fs/btrfs/ordered-data.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/slab.h>
   7 #include <linux/blkdev.h>
   8 #include <linux/writeback.h>
   9 #include <linux/pagevec.h>
  10 #include "ctree.h"
  11 #include "transaction.h"
  12 #include "btrfs_inode.h"
  13 #include "extent_io.h"
  14 #include "disk-io.h"
  15 #include "compression.h"
  16
  17 static struct kmem_cache *btrfs_ordered_extent_cache;
  18
  19 static u64 entry_end(struct btrfs_ordered_extent *entry)
  20 {
  21         if (entry->file_offset + entry->len < entry->file_offset)
  22                 return (u64)-1;
  23         return entry->file_offset + entry->len;
  24 }
  25
  26 /* returns NULL if the insertion worked, or it returns the node it did find
  27  * in the tree
  28  */
  29 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
  30                                    struct rb_node *node)
  31 {
  32         struct rb_node **p = &root->rb_node;
  33         struct rb_node *parent = NULL;
  34         struct btrfs_ordered_extent *entry;
  35
  36         while (*p) {
  37                 parent = *p;
  38                 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
  39
  40                 if (file_offset < entry->file_offset)
  41                         p = &(*p)->rb_left;
  42                 else if (file_offset >= entry_end(entry))
  43                         p = &(*p)->rb_right;
  44                 else
  45                         return parent;
  46         }
  47
  48         rb_link_node(node, parent, p);
  49         rb_insert_color(node, root);
  50         return NULL;
  51 }
  52
  53 static void ordered_data_tree_panic(struct inode *inode, int errno,
  54                                                u64 offset)
  55 {
  56         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
  57         btrfs_panic(fs_info, errno,
  58                     "Inconsistency in ordered tree at offset %llu", offset);
  59 }
  60
  61 /*
  62  * look for a given offset in the tree, and if it can't be found return the
  63  * first lesser offset
  64  */
  65 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
  66                                      struct rb_node **prev_ret)
  67 {
  68         struct rb_node *n = root->rb_node;
  69         struct rb_node *prev = NULL;
  70         struct rb_node *test;
  71         struct btrfs_ordered_extent *entry;
  72         struct btrfs_ordered_extent *prev_entry = NULL;
  73
  74         while (n) {
  75                 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
  76                 prev = n;
  77                 prev_entry = entry;
  78
  79                 if (file_offset < entry->file_offset)
  80                         n = n->rb_left;
  81                 else if (file_offset >= entry_end(entry))
  82                         n = n->rb_right;
  83                 else
  84                         return n;
  85         }
  86         if (!prev_ret)
  87                 return NULL;
  88
  89         while (prev && file_offset >= entry_end(prev_entry)) {
  90                 test = rb_next(prev);
  91                 if (!test)
  92                         break;
  93                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
  94                                       rb_node);
  95                 if (file_offset < entry_end(prev_entry))
  96                         break;
  97
  98                 prev = test;
  99         }
 100         if (prev)
 101                 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
 102                                       rb_node);
 103         while (prev && file_offset < entry_end(prev_entry)) {
 104                 test = rb_prev(prev);
 105                 if (!test)
 106                         break;
 107                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
 108                                       rb_node);
 109                 prev = test;
 110         }
 111         *prev_ret = prev;
 112         return NULL;
 113 }
 114
 115 /*
 116  * helper to check if a given offset is inside a given entry
 117  */
 118 static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
 119 {
 120         if (file_offset < entry->file_offset ||
 121             entry->file_offset + entry->len <= file_offset)
 122                 return 0;
 123         return 1;
 124 }
 125
 126 static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
 127                           u64 len)
 128 {
 129         if (file_offset + len <= entry->file_offset ||
 130             entry->file_offset + entry->len <= file_offset)
 131                 return 0;
 132         return 1;
 133 }
 134
 135 /*
 136  * look find the first ordered struct that has this offset, otherwise
 137  * the first one less than this offset
 138  */
 139 static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
 140                                           u64 file_offset)
 141 {
 142         struct rb_root *root = &tree->tree;
 143         struct rb_node *prev = NULL;
 144         struct rb_node *ret;
 145         struct btrfs_ordered_extent *entry;
 146
 147         if (tree->last) {
 148                 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
 149                                  rb_node);
 150                 if (offset_in_entry(entry, file_offset))
 151                         return tree->last;
 152         }
 153         ret = __tree_search(root, file_offset, &prev);
 154         if (!ret)
 155                 ret = prev;
 156         if (ret)
 157                 tree->last = ret;
 158         return ret;
 159 }
 160
 161 /* allocate and add a new ordered_extent into the per-inode tree.
 162  * file_offset is the logical offset in the file
 163  *
 164  * start is the disk block number of an extent already reserved in the
 165  * extent allocation tree
 166  *
 167  * len is the length of the extent
 168  *
 169  * The tree is given a single reference on the ordered extent that was
 170  * inserted.
