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/sched/mm.h>
  10 #include "misc.h"
  11 #include "ctree.h"
  12 #include "transaction.h"
  13 #include "btrfs_inode.h"
  14 #include "extent_io.h"
  15 #include "disk-io.h"
  16 #include "compression.h"
  17 #include "delalloc-space.h"
  18 #include "qgroup.h"
  19
  20 static struct kmem_cache *btrfs_ordered_extent_cache;
  21
  22 static u64 entry_end(struct btrfs_ordered_extent *entry)
  23 {
  24         if (entry->file_offset + entry->num_bytes < entry->file_offset)
  25                 return (u64)-1;
  26         return entry->file_offset + entry->num_bytes;
  27 }
  28
  29 /* returns NULL if the insertion worked, or it returns the node it did find
  30  * in the tree
  31  */
  32 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
  33                                    struct rb_node *node)
  34 {
  35         struct rb_node **p = &root->rb_node;
  36         struct rb_node *parent = NULL;
  37         struct btrfs_ordered_extent *entry;
  38
  39         while (*p) {
  40                 parent = *p;
  41                 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
  42
  43                 if (file_offset < entry->file_offset)
  44                         p = &(*p)->rb_left;
  45                 else if (file_offset >= entry_end(entry))
  46                         p = &(*p)->rb_right;
  47                 else
  48                         return parent;
  49         }
  50
  51         rb_link_node(node, parent, p);
  52         rb_insert_color(node, root);
  53         return NULL;
  54 }
  55
  56 /*
  57  * look for a given offset in the tree, and if it can't be found return the
  58  * first lesser offset
  59  */
  60 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
  61                                      struct rb_node **prev_ret)
  62 {
  63         struct rb_node *n = root->rb_node;
  64         struct rb_node *prev = NULL;
  65         struct rb_node *test;
  66         struct btrfs_ordered_extent *entry;
  67         struct btrfs_ordered_extent *prev_entry = NULL;
  68
  69         while (n) {
  70                 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
  71                 prev = n;
  72                 prev_entry = entry;
  73
  74                 if (file_offset < entry->file_offset)
  75                         n = n->rb_left;
  76                 else if (file_offset >= entry_end(entry))
  77                         n = n->rb_right;
  78                 else
  79                         return n;
  80         }
  81         if (!prev_ret)
  82                 return NULL;
  83
  84         while (prev && file_offset >= entry_end(prev_entry)) {
  85                 test = rb_next(prev);
  86                 if (!test)
  87                         break;
  88                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
  89                                       rb_node);
  90                 if (file_offset < entry_end(prev_entry))
  91                         break;
  92
  93                 prev = test;
  94         }
  95         if (prev)
  96                 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
  97                                       rb_node);
  98         while (prev && file_offset < entry_end(prev_entry)) {
  99                 test = rb_prev(prev);
 100                 if (!test)
 101                         break;
 102                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
 103                                       rb_node);
 104                 prev = test;
 105         }
 106         *prev_ret = prev;
 107         return NULL;
 108 }
 109
 110 static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
 111                           u64 len)
 112 {
 113         if (file_offset + len <= entry->file_offset ||
 114             entry->file_offset + entry->num_bytes <= file_offset)
 115                 return 0;
 116         return 1;
 117 }
 118
 119 /*
 120  * look find the first ordered struct that has this offset, otherwise
 121  * the first one less than this offset
 122  */
 123 static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
 124                                           u64 file_offset)
 125 {
 126         struct rb_root *root = &tree->tree;
 127         struct rb_node *prev = NULL;
 128         struct rb_node *ret;
 129         struct btrfs_ordered_extent *entry;
 130
 131         if (tree->last) {
 132                 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
 133                                  rb_node);
 134                 if (in_range(file_offset, entry->file_offset, entry->num_bytes))
 135                         return tree->last;
 136         }
 137         ret = __tree_search(root, file_offset, &prev);
 138         if (!ret)
 139                 ret = prev;
 140         if (ret)
 141                 tree->last = ret;
 142         return ret;
 143 }
 144
 145 /*
 146  * Allocate and add a new ordered_extent into the per-inode tree.
 147  *
 148  * The tree is given a single reference on the ordered extent that was
 149  * inserted.
