __btrfs_releasepage(page, gfp_flags);
7899 }
7900
7901 #ifdef CONFIG_MIGRATION
7902 static int btrfs_migratepage(struct address_space *mapping,
7903                              struct page *newpage, struct page *page,
7904                              enum migrate_mode mode)
7905 {
7906         int ret;
7907
7908         ret = migrate_page_move_mapping(mapping, newpage, page, 0);
7909         if (ret != MIGRATEPAGE_SUCCESS)
7910                 return ret;
7911
7912         if (page_has_private(page))
7913                 attach_page_private(newpage, detach_page_private(page));
7914
7915         if (PagePrivate2(page)) {
7916                 ClearPagePrivate2(page);
7917                 SetPagePrivate2(newpage);
7918         }
7919
7920         if (mode != MIGRATE_SYNC_NO_COPY)
7921                 migrate_page_copy(newpage, page);
7922         else
7923                 migrate_page_states(newpage, page);
7924         return MIGRATEPAGE_SUCCESS;
7925 }
7926 #endif
7927
7928 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
7929                                  unsigned int length)
7930 {
7931         struct inode *inode = page->mapping->host;
7932         struct extent_io_tree *tree;
7933         struct btrfs_ordered_extent *ordered;
7934         struct extent_state *cached_state = NULL;
7935         u64 page_start = page_offset(page);
7936         u64 page_end = page_start + PAGE_SIZE - 1;
7937         u64 start;
7938         u64 end;
7939         int inode_evicting = inode->i_state & I_FREEING;
7940
7941         /*
7942          * we have the page locked, so new writeback can't start,
7943          * and the dirty bit won't be cleared while we are here.
7944          *
7945          * Wait for IO on this page so that we can safely clear
7946          * the PagePrivate2 bit and do ordered accounting
7947          */
7948         wait_on_page_writeback(page);
7949
7950         tree = &BTRFS_I(inode)->io_tree;
7951         if (offset) {
7952                 btrfs_releasepage(page, GFP_NOFS);
7953                 return;
7954         }
7955
7956         if (!inode_evicting)
7957                 lock_extent_bits(tree, page_start, page_end, &cached_state);
7958 again:
7959         start = page_start;
7960         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
7961                                         page_end - start + 1);
7962         if (ordered) {
7963                 end = min(page_end,
7964                           ordered->file_offset + ordered->num_bytes - 1);
7965                 /*
7966                  * IO on this page will never be started, so we need
7967                  * to account for any ordered extents now
7968                  */
7969                 if (!inode_evicting)
7970                         clear_extent_bit(tree, start, end,
7971                                          EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
7972                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
7973                                          EXTENT_DEFRAG, 1, 0, &cached_state);
7974                 /*
7975                  * whoever cleared the private bit is responsible
7976                  * for the finish_ordered_io
7977                  */
7978                 if (TestClearPagePrivate2(page)) {
7979                         struct btrfs_ordered_inode_tree *tree;
7980                         u64 new_len;
7981
7982                         tree = &BTRFS_I(inode)->ordered_tree;
7983
7984                         spin_lock_irq(&tree->lock);
7985                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
7986                         new_len = start - ordered->file_offset;
7987                         if (new_len < ordered->truncated_len)
7988                                 ordered->truncated_len = new_len;
7989                         spin_unlock_irq(&tree->lock);
7990
7991                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
7992                                                            start,
7993                                                            end - start + 1, 1))
7994                                 btrfs_finish_ordered_io(ordered);
7995                 }
7996                 btrfs_put_ordered_extent(ordered);
7997                 if (!inode_evicting) {
7998                         cached_state = NULL;
7999                         lock_extent_bits(tree, start, end,
8000                                          &cached_state);
8001                 }
8002
8003                 start = end + 1;
8004                 if (start < page_end)
8005                         goto again;
8006         }
8007
8008         /*
8009          * Qgroup reserved space handler
8010          * Page here will be either
8011          * 1) Already written to disk
8012          *    In this case, its reserved space is released from data rsv map
8013          *    and will be freed by delayed_ref handler finally.
8014          *    So even we call qgroup_free_data(), it won't decrease reserved
8015          *    space.
8016          * 2) Not written to disk
8017          *    This means the reserved space should be freed here. However,
8018          *    if a truncate invalidates the page (by clearing PageDirty)
8019          *    and the page is accounted for while allocating extent
8020          *    in btrfs_check_data_free_space() we let delayed_ref to
8021          *    free the entire extent.
8022          */
8023         if (PageDirty(page))
8024                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8025         if (!inode_evicting) {
8026                 clear_extent_bit(tree, page_start, page_end, EXTENT_LOCKED |
8027                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8028                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8029                                  &cached_state);
8030
8031                 __btrfs_releasepage(page, GFP_NOFS);
8032         }
8033
8034         ClearPageChecked(page);
8035         detach_page_private(page);
8036 }
8037
8038 /*
8039  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8040  * called from a page fault handler when a page is first dirtied. Hence we must
8041  * be careful to check for EOF conditions here. We set the page up correctly
8042  * for a written page which means we get ENOSPC checking when writing into
8043  * holes and correct delalloc and unwritten extent mapping on filesystems that
8044  * support these features.
8045  *
8046  * We are not allowed to take the i_mutex here so we have to play games to
8047  * protect against truncate races as the page could now be beyond EOF.  Because
8048  * truncate_setsize() writes the inode size before removing pages, once we have
8049  * the page lock we can determine safely if the page is beyond EOF. If it is not
8050  * beyond EOF, then the page is guaranteed safe against truncation until we
8051  * unlock the page.
8052  */
8053 vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
8054 {
8055         struct page *page = vmf->page;
8056         struct inode *inode = file_inode(vmf->vma->vm_file);
8057         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8058         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8059         struct btrfs_ordered_extent *ordered;
8060         struct extent_state *cached_state = NULL;
8061         struct extent_changeset *data_reserved = NULL;
8062         char *kaddr;
8063         unsigned long zero_start;
8064         loff_t size;
8065         vm_fault_t ret;
8066         int ret2;
8067         int reserved = 0;
8068         u64 reserved_space;
8069         u64 page_start;
8070         u64 page_end;
8071         u64 end;
8072
8073         reserved_space = PAGE_SIZE;
8074
8075         sb_start_pagefault(inode->i_sb);
8076         page_start = page_offset(page);
8077         page_end = page_start + PAGE_SIZE - 1;
8078         end = page_end;
8079
8080         /*
8081          * Reserving delalloc space after obtaining the page lock can lead to
8082          * deadlock. For example, if a dirty page is locked by this function
8083          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8084          * dirty page write out, then the btrfs_writepage() function could
8085          * end up waiting indefinitely to get a lock on the page currently
8086          * being processed by btrfs_page_mkwrite() function.
8087          */
8088         ret2 = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
8089                                            reserved_space);
8090         if (!ret2) {
8091                 ret2 = file_update_time(vmf->vma->vm_file);
8092                 reserved = 1;
8093         }
8094         if (ret2) {
8095                 ret = vmf_error(ret2);
8096                 if (reserved)
8097                         goto out;
8098                 goto out_noreserve;
8099         }
8100
8101         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8102 again:
8103         lock_page(page);
8104         size = i_size_read(inode);
8105
8106         if ((page->mapping != inode->i_mapping) ||
8107             (page_start >= size)) {
8108                 /* page got truncated out from underneath us */
8109                 goto out_unlock;
8110         }
8111         wait_on_page_writeback(page);
8112
8113         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8114         set_page_extent_mapped(page);
8115
8116         /*
8117          * we can't set the delalloc bits if there are pending ordered
8118          * extents.  Drop our locks and wait for them to finish
8119          */
8120         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
8121                         PAGE_SIZE);
8122         if (ordered) {
8123                 unlock_extent_cached(io_tree, page_start, page_end,
8124                                      &cached_state);
8125                 unlock_page(page);
8126                 btrfs_start_ordered_extent(inode, ordered, 1);
8127                 btrfs_put_ordered_extent(ordered);
8128                 goto again;
8129         }
8130
8131         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8132                 reserved_space = round_up(size - page_start,
8133                                           fs_info->sectorsize);
8134                 if (reserved_space < PAGE_SIZE) {
8135                         end = page_start + reserved_space - 1;
8136                         btrfs_delalloc_release_space(inode, data_reserved,
8137                                         page_start, PAGE_SIZE - reserved_space,
8138                                         true);
8139                 }
8140         }
8141
8142         /*
8143          * page_mkwrite gets called when the page is firstly dirtied after it's
8144          * faulted in, but write(2) could also dirty a page and set delalloc
8145          * bits, thus in this case for space account reason, we still need to
8146          * clear any delalloc bits within this page range since we have to
8147          * reserve data&meta space before lock_page() (see above comments).
8148          */
8149         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
8150                           EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8151                           EXTENT_DEFRAG, 0, 0, &cached_state);
8152
8153         ret2 = btrfs_set_extent_delalloc(inode, page_start, end, 0,
8154                                         &cached_state);
8155         if (ret2) {
8156                 unlock_extent_cached(io_tree, page_start, page_end,
8157                                      &cached_state);
8158                 ret = VM_FAULT_SIGBUS;
8159                 goto out_unlock;
8160         }
8161
8162         /* page is wholly or partially inside EOF */
8163         if (page_start + PAGE_SIZE > size)
8164                 zero_start = offset_in_page(size);
8165         else
8166                 zero_start = PAGE_SIZE;
8167
8168         if (zero_start != PAGE_SIZE) {
8169                 kaddr = kmap(page);
8170                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
8171                 flush_dcache_page(page);
8172                 kunmap(page);
8173         }
8174         ClearPageChecked(page);
8175         set_page_dirty(page);
8176         SetPageUptodate(page);
8177
8178         BTRFS_I(inode)->last_trans = fs_info->generation;
8179         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8180         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8181
8182         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
8183
8184         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
8185         sb_end_pagefault(inode->i_sb);
8186         extent_changeset_free(data_reserved);
8187         return VM_FAULT_LOCKED;
8188
8189 out_unlock:
8190         unlock_page(page);
8191 out:
8192         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
8193         btrfs_delalloc_release_space(inode, data_reserved, page_start,
8194                                      reserved_space, (ret != 0));
8195 out_noreserve:
8196         sb_end_pagefault(inode->i_sb);
8197         extent_changeset_free(data_reserved);
8198         return ret;
8199 }
8200
8201 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
8202 {
8203         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8204         struct btrfs_root *root = BTRFS_I(inode)->root;
8205         struct btrfs_block_rsv *rsv;
8206         int ret;
8207         struct btrfs_trans_handle *trans;
8208         u64 mask = fs_info->sectorsize - 1;
8209         u64 min_size = btrfs_calc_metadata_size(fs_info, 1);
8210
8211         if (!skip_writeback) {
8212                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8213                                                (u64)-1);
8214                 if (ret)
8215                         return ret;
8216         }
8217
8218         /*
8219          * Yes ladies and gentlemen, this is indeed ugly.  We have a couple of
8220          * things going on here:
8221          *
8222          * 1) We need to reserve space to update our inode.