 171  */
 172 static int __btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
 173                                       u64 start, u64 len, u64 disk_len,
 174                                       int type, int dio, int compress_type)
 175 {
 176         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 177         struct btrfs_root *root = BTRFS_I(inode)->root;
 178         struct btrfs_ordered_inode_tree *tree;
 179         struct rb_node *node;
 180         struct btrfs_ordered_extent *entry;
 181
 182         tree = &BTRFS_I(inode)->ordered_tree;
 183         entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
 184         if (!entry)
 185                 return -ENOMEM;
 186
 187         entry->file_offset = file_offset;
 188         entry->start = start;
 189         entry->len = len;
 190         entry->disk_len = disk_len;
 191         entry->bytes_left = len;
 192         entry->inode = igrab(inode);
 193         entry->compress_type = compress_type;
 194         entry->truncated_len = (u64)-1;
 195         if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
 196                 set_bit(type, &entry->flags);
 197
 198         if (dio)
 199                 set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
 200
 201         /* one ref for the tree */
 202         refcount_set(&entry->refs, 1);
 203         init_waitqueue_head(&entry->wait);
 204         INIT_LIST_HEAD(&entry->list);
 205         INIT_LIST_HEAD(&entry->root_extent_list);
 206         INIT_LIST_HEAD(&entry->work_list);
 207         init_completion(&entry->completion);
 208         INIT_LIST_HEAD(&entry->log_list);
 209         INIT_LIST_HEAD(&entry->trans_list);
 210
 211         trace_btrfs_ordered_extent_add(inode, entry);
 212
 213         spin_lock_irq(&tree->lock);
 214         node = tree_insert(&tree->tree, file_offset,
 215                            &entry->rb_node);
 216         if (node)
 217                 ordered_data_tree_panic(inode, -EEXIST, file_offset);
 218         spin_unlock_irq(&tree->lock);
 219
 220         spin_lock(&root->ordered_extent_lock);
 221         list_add_tail(&entry->root_extent_list,
 222                       &root->ordered_extents);
 223         root->nr_ordered_extents++;
 224         if (root->nr_ordered_extents == 1) {
 225                 spin_lock(&fs_info->ordered_root_lock);
 226                 BUG_ON(!list_empty(&root->ordered_root));
 227                 list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
 228                 spin_unlock(&fs_info->ordered_root_lock);
 229         }
 230         spin_unlock(&root->ordered_extent_lock);
 231
 232         /*
 233          * We don't need the count_max_extents here, we can assume that all of
 234          * that work has been done at higher layers, so this is truly the
 235          * smallest the extent is going to get.
 236          */
 237         spin_lock(&BTRFS_I(inode)->lock);
 238         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
 239         spin_unlock(&BTRFS_I(inode)->lock);
 240
 241         return 0;
 242 }
 243
 244 int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
 245                              u64 start, u64 len, u64 disk_len, int type)
 246 {
 247         return __btrfs_add_ordered_extent(inode, file_offset, start, len,
 248                                           disk_len, type, 0,
 249                                           BTRFS_COMPRESS_NONE);
 250 }
 251
 252 int btrfs_add_ordered_extent_dio(struct inode *inode, u64 file_offset,
 253                                  u64 start, u64 len, u64 disk_len, int type)
 254 {
 255         return __btrfs_add_ordered_extent(inode, file_offset, start, len,
 256                                           disk_len, type, 1,
 257                                           BTRFS_COMPRESS_NONE);
 258 }
 259
 260 int btrfs_add_ordered_extent_compress(struct inode *inode, u64 file_offset,
 261                                       u64 start, u64 len, u64 disk_len,
 262                                       int type, int compress_type)
 263 {
 264         return __btrfs_add_ordered_extent(inode, file_offset, start, len,
 265                                           disk_len, type, 0,
 266                                           compress_type);
 267 }
 268
 269 /*
 270  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
 271  * when an ordered extent is finished.  If the list covers more than one
 272  * ordered extent, it is split across multiples.
 273  */
 274 void btrfs_add_ordered_sum(struct inode *inode,
 275                            struct btrfs_ordered_extent *entry,
 276                            struct btrfs_ordered_sum *sum)
 277 {
 278         struct btrfs_ordered_inode_tree *tree;
 279
 280         tree = &BTRFS_I(inode)->ordered_tree;
 281         spin_lock_irq(&tree->lock);
 282         list_add_tail(&sum->list, &entry->list);
 283         spin_unlock_irq(&tree->lock);
 284 }
 285
 286 /*
 287  * this is used to account for finished IO across a given range
 288  * of the file.  The IO may span ordered extents.  If
 289  * a given ordered_extent is completely done, 1 is returned, otherwise
 290  * 0.
 291  *
 292  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
 293  * to make sure this function only returns 1 once for a given ordered extent.
 294  *
 295  * file_offset is updated to one byte past the range that is recorded as
 296  * complete.  This allows you to walk forward in the file.