 150  */
 151 static int __btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
 152                                       u64 disk_bytenr, u64 num_bytes,
 153                                       u64 disk_num_bytes, int type, int dio,
 154                                       int compress_type)
 155 {
 156         struct btrfs_root *root = inode->root;
 157         struct btrfs_fs_info *fs_info = root->fs_info;
 158         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 159         struct rb_node *node;
 160         struct btrfs_ordered_extent *entry;
 161         int ret;
 162
 163         if (type == BTRFS_ORDERED_NOCOW || type == BTRFS_ORDERED_PREALLOC) {
 164                 /* For nocow write, we can release the qgroup rsv right now */
 165                 ret = btrfs_qgroup_free_data(inode, NULL, file_offset, num_bytes);
 166                 if (ret < 0)
 167                         return ret;
 168                 ret = 0;
 169         } else {
 170                 /*
 171                  * The ordered extent has reserved qgroup space, release now
 172                  * and pass the reserved number for qgroup_record to free.
 173                  */
 174                 ret = btrfs_qgroup_release_data(inode, file_offset, num_bytes);
 175                 if (ret < 0)
 176                         return ret;
 177         }
 178         entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
 179         if (!entry)
 180                 return -ENOMEM;
 181
 182         entry->file_offset = file_offset;
 183         entry->disk_bytenr = disk_bytenr;
 184         entry->num_bytes = num_bytes;
 185         entry->disk_num_bytes = disk_num_bytes;
 186         entry->bytes_left = num_bytes;
 187         entry->inode = igrab(&inode->vfs_inode);
 188         entry->compress_type = compress_type;
 189         entry->truncated_len = (u64)-1;
 190         entry->qgroup_rsv = ret;
 191         entry->physical = (u64)-1;
 192         entry->disk = NULL;
 193         entry->partno = (u8)-1;
 194
 195         ASSERT(type == BTRFS_ORDERED_REGULAR ||
 196                type == BTRFS_ORDERED_NOCOW ||
 197                type == BTRFS_ORDERED_PREALLOC ||
 198                type == BTRFS_ORDERED_COMPRESSED);
 199         set_bit(type, &entry->flags);
 200
 201         percpu_counter_add_batch(&fs_info->ordered_bytes, num_bytes,
 202                                  fs_info->delalloc_batch);
 203
 204         if (dio)
 205                 set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
 206
 207         /* one ref for the tree */
 208         refcount_set(&entry->refs, 1);
 209         init_waitqueue_head(&entry->wait);
 210         INIT_LIST_HEAD(&entry->list);
 211         INIT_LIST_HEAD(&entry->log_list);
 212         INIT_LIST_HEAD(&entry->root_extent_list);
 213         INIT_LIST_HEAD(&entry->work_list);
 214         init_completion(&entry->completion);
 215
 216         trace_btrfs_ordered_extent_add(inode, entry);
 217
 218         spin_lock_irq(&tree->lock);
 219         node = tree_insert(&tree->tree, file_offset,
 220                            &entry->rb_node);
 221         if (node)
 222                 btrfs_panic(fs_info, -EEXIST,
 223                                 "inconsistency in ordered tree at offset %llu",
 224                                 file_offset);
 225         spin_unlock_irq(&tree->lock);
 226
 227         spin_lock(&root->ordered_extent_lock);
 228         list_add_tail(&entry->root_extent_list,
 229                       &root->ordered_extents);
 230         root->nr_ordered_extents++;
 231         if (root->nr_ordered_extents == 1) {
 232                 spin_lock(&fs_info->ordered_root_lock);
 233                 BUG_ON(!list_empty(&root->ordered_root));
 234                 list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
 235                 spin_unlock(&fs_info->ordered_root_lock);
 236         }
 237         spin_unlock(&root->ordered_extent_lock);
 238
 239         /*
 240          * We don't need the count_max_extents here, we can assume that all of
 241          * that work has been done at higher layers, so this is truly the
 242          * smallest the extent is going to get.
 243          */
 244         spin_lock(&inode->lock);
 245         btrfs_mod_outstanding_extents(inode, 1);
 246         spin_unlock(&inode->lock);
 247
 248         return 0;
 249 }
 250
 251 int btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
 252                              u64 disk_bytenr, u64 num_bytes, u64 disk_num_bytes,
 253                              int type)
 254 {
 255         ASSERT(type == BTRFS_ORDERED_REGULAR ||
 256                type == BTRFS_ORDERED_NOCOW ||
 257                type == BTRFS_ORDERED_PREALLOC);
 258         return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
 259                                           num_bytes, disk_num_bytes, type, 0,
 260                                           BTRFS_COMPRESS_NONE);
 261 }
 262
 263 int btrfs_add_ordered_extent_dio(struct btrfs_inode *inode, u64 file_offset,
 264                                  u64 disk_bytenr, u64 num_bytes,
 265                                  u64 disk_num_bytes, int type)
 266 {
 267         ASSERT(type == BTRFS_ORDERED_REGULAR ||
 268                type == BTRFS_ORDERED_NOCOW ||
 269                type == BTRFS_ORDERED_PREALLOC);
 270         return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
 271                                           num_bytes, disk_num_bytes, type, 1,
 272                                           BTRFS_COMPRESS_NONE);
 273 }
 274
 275 int btrfs_add_ordered_extent_compress(struct btrfs_inode *inode, u64 file_offset,
 276                                       u64 disk_bytenr, u64 num_bytes,
 277                                       u64 disk_num_bytes, int compress_type)
 278 {
 279         ASSERT(compress_type != BTRFS_COMPRESS_NONE);
 280         return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
 281                                           num_bytes, disk_num_bytes,
 282                                           BTRFS_ORDERED_COMPRESSED, 0,
 283                                           compress_type);
 284 }
 285
 286 /*
 287  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
 288  * when an ordered extent is finished.  If the list covers more than one
 289  * ordered extent, it is split across multiples.
