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 "messages.h"
  11 #include "misc.h"
  12 #include "ctree.h"
  13 #include "transaction.h"
  14 #include "btrfs_inode.h"
  15 #include "extent_io.h"
  16 #include "disk-io.h"
  17 #include "compression.h"
  18 #include "delalloc-space.h"
  19 #include "qgroup.h"
  20 #include "subpage.h"
  21 #include "file.h"
  22 #include "super.h"
  23
  24 static struct kmem_cache *btrfs_ordered_extent_cache;
  25
  26 static u64 entry_end(struct btrfs_ordered_extent *entry)
  27 {
  28         if (entry->file_offset + entry->num_bytes < entry->file_offset)
  29                 return (u64)-1;
  30         return entry->file_offset + entry->num_bytes;
  31 }
  32
  33 /* returns NULL if the insertion worked, or it returns the node it did find
  34  * in the tree
  35  */
  36 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
  37                                    struct rb_node *node)
  38 {
  39         struct rb_node **p = &root->rb_node;
  40         struct rb_node *parent = NULL;
  41         struct btrfs_ordered_extent *entry;
  42
  43         while (*p) {
  44                 parent = *p;
  45                 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
  46
  47                 if (file_offset < entry->file_offset)
  48                         p = &(*p)->rb_left;
  49                 else if (file_offset >= entry_end(entry))
  50                         p = &(*p)->rb_right;
  51                 else
  52                         return parent;
  53         }
  54
  55         rb_link_node(node, parent, p);
  56         rb_insert_color(node, root);
  57         return NULL;
  58 }
  59
  60 /*
  61  * look for a given offset in the tree, and if it can't be found return the
  62  * first lesser offset
  63  */
  64 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
  65                                      struct rb_node **prev_ret)
  66 {
  67         struct rb_node *n = root->rb_node;
  68         struct rb_node *prev = NULL;
  69         struct rb_node *test;
  70         struct btrfs_ordered_extent *entry;
  71         struct btrfs_ordered_extent *prev_entry = NULL;
  72
  73         while (n) {
  74                 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
  75                 prev = n;
  76                 prev_entry = entry;
  77
  78                 if (file_offset < entry->file_offset)
  79                         n = n->rb_left;
  80                 else if (file_offset >= entry_end(entry))
  81                         n = n->rb_right;
  82                 else
  83                         return n;
  84         }
  85         if (!prev_ret)
  86                 return NULL;
  87
  88         while (prev && file_offset >= entry_end(prev_entry)) {
  89                 test = rb_next(prev);
  90                 if (!test)
  91                         break;
  92                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
  93                                       rb_node);
  94                 if (file_offset < entry_end(prev_entry))
  95                         break;
  96
  97                 prev = test;
  98         }
  99         if (prev)
 100                 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
 101                                       rb_node);
 102         while (prev && file_offset < entry_end(prev_entry)) {
 103                 test = rb_prev(prev);
 104                 if (!test)
 105                         break;
 106                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
 107                                       rb_node);
 108                 prev = test;
 109         }
 110         *prev_ret = prev;
 111         return NULL;
 112 }
 113
 114 static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
 115                           u64 len)
 116 {
 117         if (file_offset + len <= entry->file_offset ||
 118             entry->file_offset + entry->num_bytes <= file_offset)
 119                 return 0;
 120         return 1;
 121 }
 122
 123 /*
 124  * look find the first ordered struct that has this offset, otherwise
 125  * the first one less than this offset
 126  */
 127 static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
 128                                           u64 file_offset)
 129 {
 130         struct rb_root *root = &tree->tree;
 131         struct rb_node *prev = NULL;
 132         struct rb_node *ret;
 133         struct btrfs_ordered_extent *entry;
 134
 135         if (tree->last) {
 136                 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
 137                                  rb_node);
 138                 if (in_range(file_offset, entry->file_offset, entry->num_bytes))
 139                         return tree->last;
 140         }
 141         ret = __tree_search(root, file_offset, &prev);
 142         if (!ret)
 143                 ret = prev;
 144         if (ret)
 145                 tree->last = ret;
 146         return ret;
 147 }
 148
 149 static struct btrfs_ordered_extent *alloc_ordered_extent(
 150                         struct btrfs_inode *inode, u64 file_offset, u64 num_bytes,
 151                         u64 ram_bytes, u64 disk_bytenr, u64 disk_num_bytes,
 152                         u64 offset, unsigned long flags, int compress_type)
 153 {
 154         struct btrfs_ordered_extent *entry;
 155         int ret;
 156
 157         if (flags &
 158             ((1 << BTRFS_ORDERED_NOCOW) | (1 << BTRFS_ORDERED_PREALLOC))) {
 159                 /* For nocow write, we can release the qgroup rsv right now */
 160                 ret = btrfs_qgroup_free_data(inode, NULL, file_offset, num_bytes);
 161                 if (ret < 0)
 162                         return ERR_PTR(ret);
 163         } else {
 164                 /*
 165                  * The ordered extent has reserved qgroup space, release now
 166                  * and pass the reserved number for qgroup_record to free.