 297  */
 298 int btrfs_dec_test_first_ordered_pending(struct inode *inode,
 299                                    struct btrfs_ordered_extent **cached,
 300                                    u64 *file_offset, u64 io_size, int uptodate)
 301 {
 302         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 303         struct btrfs_ordered_inode_tree *tree;
 304         struct rb_node *node;
 305         struct btrfs_ordered_extent *entry = NULL;
 306         int ret;
 307         unsigned long flags;
 308         u64 dec_end;
 309         u64 dec_start;
 310         u64 to_dec;
 311
 312         tree = &BTRFS_I(inode)->ordered_tree;
 313         spin_lock_irqsave(&tree->lock, flags);
 314         node = tree_search(tree, *file_offset);
 315         if (!node) {
 316                 ret = 1;
 317                 goto out;
 318         }
 319
 320         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 321         if (!offset_in_entry(entry, *file_offset)) {
 322                 ret = 1;
 323                 goto out;
 324         }
 325
 326         dec_start = max(*file_offset, entry->file_offset);
 327         dec_end = min(*file_offset + io_size, entry->file_offset +
 328                       entry->len);
 329         *file_offset = dec_end;
 330         if (dec_start > dec_end) {
 331                 btrfs_crit(fs_info, "bad ordering dec_start %llu end %llu",
 332                            dec_start, dec_end);
 333         }
 334         to_dec = dec_end - dec_start;
 335         if (to_dec > entry->bytes_left) {
 336                 btrfs_crit(fs_info,
 337                            "bad ordered accounting left %llu size %llu",
 338                            entry->bytes_left, to_dec);
 339         }
 340         entry->bytes_left -= to_dec;
 341         if (!uptodate)
 342                 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
 343
 344         if (entry->bytes_left == 0) {
 345                 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 346                 /*
 347                  * Implicit memory barrier after test_and_set_bit
 348                  */
 349                 if (waitqueue_active(&entry->wait))
 350                         wake_up(&entry->wait);
 351         } else {
 352                 ret = 1;
 353         }
 354 out:
 355         if (!ret && cached && entry) {
 356                 *cached = entry;
 357                 refcount_inc(&entry->refs);
 358         }
 359         spin_unlock_irqrestore(&tree->lock, flags);
 360         return ret == 0;
 361 }
 362
 363 /*
 364  * this is used to account for finished IO across a given range
 365  * of the file.  The IO should not span ordered extents.  If
 366  * a given ordered_extent is completely done, 1 is returned, otherwise
 367  * 0.
 368  *
 369  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
 370  * to make sure this function only returns 1 once for a given ordered extent.
 371  */
 372 int btrfs_dec_test_ordered_pending(struct inode *inode,
 373                                    struct btrfs_ordered_extent **cached,
 374                                    u64 file_offset, u64 io_size, int uptodate)
 375 {
 376         struct btrfs_ordered_inode_tree *tree;
 377         struct rb_node *node;
 378         struct btrfs_ordered_extent *entry = NULL;
 379         unsigned long flags;
 380         int ret;
 381
 382         tree = &BTRFS_I(inode)->ordered_tree;
 383         spin_lock_irqsave(&tree->lock, flags);
 384         if (cached && *cached) {
 385                 entry = *cached;
 386                 goto have_entry;
 387         }
 388
 389         node = tree_search(tree, file_offset);
 390         if (!node) {
 391                 ret = 1;
 392                 goto out;
 393         }
 394
 395         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 396 have_entry:
 397         if (!offset_in_entry(entry, file_offset)) {
 398                 ret = 1;
 399                 goto out;
 400         }
 401
 402         if (io_size > entry->bytes_left) {
 403                 btrfs_crit(BTRFS_I(inode)->root->fs_info,
 404                            "bad ordered accounting left %llu size %llu",
 405                        entry->bytes_left, io_size);
 406         }
 407         entry->bytes_left -= io_size;
 408         if (!uptodate)
 409                 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
 410
 411         if (entry->bytes_left == 0) {
 412                 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 413                 /*
 414                  * Implicit memory barrier after test_and_set_bit
 415                  */
 416                 if (waitqueue_active(&entry->wait))
 417                         wake_up(&entry->wait);
 418         } else {
 419                 ret = 1;
 420         }
 421 out:
 422         if (!ret && cached && entry) {
 423                 *cached = entry;
 424                 refcount_inc(&entry->refs);
 425         }
 426         spin_unlock_irqrestore(&tree->lock, flags);
 427         return ret == 0;
 428 }
 429
 430 /* Needs to either be called under a log transaction or the log_mutex */
 431 void btrfs_get_logged_extents(struct btrfs_inode *inode,
 432                               struct list_head *logged_list,
 433                               const loff_t start,
 434                               const loff_t end)
 435 {
 436         struct btrfs_ordered_inode_tree *tree;
 437         struct btrfs_ordered_extent *ordered;
 438         struct rb_node *n;
 439         struct rb_node *prev;
 440
 441         tree = &inode->ordered_tree;
 442         spin_lock_irq(&tree->lock);
 443         n = __tree_search(&tree->tree, end, &prev);
 444         if (!n)
 445                 n = prev;
 446         for (; n; n = rb_prev(n)) {
 447                 ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
 448                 if (ordered->file_offset > end)
 449                         continue;
 450                 if (entry_end(ordered) <= start)
 451                         break;
