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 #include "subpage.h"
  20
  21 static struct kmem_cache *btrfs_ordered_extent_cache;
  22
  23 static u64 entry_end(struct btrfs_ordered_extent *entry)
  24 {
  25         if (entry->file_offset + entry->num_bytes < entry->file_offset)
  26                 return (u64)-1;
  27         return entry->file_offset + entry->num_bytes;
  28 }
  29
  30 /* returns NULL if the insertion worked, or it returns the node it did find
  31  * in the tree
  32  */
  33 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
  34                                    struct rb_node *node)
  35 {
  36         struct rb_node **p = &root->rb_node;
  37         struct rb_node *parent = NULL;
  38         struct btrfs_ordered_extent *entry;
  39
  40         while (*p) {
  41                 parent = *p;
  42                 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
  43
  44                 if (file_offset < entry->file_offset)
  45                         p = &(*p)->rb_left;
  46                 else if (file_offset >= entry_end(entry))
  47                         p = &(*p)->rb_right;
  48                 else
  49                         return parent;
  50         }
  51
  52         rb_link_node(node, parent, p);
  53         rb_insert_color(node, root);
  54         return NULL;
  55 }
  56
  57 /*
  58  * look for a given offset in the tree, and if it can't be found return the
  59  * first lesser offset
  60  */
  61 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
  62                                      struct rb_node **prev_ret)
  63 {
  64         struct rb_node *n = root->rb_node;
  65         struct rb_node *prev = NULL;
  66         struct rb_node *test;
  67         struct btrfs_ordered_extent *entry;
  68         struct btrfs_ordered_extent *prev_entry = NULL;
  69
  70         while (n) {
  71                 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
  72                 prev = n;
  73                 prev_entry = entry;
  74
  75                 if (file_offset < entry->file_offset)
  76                         n = n->rb_left;
  77                 else if (file_offset >= entry_end(entry))
  78                         n = n->rb_right;
  79                 else
  80                         return n;
  81         }
  82         if (!prev_ret)
  83                 return NULL;
  84
  85         while (prev && file_offset >= entry_end(prev_entry)) {
  86                 test = rb_next(prev);
  87                 if (!test)
  88                         break;
  89                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
  90                                       rb_node);
  91                 if (file_offset < entry_end(prev_entry))
  92                         break;
  93
  94                 prev = test;
  95         }
  96         if (prev)
  97                 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
  98                                       rb_node);
  99         while (prev && file_offset < entry_end(prev_entry)) {
 100                 test = rb_prev(prev);
 101                 if (!test)
 102                         break;
 103                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
 104                                       rb_node);
 105                 prev = test;
 106         }
 107         *prev_ret = prev;
 108         return NULL;
 109 }
 110
 111 static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
 112                           u64 len)
 113 {
 114         if (file_offset + len <= entry->file_offset ||
 115             entry->file_offset + entry->num_bytes <= file_offset)
 116                 return 0;
 117         return 1;
 118 }
 119
 120 /*
 121  * look find the first ordered struct that has this offset, otherwise
 122  * the first one less than this offset
 123  */
 124 static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
 125                                           u64 file_offset)
 126 {
 127         struct rb_root *root = &tree->tree;
 128         struct rb_node *prev = NULL;
 129         struct rb_node *ret;
 130         struct btrfs_ordered_extent *entry;
 131
 132         if (tree->last) {
 133                 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
 134                                  rb_node);
 135                 if (in_range(file_offset, entry->file_offset, entry->num_bytes))
 136                         return tree->last;
 137         }
 138         ret = __tree_search(root, file_offset, &prev);
 139         if (!ret)
 140                 ret = prev;
 141         if (ret)
 142                 tree->last = ret;
 143         return ret;
 144 }
 145
 146 /**
 147  * Add an ordered extent to the per-inode tree.
 148  *
 149  * @inode:           Inode that this extent is for.
 150  * @file_offset:     Logical offset in file where the extent starts.
 151  * @num_bytes:       Logical length of extent in file.
 152  * @ram_bytes:       Full length of unencoded data.
 153  * @disk_bytenr:     Offset of extent on disk.
 154  * @disk_num_bytes:  Size of extent on disk.
 155  * @offset:          Offset into unencoded data where file data starts.
 156  * @flags:           Flags specifying type of extent (1 << BTRFS_ORDERED_*).
 157  * @compress_type:   Compression algorithm used for data.
 158  *
 159  * Most of these parameters correspond to &struct btrfs_file_extent_item. The
 160  * tree is given a single reference on the ordered extent that was inserted.
 161  *
 162  * Return: 0 or -ENOMEM.
