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         bool freespace_inode;
 528
 529         /*
 530          * If this is a free space inode the thread has not acquired the ordered
 531          * extents lockdep map.
 532          */
 533         freespace_inode = btrfs_is_free_space_inode(btrfs_inode);
 534
 535         btrfs_lockdep_acquire(fs_info, btrfs_trans_pending_ordered);
 536         /* This is paired with btrfs_add_ordered_extent. */
 537         spin_lock(&btrfs_inode->lock);
 538         btrfs_mod_outstanding_extents(btrfs_inode, -1);
 539         spin_unlock(&btrfs_inode->lock);
 540         if (root != fs_info->tree_root) {
 541                 u64 release;
 542
 543                 if (test_bit(BTRFS_ORDERED_ENCODED, &entry->flags))
 544                         release = entry->disk_num_bytes;
 545                 else
 546                         release = entry->num_bytes;
 547                 btrfs_delalloc_release_metadata(btrfs_inode, release, false);
 548         }
 549
 550         percpu_counter_add_batch(&fs_info->ordered_bytes, -entry->num_bytes,
 551                                  fs_info->delalloc_batch);
 552
 553         tree = &btrfs_inode->ordered_tree;
 554         spin_lock_irq(&tree->lock);
 555         node = &entry->rb_node;
 556         rb_erase(node, &tree->tree);
 557         RB_CLEAR_NODE(node);
 558         if (tree->last == node)
 559                 tree->last = NULL;
 560         set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
 561         pending = test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags);
 562         spin_unlock_irq(&tree->lock);
 563
 564         /*
 565          * The current running transaction is waiting on us, we need to let it
 566          * know that we're complete and wake it up.
 567          */
 568         if (pending) {
 569                 struct btrfs_transaction *trans;
 570
 571                 /*
 572                  * The checks for trans are just a formality, it should be set,
 573                  * but if it isn't we don't want to deref/assert under the spin
 574                  * lock, so be nice and check if trans is set, but ASSERT() so
 575                  * if it isn't set a developer will notice.
 576                  */
 577                 spin_lock(&fs_info->trans_lock);
 578                 trans = fs_info->running_transaction;
 579                 if (trans)
 580                         refcount_inc(&trans->use_count);
 581                 spin_unlock(&fs_info->trans_lock);
 582
 583                 ASSERT(trans);
 584                 if (trans) {
 585                         if (atomic_dec_and_test(&trans->pending_ordered))
 586                                 wake_up(&trans->pending_wait);
 587                         btrfs_put_transaction(trans);
 588                 }
 589         }
 590
 591         btrfs_lockdep_release(fs_info, btrfs_trans_pending_ordered);
 592
 593         spin_lock(&root->ordered_extent_lock);
 594         list_del_init(&entry->root_extent_list);
 595         root->nr_ordered_extents--;
 596
 597         trace_btrfs_ordered_extent_remove(btrfs_inode, entry);
 598
 599         if (!root->nr_ordered_extents) {
 600                 spin_lock(&fs_info->ordered_root_lock);
 601                 BUG_ON(list_empty(&root->ordered_root));
 602                 list_del_init(&root->ordered_root);
 603                 spin_unlock(&fs_info->ordered_root_lock);
 604         }
 605         spin_unlock(&root->ordered_extent_lock);
 606         wake_up(&entry->wait);
 607         if (!freespace_inode)
 608                 btrfs_lockdep_release(fs_info, btrfs_ordered_extent);
 609 }
 610
 611 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
 612 {
 613         struct btrfs_ordered_extent *ordered;
 614
 615         ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
 616         btrfs_start_ordered_extent(ordered, 1);
 617         complete(&ordered->completion);
 618 }
 619
 620 /*
 621  * wait for all the ordered extents in a root.  This is done when balancing
 622  * space between drives.