 290  */
 291 void btrfs_add_ordered_sum(struct btrfs_ordered_extent *entry,
 292                            struct btrfs_ordered_sum *sum)
 293 {
 294         struct btrfs_ordered_inode_tree *tree;
 295
 296         tree = &BTRFS_I(entry->inode)->ordered_tree;
 297         spin_lock_irq(&tree->lock);
 298         list_add_tail(&sum->list, &entry->list);
 299         spin_unlock_irq(&tree->lock);
 300 }
 301
 302 /*
 303  * Finish IO for one ordered extent across a given range.  The range can
 304  * contain several ordered extents.
 305  *
 306  * @found_ret:   Return the finished ordered extent
 307  * @file_offset: File offset for the finished IO
 308  *               Will also be updated to one byte past the range that is
 309  *               recordered as finished. This allows caller to walk forward.
 310  * @io_size:     Length of the finish IO range
 311  * @uptodate:    If the IO finished without problem
 312  *
 313  * Return true if any ordered extent is finished in the range, and update
 314  * @found_ret and @file_offset.
 315  * Return false otherwise.
 316  *
 317  * NOTE: Although The range can cross multiple ordered extents, only one
 318  * ordered extent will be updated during one call. The caller is responsible to
 319  * iterate all ordered extents in the range.
 320  */
 321 bool btrfs_dec_test_first_ordered_pending(struct btrfs_inode *inode,
 322                                    struct btrfs_ordered_extent **finished_ret,
 323                                    u64 *file_offset, u64 io_size, int uptodate)
 324 {
 325         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 326         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 327         struct rb_node *node;
 328         struct btrfs_ordered_extent *entry = NULL;
 329         bool finished = false;
 330         unsigned long flags;
 331         u64 dec_end;
 332         u64 dec_start;
 333         u64 to_dec;
 334
 335         spin_lock_irqsave(&tree->lock, flags);
 336         node = tree_search(tree, *file_offset);
 337         if (!node)
 338                 goto out;
 339
 340         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 341         if (!in_range(*file_offset, entry->file_offset, entry->num_bytes))
 342                 goto out;
 343
 344         dec_start = max(*file_offset, entry->file_offset);
 345         dec_end = min(*file_offset + io_size,
 346                       entry->file_offset + entry->num_bytes);
 347         *file_offset = dec_end;
 348         if (dec_start > dec_end) {
 349                 btrfs_crit(fs_info, "bad ordering dec_start %llu end %llu",
 350                            dec_start, dec_end);
 351         }
 352         to_dec = dec_end - dec_start;
 353         if (to_dec > entry->bytes_left) {
 354                 btrfs_crit(fs_info,
 355                            "bad ordered accounting left %llu size %llu",
 356                            entry->bytes_left, to_dec);
 357         }
 358         entry->bytes_left -= to_dec;
 359         if (!uptodate)
 360                 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
 361
 362         if (entry->bytes_left == 0) {
 363                 /*
 364                  * Ensure only one caller can set the flag and finished_ret
 365                  * accordingly
 366                  */
 367                 finished = !test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 368                 /* test_and_set_bit implies a barrier */
 369                 cond_wake_up_nomb(&entry->wait);
 370         }
 371 out:
 372         if (finished && finished_ret && entry) {
 373                 *finished_ret = entry;
 374                 refcount_inc(&entry->refs);
 375         }
 376         spin_unlock_irqrestore(&tree->lock, flags);
 377         return finished;
 378 }
 379
 380 /*
 381  * Finish IO for one ordered extent across a given range.  The range can only
 382  * contain one ordered extent.
 383  *
 384  * @cached:      The cached ordered extent. If not NULL, we can skip the tree
 385  *               search and use the ordered extent directly.