 163  */
 164 int btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
 165                              u64 num_bytes, u64 ram_bytes, u64 disk_bytenr,
 166                              u64 disk_num_bytes, u64 offset, unsigned flags,
 167                              int compress_type)
 168 {
 169         struct btrfs_root *root = inode->root;
 170         struct btrfs_fs_info *fs_info = root->fs_info;
 171         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 172         struct rb_node *node;
 173         struct btrfs_ordered_extent *entry;
 174         int ret;
 175
 176         if (flags &
 177             ((1 << BTRFS_ORDERED_NOCOW) | (1 << BTRFS_ORDERED_PREALLOC))) {
 178                 /* For nocow write, we can release the qgroup rsv right now */
 179                 ret = btrfs_qgroup_free_data(inode, NULL, file_offset, num_bytes);
 180                 if (ret < 0)
 181                         return ret;
 182                 ret = 0;
 183         } else {
 184                 /*
 185                  * The ordered extent has reserved qgroup space, release now
 186                  * and pass the reserved number for qgroup_record to free.
 187                  */
 188                 ret = btrfs_qgroup_release_data(inode, file_offset, num_bytes);
 189                 if (ret < 0)
 190                         return ret;
 191         }
 192         entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
 193         if (!entry)
 194                 return -ENOMEM;
 195
 196         entry->file_offset = file_offset;
 197         entry->num_bytes = num_bytes;
 198         entry->ram_bytes = ram_bytes;
 199         entry->disk_bytenr = disk_bytenr;
 200         entry->disk_num_bytes = disk_num_bytes;
 201         entry->offset = offset;
 202         entry->bytes_left = num_bytes;
 203         entry->inode = igrab(&inode->vfs_inode);
 204         entry->compress_type = compress_type;
 205         entry->truncated_len = (u64)-1;
 206         entry->qgroup_rsv = ret;
 207         entry->physical = (u64)-1;
 208
 209         ASSERT((flags & ~BTRFS_ORDERED_TYPE_FLAGS) == 0);
 210         entry->flags = flags;
 211
 212         percpu_counter_add_batch(&fs_info->ordered_bytes, num_bytes,
 213                                  fs_info->delalloc_batch);
 214
 215         /* one ref for the tree */
 216         refcount_set(&entry->refs, 1);
 217         init_waitqueue_head(&entry->wait);
 218         INIT_LIST_HEAD(&entry->list);
 219         INIT_LIST_HEAD(&entry->log_list);
 220         INIT_LIST_HEAD(&entry->root_extent_list);
 221         INIT_LIST_HEAD(&entry->work_list);
 222         init_completion(&entry->completion);
 223
 224         trace_btrfs_ordered_extent_add(inode, entry);
 225
 226         spin_lock_irq(&tree->lock);
 227         node = tree_insert(&tree->tree, file_offset,
 228                            &entry->rb_node);
 229         if (node)
 230                 btrfs_panic(fs_info, -EEXIST,
 231                                 "inconsistency in ordered tree at offset %llu",
 232                                 file_offset);
 233         spin_unlock_irq(&tree->lock);
 234
 235         spin_lock(&root->ordered_extent_lock);
 236         list_add_tail(&entry->root_extent_list,
 237                       &root->ordered_extents);
 238         root->nr_ordered_extents++;
 239         if (root->nr_ordered_extents == 1) {
 240                 spin_lock(&fs_info->ordered_root_lock);
 241                 BUG_ON(!list_empty(&root->ordered_root));
 242                 list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
 243                 spin_unlock(&fs_info->ordered_root_lock);
 244         }
 245         spin_unlock(&root->ordered_extent_lock);
 246
 247         /*
 248          * We don't need the count_max_extents here, we can assume that all of
 249          * that work has been done at higher layers, so this is truly the
 250          * smallest the extent is going to get.
 251          */
 252         spin_lock(&inode->lock);
 253         btrfs_mod_outstanding_extents(inode, 1);
 254         spin_unlock(&inode->lock);
 255
 256         return 0;
 257 }
 258
 259 /*
 260  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
 261  * when an ordered extent is finished.  If the list covers more than one
 262  * ordered extent, it is split across multiples.
 263  */
 264 void btrfs_add_ordered_sum(struct btrfs_ordered_extent *entry,
 265                            struct btrfs_ordered_sum *sum)
 266 {
 267         struct btrfs_ordered_inode_tree *tree;
 268
 269         tree = &BTRFS_I(entry->inode)->ordered_tree;
 270         spin_lock_irq(&tree->lock);
 271         list_add_tail(&sum->list, &entry->list);
 272         spin_unlock_irq(&tree->lock);
 273 }
 274
 275 static void finish_ordered_fn(struct btrfs_work *work)
 276 {
 277         struct btrfs_ordered_extent *ordered_extent;
 278
 279         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
 280         btrfs_finish_ordered_io(ordered_extent);
 281 }
 282
 283 /*
 284  * Mark all ordered extents io inside the specified range finished.
 285  *
 286  * @page:        The involved page for the operation.
 287  *               For uncompressed buffered IO, the page status also needs to be
 288  *               updated to indicate whether the pending ordered io is finished.
 289  *               Can be NULL for direct IO and compressed write.
 290  *               For these cases, callers are ensured they won't execute the
 291  *               endio function twice.
 292  *
 293  * This function is called for endio, thus the range must have ordered
 294  * extent(s) covering it.