 623  */
 624 u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
 625                                const u64 range_start, const u64 range_len)
 626 {
 627         struct btrfs_fs_info *fs_info = root->fs_info;
 628         LIST_HEAD(splice);
 629         LIST_HEAD(skipped);
 630         LIST_HEAD(works);
 631         struct btrfs_ordered_extent *ordered, *next;
 632         u64 count = 0;
 633         const u64 range_end = range_start + range_len;
 634
 635         mutex_lock(&root->ordered_extent_mutex);
 636         spin_lock(&root->ordered_extent_lock);
 637         list_splice_init(&root->ordered_extents, &splice);
 638         while (!list_empty(&splice) && nr) {
 639                 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
 640                                            root_extent_list);
 641
 642                 if (range_end <= ordered->disk_bytenr ||
 643                     ordered->disk_bytenr + ordered->disk_num_bytes <= range_start) {
 644                         list_move_tail(&ordered->root_extent_list, &skipped);
 645                         cond_resched_lock(&root->ordered_extent_lock);
 646                         continue;
 647                 }
 648
 649                 list_move_tail(&ordered->root_extent_list,
 650                                &root->ordered_extents);
 651                 refcount_inc(&ordered->refs);
 652                 spin_unlock(&root->ordered_extent_lock);
 653
 654                 btrfs_init_work(&ordered->flush_work,
 655                                 btrfs_run_ordered_extent_work, NULL, NULL);
 656                 list_add_tail(&ordered->work_list, &works);
 657                 btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
 658
 659                 cond_resched();
 660                 spin_lock(&root->ordered_extent_lock);
 661                 if (nr != U64_MAX)
 662                         nr--;
 663                 count++;
 664         }
 665         list_splice_tail(&skipped, &root->ordered_extents);
 666         list_splice_tail(&splice, &root->ordered_extents);
 667         spin_unlock(&root->ordered_extent_lock);
 668
 669         list_for_each_entry_safe(ordered, next, &works, work_list) {
 670                 list_del_init(&ordered->work_list);
 671                 wait_for_completion(&ordered->completion);
 672                 btrfs_put_ordered_extent(ordered);
 673                 cond_resched();
 674         }
 675         mutex_unlock(&root->ordered_extent_mutex);
 676
 677         return count;
 678 }
 679
 680 void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
 681                              const u64 range_start, const u64 range_len)
 682 {
 683         struct btrfs_root *root;
 684         struct list_head splice;
 685         u64 done;
 686
 687         INIT_LIST_HEAD(&splice);
 688
 689         mutex_lock(&fs_info->ordered_operations_mutex);
 690         spin_lock(&fs_info->ordered_root_lock);
 691         list_splice_init(&fs_info->ordered_roots, &splice);
 692         while (!list_empty(&splice) && nr) {
 693                 root = list_first_entry(&splice, struct btrfs_root,
 694                                         ordered_root);
 695                 root = btrfs_grab_root(root);
 696                 BUG_ON(!root);
 697                 list_move_tail(&root->ordered_root,
 698                                &fs_info->ordered_roots);
 699                 spin_unlock(&fs_info->ordered_root_lock);
 700
 701                 done = btrfs_wait_ordered_extents(root, nr,
 702                                                   range_start, range_len);
 703                 btrfs_put_root(root);
 704
 705                 spin_lock(&fs_info->ordered_root_lock);
 706                 if (nr != U64_MAX) {
 707                         nr -= done;
 708                 }
 709         }
 710         list_splice_tail(&splice, &fs_info->ordered_roots);
 711         spin_unlock(&fs_info->ordered_root_lock);
 712         mutex_unlock(&fs_info->ordered_operations_mutex);
 713 }
 714
 715 /*
 716  * Used to start IO or wait for a given ordered extent to finish.
 717  *
 718  * If wait is one, this effectively waits on page writeback for all the pages
 719  * in the extent, and it waits on the io completion code to insert
 720  * metadata into the btree corresponding to the extent
 721  */
 722 void btrfs_start_ordered_extent(struct btrfs_ordered_extent *entry, int wait)
 723 {
 724         u64 start = entry->file_offset;
 725         u64 end = start + entry->num_bytes - 1;
 726         struct btrfs_inode *inode = BTRFS_I(entry->inode);
 727         bool freespace_inode;
 728
 729         trace_btrfs_ordered_extent_start(inode, entry);
 730
 731         /*
 732          * If this is a free space inode do not take the ordered extents lockdep
 733          * map.