8223          *
8224          * 2) We need to have something to cache all the space that is going to
8225          * be free'd up by the truncate operation, but also have some slack
8226          * space reserved in case it uses space during the truncate (thank you
8227          * very much snapshotting).
8228          *
8229          * And we need these to be separate.  The fact is we can use a lot of
8230          * space doing the truncate, and we have no earthly idea how much space
8231          * we will use, so we need the truncate reservation to be separate so it
8232          * doesn't end up using space reserved for updating the inode.  We also
8233          * need to be able to stop the transaction and start a new one, which
8234          * means we need to be able to update the inode several times, and we
8235          * have no idea of knowing how many times that will be, so we can't just
8236          * reserve 1 item for the entirety of the operation, so that has to be
8237          * done separately as well.
8238          *
8239          * So that leaves us with
8240          *
8241          * 1) rsv - for the truncate reservation, which we will steal from the
8242          * transaction reservation.
8243          * 2) fs_info->trans_block_rsv - this will have 1 items worth left for
8244          * updating the inode.
8245          */
8246         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
8247         if (!rsv)
8248                 return -ENOMEM;
8249         rsv->size = min_size;
8250         rsv->failfast = 1;
8251
8252         /*
8253          * 1 for the truncate slack space
8254          * 1 for updating the inode.
8255          */
8256         trans = btrfs_start_transaction(root, 2);
8257         if (IS_ERR(trans)) {
8258                 ret = PTR_ERR(trans);
8259                 goto out;
8260         }
8261
8262         /* Migrate the slack space for the truncate to our reserve */
8263         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
8264                                       min_size, false);
8265         BUG_ON(ret);
8266
8267         /*
8268          * So if we truncate and then write and fsync we normally would just
8269          * write the extents that changed, which is a problem if we need to
8270          * first truncate that entire inode.  So set this flag so we write out
8271          * all of the extents in the inode to the sync log so we're completely
8272          * safe.
8273          */
8274         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
8275         trans->block_rsv = rsv;
8276
8277         while (1) {
8278                 ret = btrfs_truncate_inode_items(trans, root, inode,
8279                                                  inode->i_size,
8280                                                  BTRFS_EXTENT_DATA_KEY);
8281                 trans->block_rsv = &fs_info->trans_block_rsv;
8282                 if (ret != -ENOSPC && ret != -EAGAIN)
8283                         break;
8284
8285                 ret = btrfs_update_inode(trans, root, inode);
8286                 if (ret)
8287                         break;
8288
8289                 btrfs_end_transaction(trans);
8290                 btrfs_btree_balance_dirty(fs_info);
8291
8292                 trans = btrfs_start_transaction(root, 2);
8293                 if (IS_ERR(trans)) {
8294                         ret = PTR_ERR(trans);
8295                         trans = NULL;
8296                         break;
8297                 }
8298
8299                 btrfs_block_rsv_release(fs_info, rsv, -1, NULL);
8300                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
8301                                               rsv, min_size, false);
8302                 BUG_ON(ret);    /* shouldn't happen */
8303                 trans->block_rsv = rsv;
8304         }
8305
8306         /*
8307          * We can't call btrfs_truncate_block inside a trans handle as we could
8308          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
8309          * we've truncated everything except the last little bit, and can do
8310          * btrfs_truncate_block and then update the disk_i_size.
8311          */
8312         if (ret == NEED_TRUNCATE_BLOCK) {
8313                 btrfs_end_transaction(trans);
8314                 btrfs_btree_balance_dirty(fs_info);
8315
8316                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
8317                 if (ret)
8318                         goto out;
8319                 trans = btrfs_start_transaction(root, 1);
8320                 if (IS_ERR(trans)) {
8321                         ret = PTR_ERR(trans);
8322                         goto out;
8323                 }
8324                 btrfs_inode_safe_disk_i_size_write(inode, 0);
8325         }
8326
8327         if (trans) {
8328                 int ret2;
8329
8330                 trans->block_rsv = &fs_info->trans_block_rsv;
8331                 ret2 = btrfs_update_inode(trans, root, inode);
8332                 if (ret2 && !ret)
8333                         ret = ret2;
8334
8335                 ret2 = btrfs_end_transaction(trans);
8336                 if (ret2 && !ret)
8337                         ret = ret2;
8338                 btrfs_btree_balance_dirty(fs_info);
8339         }
8340 out:
8341         btrfs_free_block_rsv(fs_info, rsv);
8342
8343         return ret;
8344 }
8345
8346 /*
8347  * create a new subvolume directory/inode (helper for the ioctl).
8348  */
8349 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
8350                              struct btrfs_root *new_root,
8351                              struct btrfs_root *parent_root,
8352                              u64 new_dirid)
8353 {
8354         struct inode *inode;
8355         int err;
8356         u64 index = 0;
8357
8358         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
8359                                 new_dirid, new_dirid,
8360                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
8361                                 &index);
8362         if (IS_ERR(inode))
8363                 return PTR_ERR(inode);
8364         inode->i_op = &btrfs_dir_inode_operations;
8365         inode->i_fop = &btrfs_dir_file_operations;
8366
8367         set_nlink(inode, 1);
8368         btrfs_i_size_write(BTRFS_I(inode), 0);
8369         unlock_new_inode(inode);
8370
8371         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
8372         if (err)
8373                 btrfs_err(new_root->fs_info,
8374                           "error inheriting subvolume %llu properties: %d",
8375                           new_root->root_key.objectid, err);
8376
8377         err = btrfs_update_inode(trans, new_root, inode);
8378
8379         iput(inode);
8380         return err;
8381 }
8382
8383 struct inode *btrfs_alloc_inode(struct super_block *sb)
8384 {
8385         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
8386         struct btrfs_inode *ei;
8387         struct inode *inode;
8388
8389         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
8390         if (!ei)
8391                 return NULL;
8392
8393         ei->root = NULL;
8394         ei->generation = 0;
8395         ei->last_trans = 0;
8396         ei->last_sub_trans = 0;
8397         ei->logged_trans = 0;
8398         ei->delalloc_bytes = 0;
8399         ei->new_delalloc_bytes = 0;
8400         ei->defrag_bytes = 0;
8401         ei->disk_i_size = 0;
8402         ei->flags = 0;
8403         ei->csum_bytes = 0;
8404         ei->index_cnt = (u64)-1;
8405         ei->dir_index = 0;
8406         ei->last_unlink_trans = 0;
8407         ei->last_log_commit = 0;
8408
8409         spin_lock_init(&ei->lock);
8410         ei->outstanding_extents = 0;
8411         if (sb->s_magic != BTRFS_TEST_MAGIC)
8412                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
8413                                               BTRFS_BLOCK_RSV_DELALLOC);
8414         ei->runtime_flags = 0;
8415         ei->prop_compress = BTRFS_COMPRESS_NONE;
8416         ei->defrag_compress = BTRFS_COMPRESS_NONE;
8417
8418         ei->delayed_node = NULL;
8419
8420         ei->i_otime.tv_sec = 0;
8421         ei->i_otime.tv_nsec = 0;
8422
8423         inode = &ei->vfs_inode;
8424         extent_map_tree_init(&ei->extent_tree);
8425         extent_io_tree_init(fs_info, &ei->io_tree, IO_TREE_INODE_IO, inode);
8426         extent_io_tree_init(fs_info, &ei->io_failure_tree,
8427                             IO_TREE_INODE_IO_FAILURE, inode);
8428         extent_io_tree_init(fs_info, &ei->file_extent_tree,
8429                             IO_TREE_INODE_FILE_EXTENT, inode);
8430         ei->io_tree.track_uptodate = true;
8431         ei->io_failure_tree.track_uptodate = true;
8432         atomic_set(&ei->sync_writers, 0);
8433         mutex_init(&ei->log_mutex);
8434         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
8435         INIT_LIST_HEAD(&ei->delalloc_inodes);
8436         INIT_LIST_HEAD(&ei->delayed_iput);
8437         RB_CLEAR_NODE(&ei->rb_node);
8438         init_rwsem(&ei->dio_sem);
8439
8440         return inode;
8441 }
8442
8443 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8444 void btrfs_test_destroy_inode(struct inode *inode)
8445 {
8446         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
8447         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
8448 }
8449 #endif
8450
8451 void btrfs_free_inode(struct inode *inode)
8452 {
8453         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
8454 }
8455
8456 void btrfs_destroy_inode(struct inode *inode)
8457 {
8458         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8459         struct btrfs_ordered_extent *ordered;
8460         struct btrfs_root *root = BTRFS_I(inode)->root;
8461
8462         WARN_ON(!hlist_empty(&inode->i_dentry));
8463         WARN_ON(inode->i_data.nrpages);
8464         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
8465         WARN_ON(BTRFS_I(inode)->block_rsv.size);
8466         WARN_ON(BTRFS_I(inode)->outstanding_extents);
8467         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
8468         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
8469         WARN_ON(BTRFS_I(inode)->csum_bytes);
8470         WARN_ON(BTRFS_I(inode)->defrag_bytes);
8471
8472         /*
8473          * This can happen where we create an inode, but somebody else also
8474          * created the same inode and we need to destroy the one we already
8475          * created.