 295  */
 296 void btrfs_mark_ordered_io_finished(struct btrfs_inode *inode,
 297                                     struct page *page, u64 file_offset,
 298                                     u64 num_bytes, bool uptodate)
 299 {
 300         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 301         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 302         struct btrfs_workqueue *wq;
 303         struct rb_node *node;
 304         struct btrfs_ordered_extent *entry = NULL;
 305         unsigned long flags;
 306         u64 cur = file_offset;
 307
 308         if (btrfs_is_free_space_inode(inode))
 309                 wq = fs_info->endio_freespace_worker;
 310         else
 311                 wq = fs_info->endio_write_workers;
 312
 313         if (page)
 314                 ASSERT(page->mapping && page_offset(page) <= file_offset &&
 315                        file_offset + num_bytes <= page_offset(page) + PAGE_SIZE);
 316
 317         spin_lock_irqsave(&tree->lock, flags);
 318         while (cur < file_offset + num_bytes) {
 319                 u64 entry_end;
 320                 u64 end;
 321                 u32 len;
 322
 323                 node = tree_search(tree, cur);
 324                 /* No ordered extents at all */
 325                 if (!node)
 326                         break;
 327
 328                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 329                 entry_end = entry->file_offset + entry->num_bytes;
 330                 /*
 331                  * |<-- OE --->|  |
 332                  *                cur
 333                  * Go to next OE.
 334                  */
 335                 if (cur >= entry_end) {
 336                         node = rb_next(node);
 337                         /* No more ordered extents, exit */
 338                         if (!node)
 339                                 break;
 340                         entry = rb_entry(node, struct btrfs_ordered_extent,
 341                                          rb_node);
 342
 343                         /* Go to next ordered extent and continue */
 344                         cur = entry->file_offset;
 345                         continue;
 346                 }
 347                 /*
 348                  * |    |<--- OE --->|
 349                  * cur
 350                  * Go to the start of OE.
 351                  */
 352                 if (cur < entry->file_offset) {
 353                         cur = entry->file_offset;
 354                         continue;
 355                 }
 356
 357                 /*
 358                  * Now we are definitely inside one ordered extent.
 359                  *
 360                  * |<--- OE --->|
 361                  *      |
 362                  *      cur
 363                  */
 364                 end = min(entry->file_offset + entry->num_bytes,
 365                           file_offset + num_bytes) - 1;
 366                 ASSERT(end + 1 - cur < U32_MAX);
 367                 len = end + 1 - cur;
 368
 369                 if (page) {
 370                         /*
 371                          * Ordered (Private2) bit indicates whether we still
 372                          * have pending io unfinished for the ordered extent.
 373                          *
 374                          * If there's no such bit, we need to skip to next range.
 375                          */
 376                         if (!btrfs_page_test_ordered(fs_info, page, cur, len)) {
 377                                 cur += len;
 378                                 continue;
 379                         }
 380                         btrfs_page_clear_ordered(fs_info, page, cur, len);
 381                 }
 382
 383                 /* Now we're fine to update the accounting */
 384                 if (unlikely(len > entry->bytes_left)) {
 385                         WARN_ON(1);
 386                         btrfs_crit(fs_info,
 387 "bad ordered extent accounting, root=%llu ino=%llu OE offset=%llu OE len=%llu to_dec=%u left=%llu",
 388                                    inode->root->root_key.objectid,
 389                                    btrfs_ino(inode),
 390                                    entry->file_offset,
 391                                    entry->num_bytes,
 392                                    len, entry->bytes_left);
 393                         entry->bytes_left = 0;
 394                 } else {
 395                         entry->bytes_left -= len;
 396                 }
 397
 398                 if (!uptodate)
 399                         set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
 400
 401                 /*
 402                  * All the IO of the ordered extent is finished, we need to queue
 403                  * the finish_func to be executed.
 404                  */
 405                 if (entry->bytes_left == 0) {
 406                         set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 407                         cond_wake_up(&entry->wait);
 408                         refcount_inc(&entry->refs);
 409                         trace_btrfs_ordered_extent_mark_finished(inode, entry);
 410                         spin_unlock_irqrestore(&tree->lock, flags);
 411                         btrfs_init_work(&entry->work, finish_ordered_fn, NULL, NULL);
 412                         btrfs_queue_work(wq, &entry->work);
 413                         spin_lock_irqsave(&tree->lock, flags);
 414                 }
 415                 cur += len;
 416         }
 417         spin_unlock_irqrestore(&tree->lock, flags);
 418 }
 419
 420 /*
 421  * Finish IO for one ordered extent across a given range.  The range can only
 422  * contain one ordered extent.
 423  *
 424  * @cached:      The cached ordered extent. If not NULL, we can skip the tree
 425  *               search and use the ordered extent directly.
 426  *               Will be also used to store the finished ordered extent.
 427  * @file_offset: File offset for the finished IO
 428  * @io_size:     Length of the finish IO range
 429  *
 430  * Return true if the ordered extent is finished in the range, and update
 431  * @cached.
 432  * Return false otherwise.
 433  *
 434  * NOTE: The range can NOT cross multiple ordered extents.
 435  * Thus caller should ensure the range doesn't cross ordered extents.