 452                 if (test_and_set_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
 453                         continue;
 454                 list_add(&ordered->log_list, logged_list);
 455                 refcount_inc(&ordered->refs);
 456         }
 457         spin_unlock_irq(&tree->lock);
 458 }
 459
 460 void btrfs_put_logged_extents(struct list_head *logged_list)
 461 {
 462         struct btrfs_ordered_extent *ordered;
 463
 464         while (!list_empty(logged_list)) {
 465                 ordered = list_first_entry(logged_list,
 466                                            struct btrfs_ordered_extent,
 467                                            log_list);
 468                 list_del_init(&ordered->log_list);
 469                 btrfs_put_ordered_extent(ordered);
 470         }
 471 }
 472
 473 void btrfs_submit_logged_extents(struct list_head *logged_list,
 474                                  struct btrfs_root *log)
 475 {
 476         int index = log->log_transid % 2;
 477
 478         spin_lock_irq(&log->log_extents_lock[index]);
 479         list_splice_tail(logged_list, &log->logged_list[index]);
 480         spin_unlock_irq(&log->log_extents_lock[index]);
 481 }
 482
 483 void btrfs_wait_logged_extents(struct btrfs_trans_handle *trans,
 484                                struct btrfs_root *log, u64 transid)
 485 {
 486         struct btrfs_ordered_extent *ordered;
 487         int index = transid % 2;
 488
 489         spin_lock_irq(&log->log_extents_lock[index]);
 490         while (!list_empty(&log->logged_list[index])) {
 491                 struct inode *inode;
 492                 ordered = list_first_entry(&log->logged_list[index],
 493                                            struct btrfs_ordered_extent,
 494                                            log_list);
 495                 list_del_init(&ordered->log_list);
 496                 inode = ordered->inode;
 497                 spin_unlock_irq(&log->log_extents_lock[index]);
 498
 499                 if (!test_bit(BTRFS_ORDERED_IO_DONE, &ordered->flags) &&
 500                     !test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags)) {
 501                         u64 start = ordered->file_offset;
 502                         u64 end = ordered->file_offset + ordered->len - 1;
 503
 504                         WARN_ON(!inode);
 505                         filemap_fdatawrite_range(inode->i_mapping, start, end);
 506                 }
 507                 wait_event(ordered->wait, test_bit(BTRFS_ORDERED_IO_DONE,
 508                                                    &ordered->flags));
 509
 510                 /*
 511                  * In order to keep us from losing our ordered extent
 512                  * information when committing the transaction we have to make
 513                  * sure that any logged extents are completed when we go to
 514                  * commit the transaction.  To do this we simply increase the
 515                  * current transactions pending_ordered counter and decrement it
 516                  * when the ordered extent completes.
 517                  */
 518                 if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
 519                         struct btrfs_ordered_inode_tree *tree;
 520
 521                         tree = &BTRFS_I(inode)->ordered_tree;
 522                         spin_lock_irq(&tree->lock);
 523                         if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
 524                                 set_bit(BTRFS_ORDERED_PENDING, &ordered->flags);
 525                                 atomic_inc(&trans->transaction->pending_ordered);
 526                         }
 527                         spin_unlock_irq(&tree->lock);
 528                 }
 529                 btrfs_put_ordered_extent(ordered);
 530                 spin_lock_irq(&log->log_extents_lock[index]);
 531         }
 532         spin_unlock_irq(&log->log_extents_lock[index]);
 533 }
 534
 535 void btrfs_free_logged_extents(struct btrfs_root *log, u64 transid)
 536 {
 537         struct btrfs_ordered_extent *ordered;
 538         int index = transid % 2;
 539
 540         spin_lock_irq(&log->log_extents_lock[index]);
 541         while (!list_empty(&log->logged_list[index])) {
 542                 ordered = list_first_entry(&log->logged_list[index],
 543                                            struct btrfs_ordered_extent,
 544                                            log_list);
 545                 list_del_init(&ordered->log_list);
 546                 spin_unlock_irq(&log->log_extents_lock[index]);
 547                 btrfs_put_ordered_extent(ordered);
 548                 spin_lock_irq(&log->log_extents_lock[index]);
 549         }
 550         spin_unlock_irq(&log->log_extents_lock[index]);
 551 }
 552
 553 /*
 554  * used to drop a reference on an ordered extent.  This will free
 555  * the extent if the last reference is dropped
 556  */
 557 void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
 558 {
 559         struct list_head *cur;
 560         struct btrfs_ordered_sum *sum;
 561
 562         trace_btrfs_ordered_extent_put(entry->inode, entry);
 563
 564         if (refcount_dec_and_test(&entry->refs)) {
 565                 ASSERT(list_empty(&entry->log_list));
 566                 ASSERT(list_empty(&entry->trans_list));
 567                 ASSERT(list_empty(&entry->root_extent_list));
 568                 ASSERT(RB_EMPTY_NODE(&entry->rb_node));
 569                 if (entry->inode)
 570                         btrfs_add_delayed_iput(entry->inode);
 571                 while (!list_empty(&entry->list)) {
 572                         cur = entry->list.next;
 573                         sum = list_entry(cur, struct btrfs_ordered_sum, list);
 574                         list_del(&sum->list);
 575                         kfree(sum);
 576                 }
 577                 kmem_cache_free(btrfs_ordered_extent_cache, entry);
 578         }
 579 }
 580
 581 /*
 582  * remove an ordered extent from the tree.  No references are dropped
 583  * and waiters are woken up.