 386  *               Will be also used to store the finished ordered extent.
 387  * @file_offset: File offset for the finished IO
 388  * @io_size:     Length of the finish IO range
 389  * @uptodate:    If the IO finishes without problem
 390  *
 391  * Return true if the ordered extent is finished in the range, and update
 392  * @cached.
 393  * Return false otherwise.
 394  *
 395  * NOTE: The range can NOT cross multiple ordered extents.
 396  * Thus caller should ensure the range doesn't cross ordered extents.
 397  */
 398 bool btrfs_dec_test_ordered_pending(struct btrfs_inode *inode,
 399                                     struct btrfs_ordered_extent **cached,
 400                                     u64 file_offset, u64 io_size, int uptodate)
 401 {
 402         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 403         struct rb_node *node;
 404         struct btrfs_ordered_extent *entry = NULL;
 405         unsigned long flags;
 406         bool finished = false;
 407
 408         spin_lock_irqsave(&tree->lock, flags);
 409         if (cached && *cached) {
 410                 entry = *cached;
 411                 goto have_entry;
 412         }
 413
 414         node = tree_search(tree, file_offset);
 415         if (!node)
 416                 goto out;
 417
 418         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 419 have_entry:
 420         if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
 421                 goto out;
 422
 423         if (io_size > entry->bytes_left)
 424                 btrfs_crit(inode->root->fs_info,
 425                            "bad ordered accounting left %llu size %llu",
 426                        entry->bytes_left, io_size);
 427
 428         entry->bytes_left -= io_size;
 429         if (!uptodate)
 430                 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
 431
 432         if (entry->bytes_left == 0) {
 433                 /*
 434                  * Ensure only one caller can set the flag and finished_ret
 435                  * accordingly
 436                  */
 437                 finished = !test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 438                 /* test_and_set_bit implies a barrier */
 439                 cond_wake_up_nomb(&entry->wait);
 440         }
 441 out:
 442         if (finished && cached && entry) {
 443                 *cached = entry;
 444                 refcount_inc(&entry->refs);
 445         }
 446         spin_unlock_irqrestore(&tree->lock, flags);
 447         return finished;
 448 }
 449
 450 /*
 451  * used to drop a reference on an ordered extent.  This will free
 452  * the extent if the last reference is dropped
 453  */
 454 void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
 455 {
 456         struct list_head *cur;
 457         struct btrfs_ordered_sum *sum;
 458
 459         trace_btrfs_ordered_extent_put(BTRFS_I(entry->inode), entry);
 460
 461         if (refcount_dec_and_test(&entry->refs)) {
 462                 ASSERT(list_empty(&entry->root_extent_list));
 463                 ASSERT(list_empty(&entry->log_list));
 464                 ASSERT(RB_EMPTY_NODE(&entry->rb_node));
 465                 if (entry->inode)
 466                         btrfs_add_delayed_iput(entry->inode);
 467                 while (!list_empty(&entry->list)) {
 468                         cur = entry->list.next;
 469                         sum = list_entry(cur, struct btrfs_ordered_sum, list);
 470                         list_del(&sum->list);
 471                         kvfree(sum);
 472                 }
 473                 kmem_cache_free(btrfs_ordered_extent_cache, entry);
 474         }
 475 }
 476
 477 /*
 478  * remove an ordered extent from the tree.  No references are dropped
 479  * and waiters are woken up.
 480  */
 481 void btrfs_remove_ordered_extent(struct btrfs_inode *btrfs_inode,
 482                                  struct btrfs_ordered_extent *entry)
 483 {
 484         struct btrfs_ordered_inode_tree *tree;
 485         struct btrfs_root *root = btrfs_inode->root;
 486         struct btrfs_fs_info *fs_info = root->fs_info;
 487         struct rb_node *node;
 488         bool pending;
 489
 490         /* This is paired with btrfs_add_ordered_extent. */
 491         spin_lock(&btrfs_inode->lock);
 492         btrfs_mod_outstanding_extents(btrfs_inode, -1);
 493         spin_unlock(&btrfs_inode->lock);
 494         if (root != fs_info->tree_root)
 495                 btrfs_delalloc_release_metadata(btrfs_inode, entry->num_bytes,
 496                                                 false);
 497
 498         percpu_counter_add_batch(&fs_info->ordered_bytes, -entry->num_bytes,
 499                                  fs_info->delalloc_batch);
 500
 501         tree = &btrfs_inode->ordered_tree;
 502         spin_lock_irq(&tree->lock);
 503         node = &entry->rb_node;
 504         rb_erase(node, &tree->tree);
 505         RB_CLEAR_NODE(node);
 506         if (tree->last == node)
 507                 tree->last = NULL;
 508         set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
 509         pending = test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags);
 510         spin_unlock_irq(&tree->lock);
 511
 512         /*
 513          * The current running transaction is waiting on us, we need to let it
 514          * know that we're complete and wake it up.