8476          */
8477         if (!root)
8478                 return;
8479
8480         while (1) {
8481                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
8482                 if (!ordered)
8483                         break;
8484                 else {
8485                         btrfs_err(fs_info,
8486                                   "found ordered extent %llu %llu on inode cleanup",
8487                                   ordered->file_offset, ordered->num_bytes);
8488                         btrfs_remove_ordered_extent(inode, ordered);
8489                         btrfs_put_ordered_extent(ordered);
8490                         btrfs_put_ordered_extent(ordered);
8491                 }
8492         }
8493         btrfs_qgroup_check_reserved_leak(inode);
8494         inode_tree_del(inode);
8495         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
8496         btrfs_inode_clear_file_extent_range(BTRFS_I(inode), 0, (u64)-1);
8497         btrfs_put_root(BTRFS_I(inode)->root);
8498 }
8499
8500 int btrfs_drop_inode(struct inode *inode)
8501 {
8502         struct btrfs_root *root = BTRFS_I(inode)->root;
8503
8504         if (root == NULL)
8505                 return 1;
8506
8507         /* the snap/subvol tree is on deleting */
8508         if (btrfs_root_refs(&root->root_item) == 0)
8509                 return 1;
8510         else
8511                 return generic_drop_inode(inode);
8512 }
8513
8514 static void init_once(void *foo)
8515 {
8516         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
8517
8518         inode_init_once(&ei->vfs_inode);
8519 }
8520
8521 void __cold btrfs_destroy_cachep(void)
8522 {
8523         /*
8524          * Make sure all delayed rcu free inodes are flushed before we
8525          * destroy cache.
8526          */
8527         rcu_barrier();
8528         kmem_cache_destroy(btrfs_inode_cachep);
8529         kmem_cache_destroy(btrfs_trans_handle_cachep);
8530         kmem_cache_destroy(btrfs_path_cachep);
8531         kmem_cache_destroy(btrfs_free_space_cachep);
8532         kmem_cache_destroy(btrfs_free_space_bitmap_cachep);
8533 }
8534
8535 int __init btrfs_init_cachep(void)
8536 {
8537         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
8538                         sizeof(struct btrfs_inode), 0,
8539                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
8540                         init_once);
8541         if (!btrfs_inode_cachep)
8542                 goto fail;
8543
8544         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
8545                         sizeof(struct btrfs_trans_handle), 0,
8546                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
8547         if (!btrfs_trans_handle_cachep)
8548                 goto fail;
8549
8550         btrfs_path_cachep = kmem_cache_create("btrfs_path",
8551                         sizeof(struct btrfs_path), 0,
8552                         SLAB_MEM_SPREAD, NULL);
8553         if (!btrfs_path_cachep)
8554                 goto fail;
8555
8556         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
8557                         sizeof(struct btrfs_free_space), 0,
8558                         SLAB_MEM_SPREAD, NULL);
8559         if (!btrfs_free_space_cachep)
8560                 goto fail;
8561
8562         btrfs_free_space_bitmap_cachep = kmem_cache_create("btrfs_free_space_bitmap",
8563                                                         PAGE_SIZE, PAGE_SIZE,
8564                                                         SLAB_RED_ZONE, NULL);
8565         if (!btrfs_free_space_bitmap_cachep)
8566                 goto fail;
8567
8568         return 0;
8569 fail:
8570         btrfs_destroy_cachep();
8571         return -ENOMEM;
8572 }
8573
8574 static int btrfs_getattr(const struct path *path, struct kstat *stat,
8575                          u32 request_mask, unsigned int flags)
8576 {
8577         u64 delalloc_bytes;
8578         struct inode *inode = d_inode(path->dentry);
8579         u32 blocksize = inode->i_sb->s_blocksize;
8580         u32 bi_flags = BTRFS_I(inode)->flags;
8581
8582         stat->result_mask |= STATX_BTIME;
8583         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
8584         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
8585         if (bi_flags & BTRFS_INODE_APPEND)
8586                 stat->attributes |= STATX_ATTR_APPEND;
8587         if (bi_flags & BTRFS_INODE_COMPRESS)
8588                 stat->attributes |= STATX_ATTR_COMPRESSED;
8589         if (bi_flags & BTRFS_INODE_IMMUTABLE)
8590                 stat->attributes |= STATX_ATTR_IMMUTABLE;
8591         if (bi_flags & BTRFS_INODE_NODUMP)
8592                 stat->attributes |= STATX_ATTR_NODUMP;
8593
8594         stat->attributes_mask |= (STATX_ATTR_APPEND |
8595                                   STATX_ATTR_COMPRESSED |
8596                                   STATX_ATTR_IMMUTABLE |
8597                                   STATX_ATTR_NODUMP);
8598
8599         generic_fillattr(inode, stat);
8600         stat->dev = BTRFS_I(inode)->root->anon_dev;
8601
8602         spin_lock(&BTRFS_I(inode)->lock);
8603         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
8604         spin_unlock(&BTRFS_I(inode)->lock);
8605         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
8606                         ALIGN(delalloc_bytes, blocksize)) >> 9;
8607         return 0;
8608 }
8609
8610 static int btrfs_rename_exchange(struct inode *old_dir,
8611                               struct dentry *old_dentry,
8612                               struct inode *new_dir,
8613                               struct dentry *new_dentry)
8614 {
8615         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
8616         struct btrfs_trans_handle *trans;
8617         struct btrfs_root *root = BTRFS_I(old_dir)->root;
8618         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
8619         struct inode *new_inode = new_dentry->d_inode;
8620         struct inode *old_inode = old_dentry->d_inode;
8621         struct timespec64 ctime = current_time(old_inode);
8622         struct dentry *parent;
8623         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
8624         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
8625         u64 old_idx = 0;
8626         u64 new_idx = 0;
8627         int ret;
8628         bool root_log_pinned = false;
8629         bool dest_log_pinned = false;
8630         struct btrfs_log_ctx ctx_root;
8631         struct btrfs_log_ctx ctx_dest;
8632         bool sync_log_root = false;
8633         bool sync_log_dest = false;
8634         bool commit_transaction = false;
8635
8636         /* we only allow rename subvolume link between subvolumes */
8637         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
8638                 return -EXDEV;
8639
8640         btrfs_init_log_ctx(&ctx_root, old_inode);
8641         btrfs_init_log_ctx(&ctx_dest, new_inode);
8642
8643         /* close the race window with snapshot create/destroy ioctl */
8644         if (old_ino == BTRFS_FIRST_FREE_OBJECTID ||
8645             new_ino == BTRFS_FIRST_FREE_OBJECTID)
8646                 down_read(&fs_info->subvol_sem);
8647
8648         /*
8649          * We want to reserve the absolute worst case amount of items.  So if
8650          * both inodes are subvols and we need to unlink them then that would
8651          * require 4 item modifications, but if they are both normal inodes it
8652          * would require 5 item modifications, so we'll assume their normal
8653          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
8654          * should cover the worst case number of items we'll modify.
8655          */
8656         trans = btrfs_start_transaction(root, 12);
8657         if (IS_ERR(trans)) {
8658                 ret = PTR_ERR(trans);
8659                 goto out_notrans;
8660         }
8661
8662         if (dest != root)
8663                 btrfs_record_root_in_trans(trans, dest);
8664
8665         /*
8666          * We need to find a free sequence number both in the source and
8667          * in the destination directory for the exchange.