 167                  */
 168                 ret = btrfs_qgroup_release_data(inode, file_offset, num_bytes);
 169                 if (ret < 0)
 170                         return ERR_PTR(ret);
 171         }
 172         entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
 173         if (!entry)
 174                 return ERR_PTR(-ENOMEM);
 175
 176         entry->file_offset = file_offset;
 177         entry->num_bytes = num_bytes;
 178         entry->ram_bytes = ram_bytes;
 179         entry->disk_bytenr = disk_bytenr;
 180         entry->disk_num_bytes = disk_num_bytes;
 181         entry->offset = offset;
 182         entry->bytes_left = num_bytes;
 183         entry->inode = igrab(&inode->vfs_inode);
 184         entry->compress_type = compress_type;
 185         entry->truncated_len = (u64)-1;
 186         entry->qgroup_rsv = ret;
 187         entry->flags = flags;
 188         refcount_set(&entry->refs, 1);
 189         init_waitqueue_head(&entry->wait);
 190         INIT_LIST_HEAD(&entry->list);
 191         INIT_LIST_HEAD(&entry->log_list);
 192         INIT_LIST_HEAD(&entry->root_extent_list);
 193         INIT_LIST_HEAD(&entry->work_list);
 194         init_completion(&entry->completion);
 195
 196         /*
 197          * We don't need the count_max_extents here, we can assume that all of
 198          * that work has been done at higher layers, so this is truly the
 199          * smallest the extent is going to get.
 200          */
 201         spin_lock(&inode->lock);
 202         btrfs_mod_outstanding_extents(inode, 1);
 203         spin_unlock(&inode->lock);
 204
 205         return entry;
 206 }
 207
 208 static void insert_ordered_extent(struct btrfs_ordered_extent *entry)
 209 {
 210         struct btrfs_inode *inode = BTRFS_I(entry->inode);
 211         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 212         struct btrfs_root *root = inode->root;
 213         struct btrfs_fs_info *fs_info = root->fs_info;
 214         struct rb_node *node;
 215
 216         trace_btrfs_ordered_extent_add(inode, entry);
 217
 218         percpu_counter_add_batch(&fs_info->ordered_bytes, entry->num_bytes,
 219                                  fs_info->delalloc_batch);
 220
 221         /* One ref for the tree. */
 222         refcount_inc(&entry->refs);
 223
 224         spin_lock_irq(&tree->lock);
 225         node = tree_insert(&tree->tree, entry->file_offset, &entry->rb_node);
 226         if (node)
 227                 btrfs_panic(fs_info, -EEXIST,
 228                                 "inconsistency in ordered tree at offset %llu",
 229                                 entry->file_offset);
 230         spin_unlock_irq(&tree->lock);
 231
 232         spin_lock(&root->ordered_extent_lock);
 233         list_add_tail(&entry->root_extent_list,
 234                       &root->ordered_extents);
 235         root->nr_ordered_extents++;
 236         if (root->nr_ordered_extents == 1) {
 237                 spin_lock(&fs_info->ordered_root_lock);
 238                 BUG_ON(!list_empty(&root->ordered_root));
 239                 list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
 240                 spin_unlock(&fs_info->ordered_root_lock);
 241         }
 242         spin_unlock(&root->ordered_extent_lock);
 243 }
 244
 245 /*
 246  * Add an ordered extent to the per-inode tree.