 515          */
 516         if (pending) {
 517                 struct btrfs_transaction *trans;
 518
 519                 /*
 520                  * The checks for trans are just a formality, it should be set,
 521                  * but if it isn't we don't want to deref/assert under the spin
 522                  * lock, so be nice and check if trans is set, but ASSERT() so
 523                  * if it isn't set a developer will notice.
 524                  */
 525                 spin_lock(&fs_info->trans_lock);
 526                 trans = fs_info->running_transaction;
 527                 if (trans)
 528                         refcount_inc(&trans->use_count);
 529                 spin_unlock(&fs_info->trans_lock);
 530
 531                 ASSERT(trans);
 532                 if (trans) {
 533                         if (atomic_dec_and_test(&trans->pending_ordered))
 534                                 wake_up(&trans->pending_wait);
 535                         btrfs_put_transaction(trans);
 536                 }
 537         }
 538
 539         spin_lock(&root->ordered_extent_lock);
 540         list_del_init(&entry->root_extent_list);
 541         root->nr_ordered_extents--;
 542
 543         trace_btrfs_ordered_extent_remove(btrfs_inode, entry);
 544
 545         if (!root->nr_ordered_extents) {
 546                 spin_lock(&fs_info->ordered_root_lock);
 547                 BUG_ON(list_empty(&root->ordered_root));
 548                 list_del_init(&root->ordered_root);
 549                 spin_unlock(&fs_info->ordered_root_lock);
 550         }
 551         spin_unlock(&root->ordered_extent_lock);
 552         wake_up(&entry->wait);
 553 }
 554
 555 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
 556 {
 557         struct btrfs_ordered_extent *ordered;
 558
 559         ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
 560         btrfs_start_ordered_extent(ordered, 1);
 561         complete(&ordered->completion);
 562 }
 563
 564 /*
 565  * wait for all the ordered extents in a root.  This is done when balancing
 566  * space between drives.
 567  */
 568 u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
 569                                const u64 range_start, const u64 range_len)
 570 {
 571         struct btrfs_fs_info *fs_info = root->fs_info;
 572         LIST_HEAD(splice);
 573         LIST_HEAD(skipped);
 574         LIST_HEAD(works);
 575         struct btrfs_ordered_extent *ordered, *next;
 576         u64 count = 0;
 577         const u64 range_end = range_start + range_len;
 578
 579         mutex_lock(&root->ordered_extent_mutex);
 580         spin_lock(&root->ordered_extent_lock);
 581         list_splice_init(&root->ordered_extents, &splice);
 582         while (!list_empty(&splice) && nr) {
 583                 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
 584                                            root_extent_list);
 585
 586                 if (range_end <= ordered->disk_bytenr ||
 587                     ordered->disk_bytenr + ordered->disk_num_bytes <= range_start) {
 588                         list_move_tail(&ordered->root_extent_list, &skipped);
 589                         cond_resched_lock(&root->ordered_extent_lock);
 590                         continue;
 591                 }
 592
 593                 list_move_tail(&ordered->root_extent_list,
 594                                &root->ordered_extents);
 595                 refcount_inc(&ordered->refs);
 596                 spin_unlock(&root->ordered_extent_lock);
 597
 598                 btrfs_init_work(&ordered->flush_work,
 599                                 btrfs_run_ordered_extent_work, NULL, NULL);
 600                 list_add_tail(&ordered->work_list, &works);
 601                 btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
 602
 603                 cond_resched();
 604                 spin_lock(&root->ordered_extent_lock);
 605                 if (nr != U64_MAX)
 606                         nr--;
 607                 count++;
 608         }
 609         list_splice_tail(&skipped, &root->ordered_extents);
 610         list_splice_tail(&splice, &root->ordered_extents);
 611         spin_unlock(&root->ordered_extent_lock);
 612
 613         list_for_each_entry_safe(ordered, next, &works, work_list) {
 614                 list_del_init(&ordered->work_list);
 615                 wait_for_completion(&ordered->completion);
 616                 btrfs_put_ordered_extent(ordered);
 617                 cond_resched();
 618         }
 619         mutex_unlock(&root->ordered_extent_mutex);
 620
 621         return count;
 622 }
 623
 624 void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
 625                              const u64 range_start, const u64 range_len)
 626 {
 627         struct btrfs_root *root;
 628         struct list_head splice;
 629         u64 done;
 630
 631         INIT_LIST_HEAD(&splice);
 632
 633         mutex_lock(&fs_info->ordered_operations_mutex);
 634         spin_lock(&fs_info->ordered_root_lock);
 635         list_splice_init(&fs_info->ordered_roots, &splice);
 636         while (!list_empty(&splice) && nr) {
 637                 root = list_first_entry(&splice, struct btrfs_root,
 638                                         ordered_root);
 639                 root = btrfs_grab_root(root);
 640                 BUG_ON(!root);
 641                 list_move_tail(&root->ordered_root,
 642                                &fs_info->ordered_roots);
 643                 spin_unlock(&fs_info->ordered_root_lock);
 644
 645                 done = btrfs_wait_ordered_extents(root, nr,
 646                                                   range_start, range_len);
 647                 btrfs_put_root(root);
 648
 649                 spin_lock(&fs_info->ordered_root_lock);
 650                 if (nr != U64_MAX) {
 651                         nr -= done;
 652                 }
 653         }
 654         list_splice_tail(&splice, &fs_info->ordered_roots);
 655         spin_unlock(&fs_info->ordered_root_lock);
 656         mutex_unlock(&fs_info->ordered_operations_mutex);
 657 }
 658
 659 /*
 660  * Used to start IO or wait for a given ordered extent to finish.