8668          */
8669         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
8670         if (ret)
8671                 goto out_fail;
8672         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
8673         if (ret)
8674                 goto out_fail;
8675
8676         BTRFS_I(old_inode)->dir_index = 0ULL;
8677         BTRFS_I(new_inode)->dir_index = 0ULL;
8678
8679         /* Reference for the source. */
8680         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
8681                 /* force full log commit if subvolume involved. */
8682                 btrfs_set_log_full_commit(trans);
8683         } else {
8684                 btrfs_pin_log_trans(root);
8685                 root_log_pinned = true;
8686                 ret = btrfs_insert_inode_ref(trans, dest,
8687                                              new_dentry->d_name.name,
8688                                              new_dentry->d_name.len,
8689                                              old_ino,
8690                                              btrfs_ino(BTRFS_I(new_dir)),
8691                                              old_idx);
8692                 if (ret)
8693                         goto out_fail;
8694         }
8695
8696         /* And now for the dest. */
8697         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
8698                 /* force full log commit if subvolume involved. */
8699                 btrfs_set_log_full_commit(trans);
8700         } else {
8701                 btrfs_pin_log_trans(dest);
8702                 dest_log_pinned = true;
8703                 ret = btrfs_insert_inode_ref(trans, root,
8704                                              old_dentry->d_name.name,
8705                                              old_dentry->d_name.len,
8706                                              new_ino,
8707                                              btrfs_ino(BTRFS_I(old_dir)),
8708                                              new_idx);
8709                 if (ret)
8710                         goto out_fail;
8711         }
8712
8713         /* Update inode version and ctime/mtime. */
8714         inode_inc_iversion(old_dir);
8715         inode_inc_iversion(new_dir);
8716         inode_inc_iversion(old_inode);
8717         inode_inc_iversion(new_inode);
8718         old_dir->i_ctime = old_dir->i_mtime = ctime;
8719         new_dir->i_ctime = new_dir->i_mtime = ctime;
8720         old_inode->i_ctime = ctime;
8721         new_inode->i_ctime = ctime;
8722
8723         if (old_dentry->d_parent != new_dentry->d_parent) {
8724                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
8725                                 BTRFS_I(old_inode), 1);
8726                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
8727                                 BTRFS_I(new_inode), 1);
8728         }
8729
8730         /* src is a subvolume */
8731         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
8732                 ret = btrfs_unlink_subvol(trans, old_dir, old_dentry);
8733         } else { /* src is an inode */
8734                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
8735                                            BTRFS_I(old_dentry->d_inode),
8736                                            old_dentry->d_name.name,
8737                                            old_dentry->d_name.len);
8738                 if (!ret)
8739                         ret = btrfs_update_inode(trans, root, old_inode);
8740         }
8741         if (ret) {
8742                 btrfs_abort_transaction(trans, ret);
8743                 goto out_fail;
8744         }
8745
8746         /* dest is a subvolume */
8747         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
8748                 ret = btrfs_unlink_subvol(trans, new_dir, new_dentry);
8749         } else { /* dest is an inode */
8750                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
8751                                            BTRFS_I(new_dentry->d_inode),
8752                                            new_dentry->d_name.name,
8753                                            new_dentry->d_name.len);
8754                 if (!ret)
8755                         ret = btrfs_update_inode(trans, dest, new_inode);
8756         }
8757         if (ret) {
8758                 btrfs_abort_transaction(trans, ret);
8759                 goto out_fail;
8760         }
8761
8762         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
8763                              new_dentry->d_name.name,
8764                              new_dentry->d_name.len, 0, old_idx);
8765         if (ret) {
8766                 btrfs_abort_transaction(trans, ret);
8767                 goto out_fail;
8768         }
8769
8770         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
8771                              old_dentry->d_name.name,
8772                              old_dentry->d_name.len, 0, new_idx);
8773         if (ret) {
8774                 btrfs_abort_transaction(trans, ret);
8775                 goto out_fail;
8776         }
8777
8778         if (old_inode->i_nlink == 1)
8779                 BTRFS_I(old_inode)->dir_index = old_idx;
8780         if (new_inode->i_nlink == 1)
8781                 BTRFS_I(new_inode)->dir_index = new_idx;
8782
8783         if (root_log_pinned) {
8784                 parent = new_dentry->d_parent;
8785                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
8786                                          BTRFS_I(old_dir), parent,
8787                                          false, &ctx_root);
8788                 if (ret == BTRFS_NEED_LOG_SYNC)
8789                         sync_log_root = true;
8790                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
8791                         commit_transaction = true;
8792                 ret = 0;
8793                 btrfs_end_log_trans(root);
8794                 root_log_pinned = false;
8795         }
8796         if (dest_log_pinned) {
8797                 if (!commit_transaction) {
8798                         parent = old_dentry->d_parent;
8799                         ret = btrfs_log_new_name(trans, BTRFS_I(new_inode),
8800                                                  BTRFS_I(new_dir), parent,
8801                                                  false, &ctx_dest);
8802                         if (ret == BTRFS_NEED_LOG_SYNC)
8803                                 sync_log_dest = true;
8804                         else if (ret == BTRFS_NEED_TRANS_COMMIT)
8805                                 commit_transaction = true;
8806                         ret = 0;
8807                 }
8808                 btrfs_end_log_trans(dest);
8809                 dest_log_pinned = false;
8810         }
8811 out_fail:
8812         /*
8813          * If we have pinned a log and an error happened, we unpin tasks
8814          * trying to sync the log and force them to fallback to a transaction
8815          * commit if the log currently contains any of the inodes involved in
8816          * this rename operation (to ensure we do not persist a log with an
8817          * inconsistent state for any of these inodes or leading to any
8818          * inconsistencies when replayed). If the transaction was aborted, the
8819          * abortion reason is propagated to userspace when attempting to commit
8820          * the transaction. If the log does not contain any of these inodes, we
8821          * allow the tasks to sync it.
8822          */
8823         if (ret && (root_log_pinned || dest_log_pinned)) {
8824                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
8825                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
8826                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
8827                     (new_inode &&
8828                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
8829                         btrfs_set_log_full_commit(trans);
8830
8831                 if (root_log_pinned) {
8832                         btrfs_end_log_trans(root);
8833                         root_log_pinned = false;
8834                 }
8835                 if (dest_log_pinned) {
8836                         btrfs_end_log_trans(dest);
8837                         dest_log_pinned = false;
8838                 }
8839         }
8840         if (!ret && sync_log_root && !commit_transaction) {
8841                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root,
8842                                      &ctx_root);
8843                 if (ret)
8844                         commit_transaction = true;
8845         }
8846         if (!ret && sync_log_dest && !commit_transaction) {
8847                 ret = btrfs_sync_log(trans, BTRFS_I(new_inode)->root,
8848                                      &ctx_dest);
8849                 if (ret)
8850                         commit_transaction = true;
8851         }
8852         if (commit_transaction) {
8853                 /*
8854                  * We may have set commit_transaction when logging the new name
8855                  * in the destination root, in which case we left the source
8856                  * root context in the list of log contextes. So make sure we
8857                  * remove it to avoid invalid memory accesses, since the context
8858                  * was allocated in our stack frame.
8859                  */
8860                 if (sync_log_root) {
8861                         mutex_lock(&root->log_mutex);
8862                         list_del_init(&ctx_root.list);
8863                         mutex_unlock(&root->log_mutex);
8864                 }
8865                 ret = btrfs_commit_transaction(trans);
8866         } else {
8867                 int ret2;
8868
8869                 ret2 = btrfs_end_transaction(trans);
8870                 ret = ret ? ret : ret2;
8871         }
8872 out_notrans:
8873         if (new_ino == BTRFS_FIRST_FREE_OBJECTID ||
8874             old_ino == BTRFS_FIRST_FREE_OBJECTID)
8875                 up_read(&fs_info->subvol_sem);
8876
8877         ASSERT(list_empty(&ctx_root.list));
8878         ASSERT(list_empty(&ctx_dest.list));
8879
8880         return ret;
8881 }
8882
8883 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
8884                                      struct btrfs_root *root,
8885                                      struct inode *dir,
8886                                      struct dentry *dentry)
8887 {
8888         int ret;
8889         struct inode *inode;
8890         u64 objectid;
8891         u64 index;
8892
8893         ret = btrfs_find_free_ino(root, &objectid);
8894         if (ret)
8895                 return ret;
8896
8897         inode = btrfs_new_inode(trans, root, dir,
8898                                 dentry->d_name.name,
8899                                 dentry->d_name.len,
8900                                 btrfs_ino(BTRFS_I(dir)),
8901                                 objectid,
8902                                 S_IFCHR | WHITEOUT_MODE,
8903                                 &index);
8904
8905         if (IS_ERR(inode)) {
8906                 ret = PTR_ERR(inode);
8907                 return ret;
8908         }
8909
8910         inode->i_op = &btrfs_special_inode_operations;
8911         init_special_inode(inode, inode->i_mode,
8912                 WHITEOUT_DEV);
8913
8914         ret = btrfs_init_inode_security(trans, inode, dir,
8915                                 &dentry->d_name);
8916         if (ret)
8917                 goto out;
8918
8919         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
8920                                 BTRFS_I(inode), 0, index);
8921         if (ret)
8922                 goto out;
8923
8924         ret = btrfs_update_inode(trans, root, inode);
8925 out:
8926         unlock_new_inode(inode);
8927         if (ret)
8928                 inode_dec_link_count(inode);
8929         iput(inode);
8930
8931         return ret;
8932 }
8933
8934 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
8935                            struct inode *new_dir, struct dentry *new_dentry,
8936                            unsigned int flags)
8937 {
8938         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
8939         struct btrfs_trans_handle *trans;
8940         unsigned int trans_num_items;
8941         struct btrfs_root *root = BTRFS_I(old_dir)->root;
8942         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
8943         struct inode *new_inode = d_inode(new_dentry);
8944         struct inode *old_inode = d_inode(old_dentry);
8945         u64 index = 0;
8946         int ret;
8947         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
8948         bool log_pinned = false;
8949         struct btrfs_log_ctx ctx;
8950         bool sync_log = false;
8951         bool commit_transaction = false;
8952
8953         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
8954                 return -EPERM;
8955
8956         /* we only allow rename subvolume link between subvolumes */
8957         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
8958                 return -EXDEV;
8959
8960         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
8961             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
8962                 return -ENOTEMPTY;
8963
8964         if (S_ISDIR(old_inode->i_mode) && new_inode &&
8965             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
8966                 return -ENOTEMPTY;
8967
8968
8969         /* check for collisions, even if the  name isn't there */
8970         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
8971                              new_dentry->d_name.name,
8972                              new_dentry->d_name.len);
8973
8974         if (ret) {
8975                 if (ret == -EEXIST) {
8976                         /* we shouldn't get
8977                          * eexist without a new_inode */
8978                         if (WARN_ON(!new_inode)) {
8979                                 return ret;
8980                         }
8981                 } else {
8982                         /* maybe -EOVERFLOW */
8983                         return ret;
8984                 }
8985         }
8986         ret = 0;
8987
8988         /*
8989          * we're using rename to replace one file with another.  Start IO on it
8990          * now so  we don't add too much work to the end of the transaction
8991          */
8992         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
8993                 filemap_flush(old_inode->i_mapping);
8994
8995         /* close the racy window with snapshot create/destroy ioctl */
8996         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
8997                 down_read(&fs_info->subvol_sem);
8998         /*
8999          * We want to reserve the absolute worst case amount of items.  So if
9000          * both inodes are subvols and we need to unlink them then that would
9001          * require 4 item modifications, but if they are both normal inodes it
9002          * would require 5 item modifications, so we'll assume they are normal
9003          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9004          * should cover the worst case number of items we'll modify.
9005          * If our rename has the whiteout flag, we need more 5 units for the
9006          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9007          * when selinux is enabled).