 584  */
 585 void btrfs_remove_ordered_extent(struct inode *inode,
 586                                  struct btrfs_ordered_extent *entry)
 587 {
 588         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 589         struct btrfs_ordered_inode_tree *tree;
 590         struct btrfs_inode *btrfs_inode = BTRFS_I(inode);
 591         struct btrfs_root *root = btrfs_inode->root;
 592         struct rb_node *node;
 593         bool dec_pending_ordered = false;
 594
 595         /* This is paired with btrfs_add_ordered_extent. */
 596         spin_lock(&btrfs_inode->lock);
 597         btrfs_mod_outstanding_extents(btrfs_inode, -1);
 598         spin_unlock(&btrfs_inode->lock);
 599         if (root != fs_info->tree_root)
 600                 btrfs_delalloc_release_metadata(btrfs_inode, entry->len, false);
 601
 602         tree = &btrfs_inode->ordered_tree;
 603         spin_lock_irq(&tree->lock);
 604         node = &entry->rb_node;
 605         rb_erase(node, &tree->tree);
 606         RB_CLEAR_NODE(node);
 607         if (tree->last == node)
 608                 tree->last = NULL;
 609         set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
 610         if (test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags))
 611                 dec_pending_ordered = true;
 612         spin_unlock_irq(&tree->lock);
 613
 614         /*
 615          * The current running transaction is waiting on us, we need to let it
 616          * know that we're complete and wake it up.
 617          */
 618         if (dec_pending_ordered) {
 619                 struct btrfs_transaction *trans;
 620
 621                 /*
 622                  * The checks for trans are just a formality, it should be set,
 623                  * but if it isn't we don't want to deref/assert under the spin
 624                  * lock, so be nice and check if trans is set, but ASSERT() so
 625                  * if it isn't set a developer will notice.
 626                  */
 627                 spin_lock(&fs_info->trans_lock);
 628                 trans = fs_info->running_transaction;
 629                 if (trans)
 630                         refcount_inc(&trans->use_count);
 631                 spin_unlock(&fs_info->trans_lock);
 632
 633                 ASSERT(trans);
 634                 if (trans) {
 635                         if (atomic_dec_and_test(&trans->pending_ordered))
 636                                 wake_up(&trans->pending_wait);
 637                         btrfs_put_transaction(trans);
 638                 }
 639         }
 640
 641         spin_lock(&root->ordered_extent_lock);
 642         list_del_init(&entry->root_extent_list);
 643         root->nr_ordered_extents--;
 644
 645         trace_btrfs_ordered_extent_remove(inode, entry);
 646
 647         if (!root->nr_ordered_extents) {
 648                 spin_lock(&fs_info->ordered_root_lock);
 649                 BUG_ON(list_empty(&root->ordered_root));
 650                 list_del_init(&root->ordered_root);
 651                 spin_unlock(&fs_info->ordered_root_lock);
 652         }
 653         spin_unlock(&root->ordered_extent_lock);
 654         wake_up(&entry->wait);
 655 }
 656
 657 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
 658 {
 659         struct btrfs_ordered_extent *ordered;
 660
 661         ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
 662         btrfs_start_ordered_extent(ordered->inode, ordered, 1);
 663         complete(&ordered->completion);
 664 }
 665
 666 /*
 667  * wait for all the ordered extents in a root.  This is done when balancing
 668  * space between drives.
 669  */
 670 u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
 671                                const u64 range_start, const u64 range_len)
 672 {
 673         struct btrfs_fs_info *fs_info = root->fs_info;
 674         LIST_HEAD(splice);
 675         LIST_HEAD(skipped);
 676         LIST_HEAD(works);
 677         struct btrfs_ordered_extent *ordered, *next;
 678         u64 count = 0;
 679         const u64 range_end = range_start + range_len;
 680
 681         mutex_lock(&root->ordered_extent_mutex);
 682         spin_lock(&root->ordered_extent_lock);
 683         list_splice_init(&root->ordered_extents, &splice);
 684         while (!list_empty(&splice) && nr) {
 685                 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
 686                                            root_extent_list);
 687
 688                 if (range_end <= ordered->start ||
 689                     ordered->start + ordered->disk_len <= range_start) {
 690                         list_move_tail(&ordered->root_extent_list, &skipped);
 691                         cond_resched_lock(&root->ordered_extent_lock);
 692                         continue;
 693                 }
 694
 695                 list_move_tail(&ordered->root_extent_list,
 696                                &root->ordered_extents);
 697                 refcount_inc(&ordered->refs);
 698                 spin_unlock(&root->ordered_extent_lock);
 699
 700                 btrfs_init_work(&ordered->flush_work,
 701                                 btrfs_flush_delalloc_helper,