 247  *
 248  * @inode:           Inode that this extent is for.
 249  * @file_offset:     Logical offset in file where the extent starts.
 250  * @num_bytes:       Logical length of extent in file.
 251  * @ram_bytes:       Full length of unencoded data.
 252  * @disk_bytenr:     Offset of extent on disk.
 253  * @disk_num_bytes:  Size of extent on disk.
 254  * @offset:          Offset into unencoded data where file data starts.
 255  * @flags:           Flags specifying type of extent (1 << BTRFS_ORDERED_*).
 256  * @compress_type:   Compression algorithm used for data.
 257  *
 258  * Most of these parameters correspond to &struct btrfs_file_extent_item. The
 259  * tree is given a single reference on the ordered extent that was inserted, and
 260  * the returned pointer is given a second reference.
 261  *
 262  * Return: the new ordered extent or error pointer.
 263  */
 264 struct btrfs_ordered_extent *btrfs_alloc_ordered_extent(
 265                         struct btrfs_inode *inode, u64 file_offset,
 266                         u64 num_bytes, u64 ram_bytes, u64 disk_bytenr,
 267                         u64 disk_num_bytes, u64 offset, unsigned long flags,
 268                         int compress_type)
 269 {
 270         struct btrfs_ordered_extent *entry;
 271
 272         ASSERT((flags & ~BTRFS_ORDERED_TYPE_FLAGS) == 0);
 273
 274         entry = alloc_ordered_extent(inode, file_offset, num_bytes, ram_bytes,
 275                                      disk_bytenr, disk_num_bytes, offset, flags,
 276                                      compress_type);
 277         if (!IS_ERR(entry))
 278                 insert_ordered_extent(entry);
 279         return entry;
 280 }
 281
 282 /*
 283  * Add a new btrfs_ordered_extent for the range, but drop the reference instead
 284  * of returning it to the caller.
 285  */
 286 int btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
 287                              u64 num_bytes, u64 ram_bytes, u64 disk_bytenr,
 288                              u64 disk_num_bytes, u64 offset, unsigned long flags,
 289                              int compress_type)
 290 {
 291         struct btrfs_ordered_extent *ordered;
 292
 293         ordered = btrfs_alloc_ordered_extent(inode, file_offset, num_bytes,
 294                                              ram_bytes, disk_bytenr,
 295                                              disk_num_bytes, offset, flags,
 296                                              compress_type);
 297
 298         if (IS_ERR(ordered))
 299                 return PTR_ERR(ordered);
 300         btrfs_put_ordered_extent(ordered);
 301
 302         return 0;
 303 }
 304
 305 /*
 306  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
 307  * when an ordered extent is finished.  If the list covers more than one
 308  * ordered extent, it is split across multiples.
 309  */
 310 void btrfs_add_ordered_sum(struct btrfs_ordered_extent *entry,
 311                            struct btrfs_ordered_sum *sum)
 312 {
 313         struct btrfs_ordered_inode_tree *tree;
 314
 315         tree = &BTRFS_I(entry->inode)->ordered_tree;
 316         spin_lock_irq(&tree->lock);
 317         list_add_tail(&sum->list, &entry->list);
 318         spin_unlock_irq(&tree->lock);
 319 }
 320
 321 static void finish_ordered_fn(struct btrfs_work *work)
 322 {
 323         struct btrfs_ordered_extent *ordered_extent;
 324
 325         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
 326         btrfs_finish_ordered_io(ordered_extent);
 327 }
 328
 329 /*
 330  * Mark all ordered extents io inside the specified range finished.