9008          */
9009         trans_num_items = 11;
9010         if (flags & RENAME_WHITEOUT)
9011                 trans_num_items += 5;
9012         trans = btrfs_start_transaction(root, trans_num_items);
9013         if (IS_ERR(trans)) {
9014                 ret = PTR_ERR(trans);
9015                 goto out_notrans;
9016         }
9017
9018         if (dest != root)
9019                 btrfs_record_root_in_trans(trans, dest);
9020
9021         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9022         if (ret)
9023                 goto out_fail;
9024
9025         BTRFS_I(old_inode)->dir_index = 0ULL;
9026         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9027                 /* force full log commit if subvolume involved. */
9028                 btrfs_set_log_full_commit(trans);
9029         } else {
9030                 btrfs_pin_log_trans(root);
9031                 log_pinned = true;
9032                 ret = btrfs_insert_inode_ref(trans, dest,
9033                                              new_dentry->d_name.name,
9034                                              new_dentry->d_name.len,
9035                                              old_ino,
9036                                              btrfs_ino(BTRFS_I(new_dir)), index);
9037                 if (ret)
9038                         goto out_fail;
9039         }
9040
9041         inode_inc_iversion(old_dir);
9042         inode_inc_iversion(new_dir);
9043         inode_inc_iversion(old_inode);
9044         old_dir->i_ctime = old_dir->i_mtime =
9045         new_dir->i_ctime = new_dir->i_mtime =
9046         old_inode->i_ctime = current_time(old_dir);
9047
9048         if (old_dentry->d_parent != new_dentry->d_parent)
9049                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9050                                 BTRFS_I(old_inode), 1);
9051
9052         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9053                 ret = btrfs_unlink_subvol(trans, old_dir, old_dentry);
9054         } else {
9055                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9056                                         BTRFS_I(d_inode(old_dentry)),
9057                                         old_dentry->d_name.name,
9058                                         old_dentry->d_name.len);
9059                 if (!ret)
9060                         ret = btrfs_update_inode(trans, root, old_inode);
9061         }
9062         if (ret) {
9063                 btrfs_abort_transaction(trans, ret);
9064                 goto out_fail;
9065         }
9066
9067         if (new_inode) {
9068                 inode_inc_iversion(new_inode);
9069                 new_inode->i_ctime = current_time(new_inode);
9070                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9071                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9072                         ret = btrfs_unlink_subvol(trans, new_dir, new_dentry);
9073                         BUG_ON(new_inode->i_nlink == 0);
9074                 } else {
9075                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9076                                                  BTRFS_I(d_inode(new_dentry)),
9077                                                  new_dentry->d_name.name,
9078                                                  new_dentry->d_name.len);
9079                 }
9080                 if (!ret && new_inode->i_nlink == 0)
9081                         ret = btrfs_orphan_add(trans,
9082                                         BTRFS_I(d_inode(new_dentry)));
9083                 if (ret) {
9084                         btrfs_abort_transaction(trans, ret);
9085                         goto out_fail;
9086                 }
9087         }
9088
9089         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9090                              new_dentry->d_name.name,
9091                              new_dentry->d_name.len, 0, index);
9092         if (ret) {
9093                 btrfs_abort_transaction(trans, ret);
9094                 goto out_fail;
9095         }
9096
9097         if (old_inode->i_nlink == 1)
9098                 BTRFS_I(old_inode)->dir_index = index;
9099
9100         if (log_pinned) {
9101                 struct dentry *parent = new_dentry->d_parent;
9102
9103                 btrfs_init_log_ctx(&ctx, old_inode);
9104                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9105                                          BTRFS_I(old_dir), parent,
9106                                          false, &ctx);
9107                 if (ret == BTRFS_NEED_LOG_SYNC)
9108                         sync_log = true;
9109                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9110                         commit_transaction = true;
9111                 ret = 0;
9112                 btrfs_end_log_trans(root);
9113                 log_pinned = false;
9114         }
9115
9116         if (flags & RENAME_WHITEOUT) {
9117                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9118                                                 old_dentry);
9119
9120                 if (ret) {
9121                         btrfs_abort_transaction(trans, ret);
9122                         goto out_fail;
9123                 }
9124         }
9125 out_fail:
9126         /*
9127          * If we have pinned the log and an error happened, we unpin tasks
9128          * trying to sync the log and force them to fallback to a transaction
9129          * commit if the log currently contains any of the inodes involved in
9130          * this rename operation (to ensure we do not persist a log with an
9131          * inconsistent state for any of these inodes or leading to any
9132          * inconsistencies when replayed). If the transaction was aborted, the
9133          * abortion reason is propagated to userspace when attempting to commit
9134          * the transaction. If the log does not contain any of these inodes, we
9135          * allow the tasks to sync it.
9136          */
9137         if (ret && log_pinned) {
9138                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9139                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9140                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9141                     (new_inode &&
9142                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9143                         btrfs_set_log_full_commit(trans);
9144
9145                 btrfs_end_log_trans(root);
9146                 log_pinned = false;
9147         }
9148         if (!ret && sync_log) {
9149                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root, &ctx);
9150                 if (ret)
9151                         commit_transaction = true;
9152         } else if (sync_log) {
9153                 mutex_lock(&root->log_mutex);
9154                 list_del(&ctx.list);
9155                 mutex_unlock(&root->log_mutex);
9156         }
9157         if (commit_transaction) {
9158                 ret = btrfs_commit_transaction(trans);
9159         } else {
9160                 int ret2;
9161
9162                 ret2 = btrfs_end_transaction(trans);
9163                 ret = ret ? ret : ret2;
9164         }
9165 out_notrans:
9166         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9167                 up_read(&fs_info->subvol_sem);
9168
9169         return ret;
9170 }
9171
9172 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9173                          struct inode *new_dir, struct dentry *new_dentry,
9174                          unsigned int flags)
9175 {
9176         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9177                 return -EINVAL;
9178
9179         if (flags & RENAME_EXCHANGE)
9180                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9181                                           new_dentry);
9182
9183         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9184 }
9185
9186 struct btrfs_delalloc_work {
9187         struct inode *inode;
9188         struct completion completion;
9189         struct list_head list;
9190         struct btrfs_work work;
9191 };
9192
9193 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9194 {
9195         struct btrfs_delalloc_work *delalloc_work;
9196         struct inode *inode;
9197
9198         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9199                                      work);
9200         inode = delalloc_work->inode;
9201         filemap_flush(inode->i_mapping);
9202         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9203                                 &BTRFS_I(inode)->runtime_flags))
9204                 filemap_flush(inode->i_mapping);
9205
9206         iput(inode);
9207         complete(&delalloc_work->completion);
9208 }
9209
9210 static struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode)
9211 {
9212         struct btrfs_delalloc_work *work;
9213
9214         work = kmalloc(sizeof(*work), GFP_NOFS);
9215         if (!work)
9216                 return NULL;
9217
9218         init_completion(&work->completion);
9219         INIT_LIST_HEAD(&work->list);
9220         work->inode = inode;
9221         btrfs_init_work(&work->work, btrfs_run_delalloc_work, NULL, NULL);
9222
9223         return work;
9224 }
9225
9226 /*
9227  * some fairly slow code that needs optimization. This walks the list
9228  * of all the inodes with pending delalloc and forces them to disk.
9229  */
9230 static int start_delalloc_inodes(struct btrfs_root *root, int nr, bool snapshot)
9231 {
9232         struct btrfs_inode *binode;
9233         struct inode *inode;
9234         struct btrfs_delalloc_work *work, *next;
9235         struct list_head works;
9236         struct list_head splice;
9237         int ret = 0;
9238
9239         INIT_LIST_HEAD(&works);
9240         INIT_LIST_HEAD(&splice);
9241
9242         mutex_lock(&root->delalloc_mutex);
9243         spin_lock(&root->delalloc_lock);
9244         list_splice_init(&root->delalloc_inodes, &splice);
9245         while (!list_empty(&splice)) {
9246                 binode = list_entry(splice.next, struct btrfs_inode,
9247                                     delalloc_inodes);
9248
9249                 list_move_tail(&binode->delalloc_inodes,
9250                                &root->delalloc_inodes);
9251                 inode = igrab(&binode->vfs_inode);
9252                 if (!inode) {
9253                         cond_resched_lock(&root->delalloc_lock);
9254                         continue;
9255                 }
9256                 spin_unlock(&root->delalloc_lock);
9257
9258                 if (snapshot)
9259                         set_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
9260                                 &binode->runtime_flags);
9261                 work = btrfs_alloc_delalloc_work(inode);
9262                 if (!work) {
9263                         iput(inode);
9264                         ret = -ENOMEM;
9265                         goto out;
9266                 }
9267                 list_add_tail(&work->list, &works);
9268                 btrfs_queue_work(root->fs_info->flush_workers,
9269                                  &work->work);
9270                 ret++;
9271                 if (nr != -1 && ret >= nr)
9272                         goto out;
9273                 cond_resched();
9274                 spin_lock(&root->delalloc_lock);
9275         }
9276         spin_unlock(&root->delalloc_lock);
9277
9278 out:
9279         list_for_each_entry_safe(work, next, &works, list) {
9280                 list_del_init(&work->list);
9281                 wait_for_completion(&work->completion);
9282                 kfree(work);
9283         }
9284
9285         if (!list_empty(&splice)) {
9286                 spin_lock(&root->delalloc_lock);
9287                 list_splice_tail(&splice, &root->delalloc_inodes);
9288                 spin_unlock(&root->delalloc_lock);
9289         }
9290         mutex_unlock(&root->delalloc_mutex);
9291         return ret;
9292 }
9293
9294 int btrfs_start_delalloc_snapshot(struct btrfs_root *root)
9295 {
9296         struct btrfs_fs_info *fs_info = root->fs_info;
9297         int ret;
9298
9299         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9300                 return -EROFS;
9301
9302         ret = start_delalloc_inodes(root, -1, true);
9303         if (ret > 0)
9304                 ret = 0;
9305         return ret;
9306 }
9307
9308 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int nr)
9309 {
9310         struct btrfs_root *root;
9311         struct list_head splice;
9312         int ret;
9313
9314         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9315                 return -EROFS;
9316
9317         INIT_LIST_HEAD(&splice);
9318
9319         mutex_lock(&fs_info->delalloc_root_mutex);
9320         spin_lock(&fs_info->delalloc_root_lock);
9321         list_splice_init(&fs_info->delalloc_roots, &splice);
9322         while (!list_empty(&splice) && nr) {
9323                 root = list_first_entry(&splice, struct btrfs_root,
9324                                         delalloc_root);
9325                 root = btrfs_grab_root(root);
9326                 BUG_ON(!root);
9327                 list_move_tail(&root->delalloc_root,
9328                                &fs_info->delalloc_roots);
9329                 spin_unlock(&fs_info->delalloc_root_lock);
9330
9331                 ret = start_delalloc_inodes(root, nr, false);
9332                 btrfs_put_root(root);
9333                 if (ret < 0)
9334                         goto out;
9335
9336                 if (nr != -1) {
9337                         nr -= ret;
9338                         WARN_ON(nr < 0);
9339                 }
9340                 spin_lock(&fs_info->delalloc_root_lock);
9341         }
9342         spin_unlock(&fs_info->delalloc_root_lock);
9343
9344         ret = 0;
9345 out:
9346         if (!list_empty(&splice)) {
9347                 spin_lock(&fs_info->delalloc_root_lock);
9348                 list_splice_tail(&splice, &fs_info->delalloc_roots);
9349                 spin_unlock(&fs_info->delalloc_root_lock);
9350         }
9351         mutex_unlock(&fs_info->delalloc_root_mutex);
9352         return ret;
9353 }
9354
9355 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
9356                          const char *symname)
9357 {
9358         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
9359         struct btrfs_trans_handle *trans;
9360         struct btrfs_root *root = BTRFS_I(dir)->root;
9361         struct btrfs_path *path;
9362         struct btrfs_key key;
9363         struct inode *inode = NULL;
9364         int err;
9365         u64 objectid;
9366         u64 index = 0;
9367         int name_len;
9368         int datasize;
9369         unsigned long ptr;
9370         struct btrfs_file_extent_item *ei;
9371         struct extent_buffer *leaf;
9372
9373         name_len = strlen(symname);
9374         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
9375                 return -ENAMETOOLONG;
9376
9377         /*
9378          * 2 items for inode item and ref
9379          * 2 items for dir items
9380          * 1 item for updating parent inode item
9381          * 1 item for the inline extent item
9382          * 1 item for xattr if selinux is on
9383          */
9384         trans = btrfs_start_transaction(root, 7);
9385         if (IS_ERR(trans))
9386                 return PTR_ERR(trans);
9387
9388         err = btrfs_find_free_ino(root, &objectid);
9389         if (err)
9390                 goto out_unlock;
9391
9392         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
9393                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
9394                                 objectid, S_IFLNK|S_IRWXUGO, &index);
9395         if (IS_ERR(inode)) {
9396                 err = PTR_ERR(inode);
9397                 inode = NULL;
9398                 goto out_unlock;
9399         }
9400
9401         /*
9402         * If the active LSM wants to access the inode during
9403         * d_instantiate it needs these. Smack checks to see
9404         * if the filesystem supports xattrs by looking at the
9405         * ops vector.