 702                                 btrfs_run_ordered_extent_work, NULL, NULL);
 703                 list_add_tail(&ordered->work_list, &works);
 704                 btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
 705
 706                 cond_resched();
 707                 spin_lock(&root->ordered_extent_lock);
 708                 if (nr != U64_MAX)
 709                         nr--;
 710                 count++;
 711         }
 712         list_splice_tail(&skipped, &root->ordered_extents);
 713         list_splice_tail(&splice, &root->ordered_extents);
 714         spin_unlock(&root->ordered_extent_lock);
 715
 716         list_for_each_entry_safe(ordered, next, &works, work_list) {
 717                 list_del_init(&ordered->work_list);
 718                 wait_for_completion(&ordered->completion);
 719                 btrfs_put_ordered_extent(ordered);
 720                 cond_resched();
 721         }
 722         mutex_unlock(&root->ordered_extent_mutex);
 723
 724         return count;
 725 }
 726
 727 u64 btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
 728                              const u64 range_start, const u64 range_len)
 729 {
 730         struct btrfs_root *root;
 731         struct list_head splice;
 732         u64 total_done = 0;
 733         u64 done;
 734
 735         INIT_LIST_HEAD(&splice);
 736
 737         mutex_lock(&fs_info->ordered_operations_mutex);
 738         spin_lock(&fs_info->ordered_root_lock);
 739         list_splice_init(&fs_info->ordered_roots, &splice);
 740         while (!list_empty(&splice) && nr) {
 741                 root = list_first_entry(&splice, struct btrfs_root,
 742                                         ordered_root);
 743                 root = btrfs_grab_fs_root(root);
 744                 BUG_ON(!root);
 745                 list_move_tail(&root->ordered_root,
 746                                &fs_info->ordered_roots);
 747                 spin_unlock(&fs_info->ordered_root_lock);
 748
 749                 done = btrfs_wait_ordered_extents(root, nr,
 750                                                   range_start, range_len);
 751                 btrfs_put_fs_root(root);
 752                 total_done += done;
 753
 754                 spin_lock(&fs_info->ordered_root_lock);
 755                 if (nr != U64_MAX) {
 756                         nr -= done;
 757                 }
 758         }
 759         list_splice_tail(&splice, &fs_info->ordered_roots);
 760         spin_unlock(&fs_info->ordered_root_lock);
 761         mutex_unlock(&fs_info->ordered_operations_mutex);
 762
 763         return total_done;
 764 }
 765
 766 /*
 767  * Used to start IO or wait for a given ordered extent to finish.
 768  *
 769  * If wait is one, this effectively waits on page writeback for all the pages
 770  * in the extent, and it waits on the io completion code to insert
 771  * metadata into the btree corresponding to the extent
 772  */
 773 void btrfs_start_ordered_extent(struct inode *inode,
 774                                        struct btrfs_ordered_extent *entry,
 775                                        int wait)
 776 {
 777         u64 start = entry->file_offset;
 778         u64 end = start + entry->len - 1;
 779
 780         trace_btrfs_ordered_extent_start(inode, entry);
 781
 782         /*
 783          * pages in the range can be dirty, clean or writeback.  We
 784          * start IO on any dirty ones so the wait doesn't stall waiting
 785          * for the flusher thread to find them
 786          */
 787         if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
 788                 filemap_fdatawrite_range(inode->i_mapping, start, end);
 789         if (wait) {
 790                 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
 791                                                  &entry->flags));
 792         }
 793 }
 794
 795 /*
 796  * Used to wait on ordered extents across a large range of bytes.
 797  */
 798 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
 799 {
 800         int ret = 0;
 801         int ret_wb = 0;
 802         u64 end;
 803         u64 orig_end;
 804         struct btrfs_ordered_extent *ordered;
 805
 806         if (start + len < start) {
 807                 orig_end = INT_LIMIT(loff_t);
 808         } else {
 809                 orig_end = start + len - 1;
 810                 if (orig_end > INT_LIMIT(loff_t))
 811                         orig_end = INT_LIMIT(loff_t);
 812         }
 813
 814         /* start IO across the range first to instantiate any delalloc
 815          * extents
 816          */
 817         ret = btrfs_fdatawrite_range(inode, start, orig_end);
 818         if (ret)
 819                 return ret;
 820
 821         /*
 822          * If we have a writeback error don't return immediately. Wait first
 823          * for any ordered extents that haven't completed yet. This is to make
 824          * sure no one can dirty the same page ranges and call writepages()
 825          * before the ordered extents complete - to avoid failures (-EEXIST)
 826          * when adding the new ordered extents to the ordered tree.