 661  *
 662  * If wait is one, this effectively waits on page writeback for all the pages
 663  * in the extent, and it waits on the io completion code to insert
 664  * metadata into the btree corresponding to the extent
 665  */
 666 void btrfs_start_ordered_extent(struct btrfs_ordered_extent *entry, int wait)
 667 {
 668         u64 start = entry->file_offset;
 669         u64 end = start + entry->num_bytes - 1;
 670         struct btrfs_inode *inode = BTRFS_I(entry->inode);
 671
 672         trace_btrfs_ordered_extent_start(inode, entry);
 673
 674         /*
 675          * pages in the range can be dirty, clean or writeback.  We
 676          * start IO on any dirty ones so the wait doesn't stall waiting
 677          * for the flusher thread to find them
 678          */
 679         if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
 680                 filemap_fdatawrite_range(inode->vfs_inode.i_mapping, start, end);
 681         if (wait) {
 682                 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
 683                                                  &entry->flags));
 684         }
 685 }
 686
 687 /*
 688  * Used to wait on ordered extents across a large range of bytes.
 689  */
 690 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
 691 {
 692         int ret = 0;
 693         int ret_wb = 0;
 694         u64 end;
 695         u64 orig_end;
 696         struct btrfs_ordered_extent *ordered;
 697
 698         if (start + len < start) {
 699                 orig_end = INT_LIMIT(loff_t);
 700         } else {
 701                 orig_end = start + len - 1;
 702                 if (orig_end > INT_LIMIT(loff_t))
 703                         orig_end = INT_LIMIT(loff_t);
 704         }
 705
 706         /* start IO across the range first to instantiate any delalloc
 707          * extents
 708          */
 709         ret = btrfs_fdatawrite_range(inode, start, orig_end);
 710         if (ret)
 711                 return ret;
 712
 713         /*
 714          * If we have a writeback error don't return immediately. Wait first
 715          * for any ordered extents that haven't completed yet. This is to make
 716          * sure no one can dirty the same page ranges and call writepages()
 717          * before the ordered extents complete - to avoid failures (-EEXIST)
 718          * when adding the new ordered extents to the ordered tree.
 719          */
 720         ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
 721
 722         end = orig_end;
 723         while (1) {
 724                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode), end);
 725                 if (!ordered)
 726                         break;
 727                 if (ordered->file_offset > orig_end) {
 728                         btrfs_put_ordered_extent(ordered);
 729                         break;
 730                 }
 731                 if (ordered->file_offset + ordered->num_bytes <= start) {
 732                         btrfs_put_ordered_extent(ordered);
 733                         break;
 734                 }
 735                 btrfs_start_ordered_extent(ordered, 1);
 736                 end = ordered->file_offset;
 737                 /*
 738                  * If the ordered extent had an error save the error but don't
 739                  * exit without waiting first for all other ordered extents in
 740                  * the range to complete.