9406         */
9407         inode->i_fop = &btrfs_file_operations;
9408         inode->i_op = &btrfs_file_inode_operations;
9409         inode->i_mapping->a_ops = &btrfs_aops;
9410         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9411
9412         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
9413         if (err)
9414                 goto out_unlock;
9415
9416         path = btrfs_alloc_path();
9417         if (!path) {
9418                 err = -ENOMEM;
9419                 goto out_unlock;
9420         }
9421         key.objectid = btrfs_ino(BTRFS_I(inode));
9422         key.offset = 0;
9423         key.type = BTRFS_EXTENT_DATA_KEY;
9424         datasize = btrfs_file_extent_calc_inline_size(name_len);
9425         err = btrfs_insert_empty_item(trans, root, path, &key,
9426                                       datasize);
9427         if (err) {
9428                 btrfs_free_path(path);
9429                 goto out_unlock;
9430         }
9431         leaf = path->nodes[0];
9432         ei = btrfs_item_ptr(leaf, path->slots[0],
9433                             struct btrfs_file_extent_item);
9434         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
9435         btrfs_set_file_extent_type(leaf, ei,
9436                                    BTRFS_FILE_EXTENT_INLINE);
9437         btrfs_set_file_extent_encryption(leaf, ei, 0);
9438         btrfs_set_file_extent_compression(leaf, ei, 0);
9439         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
9440         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
9441
9442         ptr = btrfs_file_extent_inline_start(ei);
9443         write_extent_buffer(leaf, symname, ptr, name_len);
9444         btrfs_mark_buffer_dirty(leaf);
9445         btrfs_free_path(path);
9446
9447         inode->i_op = &btrfs_symlink_inode_operations;
9448         inode_nohighmem(inode);
9449         inode_set_bytes(inode, name_len);
9450         btrfs_i_size_write(BTRFS_I(inode), name_len);
9451         err = btrfs_update_inode(trans, root, inode);
9452         /*
9453          * Last step, add directory indexes for our symlink inode. This is the
9454          * last step to avoid extra cleanup of these indexes if an error happens
9455          * elsewhere above.
9456          */
9457         if (!err)
9458                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9459                                 BTRFS_I(inode), 0, index);
9460         if (err)
9461                 goto out_unlock;
9462
9463         d_instantiate_new(dentry, inode);
9464
9465 out_unlock:
9466         btrfs_end_transaction(trans);
9467         if (err && inode) {
9468                 inode_dec_link_count(inode);
9469                 discard_new_inode(inode);
9470         }
9471         btrfs_btree_balance_dirty(fs_info);
9472         return err;
9473 }
9474
9475 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
9476                                        u64 start, u64 num_bytes, u64 min_size,
9477                                        loff_t actual_len, u64 *alloc_hint,
9478                                        struct btrfs_trans_handle *trans)
9479 {
9480         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9481         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
9482         struct extent_map *em;
9483         struct btrfs_root *root = BTRFS_I(inode)->root;
9484         struct btrfs_key ins;
9485         u64 cur_offset = start;
9486         u64 clear_offset = start;
9487         u64 i_size;
9488         u64 cur_bytes;
9489         u64 last_alloc = (u64)-1;
9490         int ret = 0;
9491         bool own_trans = true;
9492         u64 end = start + num_bytes - 1;
9493
9494         if (trans)
9495                 own_trans = false;
9496         while (num_bytes > 0) {
9497                 if (own_trans) {
9498                         trans = btrfs_start_transaction(root, 3);
9499                         if (IS_ERR(trans)) {
9500                                 ret = PTR_ERR(trans);
9501                                 break;
9502                         }
9503                 }
9504
9505                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
9506                 cur_bytes = max(cur_bytes, min_size);
9507                 /*
9508                  * If we are severely fragmented we could end up with really
9509                  * small allocations, so if the allocator is returning small
9510                  * chunks lets make its job easier by only searching for those
9511                  * sized chunks.
9512                  */
9513                 cur_bytes = min(cur_bytes, last_alloc);
9514                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
9515                                 min_size, 0, *alloc_hint, &ins, 1, 0);
9516                 if (ret) {
9517                         if (own_trans)
9518                                 btrfs_end_transaction(trans);
9519                         break;
9520                 }
9521
9522                 /*
9523                  * We've reserved this space, and thus converted it from
9524                  * ->bytes_may_use to ->bytes_reserved.  Any error that happens
9525                  * from here on out we will only need to clear our reservation
9526                  * for the remaining unreserved area, so advance our
9527                  * clear_offset by our extent size.
9528                  */
9529                 clear_offset += ins.offset;
9530                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
9531
9532                 last_alloc = ins.offset;
9533                 ret = insert_reserved_file_extent(trans, inode,
9534                                                   cur_offset, ins.objectid,
9535                                                   ins.offset, ins.offset,
9536                                                   ins.offset, 0, 0, 0,
9537                                                   BTRFS_FILE_EXTENT_PREALLOC);
9538                 if (ret) {
9539                         btrfs_free_reserved_extent(fs_info, ins.objectid,
9540                                                    ins.offset, 0);
9541                         btrfs_abort_transaction(trans, ret);
9542                         if (own_trans)
9543                                 btrfs_end_transaction(trans);
9544                         break;
9545                 }
9546
9547                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
9548                                         cur_offset + ins.offset -1, 0);
9549
9550                 em = alloc_extent_map();
9551                 if (!em) {
9552                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
9553                                 &BTRFS_I(inode)->runtime_flags);
9554                         goto next;
9555                 }
9556
9557                 em->start = cur_offset;
9558                 em->orig_start = cur_offset;
9559                 em->len = ins.offset;
9560                 em->block_start = ins.objectid;
9561                 em->block_len = ins.offset;
9562                 em->orig_block_len = ins.offset;
9563                 em->ram_bytes = ins.offset;
9564                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
9565                 em->generation = trans->transid;
9566
9567                 while (1) {
9568                         write_lock(&em_tree->lock);
9569                         ret = add_extent_mapping(em_tree, em, 1);
9570                         write_unlock(&em_tree->lock);
9571                         if (ret != -EEXIST)
9572                                 break;
9573                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
9574                                                 cur_offset + ins.offset - 1,
9575                                                 0);
9576                 }
9577                 free_extent_map(em);
9578 next:
9579                 num_bytes -= ins.offset;
9580                 cur_offset += ins.offset;
9581                 *alloc_hint = ins.objectid + ins.offset;
9582
9583                 inode_inc_iversion(inode);
9584                 inode->i_ctime = current_time(inode);
9585                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
9586                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
9587                     (actual_len > inode->i_size) &&
9588                     (cur_offset > inode->i_size)) {
9589                         if (cur_offset > actual_len)
9590                                 i_size = actual_len;
9591                         else
9592                                 i_size = cur_offset;
9593                         i_size_write(inode, i_size);
9594                         btrfs_inode_safe_disk_i_size_write(inode, 0);
9595                 }
9596
9597                 ret = btrfs_update_inode(trans, root, inode);
9598
9599                 if (ret) {
9600                         btrfs_abort_transaction(trans, ret);
9601                         if (own_trans)
9602                                 btrfs_end_transaction(trans);
9603                         break;
9604                 }
9605
9606                 if (own_trans)
9607                         btrfs_end_transaction(trans);
9608         }
9609         if (clear_offset < end)
9610                 btrfs_free_reserved_data_space(inode, NULL, clear_offset,
9611                         end - clear_offset + 1);
9612         return ret;
9613 }
9614
9615 int btrfs_prealloc_file_range(struct inode *inode, int mode,
9616                               u64 start, u64 num_bytes, u64 min_size,
9617                               loff_t actual_len, u64 *alloc_hint)
9618 {
9619         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9620                                            min_size, actual_len, alloc_hint,
9621                                            NULL);
9622 }
9623
9624 int btrfs_prealloc_file_range_trans(struct inode *inode,
9625                                     struct btrfs_trans_handle *trans, int mode,
9626                                     u64 start, u64 num_bytes, u64 min_size,
9627                                     loff_t actual_len, u64 *alloc_hint)
9628 {
9629         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9630                                            min_size, actual_len, alloc_hint, trans);
9631 }
9632
9633 static int btrfs_set_page_dirty(struct page *page)
9634 {
9635         return __set_page_dirty_nobuffers(page);
9636 }
9637
9638 static int btrfs_permission(struct inode *inode, int mask)
9639 {
9640         struct btrfs_root *root = BTRFS_I(inode)->root;
9641         umode_t mode = inode->i_mode;