 827          */
 828         ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
 829
 830         end = orig_end;
 831         while (1) {
 832                 ordered = btrfs_lookup_first_ordered_extent(inode, end);
 833                 if (!ordered)
 834                         break;
 835                 if (ordered->file_offset > orig_end) {
 836                         btrfs_put_ordered_extent(ordered);
 837                         break;
 838                 }
 839                 if (ordered->file_offset + ordered->len <= start) {
 840                         btrfs_put_ordered_extent(ordered);
 841                         break;
 842                 }
 843                 btrfs_start_ordered_extent(inode, ordered, 1);
 844                 end = ordered->file_offset;
 845                 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
 846                         ret = -EIO;
 847                 btrfs_put_ordered_extent(ordered);
 848                 if (ret || end == 0 || end == start)
 849                         break;
 850                 end--;
 851         }
 852         return ret_wb ? ret_wb : ret;
 853 }
 854
 855 /*
 856  * find an ordered extent corresponding to file_offset.  return NULL if
 857  * nothing is found, otherwise take a reference on the extent and return it
 858  */
 859 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
 860                                                          u64 file_offset)
 861 {
 862         struct btrfs_ordered_inode_tree *tree;
 863         struct rb_node *node;
 864         struct btrfs_ordered_extent *entry = NULL;
 865
 866         tree = &BTRFS_I(inode)->ordered_tree;
 867         spin_lock_irq(&tree->lock);
 868         node = tree_search(tree, file_offset);
 869         if (!node)
 870                 goto out;
 871
 872         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 873         if (!offset_in_entry(entry, file_offset))
 874                 entry = NULL;
 875         if (entry)
 876                 refcount_inc(&entry->refs);
 877 out:
 878         spin_unlock_irq(&tree->lock);
 879         return entry;
 880 }
 881
 882 /* Since the DIO code tries to lock a wide area we need to look for any ordered
 883  * extents that exist in the range, rather than just the start of the range.
 884  */
 885 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
 886                 struct btrfs_inode *inode, u64 file_offset, u64 len)
 887 {
 888         struct btrfs_ordered_inode_tree *tree;
 889         struct rb_node *node;
 890         struct btrfs_ordered_extent *entry = NULL;
 891
 892         tree = &inode->ordered_tree;
 893         spin_lock_irq(&tree->lock);
 894         node = tree_search(tree, file_offset);
 895         if (!node) {
 896                 node = tree_search(tree, file_offset + len);
 897                 if (!node)
 898                         goto out;
 899         }
 900
 901         while (1) {
 902                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 903                 if (range_overlaps(entry, file_offset, len))
 904                         break;
 905
 906                 if (entry->file_offset >= file_offset + len) {
 907                         entry = NULL;
 908                         break;
 909                 }
 910                 entry = NULL;
 911                 node = rb_next(node);
 912                 if (!node)
 913                         break;
 914         }
 915 out:
 916         if (entry)
 917                 refcount_inc(&entry->refs);
 918         spin_unlock_irq(&tree->lock);
 919         return entry;
 920 }
 921
 922 bool btrfs_have_ordered_extents_in_range(struct inode *inode,
 923                                          u64 file_offset,
 924                                          u64 len)
 925 {
 926         struct btrfs_ordered_extent *oe;
 927
 928         oe = btrfs_lookup_ordered_range(BTRFS_I(inode), file_offset, len);
 929         if (oe) {
 930                 btrfs_put_ordered_extent(oe);
 931                 return true;
 932         }
 933         return false;
 934 }
 935
 936 /*
 937  * lookup and return any extent before 'file_offset'.  NULL is returned
 938  * if none is found
 939  */
 940 struct btrfs_ordered_extent *
 941 btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
 942 {
 943         struct btrfs_ordered_inode_tree *tree;
 944         struct rb_node *node;
 945         struct btrfs_ordered_extent *entry = NULL;
 946
 947         tree = &BTRFS_I(inode)->ordered_tree;
 948         spin_lock_irq(&tree->lock);
 949         node = tree_search(tree, file_offset);
 950         if (!node)
 951                 goto out;
 952
 953         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 954         refcount_inc(&entry->refs);
 955 out:
 956         spin_unlock_irq(&tree->lock);
 957         return entry;
 958 }
 959
 960 /*
 961  * After an extent is done, call this to conditionally update the on disk
 962  * i_size.  i_size is updated to cover any fully written part of the file.
 963  */
 964 int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
 965                                 struct btrfs_ordered_extent *ordered)
 966 {
 967         struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
 968         u64 disk_i_size;
 969         u64 new_i_size;
 970         u64 i_size = i_size_read(inode);
 971         struct rb_node *node;
 972         struct rb_node *prev = NULL;
 973         struct btrfs_ordered_extent *test;
 974         int ret = 1;
 975         u64 orig_offset = offset;
 976
 977         spin_lock_irq(&tree->lock);
 978         if (ordered) {
 979                 offset = entry_end(ordered);
 980                 if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags))
 981                         offset = min(offset,
 982                                      ordered->file_offset +
 983                                      ordered->truncated_len);
 984         } else {
 985                 offset = ALIGN(offset, btrfs_inode_sectorsize(inode));
 986         }
 987         disk_i_size = BTRFS_I(inode)->disk_i_size;
 988
 989         /*
 990          * truncate file.
 991          * If ordered is not NULL, then this is called from endio and
 992          * disk_i_size will be updated by either truncate itself or any
 993          * in-flight IOs which are inside the disk_i_size.
 994          *
 995          * Because btrfs_setsize() may set i_size with disk_i_size if truncate
 996          * fails somehow, we need to make sure we have a precise disk_i_size by
 997          * updating it as usual.
 998          *
 999          */
1000         if (!ordered && disk_i_size > i_size) {
1001                 BTRFS_I(inode)->disk_i_size = orig_offset;
1002                 ret = 0;
1003                 goto out;
1004         }
1005
1006         /*
1007          * if the disk i_size is already at the inode->i_size, or
1008          * this ordered extent is inside the disk i_size, we're done
1009          */
1010         if (disk_i_size == i_size)
1011                 goto out;
1012
1013         /*
1014          * We still need to update disk_i_size if outstanding_isize is greater
1015          * than disk_i_size.