 436  */
 437 bool btrfs_dec_test_ordered_pending(struct btrfs_inode *inode,
 438                                     struct btrfs_ordered_extent **cached,
 439                                     u64 file_offset, u64 io_size)
 440 {
 441         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 442         struct rb_node *node;
 443         struct btrfs_ordered_extent *entry = NULL;
 444         unsigned long flags;
 445         bool finished = false;
 446
 447         spin_lock_irqsave(&tree->lock, flags);
 448         if (cached && *cached) {
 449                 entry = *cached;
 450                 goto have_entry;
 451         }
 452
 453         node = tree_search(tree, file_offset);
 454         if (!node)
 455                 goto out;
 456
 457         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 458 have_entry:
 459         if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
 460                 goto out;
 461
 462         if (io_size > entry->bytes_left)
 463                 btrfs_crit(inode->root->fs_info,
 464                            "bad ordered accounting left %llu size %llu",
 465                        entry->bytes_left, io_size);
 466
 467         entry->bytes_left -= io_size;
 468
 469         if (entry->bytes_left == 0) {
 470                 /*
 471                  * Ensure only one caller can set the flag and finished_ret
 472                  * accordingly
 473                  */
 474                 finished = !test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 475                 /* test_and_set_bit implies a barrier */
 476                 cond_wake_up_nomb(&entry->wait);
 477         }
 478 out:
 479         if (finished && cached && entry) {
 480                 *cached = entry;
 481                 refcount_inc(&entry->refs);
 482                 trace_btrfs_ordered_extent_dec_test_pending(inode, entry);
 483         }
 484         spin_unlock_irqrestore(&tree->lock, flags);
 485         return finished;
 486 }
 487
 488 /*
 489  * used to drop a reference on an ordered extent.  This will free
 490  * the extent if the last reference is dropped
 491  */
 492 void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
 493 {
 494         struct list_head *cur;
 495         struct btrfs_ordered_sum *sum;
 496
 497         trace_btrfs_ordered_extent_put(BTRFS_I(entry->inode), entry);
 498
 499         if (refcount_dec_and_test(&entry->refs)) {
 500                 ASSERT(list_empty(&entry->root_extent_list));
 501                 ASSERT(list_empty(&entry->log_list));
 502                 ASSERT(RB_EMPTY_NODE(&entry->rb_node));
 503                 if (entry->inode)
 504                         btrfs_add_delayed_iput(entry->inode);
 505                 while (!list_empty(&entry->list)) {
 506                         cur = entry->list.next;
 507                         sum = list_entry(cur, struct btrfs_ordered_sum, list);
 508                         list_del(&sum->list);
 509                         kvfree(sum);
 510                 }
 511                 kmem_cache_free(btrfs_ordered_extent_cache, entry);
 512         }
 513 }
 514
 515 /*
 516  * remove an ordered extent from the tree.  No references are dropped
 517  * and waiters are woken up.
 518  */
 519 void btrfs_remove_ordered_extent(struct btrfs_inode *btrfs_inode,
 520                                  struct btrfs_ordered_extent *entry)
 521 {
 522         struct btrfs_ordered_inode_tree *tree;
 523         struct btrfs_root *root = btrfs_inode->root;
 524         struct btrfs_fs_info *fs_info = root->fs_info;
 525         struct rb_node *node;
 526         bool pending;
 527
 528         /* This is paired with btrfs_add_ordered_extent. */
 529         spin_lock(&btrfs_inode->lock);
 530         btrfs_mod_outstanding_extents(btrfs_inode, -1);
 531         spin_unlock(&btrfs_inode->lock);
 532         if (root != fs_info->tree_root) {
 533                 u64 release;
 534
 535                 if (test_bit(BTRFS_ORDERED_ENCODED, &entry->flags))
 536                         release = entry->disk_num_bytes;
 537                 else
 538                         release = entry->num_bytes;
 539                 btrfs_delalloc_release_metadata(btrfs_inode, release, false);
 540         }
 541
 542         percpu_counter_add_batch(&fs_info->ordered_bytes, -entry->num_bytes,
 543                                  fs_info->delalloc_batch);
 544
 545         tree = &btrfs_inode->ordered_tree;
 546         spin_lock_irq(&tree->lock);
 547         node = &entry->rb_node;
 548         rb_erase(node, &tree->tree);
 549         RB_CLEAR_NODE(node);
 550         if (tree->last == node)
 551                 tree->last = NULL;
 552         set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
 553         pending = test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags);
 554         spin_unlock_irq(&tree->lock);
 555
 556         /*
 557          * The current running transaction is waiting on us, we need to let it
 558          * know that we're complete and wake it up.
 559          */
 560         if (pending) {
 561                 struct btrfs_transaction *trans;
 562
 563                 /*
 564                  * The checks for trans are just a formality, it should be set,
 565                  * but if it isn't we don't want to deref/assert under the spin
 566                  * lock, so be nice and check if trans is set, but ASSERT() so
 567                  * if it isn't set a developer will notice.