 734          */
 735         freespace_inode = btrfs_is_free_space_inode(inode);
 736
 737         /*
 738          * pages in the range can be dirty, clean or writeback.  We
 739          * start IO on any dirty ones so the wait doesn't stall waiting
 740          * for the flusher thread to find them
 741          */
 742         if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
 743                 filemap_fdatawrite_range(inode->vfs_inode.i_mapping, start, end);
 744         if (wait) {
 745                 if (!freespace_inode)
 746                         btrfs_might_wait_for_event(inode->root->fs_info, btrfs_ordered_extent);
 747                 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
 748                                                  &entry->flags));
 749         }
 750 }
 751
 752 /*
 753  * Used to wait on ordered extents across a large range of bytes.
 754  */
 755 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
 756 {
 757         int ret = 0;
 758         int ret_wb = 0;
 759         u64 end;
 760         u64 orig_end;
 761         struct btrfs_ordered_extent *ordered;
 762
 763         if (start + len < start) {
 764                 orig_end = INT_LIMIT(loff_t);
 765         } else {
 766                 orig_end = start + len - 1;
 767                 if (orig_end > INT_LIMIT(loff_t))
 768                         orig_end = INT_LIMIT(loff_t);
 769         }
 770
 771         /* start IO across the range first to instantiate any delalloc
 772          * extents
 773          */
 774         ret = btrfs_fdatawrite_range(inode, start, orig_end);
 775         if (ret)
 776                 return ret;
 777
 778         /*
 779          * If we have a writeback error don't return immediately. Wait first
 780          * for any ordered extents that haven't completed yet. This is to make
 781          * sure no one can dirty the same page ranges and call writepages()
 782          * before the ordered extents complete - to avoid failures (-EEXIST)
 783          * when adding the new ordered extents to the ordered tree.
 784          */
 785         ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
 786
 787         end = orig_end;
 788         while (1) {
 789                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode), end);
 790                 if (!ordered)
 791                         break;
 792                 if (ordered->file_offset > orig_end) {
 793                         btrfs_put_ordered_extent(ordered);
 794                         break;
 795                 }
 796                 if (ordered->file_offset + ordered->num_bytes <= start) {
 797                         btrfs_put_ordered_extent(ordered);
 798                         break;
 799                 }
 800                 btrfs_start_ordered_extent(ordered, 1);
 801                 end = ordered->file_offset;
 802                 /*
 803                  * If the ordered extent had an error save the error but don't
 804                  * exit without waiting first for all other ordered extents in
 805                  * the range to complete.
 806                  */
 807                 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
 808                         ret = -EIO;
 809                 btrfs_put_ordered_extent(ordered);
 810                 if (end == 0 || end == start)
 811                         break;
 812                 end--;
 813         }
 814         return ret_wb ? ret_wb : ret;
 815 }
 816
 817 /*
 818  * find an ordered extent corresponding to file_offset.  return NULL if
 819  * nothing is found, otherwise take a reference on the extent and return it
 820  */
 821 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct btrfs_inode *inode,
 822                                                          u64 file_offset)
 823 {
 824         struct btrfs_ordered_inode_tree *tree;
 825         struct rb_node *node;
 826         struct btrfs_ordered_extent *entry = NULL;
 827         unsigned long flags;
 828
 829         tree = &inode->ordered_tree;
 830         spin_lock_irqsave(&tree->lock, flags);
 831         node = tree_search(tree, file_offset);
 832         if (!node)
 833                 goto out;
 834
 835         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 836         if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
 837                 entry = NULL;
 838         if (entry) {
 839                 refcount_inc(&entry->refs);
 840                 trace_btrfs_ordered_extent_lookup(inode, entry);
 841         }
 842 out:
 843         spin_unlock_irqrestore(&tree->lock, flags);
 844         return entry;
 845 }
 846
 847 /* Since the DIO code tries to lock a wide area we need to look for any ordered
 848  * extents that exist in the range, rather than just the start of the range.