 331  *
 332  * @page:        The involved page for the operation.
 333  *               For uncompressed buffered IO, the page status also needs to be
 334  *               updated to indicate whether the pending ordered io is finished.
 335  *               Can be NULL for direct IO and compressed write.
 336  *               For these cases, callers are ensured they won't execute the
 337  *               endio function twice.
 338  *
 339  * This function is called for endio, thus the range must have ordered
 340  * extent(s) covering it.
 341  */
 342 void btrfs_mark_ordered_io_finished(struct btrfs_inode *inode,
 343                                     struct page *page, u64 file_offset,
 344                                     u64 num_bytes, bool uptodate)
 345 {
 346         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 347         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 348         struct btrfs_workqueue *wq;
 349         struct rb_node *node;
 350         struct btrfs_ordered_extent *entry = NULL;
 351         unsigned long flags;
 352         u64 cur = file_offset;
 353
 354         if (btrfs_is_free_space_inode(inode))
 355                 wq = fs_info->endio_freespace_worker;
 356         else
 357                 wq = fs_info->endio_write_workers;
 358
 359         if (page)
 360                 ASSERT(page->mapping && page_offset(page) <= file_offset &&
 361                        file_offset + num_bytes <= page_offset(page) + PAGE_SIZE);
 362
 363         spin_lock_irqsave(&tree->lock, flags);
 364         while (cur < file_offset + num_bytes) {
 365                 u64 entry_end;
 366                 u64 end;
 367                 u32 len;
 368
 369                 node = tree_search(tree, cur);
 370                 /* No ordered extents at all */
 371                 if (!node)
 372                         break;
 373
 374                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 375                 entry_end = entry->file_offset + entry->num_bytes;
 376                 /*
 377                  * |<-- OE --->|  |
 378                  *                cur
 379                  * Go to next OE.
 380                  */
 381                 if (cur >= entry_end) {
 382                         node = rb_next(node);
 383                         /* No more ordered extents, exit */
 384                         if (!node)
 385                                 break;
 386                         entry = rb_entry(node, struct btrfs_ordered_extent,
 387                                          rb_node);
 388
 389                         /* Go to next ordered extent and continue */
 390                         cur = entry->file_offset;
 391                         continue;
 392                 }
 393                 /*
 394                  * |    |<--- OE --->|
 395                  * cur
 396                  * Go to the start of OE.
 397                  */
 398                 if (cur < entry->file_offset) {
 399                         cur = entry->file_offset;
 400                         continue;
 401                 }
 402
 403                 /*
 404                  * Now we are definitely inside one ordered extent.
 405                  *
 406                  * |<--- OE --->|
 407                  *      |
 408                  *      cur
 409                  */
 410                 end = min(entry->file_offset + entry->num_bytes,
 411                           file_offset + num_bytes) - 1;
 412                 ASSERT(end + 1 - cur < U32_MAX);
 413                 len = end + 1 - cur;
 414
 415                 if (page) {
 416                         /*
 417                          * Ordered (Private2) bit indicates whether we still
 418                          * have pending io unfinished for the ordered extent.
 419                          *
 420                          * If there's no such bit, we need to skip to next range.
 421                          */
 422                         if (!btrfs_page_test_ordered(fs_info, page, cur, len)) {
 423                                 cur += len;
 424                                 continue;
 425                         }
 426                         btrfs_page_clear_ordered(fs_info, page, cur, len);
 427                 }
 428
 429                 /* Now we're fine to update the accounting */
 430                 if (unlikely(len > entry->bytes_left)) {
 431                         WARN_ON(1);
 432                         btrfs_crit(fs_info,
 433 "bad ordered extent accounting, root=%llu ino=%llu OE offset=%llu OE len=%llu to_dec=%u left=%llu",
 434                                    inode->root->root_key.objectid,
 435                                    btrfs_ino(inode),
 436                                    entry->file_offset,
 437                                    entry->num_bytes,
 438                                    len, entry->bytes_left);
 439                         entry->bytes_left = 0;
 440                 } else {
 441                         entry->bytes_left -= len;
 442                 }
 443
 444                 if (!uptodate)
 445                         set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
 446
 447                 /*
 448                  * All the IO of the ordered extent is finished, we need to queue
 449                  * the finish_func to be executed.