 568                  */
 569                 spin_lock(&fs_info->trans_lock);
 570                 trans = fs_info->running_transaction;
 571                 if (trans)
 572                         refcount_inc(&trans->use_count);
 573                 spin_unlock(&fs_info->trans_lock);
 574
 575                 ASSERT(trans);
 576                 if (trans) {
 577                         if (atomic_dec_and_test(&trans->pending_ordered))
 578                                 wake_up(&trans->pending_wait);
 579                         btrfs_put_transaction(trans);
 580                 }
 581         }
 582
 583         spin_lock(&root->ordered_extent_lock);
 584         list_del_init(&entry->root_extent_list);
 585         root->nr_ordered_extents--;
 586
 587         trace_btrfs_ordered_extent_remove(btrfs_inode, entry);
 588
 589         if (!root->nr_ordered_extents) {
 590                 spin_lock(&fs_info->ordered_root_lock);
 591                 BUG_ON(list_empty(&root->ordered_root));
 592                 list_del_init(&root->ordered_root);
 593                 spin_unlock(&fs_info->ordered_root_lock);
 594         }
 595         spin_unlock(&root->ordered_extent_lock);
 596         wake_up(&entry->wait);
 597 }
 598
 599 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
 600 {
 601         struct btrfs_ordered_extent *ordered;
 602
 603         ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
 604         btrfs_start_ordered_extent(ordered, 1);
 605         complete(&ordered->completion);
 606 }
 607
 608 /*
 609  * wait for all the ordered extents in a root.  This is done when balancing
 610  * space between drives.
 611  */
 612 u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
 613                                const u64 range_start, const u64 range_len)
 614 {
 615         struct btrfs_fs_info *fs_info = root->fs_info;
 616         LIST_HEAD(splice);
 617         LIST_HEAD(skipped);
 618         LIST_HEAD(works);
 619         struct btrfs_ordered_extent *ordered, *next;
 620         u64 count = 0;
 621         const u64 range_end = range_start + range_len;
 622
 623         mutex_lock(&root->ordered_extent_mutex);
 624         spin_lock(&root->ordered_extent_lock);
 625         list_splice_init(&root->ordered_extents, &splice);
 626         while (!list_empty(&splice) && nr) {
 627                 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
 628                                            root_extent_list);
 629
 630                 if (range_end <= ordered->disk_bytenr ||
 631                     ordered->disk_bytenr + ordered->disk_num_bytes <= range_start) {
 632                         list_move_tail(&ordered->root_extent_list, &skipped);
 633                         cond_resched_lock(&root->ordered_extent_lock);
 634                         continue;
 635                 }
 636
 637                 list_move_tail(&ordered->root_extent_list,
 638                                &root->ordered_extents);
 639                 refcount_inc(&ordered->refs);
 640                 spin_unlock(&root->ordered_extent_lock);
 641
 642                 btrfs_init_work(&ordered->flush_work,
 643                                 btrfs_run_ordered_extent_work, NULL, NULL);
 644                 list_add_tail(&ordered->work_list, &works);
 645                 btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
 646
 647                 cond_resched();
 648                 spin_lock(&root->ordered_extent_lock);
 649                 if (nr != U64_MAX)
 650                         nr--;
 651                 count++;
 652         }
 653         list_splice_tail(&skipped, &root->ordered_extents);
 654         list_splice_tail(&splice, &root->ordered_extents);
 655         spin_unlock(&root->ordered_extent_lock);
 656
 657         list_for_each_entry_safe(ordered, next, &works, work_list) {
 658                 list_del_init(&ordered->work_list);
 659                 wait_for_completion(&ordered->completion);
 660                 btrfs_put_ordered_extent(ordered);
 661                 cond_resched();
 662         }
 663         mutex_unlock(&root->ordered_extent_mutex);
 664
 665         return count;
 666 }
 667
 668 void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
 669                              const u64 range_start, const u64 range_len)
 670 {
 671         struct btrfs_root *root;
 672         struct list_head splice;
 673         u64 done;
 674
 675         INIT_LIST_HEAD(&splice);
 676
 677         mutex_lock(&fs_info->ordered_operations_mutex);
 678         spin_lock(&fs_info->ordered_root_lock);
 679         list_splice_init(&fs_info->ordered_roots, &splice);
 680         while (!list_empty(&splice) && nr) {
 681                 root = list_first_entry(&splice, struct btrfs_root,
 682                                         ordered_root);
 683                 root = btrfs_grab_root(root);
 684                 BUG_ON(!root);
 685                 list_move_tail(&root->ordered_root,
 686                                &fs_info->ordered_roots);
 687                 spin_unlock(&fs_info->ordered_root_lock);
 688
 689                 done = btrfs_wait_ordered_extents(root, nr,
 690                                                   range_start, range_len);
 691                 btrfs_put_root(root);
 692
 693                 spin_lock(&fs_info->ordered_root_lock);
 694                 if (nr != U64_MAX) {
 695                         nr -= done;
 696                 }
 697         }
 698         list_splice_tail(&splice, &fs_info->ordered_roots);
 699         spin_unlock(&fs_info->ordered_root_lock);
 700         mutex_unlock(&fs_info->ordered_operations_mutex);
 701 }
 702
 703 /*
 704  * Used to start IO or wait for a given ordered extent to finish.