 849  */
 850 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
 851                 struct btrfs_inode *inode, u64 file_offset, u64 len)
 852 {
 853         struct btrfs_ordered_inode_tree *tree;
 854         struct rb_node *node;
 855         struct btrfs_ordered_extent *entry = NULL;
 856
 857         tree = &inode->ordered_tree;
 858         spin_lock_irq(&tree->lock);
 859         node = tree_search(tree, file_offset);
 860         if (!node) {
 861                 node = tree_search(tree, file_offset + len);
 862                 if (!node)
 863                         goto out;
 864         }
 865
 866         while (1) {
 867                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 868                 if (range_overlaps(entry, file_offset, len))
 869                         break;
 870
 871                 if (entry->file_offset >= file_offset + len) {
 872                         entry = NULL;
 873                         break;
 874                 }
 875                 entry = NULL;
 876                 node = rb_next(node);
 877                 if (!node)
 878                         break;
 879         }
 880 out:
 881         if (entry) {
 882                 refcount_inc(&entry->refs);
 883                 trace_btrfs_ordered_extent_lookup_range(inode, entry);
 884         }
 885         spin_unlock_irq(&tree->lock);
 886         return entry;
 887 }
 888
 889 /*
 890  * Adds all ordered extents to the given list. The list ends up sorted by the
 891  * file_offset of the ordered extents.
 892  */
 893 void btrfs_get_ordered_extents_for_logging(struct btrfs_inode *inode,
 894                                            struct list_head *list)
 895 {
 896         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 897         struct rb_node *n;
 898
 899         ASSERT(inode_is_locked(&inode->vfs_inode));
 900
 901         spin_lock_irq(&tree->lock);
 902         for (n = rb_first(&tree->tree); n; n = rb_next(n)) {
 903                 struct btrfs_ordered_extent *ordered;
 904
 905                 ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
 906
 907                 if (test_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
 908                         continue;
 909
 910                 ASSERT(list_empty(&ordered->log_list));
 911                 list_add_tail(&ordered->log_list, list);
 912                 refcount_inc(&ordered->refs);
 913                 trace_btrfs_ordered_extent_lookup_for_logging(inode, ordered);
 914         }
 915         spin_unlock_irq(&tree->lock);
 916 }
 917
 918 /*
 919  * lookup and return any extent before 'file_offset'.  NULL is returned
 920  * if none is found
 921  */
 922 struct btrfs_ordered_extent *
 923 btrfs_lookup_first_ordered_extent(struct btrfs_inode *inode, u64 file_offset)
 924 {
 925         struct btrfs_ordered_inode_tree *tree;
 926         struct rb_node *node;
 927         struct btrfs_ordered_extent *entry = NULL;
 928
 929         tree = &inode->ordered_tree;
 930         spin_lock_irq(&tree->lock);
 931         node = tree_search(tree, file_offset);
 932         if (!node)
 933                 goto out;
 934
 935         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 936         refcount_inc(&entry->refs);
 937         trace_btrfs_ordered_extent_lookup_first(inode, entry);
 938 out:
 939         spin_unlock_irq(&tree->lock);
 940         return entry;
 941 }
 942
 943 /*
 944  * Lookup the first ordered extent that overlaps the range
 945  * [@file_offset, @file_offset + @len).
 946  *
 947  * The difference between this and btrfs_lookup_first_ordered_extent() is
 948  * that this one won't return any ordered extent that does not overlap the range.
 949  * And the difference against btrfs_lookup_ordered_extent() is, this function
 950  * ensures the first ordered extent gets returned.