9642
9643         if (mask & MAY_WRITE &&
9644             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
9645                 if (btrfs_root_readonly(root))
9646                         return -EROFS;
9647                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
9648                         return -EACCES;
9649         }
9650         return generic_permission(inode, mask);
9651 }
9652
9653 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
9654 {
9655         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
9656         struct btrfs_trans_handle *trans;
9657         struct btrfs_root *root = BTRFS_I(dir)->root;
9658         struct inode *inode = NULL;
9659         u64 objectid;
9660         u64 index;
9661         int ret = 0;
9662
9663         /*
9664          * 5 units required for adding orphan entry
9665          */
9666         trans = btrfs_start_transaction(root, 5);
9667         if (IS_ERR(trans))
9668                 return PTR_ERR(trans);
9669
9670         ret = btrfs_find_free_ino(root, &objectid);
9671         if (ret)
9672                 goto out;
9673
9674         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
9675                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
9676         if (IS_ERR(inode)) {
9677                 ret = PTR_ERR(inode);
9678                 inode = NULL;
9679                 goto out;
9680         }
9681
9682         inode->i_fop = &btrfs_file_operations;
9683         inode->i_op = &btrfs_file_inode_operations;
9684
9685         inode->i_mapping->a_ops = &btrfs_aops;
9686         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9687
9688         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
9689         if (ret)
9690                 goto out;
9691
9692         ret = btrfs_update_inode(trans, root, inode);
9693         if (ret)
9694                 goto out;
9695         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
9696         if (ret)
9697                 goto out;
9698
9699         /*
9700          * We set number of links to 0 in btrfs_new_inode(), and here we set
9701          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
9702          * through:
9703          *
9704          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
9705          */
9706         set_nlink(inode, 1);
9707         d_tmpfile(dentry, inode);
9708         unlock_new_inode(inode);
9709         mark_inode_dirty(inode);
9710 out:
9711         btrfs_end_transaction(trans);
9712         if (ret && inode)
9713                 discard_new_inode(inode);
9714         btrfs_btree_balance_dirty(fs_info);
9715         return ret;
9716 }
9717
9718 void btrfs_set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
9719 {
9720         struct inode *inode = tree->private_data;
9721         unsigned long index = start >> PAGE_SHIFT;
9722         unsigned long end_index = end >> PAGE_SHIFT;
9723         struct page *page;
9724
9725         while (index <= end_index) {
9726                 page = find_get_page(inode->i_mapping, index);
9727                 ASSERT(page); /* Pages should be in the extent_io_tree */
9728                 set_page_writeback(page);
9729                 put_page(page);
9730                 index++;
9731         }
9732 }
9733
9734 #ifdef CONFIG_SWAP
9735 /*
9736  * Add an entry indicating a block group or device which is pinned by a
9737  * swapfile. Returns 0 on success, 1 if there is already an entry for it, or a
9738  * negative errno on failure.
9739  */
9740 static int btrfs_add_swapfile_pin(struct inode *inode, void *ptr,
9741                                   bool is_block_group)
9742 {
9743         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
9744         struct btrfs_swapfile_pin *sp, *entry;
9745         struct rb_node **p;
9746         struct rb_node *parent = NULL;
9747
9748         sp = kmalloc(sizeof(*sp), GFP_NOFS);
9749         if (!sp)
9750                 return -ENOMEM;
9751         sp->ptr = ptr;
9752         sp->inode = inode;
9753         sp->is_block_group = is_block_group;
9754
9755         spin_lock(&fs_info->swapfile_pins_lock);
9756         p = &fs_info->swapfile_pins.rb_node;
9757         while (*p) {
9758                 parent = *p;
9759                 entry = rb_entry(parent, struct btrfs_swapfile_pin, node);
9760                 if (sp->ptr < entry->ptr ||
9761                     (sp->ptr == entry->ptr && sp->inode < entry->inode)) {
9762                         p = &(*p)->rb_left;
9763                 } else if (sp->ptr > entry->ptr ||
9764                            (sp->ptr == entry->ptr && sp->inode > entry->inode)) {
9765                         p = &(*p)->rb_right;
9766                 } else {
9767                         spin_unlock(&fs_info->swapfile_pins_lock);
9768                         kfree(sp);
9769                         return 1;
9770                 }
9771         }
9772         rb_link_node(&sp->node, parent, p);
9773         rb_insert_color(&sp->node, &fs_info->swapfile_pins);
9774         spin_unlock(&fs_info->swapfile_pins_lock);
9775         return 0;
9776 }
9777
9778 /* Free all of the entries pinned by this swapfile. */
9779 static void btrfs_free_swapfile_pins(struct inode *inode)
9780 {
9781         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
9782         struct btrfs_swapfile_pin *sp;
9783         struct rb_node *node, *next;
9784
9785         spin_lock(&fs_info->swapfile_pins_lock);
9786         node = rb_first(&fs_info->swapfile_pins);
9787         while (node) {
9788                 next = rb_next(node);
9789                 sp = rb_entry(node, struct btrfs_swapfile_pin, node);
9790                 if (sp->inode == inode) {
9791                         rb_erase(&sp->node, &fs_info->swapfile_pins);
9792                         if (sp->is_block_group)
9793                                 btrfs_put_block_group(sp->ptr);
9794                         kfree(sp);
9795                 }
9796                 node = next;
9797         }
9798         spin_unlock(&fs_info->swapfile_pins_lock);
9799 }
9800
9801 struct btrfs_swap_info {
9802         u64 start;
9803         u64 block_start;
9804         u64 block_len;
9805         u64 lowest_ppage;
9806         u64 highest_ppage;
9807         unsigned long nr_pages;
9808         int nr_extents;
9809 };
9810
9811 static int btrfs_add_swap_extent(struct swap_info_struct *sis,
9812                                  struct btrfs_swap_info *bsi)
9813 {
9814         unsigned long nr_pages;
9815         u64 first_ppage, first_ppage_reported, next_ppage;
9816         int ret;
9817
9818         first_ppage = ALIGN(bsi->block_start, PAGE_SIZE) >> PAGE_SHIFT;
9819         next_ppage = ALIGN_DOWN(bsi->block_start + bsi->block_len,
9820                                 PAGE_SIZE) >> PAGE_SHIFT;
9821
9822         if (first_ppage >= next_ppage)
9823                 return 0;
9824         nr_pages = next_ppage - first_ppage;
9825
9826         first_ppage_reported = first_ppage;
9827         if (bsi->start == 0)
9828                 first_ppage_reported++;
9829         if (bsi->lowest_ppage > first_ppage_reported)
9830                 bsi->lowest_ppage = first_ppage_reported;
9831         if (bsi->highest_ppage < (next_ppage - 1))
9832                 bsi->highest_ppage = next_ppage - 1;
9833
9834         ret = add_swap_extent(sis, bsi->nr_pages, nr_pages, first_ppage);
9835         if (ret < 0)
9836                 return ret;
9837         bsi->nr_extents += ret;
9838         bsi->nr_pages += nr_pages;
9839         return 0;
9840 }
9841
9842 static void btrfs_swap_deactivate(struct file *file)
9843 {
9844         struct inode *inode = file_inode(file);
9845
9846         btrfs_free_swapfile_pins(inode);
9847         atomic_dec(&BTRFS_I(inode)->root->nr_swapfiles);
9848 }
9849
9850 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
9851                                sector_t *span)
9852 {
9853         struct inode *inode = file_inode(file);
9854         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
9855         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
9856         struct extent_state *cached_state = NULL;
9857         struct extent_map *em = NULL;
9858         struct btrfs_device *device = NULL;
9859         struct btrfs_swap_info bsi = {
9860                 .lowest_ppage = (sector_t)-1ULL,
9861         };
9862         int ret = 0;
9863         u64 isize;
9864         u64 start;
9865
9866         /*
9867          * If the swap file was just created, make sure delalloc is done. If the
9868          * file changes again after this, the user is doing something stupid and
9869          * we don't really care.
9870          */
9871         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
9872         if (ret)
9873                 return ret;
9874
9875         /*
9876          * The inode is locked, so these flags won't change after we check them.
9877          */
9878         if (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS) {
9879                 btrfs_warn(fs_info, "swapfile must not be compressed");
9880                 return -EINVAL;
9881         }
9882         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)) {
9883                 btrfs_warn(fs_info, "swapfile must not be copy-on-write");
9884                 return -EINVAL;
9885         }
9886         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
9887                 btrfs_warn(fs_info, "swapfile must not be checksummed");
9888                 return -EINVAL;
9889         }
9890
9891         /*
9892          * Balance or device remove/replace/resize can move stuff around from
9893          * under us. The EXCL_OP flag makes sure they aren't running/won't run
9894          * concurrently while we are mapping the swap extents, and
9895          * fs_info->swapfile_pins prevents them from running while the swap file
9896          * is active and moving the extents. Note that this also prevents a
9897          * concurrent device add which isn't actually necessary, but it's not
9898          * really worth the trouble to allow it.
9899          */
9900         if (test_and_set_bit(BTRFS_FS_EXCL_OP, &fs_info->flags)) {
9901                 btrfs_warn(fs_info,
9902            "cannot activate swapfile while exclusive operation is running");
9903                 return -EBUSY;
9904         }
9905         /*
9906          * Snapshots can create extents which require COW even if NODATACOW is
9907          * set. We use this counter to prevent snapshots. We must increment it
9908          * before walking the extents because we don't want a concurrent
9909          * snapshot to run after we've already checked the extents.