 705  *
 706  * If wait is one, this effectively waits on page writeback for all the pages
 707  * in the extent, and it waits on the io completion code to insert
 708  * metadata into the btree corresponding to the extent
 709  */
 710 void btrfs_start_ordered_extent(struct btrfs_ordered_extent *entry, int wait)
 711 {
 712         u64 start = entry->file_offset;
 713         u64 end = start + entry->num_bytes - 1;
 714         struct btrfs_inode *inode = BTRFS_I(entry->inode);
 715
 716         trace_btrfs_ordered_extent_start(inode, entry);
 717
 718         /*
 719          * pages in the range can be dirty, clean or writeback.  We
 720          * start IO on any dirty ones so the wait doesn't stall waiting
 721          * for the flusher thread to find them
 722          */
 723         if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
 724                 filemap_fdatawrite_range(inode->vfs_inode.i_mapping, start, end);
 725         if (wait) {
 726                 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
 727                                                  &entry->flags));
 728         }
 729 }
 730
 731 /*
 732  * Used to wait on ordered extents across a large range of bytes.
 733  */
 734 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
 735 {
 736         int ret = 0;
 737         int ret_wb = 0;
 738         u64 end;
 739         u64 orig_end;
 740         struct btrfs_ordered_extent *ordered;
 741
 742         if (start + len < start) {
 743                 orig_end = INT_LIMIT(loff_t);
 744         } else {
 745                 orig_end = start + len - 1;
 746                 if (orig_end > INT_LIMIT(loff_t))
 747                         orig_end = INT_LIMIT(loff_t);
 748         }
 749
 750         /* start IO across the range first to instantiate any delalloc
 751          * extents
 752          */
 753         ret = btrfs_fdatawrite_range(inode, start, orig_end);
 754         if (ret)
 755                 return ret;
 756
 757         /*
 758          * If we have a writeback error don't return immediately. Wait first
 759          * for any ordered extents that haven't completed yet. This is to make
 760          * sure no one can dirty the same page ranges and call writepages()
 761          * before the ordered extents complete - to avoid failures (-EEXIST)
 762          * when adding the new ordered extents to the ordered tree.
 763          */
 764         ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
 765
 766         end = orig_end;
 767         while (1) {
 768                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode), end);
 769                 if (!ordered)
 770                         break;
 771                 if (ordered->file_offset > orig_end) {
 772                         btrfs_put_ordered_extent(ordered);
 773                         break;
 774                 }
 775                 if (ordered->file_offset + ordered->num_bytes <= start) {
 776                         btrfs_put_ordered_extent(ordered);
 777                         break;
 778                 }
 779                 btrfs_start_ordered_extent(ordered, 1);
 780                 end = ordered->file_offset;
 781                 /*
 782                  * If the ordered extent had an error save the error but don't
 783                  * exit without waiting first for all other ordered extents in
 784                  * the range to complete.
 785                  */
 786                 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
 787                         ret = -EIO;
 788                 btrfs_put_ordered_extent(ordered);
 789                 if (end == 0 || end == start)
 790                         break;
 791                 end--;
 792         }
 793         return ret_wb ? ret_wb : ret;
 794 }
 795
 796 /*
 797  * find an ordered extent corresponding to file_offset.  return NULL if
 798  * nothing is found, otherwise take a reference on the extent and return it
 799  */
 800 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct btrfs_inode *inode,
 801                                                          u64 file_offset)
 802 {
 803         struct btrfs_ordered_inode_tree *tree;
 804         struct rb_node *node;
 805         struct btrfs_ordered_extent *entry = NULL;
 806         unsigned long flags;
 807
 808         tree = &inode->ordered_tree;
 809         spin_lock_irqsave(&tree->lock, flags);
 810         node = tree_search(tree, file_offset);
 811         if (!node)
 812                 goto out;
 813
 814         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 815         if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
 816                 entry = NULL;
 817         if (entry) {
 818                 refcount_inc(&entry->refs);
 819                 trace_btrfs_ordered_extent_lookup(inode, entry);
 820         }
 821 out:
 822         spin_unlock_irqrestore(&tree->lock, flags);
 823         return entry;
 824 }
 825
 826 /* Since the DIO code tries to lock a wide area we need to look for any ordered
 827  * extents that exist in the range, rather than just the start of the range.