 450                  */
 451                 if (entry->bytes_left == 0) {
 452                         set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 453                         cond_wake_up(&entry->wait);
 454                         refcount_inc(&entry->refs);
 455                         trace_btrfs_ordered_extent_mark_finished(inode, entry);
 456                         spin_unlock_irqrestore(&tree->lock, flags);
 457                         btrfs_init_work(&entry->work, finish_ordered_fn, NULL, NULL);
 458                         btrfs_queue_work(wq, &entry->work);
 459                         spin_lock_irqsave(&tree->lock, flags);
 460                 }
 461                 cur += len;
 462         }
 463         spin_unlock_irqrestore(&tree->lock, flags);
 464 }
 465
 466 /*
 467  * Finish IO for one ordered extent across a given range.  The range can only
 468  * contain one ordered extent.
 469  *
 470  * @cached:      The cached ordered extent. If not NULL, we can skip the tree
 471  *               search and use the ordered extent directly.
 472  *               Will be also used to store the finished ordered extent.
 473  * @file_offset: File offset for the finished IO
 474  * @io_size:     Length of the finish IO range
 475  *
 476  * Return true if the ordered extent is finished in the range, and update
 477  * @cached.
 478  * Return false otherwise.
 479  *
 480  * NOTE: The range can NOT cross multiple ordered extents.
 481  * Thus caller should ensure the range doesn't cross ordered extents.
 482  */
 483 bool btrfs_dec_test_ordered_pending(struct btrfs_inode *inode,
 484                                     struct btrfs_ordered_extent **cached,
 485                                     u64 file_offset, u64 io_size)
 486 {
 487         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 488         struct rb_node *node;
 489         struct btrfs_ordered_extent *entry = NULL;
 490         unsigned long flags;
 491         bool finished = false;
 492
 493         spin_lock_irqsave(&tree->lock, flags);
 494         if (cached && *cached) {
 495                 entry = *cached;
 496                 goto have_entry;
 497         }
 498
 499         node = tree_search(tree, file_offset);
 500         if (!node)
 501                 goto out;
 502
 503         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 504 have_entry:
 505         if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
 506                 goto out;
 507
 508         if (io_size > entry->bytes_left)
 509                 btrfs_crit(inode->root->fs_info,
 510                            "bad ordered accounting left %llu size %llu",
 511                        entry->bytes_left, io_size);
 512
 513         entry->bytes_left -= io_size;
 514
 515         if (entry->bytes_left == 0) {
 516                 /*
 517                  * Ensure only one caller can set the flag and finished_ret
 518                  * accordingly
 519                  */
 520                 finished = !test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 521                 /* test_and_set_bit implies a barrier */
 522                 cond_wake_up_nomb(&entry->wait);
 523         }
 524 out:
 525         if (finished && cached && entry) {
 526                 *cached = entry;
 527                 refcount_inc(&entry->refs);
 528                 trace_btrfs_ordered_extent_dec_test_pending(inode, entry);
 529         }
 530         spin_unlock_irqrestore(&tree->lock, flags);
 531         return finished;
 532 }
 533
 534 /*
 535  * used to drop a reference on an ordered extent.  This will free
 536  * the extent if the last reference is dropped
 537  */
 538 void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
 539 {
 540         struct list_head *cur;
 541         struct btrfs_ordered_sum *sum;
 542
 543         trace_btrfs_ordered_extent_put(BTRFS_I(entry->inode), entry);
 544
 545         if (refcount_dec_and_test(&entry->refs)) {
 546                 ASSERT(list_empty(&entry->root_extent_list));
 547                 ASSERT(list_empty(&entry->log_list));
 548                 ASSERT(RB_EMPTY_NODE(&entry->rb_node));
 549                 if (entry->inode)
 550                         btrfs_add_delayed_iput(BTRFS_I(entry->inode));
 551                 while (!list_empty(&entry->list)) {
 552                         cur = entry->list.next;
 553                         sum = list_entry(cur, struct btrfs_ordered_sum, list);
 554                         list_del(&sum->list);
 555                         kvfree(sum);
 556                 }
 557                 kmem_cache_free(btrfs_ordered_extent_cache, entry);
 558         }
 559 }
 560
 561 /*
 562  * remove an ordered extent from the tree.  No references are dropped
 563  * and waiters are woken up.