 741                  */
 742                 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
 743                         ret = -EIO;
 744                 btrfs_put_ordered_extent(ordered);
 745                 if (end == 0 || end == start)
 746                         break;
 747                 end--;
 748         }
 749         return ret_wb ? ret_wb : ret;
 750 }
 751
 752 /*
 753  * find an ordered extent corresponding to file_offset.  return NULL if
 754  * nothing is found, otherwise take a reference on the extent and return it
 755  */
 756 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct btrfs_inode *inode,
 757                                                          u64 file_offset)
 758 {
 759         struct btrfs_ordered_inode_tree *tree;
 760         struct rb_node *node;
 761         struct btrfs_ordered_extent *entry = NULL;
 762         unsigned long flags;
 763
 764         tree = &inode->ordered_tree;
 765         spin_lock_irqsave(&tree->lock, flags);
 766         node = tree_search(tree, file_offset);
 767         if (!node)
 768                 goto out;
 769
 770         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 771         if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
 772                 entry = NULL;
 773         if (entry)
 774                 refcount_inc(&entry->refs);
 775 out:
 776         spin_unlock_irqrestore(&tree->lock, flags);
 777         return entry;
 778 }
 779
 780 /* Since the DIO code tries to lock a wide area we need to look for any ordered
 781  * extents that exist in the range, rather than just the start of the range.
 782  */
 783 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
 784                 struct btrfs_inode *inode, u64 file_offset, u64 len)
 785 {
 786         struct btrfs_ordered_inode_tree *tree;
 787         struct rb_node *node;
 788         struct btrfs_ordered_extent *entry = NULL;
 789
 790         tree = &inode->ordered_tree;
 791         spin_lock_irq(&tree->lock);
 792         node = tree_search(tree, file_offset);
 793         if (!node) {
 794                 node = tree_search(tree, file_offset + len);
 795                 if (!node)
 796                         goto out;
 797         }
 798
 799         while (1) {
 800                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 801                 if (range_overlaps(entry, file_offset, len))
 802                         break;
 803
 804                 if (entry->file_offset >= file_offset + len) {
 805                         entry = NULL;
 806                         break;
 807                 }
 808                 entry = NULL;
 809                 node = rb_next(node);
 810                 if (!node)
 811                         break;
 812         }
 813 out:
 814         if (entry)
 815                 refcount_inc(&entry->refs);
 816         spin_unlock_irq(&tree->lock);
 817         return entry;
 818 }
 819
 820 /*
 821  * Adds all ordered extents to the given list. The list ends up sorted by the
 822  * file_offset of the ordered extents.
 823  */
 824 void btrfs_get_ordered_extents_for_logging(struct btrfs_inode *inode,
 825                                            struct list_head *list)
 826 {
 827         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 828         struct rb_node *n;
 829
 830         ASSERT(inode_is_locked(&inode->vfs_inode));
 831
 832         spin_lock_irq(&tree->lock);
 833         for (n = rb_first(&tree->tree); n; n = rb_next(n)) {
 834                 struct btrfs_ordered_extent *ordered;
 835
 836                 ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
 837
 838                 if (test_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
 839                         continue;
 840
 841                 ASSERT(list_empty(&ordered->log_list));
 842                 list_add_tail(&ordered->log_list, list);
 843                 refcount_inc(&ordered->refs);
 844         }
 845         spin_unlock_irq(&tree->lock);
 846 }
 847
 848 /*
 849  * lookup and return any extent before 'file_offset'.  NULL is returned
 850  * if none is found
 851  */
 852 struct btrfs_ordered_extent *
 853 btrfs_lookup_first_ordered_extent(struct btrfs_inode *inode, u64 file_offset)
 854 {
 855         struct btrfs_ordered_inode_tree *tree;
 856         struct rb_node *node;
 857         struct btrfs_ordered_extent *entry = NULL;
 858
 859         tree = &inode->ordered_tree;
 860         spin_lock_irq(&tree->lock);
 861         node = tree_search(tree, file_offset);
 862         if (!node)
 863                 goto out;
 864
 865         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 866         refcount_inc(&entry->refs);
 867 out:
 868         spin_unlock_irq(&tree->lock);
 869         return entry;
 870 }
 871
 872 /*
 873  * btrfs_flush_ordered_range - Lock the passed range and ensures all pending
 874  * ordered extents in it are run to completion.
 875  *
 876  * @inode:        Inode whose ordered tree is to be searched
 877  * @start:        Beginning of range to flush
 878  * @end:          Last byte of range to lock
 879  * @cached_state: If passed, will return the extent state responsible for the
 880  * locked range. It's the caller's responsibility to free the cached state.
 881  *
 882  * This function always returns with the given range locked, ensuring after it's
 883  * called no order extent can be pending.