 828  */
 829 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
 830                 struct btrfs_inode *inode, u64 file_offset, u64 len)
 831 {
 832         struct btrfs_ordered_inode_tree *tree;
 833         struct rb_node *node;
 834         struct btrfs_ordered_extent *entry = NULL;
 835
 836         tree = &inode->ordered_tree;
 837         spin_lock_irq(&tree->lock);
 838         node = tree_search(tree, file_offset);
 839         if (!node) {
 840                 node = tree_search(tree, file_offset + len);
 841                 if (!node)
 842                         goto out;
 843         }
 844
 845         while (1) {
 846                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 847                 if (range_overlaps(entry, file_offset, len))
 848                         break;
 849
 850                 if (entry->file_offset >= file_offset + len) {
 851                         entry = NULL;
 852                         break;
 853                 }
 854                 entry = NULL;
 855                 node = rb_next(node);
 856                 if (!node)
 857                         break;
 858         }
 859 out:
 860         if (entry) {
 861                 refcount_inc(&entry->refs);
 862                 trace_btrfs_ordered_extent_lookup_range(inode, entry);
 863         }
 864         spin_unlock_irq(&tree->lock);
 865         return entry;
 866 }
 867
 868 /*
 869  * Adds all ordered extents to the given list. The list ends up sorted by the
 870  * file_offset of the ordered extents.
 871  */
 872 void btrfs_get_ordered_extents_for_logging(struct btrfs_inode *inode,
 873                                            struct list_head *list)
 874 {
 875         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 876         struct rb_node *n;
 877
 878         ASSERT(inode_is_locked(&inode->vfs_inode));
 879
 880         spin_lock_irq(&tree->lock);
 881         for (n = rb_first(&tree->tree); n; n = rb_next(n)) {
 882                 struct btrfs_ordered_extent *ordered;
 883
 884                 ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
 885
 886                 if (test_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
 887                         continue;
 888
 889                 ASSERT(list_empty(&ordered->log_list));
 890                 list_add_tail(&ordered->log_list, list);
 891                 refcount_inc(&ordered->refs);
 892                 trace_btrfs_ordered_extent_lookup_for_logging(inode, ordered);
 893         }
 894         spin_unlock_irq(&tree->lock);
 895 }
 896
 897 /*
 898  * lookup and return any extent before 'file_offset'.  NULL is returned
 899  * if none is found
 900  */
 901 struct btrfs_ordered_extent *
 902 btrfs_lookup_first_ordered_extent(struct btrfs_inode *inode, u64 file_offset)
 903 {
 904         struct btrfs_ordered_inode_tree *tree;
 905         struct rb_node *node;
 906         struct btrfs_ordered_extent *entry = NULL;
 907
 908         tree = &inode->ordered_tree;
 909         spin_lock_irq(&tree->lock);
 910         node = tree_search(tree, file_offset);
 911         if (!node)
 912                 goto out;
 913
 914         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 915         refcount_inc(&entry->refs);
 916         trace_btrfs_ordered_extent_lookup_first(inode, entry);
 917 out:
 918         spin_unlock_irq(&tree->lock);
 919         return entry;
 920 }
 921
 922 /*
 923  * Lookup the first ordered extent that overlaps the range
 924  * [@file_offset, @file_offset + @len).
 925  *
 926  * The difference between this and btrfs_lookup_first_ordered_extent() is
 927  * that this one won't return any ordered extent that does not overlap the range.
 928  * And the difference against btrfs_lookup_ordered_extent() is, this function
 929  * ensures the first ordered extent gets returned.
 930  */
 931 struct btrfs_ordered_extent *btrfs_lookup_first_ordered_range(
 932                         struct btrfs_inode *inode, u64 file_offset, u64 len)
 933 {
 934         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 935         struct rb_node *node;
 936         struct rb_node *cur;
 937         struct rb_node *prev;
 938         struct rb_node *next;
 939         struct btrfs_ordered_extent *entry = NULL;
 940
 941         spin_lock_irq(&tree->lock);
 942         node = tree->tree.rb_node;
 943         /*
 944          * Here we don't want to use tree_search() which will use tree->last
 945          * and screw up the search order.
 946          * And __tree_search() can't return the adjacent ordered extents
 947          * either, thus here we do our own search.
 948          */
 949         while (node) {
 950                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 951
 952                 if (file_offset < entry->file_offset) {
 953                         node = node->rb_left;
 954                 } else if (file_offset >= entry_end(entry)) {
 955                         node = node->rb_right;
 956                 } else {
 957                         /*
 958                          * Direct hit, got an ordered extent that starts at
 959                          * @file_offset
 960                          */
 961                         goto out;
 962                 }
 963         }
 964         if (!entry) {
 965                 /* Empty tree */
 966                 goto out;
 967         }
 968
 969         cur = &entry->rb_node;
 970         /* We got an entry around @file_offset, check adjacent entries */
 971         if (entry->file_offset < file_offset) {
 972                 prev = cur;
 973                 next = rb_next(cur);
 974         } else {
 975                 prev = rb_prev(cur);
 976                 next = cur;
 977         }
 978         if (prev) {
 979                 entry = rb_entry(prev, struct btrfs_ordered_extent, rb_node);
 980                 if (range_overlaps(entry, file_offset, len))
 981                         goto out;
 982         }
 983         if (next) {
 984                 entry = rb_entry(next, struct btrfs_ordered_extent, rb_node);
 985                 if (range_overlaps(entry, file_offset, len))
 986                         goto out;
 987         }
 988         /* No ordered extent in the range */
 989         entry = NULL;
 990 out:
 991         if (entry) {
 992                 refcount_inc(&entry->refs);
 993                 trace_btrfs_ordered_extent_lookup_first_range(inode, entry);
 994         }
 995
 996         spin_unlock_irq(&tree->lock);
 997         return entry;
 998 }
 999
1000 /*
1001  * btrfs_flush_ordered_range - Lock the passed range and ensures all pending
1002  * ordered extents in it are run to completion.