 951  */
 952 struct btrfs_ordered_extent *btrfs_lookup_first_ordered_range(
 953                         struct btrfs_inode *inode, u64 file_offset, u64 len)
 954 {
 955         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 956         struct rb_node *node;
 957         struct rb_node *cur;
 958         struct rb_node *prev;
 959         struct rb_node *next;
 960         struct btrfs_ordered_extent *entry = NULL;
 961
 962         spin_lock_irq(&tree->lock);
 963         node = tree->tree.rb_node;
 964         /*
 965          * Here we don't want to use tree_search() which will use tree->last
 966          * and screw up the search order.
 967          * And __tree_search() can't return the adjacent ordered extents
 968          * either, thus here we do our own search.
 969          */
 970         while (node) {
 971                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 972
 973                 if (file_offset < entry->file_offset) {
 974                         node = node->rb_left;
 975                 } else if (file_offset >= entry_end(entry)) {
 976                         node = node->rb_right;
 977                 } else {
 978                         /*
 979                          * Direct hit, got an ordered extent that starts at
 980                          * @file_offset
 981                          */
 982                         goto out;
 983                 }
 984         }
 985         if (!entry) {
 986                 /* Empty tree */
 987                 goto out;
 988         }
 989
 990         cur = &entry->rb_node;
 991         /* We got an entry around @file_offset, check adjacent entries */
 992         if (entry->file_offset < file_offset) {
 993                 prev = cur;
 994                 next = rb_next(cur);
 995         } else {
 996                 prev = rb_prev(cur);
 997                 next = cur;
 998         }
 999         if (prev) {
1000                 entry = rb_entry(prev, struct btrfs_ordered_extent, rb_node);
1001                 if (range_overlaps(entry, file_offset, len))
1002                         goto out;
1003         }
1004         if (next) {
1005                 entry = rb_entry(next, struct btrfs_ordered_extent, rb_node);
1006                 if (range_overlaps(entry, file_offset, len))
1007                         goto out;
1008         }
1009         /* No ordered extent in the range */
1010         entry = NULL;
1011 out:
1012         if (entry) {
1013                 refcount_inc(&entry->refs);
1014                 trace_btrfs_ordered_extent_lookup_first_range(inode, entry);
1015         }
1016
1017         spin_unlock_irq(&tree->lock);
1018         return entry;
1019 }
1020
1021 /*
1022  * btrfs_flush_ordered_range - Lock the passed range and ensures all pending
1023  * ordered extents in it are run to completion.
1024  *
1025  * @inode:        Inode whose ordered tree is to be searched
1026  * @start:        Beginning of range to flush
1027  * @end:          Last byte of range to lock
1028  * @cached_state: If passed, will return the extent state responsible for the
1029  * locked range. It's the caller's responsibility to free the cached state.
1030  *
1031  * This function always returns with the given range locked, ensuring after it's
1032  * called no order extent can be pending.
1033  */
1034 void btrfs_lock_and_flush_ordered_range(struct btrfs_inode *inode, u64 start,
1035                                         u64 end,
1036                                         struct extent_state **cached_state)
1037 {
1038         struct btrfs_ordered_extent *ordered;
1039         struct extent_state *cache = NULL;
1040         struct extent_state **cachedp = &cache;
1041
1042         if (cached_state)
1043                 cachedp = cached_state;
1044
1045         while (1) {
1046                 lock_extent(&inode->io_tree, start, end, cachedp);
1047                 ordered = btrfs_lookup_ordered_range(inode, start,
1048                                                      end - start + 1);
1049                 if (!ordered) {
1050                         /*
1051                          * If no external cached_state has been passed then
1052                          * decrement the extra ref taken for cachedp since we
1053                          * aren't exposing it outside of this function
1054                          */
1055                         if (!cached_state)
1056                                 refcount_dec(&cache->refs);
1057                         break;
1058                 }
1059                 unlock_extent(&inode->io_tree, start, end, cachedp);
1060                 btrfs_start_ordered_extent(ordered, 1);
1061                 btrfs_put_ordered_extent(ordered);
1062         }
1063 }
1064
1065 /*
1066  * Lock the passed range and ensure all pending ordered extents in it are run
1067  * to completion in nowait mode.