 564  */
 565 void btrfs_remove_ordered_extent(struct btrfs_inode *btrfs_inode,
 566                                  struct btrfs_ordered_extent *entry)
 567 {
 568         struct btrfs_ordered_inode_tree *tree;
 569         struct btrfs_root *root = btrfs_inode->root;
 570         struct btrfs_fs_info *fs_info = root->fs_info;
 571         struct rb_node *node;
 572         bool pending;
 573         bool freespace_inode;
 574
 575         /*
 576          * If this is a free space inode the thread has not acquired the ordered
 577          * extents lockdep map.
 578          */
 579         freespace_inode = btrfs_is_free_space_inode(btrfs_inode);
 580
 581         btrfs_lockdep_acquire(fs_info, btrfs_trans_pending_ordered);
 582         /* This is paired with btrfs_add_ordered_extent. */
 583         spin_lock(&btrfs_inode->lock);
 584         btrfs_mod_outstanding_extents(btrfs_inode, -1);
 585         spin_unlock(&btrfs_inode->lock);
 586         if (root != fs_info->tree_root) {
 587                 u64 release;
 588
 589                 if (test_bit(BTRFS_ORDERED_ENCODED, &entry->flags))
 590                         release = entry->disk_num_bytes;
 591                 else
 592                         release = entry->num_bytes;
 593                 btrfs_delalloc_release_metadata(btrfs_inode, release, false);
 594         }
 595
 596         percpu_counter_add_batch(&fs_info->ordered_bytes, -entry->num_bytes,
 597                                  fs_info->delalloc_batch);
 598
 599         tree = &btrfs_inode->ordered_tree;
 600         spin_lock_irq(&tree->lock);
 601         node = &entry->rb_node;
 602         rb_erase(node, &tree->tree);
 603         RB_CLEAR_NODE(node);
 604         if (tree->last == node)
 605                 tree->last = NULL;
 606         set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
 607         pending = test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags);
 608         spin_unlock_irq(&tree->lock);
 609
 610         /*
 611          * The current running transaction is waiting on us, we need to let it
 612          * know that we're complete and wake it up.
 613          */
 614         if (pending) {
 615                 struct btrfs_transaction *trans;
 616
 617                 /*
 618                  * The checks for trans are just a formality, it should be set,
 619                  * but if it isn't we don't want to deref/assert under the spin
 620                  * lock, so be nice and check if trans is set, but ASSERT() so
 621                  * if it isn't set a developer will notice.
 622                  */
 623                 spin_lock(&fs_info->trans_lock);
 624                 trans = fs_info->running_transaction;
 625                 if (trans)
 626                         refcount_inc(&trans->use_count);
 627                 spin_unlock(&fs_info->trans_lock);
 628
 629                 ASSERT(trans);
 630                 if (trans) {
 631                         if (atomic_dec_and_test(&trans->pending_ordered))
 632                                 wake_up(&trans->pending_wait);
 633                         btrfs_put_transaction(trans);
 634                 }
 635         }
 636
 637         btrfs_lockdep_release(fs_info, btrfs_trans_pending_ordered);
 638
 639         spin_lock(&root->ordered_extent_lock);
 640         list_del_init(&entry->root_extent_list);
 641         root->nr_ordered_extents--;
 642
 643         trace_btrfs_ordered_extent_remove(btrfs_inode, entry);
 644
 645         if (!root->nr_ordered_extents) {
 646                 spin_lock(&fs_info->ordered_root_lock);
 647                 BUG_ON(list_empty(&root->ordered_root));
 648                 list_del_init(&root->ordered_root);
 649                 spin_unlock(&fs_info->ordered_root_lock);
 650         }
 651         spin_unlock(&root->ordered_extent_lock);
 652         wake_up(&entry->wait);
 653         if (!freespace_inode)
 654                 btrfs_lockdep_release(fs_info, btrfs_ordered_extent);
 655 }
 656
 657 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
 658 {
 659         struct btrfs_ordered_extent *ordered;
 660
 661         ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
 662         btrfs_start_ordered_extent(ordered);
 663         complete(&ordered->completion);
 664 }
 665
 666 /*
 667  * wait for all the ordered extents in a root.  This is done when balancing
 668  * space between drives.