1003  *
1004  * @inode:        Inode whose ordered tree is to be searched
1005  * @start:        Beginning of range to flush
1006  * @end:          Last byte of range to lock
1007  * @cached_state: If passed, will return the extent state responsible for the
1008  * locked range. It's the caller's responsibility to free the cached state.
1009  *
1010  * This function always returns with the given range locked, ensuring after it's
1011  * called no order extent can be pending.
1012  */
1013 void btrfs_lock_and_flush_ordered_range(struct btrfs_inode *inode, u64 start,
1014                                         u64 end,
1015                                         struct extent_state **cached_state)
1016 {
1017         struct btrfs_ordered_extent *ordered;
1018         struct extent_state *cache = NULL;
1019         struct extent_state **cachedp = &cache;
1020
1021         if (cached_state)
1022                 cachedp = cached_state;
1023
1024         while (1) {
1025                 lock_extent_bits(&inode->io_tree, start, end, cachedp);
1026                 ordered = btrfs_lookup_ordered_range(inode, start,
1027                                                      end - start + 1);
1028                 if (!ordered) {
1029                         /*
1030                          * If no external cached_state has been passed then
1031                          * decrement the extra ref taken for cachedp since we
1032                          * aren't exposing it outside of this function
1033                          */
1034                         if (!cached_state)
1035                                 refcount_dec(&cache->refs);
1036                         break;
1037                 }
1038                 unlock_extent_cached(&inode->io_tree, start, end, cachedp);
1039                 btrfs_start_ordered_extent(ordered, 1);
1040                 btrfs_put_ordered_extent(ordered);
1041         }
1042 }
1043
1044 static int clone_ordered_extent(struct btrfs_ordered_extent *ordered, u64 pos,
1045                                 u64 len)
1046 {
1047         struct inode *inode = ordered->inode;
1048         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1049         u64 file_offset = ordered->file_offset + pos;
1050         u64 disk_bytenr = ordered->disk_bytenr + pos;
1051         unsigned long flags = ordered->flags & BTRFS_ORDERED_TYPE_FLAGS;
1052
1053         /*
1054          * The splitting extent is already counted and will be added again in
1055          * btrfs_add_ordered_extent_*(). Subtract len to avoid double counting.
1056          */
1057         percpu_counter_add_batch(&fs_info->ordered_bytes, -len,
1058                                  fs_info->delalloc_batch);
1059         WARN_ON_ONCE(flags & (1 << BTRFS_ORDERED_COMPRESSED));
1060         return btrfs_add_ordered_extent(BTRFS_I(inode), file_offset, len, len,
1061                                         disk_bytenr, len, 0, flags,
1062                                         ordered->compress_type);
1063 }
1064
1065 int btrfs_split_ordered_extent(struct btrfs_ordered_extent *ordered, u64 pre,
1066                                 u64 post)
1067 {
1068         struct inode *inode = ordered->inode;
1069         struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
1070         struct rb_node *node;
1071         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1072         int ret = 0;
1073
1074         trace_btrfs_ordered_extent_split(BTRFS_I(inode), ordered);
1075
1076         spin_lock_irq(&tree->lock);
1077         /* Remove from tree once */
1078         node = &ordered->rb_node;
1079         rb_erase(node, &tree->tree);
1080         RB_CLEAR_NODE(node);
1081         if (tree->last == node)
1082                 tree->last = NULL;
1083
1084         ordered->file_offset += pre;
1085         ordered->disk_bytenr += pre;
1086         ordered->num_bytes -= (pre + post);
1087         ordered->disk_num_bytes -= (pre + post);
1088         ordered->bytes_left -= (pre + post);
1089
1090         /* Re-insert the node */
1091         node = tree_insert(&tree->tree, ordered->file_offset, &ordered->rb_node);
1092         if (node)
1093                 btrfs_panic(fs_info, -EEXIST,
1094                         "zoned: inconsistency in ordered tree at offset %llu",
1095                             ordered->file_offset);
1096
1097         spin_unlock_irq(&tree->lock);
1098
1099         if (pre)
1100                 ret = clone_ordered_extent(ordered, 0, pre);
1101         if (ret == 0 && post)
1102                 ret = clone_ordered_extent(ordered, pre + ordered->disk_num_bytes,
1103                                            post);
1104
1105         return ret;
1106 }
1107
1108 int __init ordered_data_init(void)
1109 {
1110         btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1111                                      sizeof(struct btrfs_ordered_extent), 0,
1112                                      SLAB_MEM_SPREAD,
1113                                      NULL);
1114         if (!btrfs_ordered_extent_cache)
1115                 return -ENOMEM;
1116
1117         return 0;
1118 }
1119
1120 void __cold ordered_data_exit(void)
1121 {
1122         kmem_cache_destroy(btrfs_ordered_extent_cache);
1123 }