 884  */
 885 void btrfs_lock_and_flush_ordered_range(struct btrfs_inode *inode, u64 start,
 886                                         u64 end,
 887                                         struct extent_state **cached_state)
 888 {
 889         struct btrfs_ordered_extent *ordered;
 890         struct extent_state *cache = NULL;
 891         struct extent_state **cachedp = &cache;
 892
 893         if (cached_state)
 894                 cachedp = cached_state;
 895
 896         while (1) {
 897                 lock_extent_bits(&inode->io_tree, start, end, cachedp);
 898                 ordered = btrfs_lookup_ordered_range(inode, start,
 899                                                      end - start + 1);
 900                 if (!ordered) {
 901                         /*
 902                          * If no external cached_state has been passed then
 903                          * decrement the extra ref taken for cachedp since we
 904                          * aren't exposing it outside of this function
 905                          */
 906                         if (!cached_state)
 907                                 refcount_dec(&cache->refs);
 908                         break;
 909                 }
 910                 unlock_extent_cached(&inode->io_tree, start, end, cachedp);
 911                 btrfs_start_ordered_extent(ordered, 1);
 912                 btrfs_put_ordered_extent(ordered);
 913         }
 914 }
 915
 916 static int clone_ordered_extent(struct btrfs_ordered_extent *ordered, u64 pos,
 917                                 u64 len)
 918 {
 919         struct inode *inode = ordered->inode;
 920         u64 file_offset = ordered->file_offset + pos;
 921         u64 disk_bytenr = ordered->disk_bytenr + pos;
 922         u64 num_bytes = len;
 923         u64 disk_num_bytes = len;
 924         int type;
 925         unsigned long flags_masked = ordered->flags & ~(1 << BTRFS_ORDERED_DIRECT);
 926         int compress_type = ordered->compress_type;
 927         unsigned long weight;
 928         int ret;
 929
 930         weight = hweight_long(flags_masked);
 931         WARN_ON_ONCE(weight > 1);
 932         if (!weight)
 933                 type = 0;
 934         else
 935                 type = __ffs(flags_masked);
 936
 937         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered->flags)) {
 938                 WARN_ON_ONCE(1);
 939                 ret = btrfs_add_ordered_extent_compress(BTRFS_I(inode),
 940                                 file_offset, disk_bytenr, num_bytes,
 941                                 disk_num_bytes, compress_type);
 942         } else if (test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags)) {
 943                 ret = btrfs_add_ordered_extent_dio(BTRFS_I(inode), file_offset,
 944                                 disk_bytenr, num_bytes, disk_num_bytes, type);
 945         } else {
 946                 ret = btrfs_add_ordered_extent(BTRFS_I(inode), file_offset,
 947                                 disk_bytenr, num_bytes, disk_num_bytes, type);
 948         }
 949
 950         return ret;
 951 }
 952
 953 int btrfs_split_ordered_extent(struct btrfs_ordered_extent *ordered, u64 pre,
 954                                 u64 post)
 955 {
 956         struct inode *inode = ordered->inode;
 957         struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
 958         struct rb_node *node;
 959         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 960         int ret = 0;
 961
 962         spin_lock_irq(&tree->lock);
 963         /* Remove from tree once */
 964         node = &ordered->rb_node;
 965         rb_erase(node, &tree->tree);
 966         RB_CLEAR_NODE(node);
 967         if (tree->last == node)
 968                 tree->last = NULL;
 969
 970         ordered->file_offset += pre;
 971         ordered->disk_bytenr += pre;
 972         ordered->num_bytes -= (pre + post);
 973         ordered->disk_num_bytes -= (pre + post);
 974         ordered->bytes_left -= (pre + post);
 975
 976         /* Re-insert the node */
 977         node = tree_insert(&tree->tree, ordered->file_offset, &ordered->rb_node);
 978         if (node)
 979                 btrfs_panic(fs_info, -EEXIST,
 980                         "zoned: inconsistency in ordered tree at offset %llu",
 981                             ordered->file_offset);
 982
 983         spin_unlock_irq(&tree->lock);
 984
 985         if (pre)
 986                 ret = clone_ordered_extent(ordered, 0, pre);
 987         if (ret == 0 && post)
 988                 ret = clone_ordered_extent(ordered, pre + ordered->disk_num_bytes,
 989                                            post);
 990
 991         return ret;
 992 }
 993
 994 int __init ordered_data_init(void)
 995 {
 996         btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
 997                                      sizeof(struct btrfs_ordered_extent), 0,
 998                                      SLAB_MEM_SPREAD,
 999                                      NULL);
1000         if (!btrfs_ordered_extent_cache)
1001                 return -ENOMEM;
1002
1003         return 0;
1004 }
1005
1006 void __cold ordered_data_exit(void)
1007 {
1008         kmem_cache_destroy(btrfs_ordered_extent_cache);
1009 }