9910          */
9911         atomic_inc(&BTRFS_I(inode)->root->nr_swapfiles);
9912
9913         isize = ALIGN_DOWN(inode->i_size, fs_info->sectorsize);
9914
9915         lock_extent_bits(io_tree, 0, isize - 1, &cached_state);
9916         start = 0;
9917         while (start < isize) {
9918                 u64 logical_block_start, physical_block_start;
9919                 struct btrfs_block_group *bg;
9920                 u64 len = isize - start;
9921
9922                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len);
9923                 if (IS_ERR(em)) {
9924                         ret = PTR_ERR(em);
9925                         goto out;
9926                 }
9927
9928                 if (em->block_start == EXTENT_MAP_HOLE) {
9929                         btrfs_warn(fs_info, "swapfile must not have holes");
9930                         ret = -EINVAL;
9931                         goto out;
9932                 }
9933                 if (em->block_start == EXTENT_MAP_INLINE) {
9934                         /*
9935                          * It's unlikely we'll ever actually find ourselves
9936                          * here, as a file small enough to fit inline won't be
9937                          * big enough to store more than the swap header, but in
9938                          * case something changes in the future, let's catch it
9939                          * here rather than later.
9940                          */
9941                         btrfs_warn(fs_info, "swapfile must not be inline");
9942                         ret = -EINVAL;
9943                         goto out;
9944                 }
9945                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
9946                         btrfs_warn(fs_info, "swapfile must not be compressed");
9947                         ret = -EINVAL;
9948                         goto out;
9949                 }
9950
9951                 logical_block_start = em->block_start + (start - em->start);
9952                 len = min(len, em->len - (start - em->start));
9953                 free_extent_map(em);
9954                 em = NULL;
9955
9956                 ret = can_nocow_extent(inode, start, &len, NULL, NULL, NULL);
9957                 if (ret < 0) {
9958                         goto out;
9959                 } else if (ret) {
9960                         ret = 0;
9961                 } else {
9962                         btrfs_warn(fs_info,
9963                                    "swapfile must not be copy-on-write");
9964                         ret = -EINVAL;
9965                         goto out;
9966                 }
9967
9968                 em = btrfs_get_chunk_map(fs_info, logical_block_start, len);
9969                 if (IS_ERR(em)) {
9970                         ret = PTR_ERR(em);
9971                         goto out;
9972                 }
9973
9974                 if (em->map_lookup->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
9975                         btrfs_warn(fs_info,
9976                                    "swapfile must have single data profile");
9977                         ret = -EINVAL;
9978                         goto out;
9979                 }
9980
9981                 if (device == NULL) {
9982                         device = em->map_lookup->stripes[0].dev;
9983                         ret = btrfs_add_swapfile_pin(inode, device, false);
9984                         if (ret == 1)
9985                                 ret = 0;
9986                         else if (ret)
9987                                 goto out;
9988                 } else if (device != em->map_lookup->stripes[0].dev) {
9989                         btrfs_warn(fs_info, "swapfile must be on one device");
9990                         ret = -EINVAL;
9991                         goto out;
9992                 }
9993
9994                 physical_block_start = (em->map_lookup->stripes[0].physical +
9995                                         (logical_block_start - em->start));
9996                 len = min(len, em->len - (logical_block_start - em->start));
9997                 free_extent_map(em);
9998                 em = NULL;
9999
10000                 bg = btrfs_lookup_block_group(fs_info, logical_block_start);
10001                 if (!bg) {
10002                         btrfs_warn(fs_info,
10003                            "could not find block group containing swapfile");
10004                         ret = -EINVAL;
10005                         goto out;
10006                 }
10007
10008                 ret = btrfs_add_swapfile_pin(inode, bg, true);
10009                 if (ret) {
10010                         btrfs_put_block_group(bg);
10011                         if (ret == 1)
10012                                 ret = 0;
10013                         else
10014                                 goto out;
10015                 }
10016
10017                 if (bsi.block_len &&
10018                     bsi.block_start + bsi.block_len == physical_block_start) {
10019                         bsi.block_len += len;
10020                 } else {
10021                         if (bsi.block_len) {
10022                                 ret = btrfs_add_swap_extent(sis, &bsi);
10023                                 if (ret)
10024                                         goto out;
10025                         }
10026                         bsi.start = start;
10027                         bsi.block_start = physical_block_start;
10028                         bsi.block_len = len;
10029                 }
10030
10031                 start += len;
10032         }
10033
10034         if (bsi.block_len)
10035                 ret = btrfs_add_swap_extent(sis, &bsi);
10036
10037 out:
10038         if (!IS_ERR_OR_NULL(em))
10039                 free_extent_map(em);
10040
10041         unlock_extent_cached(io_tree, 0, isize - 1, &cached_state);
10042
10043         if (ret)
10044                 btrfs_swap_deactivate(file);
10045
10046         clear_bit(BTRFS_FS_EXCL_OP, &fs_info->flags);
10047
10048         if (ret)
10049                 return ret;
10050
10051         if (device)
10052                 sis->bdev = device->bdev;
10053         *span = bsi.highest_ppage - bsi.lowest_ppage + 1;
10054         sis->max = bsi.nr_pages;
10055         sis->pages = bsi.nr_pages - 1;
10056         sis->highest_bit = bsi.nr_pages - 1;
10057         return bsi.nr_extents;
10058 }
10059 #else
10060 static void btrfs_swap_deactivate(struct file *file)
10061 {
10062 }
10063
10064 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10065                                sector_t *span)
10066 {
10067         return -EOPNOTSUPP;
10068 }
10069 #endif
10070
10071 static const struct inode_operations btrfs_dir_inode_operations = {
10072         .getattr        = btrfs_getattr,
10073         .lookup         = btrfs_lookup,
10074         .create         = btrfs_create,
10075         .unlink         = btrfs_unlink,
10076         .link           = btrfs_link,
10077         .mkdir          = btrfs_mkdir,
10078         .rmdir          = btrfs_rmdir,
10079         .rename         = btrfs_rename2,
10080         .symlink        = btrfs_symlink,
10081         .setattr        = btrfs_setattr,
10082         .mknod          = btrfs_mknod,
10083         .listxattr      = btrfs_listxattr,
10084         .permission     = btrfs_permission,
10085         .get_acl        = btrfs_get_acl,
10086         .set_acl        = btrfs_set_acl,
10087         .update_time    = btrfs_update_time,
10088         .tmpfile        = btrfs_tmpfile,
10089 };
10090
10091 static const struct file_operations btrfs_dir_file_operations = {
10092         .llseek         = generic_file_llseek,
10093         .read           = generic_read_dir,
10094         .iterate_shared = btrfs_real_readdir,
10095         .open           = btrfs_opendir,
10096         .unlocked_ioctl = btrfs_ioctl,
10097 #ifdef CONFIG_COMPAT
10098         .compat_ioctl   = btrfs_compat_ioctl,
10099 #endif
10100         .release        = btrfs_release_file,
10101         .fsync          = btrfs_sync_file,
10102 };
10103
10104 static const struct extent_io_ops btrfs_extent_io_ops = {
10105         /* mandatory callbacks */
10106         .submit_bio_hook = btrfs_submit_bio_hook,
10107         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10108 };
10109
10110 /*
10111  * btrfs doesn't support the bmap operation because swapfiles
10112  * use bmap to make a mapping of extents in the file.  They assume
10113  * these extents won't change over the life of the file and they
10114  * use the bmap result to do IO directly to the drive.
10115  *
10116  * the btrfs bmap call would return logical addresses that aren't
10117  * suitable for IO and they also will change frequently as COW
10118  * operations happen.  So, swapfile + btrfs == corruption.
10119  *
10120  * For now we're avoiding this by dropping bmap.
10121  */
10122 static const struct address_space_operations btrfs_aops = {
10123         .readpage       = btrfs_readpage,
10124         .writepage      = btrfs_writepage,
10125         .writepages     = btrfs_writepages,
10126         .readahead      = btrfs_readahead,
10127         .direct_IO      = noop_direct_IO,
10128         .invalidatepage = btrfs_invalidatepage,
10129         .releasepage    = btrfs_releasepage,
10130 #ifdef CONFIG_MIGRATION
10131         .migratepage    = btrfs_migratepage,
10132 #endif
10133         .set_page_dirty = btrfs_set_page_dirty,
10134         .error_remove_page = generic_error_remove_page,
10135         .swap_activate  = btrfs_swap_activate,
10136         .swap_deactivate = btrfs_swap_deactivate,
10137 };
10138
10139 static const struct inode_operations btrfs_file_inode_operations = {
10140         .getattr        = btrfs_getattr,
10141         .setattr        = btrfs_setattr,
10142         .listxattr      = btrfs_listxattr,
10143         .permission     = btrfs_permission,
10144         .fiemap         = btrfs_fiemap,
10145         .get_acl        = btrfs_get_acl,
10146         .set_acl        = btrfs_set_acl,
10147         .update_time    = btrfs_update_time,
10148 };
10149 static const struct inode_operations btrfs_special_inode_operations = {
10150         .getattr        = btrfs_getattr,
10151         .setattr        = btrfs_setattr,
10152         .permission     = btrfs_permission,
10153         .listxattr      = btrfs_listxattr,
10154         .get_acl        = btrfs_get_acl,
10155         .set_acl        = btrfs_set_acl,
10156         .update_time    = btrfs_update_time,
10157 };
10158 static const struct inode_operations btrfs_symlink_inode_operations = {
10159         .get_link       = page_get_link,
10160         .getattr        = btrfs_getattr,
10161         .setattr        = btrfs_setattr,
10162         .permission     = btrfs_permission,
10163         .listxattr      = btrfs_listxattr,
10164         .update_time    = btrfs_update_time,
10165 };
10166
10167 const struct dentry_operations btrfs_dentry_operations = {
10168         .d_delete       = btrfs_dentry_delete,
10169 };