1016          */
1017         if (offset <= disk_i_size &&
1018             (!ordered || ordered->outstanding_isize <= disk_i_size))
1019                 goto out;
1020
1021         /*
1022          * walk backward from this ordered extent to disk_i_size.
1023          * if we find an ordered extent then we can't update disk i_size
1024          * yet
1025          */
1026         if (ordered) {
1027                 node = rb_prev(&ordered->rb_node);
1028         } else {
1029                 prev = tree_search(tree, offset);
1030                 /*
1031                  * we insert file extents without involving ordered struct,
1032                  * so there should be no ordered struct cover this offset
1033                  */
1034                 if (prev) {
1035                         test = rb_entry(prev, struct btrfs_ordered_extent,
1036                                         rb_node);
1037                         BUG_ON(offset_in_entry(test, offset));
1038                 }
1039                 node = prev;
1040         }
1041         for (; node; node = rb_prev(node)) {
1042                 test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
1043
1044                 /* We treat this entry as if it doesn't exist */
1045                 if (test_bit(BTRFS_ORDERED_UPDATED_ISIZE, &test->flags))
1046                         continue;
1047
1048                 if (entry_end(test) <= disk_i_size)
1049                         break;
1050                 if (test->file_offset >= i_size)
1051                         break;
1052
1053                 /*
1054                  * We don't update disk_i_size now, so record this undealt
1055                  * i_size. Or we will not know the real i_size.
1056                  */
1057                 if (test->outstanding_isize < offset)
1058                         test->outstanding_isize = offset;
1059                 if (ordered &&
1060                     ordered->outstanding_isize > test->outstanding_isize)
1061                         test->outstanding_isize = ordered->outstanding_isize;
1062                 goto out;
1063         }
1064         new_i_size = min_t(u64, offset, i_size);
1065
1066         /*
1067          * Some ordered extents may completed before the current one, and
1068          * we hold the real i_size in ->outstanding_isize.
1069          */
1070         if (ordered && ordered->outstanding_isize > new_i_size)
1071                 new_i_size = min_t(u64, ordered->outstanding_isize, i_size);
1072         BTRFS_I(inode)->disk_i_size = new_i_size;
1073         ret = 0;
1074 out:
1075         /*
1076          * We need to do this because we can't remove ordered extents until
1077          * after the i_disk_size has been updated and then the inode has been
1078          * updated to reflect the change, so we need to tell anybody who finds
1079          * this ordered extent that we've already done all the real work, we
1080          * just haven't completed all the other work.
1081          */
1082         if (ordered)
1083                 set_bit(BTRFS_ORDERED_UPDATED_ISIZE, &ordered->flags);
1084         spin_unlock_irq(&tree->lock);
1085         return ret;
1086 }
1087
1088 /*
1089  * search the ordered extents for one corresponding to 'offset' and
1090  * try to find a checksum.  This is used because we allow pages to
1091  * be reclaimed before their checksum is actually put into the btree
1092  */
1093 int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
1094                            u32 *sum, int len)
1095 {
1096         struct btrfs_ordered_sum *ordered_sum;
1097         struct btrfs_ordered_extent *ordered;
1098         struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
1099         unsigned long num_sectors;
1100         unsigned long i;
1101         u32 sectorsize = btrfs_inode_sectorsize(inode);
1102         int index = 0;
1103
1104         ordered = btrfs_lookup_ordered_extent(inode, offset);
1105         if (!ordered)
1106                 return 0;
1107
1108         spin_lock_irq(&tree->lock);
1109         list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
1110                 if (disk_bytenr >= ordered_sum->bytenr &&
1111                     disk_bytenr < ordered_sum->bytenr + ordered_sum->len) {
1112                         i = (disk_bytenr - ordered_sum->bytenr) >>
1113                             inode->i_sb->s_blocksize_bits;
1114                         num_sectors = ordered_sum->len >>
1115                                       inode->i_sb->s_blocksize_bits;
1116                         num_sectors = min_t(int, len - index, num_sectors - i);
1117                         memcpy(sum + index, ordered_sum->sums + i,
1118                                num_sectors);
1119
1120                         index += (int)num_sectors;
1121                         if (index == len)
1122                                 goto out;
1123                         disk_bytenr += num_sectors * sectorsize;
1124                 }
1125         }
1126 out:
1127         spin_unlock_irq(&tree->lock);
1128         btrfs_put_ordered_extent(ordered);
1129         return index;
1130 }
1131
1132 int __init ordered_data_init(void)
1133 {
1134         btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1135                                      sizeof(struct btrfs_ordered_extent), 0,
1136                                      SLAB_MEM_SPREAD,
1137                                      NULL);
1138         if (!btrfs_ordered_extent_cache)
1139                 return -ENOMEM;
1140
1141         return 0;
1142 }
1143
1144 void __cold ordered_data_exit(void)
1145 {
1146         kmem_cache_destroy(btrfs_ordered_extent_cache);
1147 }