1068  *
1069  * Return true if btrfs_lock_ordered_range does not return any extents,
1070  * otherwise false.
1071  */
1072 bool btrfs_try_lock_ordered_range(struct btrfs_inode *inode, u64 start, u64 end)
1073 {
1074         struct btrfs_ordered_extent *ordered;
1075
1076         if (!try_lock_extent(&inode->io_tree, start, end))
1077                 return false;
1078
1079         ordered = btrfs_lookup_ordered_range(inode, start, end - start + 1);
1080         if (!ordered)
1081                 return true;
1082
1083         btrfs_put_ordered_extent(ordered);
1084         unlock_extent(&inode->io_tree, start, end, NULL);
1085
1086         return false;
1087 }
1088
1089
1090 static int clone_ordered_extent(struct btrfs_ordered_extent *ordered, u64 pos,
1091                                 u64 len)
1092 {
1093         struct inode *inode = ordered->inode;
1094         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1095         u64 file_offset = ordered->file_offset + pos;
1096         u64 disk_bytenr = ordered->disk_bytenr + pos;
1097         unsigned long flags = ordered->flags & BTRFS_ORDERED_TYPE_FLAGS;
1098
1099         /*
1100          * The splitting extent is already counted and will be added again in
1101          * btrfs_add_ordered_extent_*(). Subtract len to avoid double counting.
1102          */
1103         percpu_counter_add_batch(&fs_info->ordered_bytes, -len,
1104                                  fs_info->delalloc_batch);
1105         WARN_ON_ONCE(flags & (1 << BTRFS_ORDERED_COMPRESSED));
1106         return btrfs_add_ordered_extent(BTRFS_I(inode), file_offset, len, len,
1107                                         disk_bytenr, len, 0, flags,
1108                                         ordered->compress_type);
1109 }
1110
1111 int btrfs_split_ordered_extent(struct btrfs_ordered_extent *ordered, u64 pre,
1112                                 u64 post)
1113 {
1114         struct inode *inode = ordered->inode;
1115         struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
1116         struct rb_node *node;
1117         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1118         int ret = 0;
1119
1120         trace_btrfs_ordered_extent_split(BTRFS_I(inode), ordered);
1121
1122         spin_lock_irq(&tree->lock);
1123         /* Remove from tree once */
1124         node = &ordered->rb_node;
1125         rb_erase(node, &tree->tree);
1126         RB_CLEAR_NODE(node);
1127         if (tree->last == node)
1128                 tree->last = NULL;
1129
1130         ordered->file_offset += pre;
1131         ordered->disk_bytenr += pre;
1132         ordered->num_bytes -= (pre + post);
1133         ordered->disk_num_bytes -= (pre + post);
1134         ordered->bytes_left -= (pre + post);
1135
1136         /* Re-insert the node */
1137         node = tree_insert(&tree->tree, ordered->file_offset, &ordered->rb_node);
1138         if (node)
1139                 btrfs_panic(fs_info, -EEXIST,
1140                         "zoned: inconsistency in ordered tree at offset %llu",
1141                             ordered->file_offset);
1142
1143         spin_unlock_irq(&tree->lock);
1144
1145         if (pre)
1146                 ret = clone_ordered_extent(ordered, 0, pre);
1147         if (ret == 0 && post)
1148                 ret = clone_ordered_extent(ordered, pre + ordered->disk_num_bytes,
1149                                            post);
1150
1151         return ret;
1152 }
1153
1154 int __init ordered_data_init(void)
1155 {
1156         btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1157                                      sizeof(struct btrfs_ordered_extent), 0,
1158                                      SLAB_MEM_SPREAD,
1159                                      NULL);
1160         if (!btrfs_ordered_extent_cache)
1161                 return -ENOMEM;
1162
1163         return 0;
1164 }
1165
1166 void __cold ordered_data_exit(void)
1167 {
1168         kmem_cache_destroy(btrfs_ordered_extent_cache);
1169 }