 669  */
 670 u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
 671                                const u64 range_start, const u64 range_len)
 672 {
 673         struct btrfs_fs_info *fs_info = root->fs_info;
 674         LIST_HEAD(splice);
 675         LIST_HEAD(skipped);
 676         LIST_HEAD(works);
 677         struct btrfs_ordered_extent *ordered, *next;
 678         u64 count = 0;
 679         const u64 range_end = range_start + range_len;
 680
 681         mutex_lock(&root->ordered_extent_mutex);
 682         spin_lock(&root->ordered_extent_lock);
 683         list_splice_init(&root->ordered_extents, &splice);
 684         while (!list_empty(&splice) && nr) {
 685                 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
 686                                            root_extent_list);
 687
 688                 if (range_end <= ordered->disk_bytenr ||
 689                     ordered->disk_bytenr + ordered->disk_num_bytes <= range_start) {
 690                         list_move_tail(&ordered->root_extent_list, &skipped);
 691                         cond_resched_lock(&root->ordered_extent_lock);
 692                         continue;
 693                 }
 694
 695                 list_move_tail(&ordered->root_extent_list,
 696                                &root->ordered_extents);
 697                 refcount_inc(&ordered->refs);
 698                 spin_unlock(&root->ordered_extent_lock);
 699
 700                 btrfs_init_work(&ordered->flush_work,
 701                                 btrfs_run_ordered_extent_work, NULL, NULL);
 702                 list_add_tail(&ordered->work_list, &works);
 703                 btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
 704
 705                 cond_resched();
 706                 spin_lock(&root->ordered_extent_lock);
 707                 if (nr != U64_MAX)
 708                         nr--;
 709                 count++;
 710         }
 711         list_splice_tail(&skipped, &root->ordered_extents);
 712         list_splice_tail(&splice, &root->ordered_extents);
 713         spin_unlock(&root->ordered_extent_lock);
 714
 715         list_for_each_entry_safe(ordered, next, &works, work_list) {
 716                 list_del_init(&ordered->work_list);
 717                 wait_for_completion(&ordered->completion);
 718                 btrfs_put_ordered_extent(ordered);
 719                 cond_resched();
 720         }
 721         mutex_unlock(&root->ordered_extent_mutex);
 722
 723         return count;
 724 }
 725
 726 void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
 727                              const u64 range_start, const u64 range_len)
 728 {
 729         struct btrfs_root *root;
 730         struct list_head splice;
 731         u64 done;
 732
 733         INIT_LIST_HEAD(&splice);
 734
 735         mutex_lock(&fs_info->ordered_operations_mutex);
 736         spin_lock(&fs_info->ordered_root_lock);
 737         list_splice_init(&fs_info->ordered_roots, &splice);
 738         while (!list_empty(&splice) && nr) {
 739                 root = list_first_entry(&splice, struct btrfs_root,
 740                                         ordered_root);
 741                 root = btrfs_grab_root(root);
 742                 BUG_ON(!root);
 743                 list_move_tail(&root->ordered_root,
 744                                &fs_info->ordered_roots);
 745                 spin_unlock(&fs_info->ordered_root_lock);
 746
 747                 done = btrfs_wait_ordered_extents(root, nr,
 748                                                   range_start, range_len);
 749                 btrfs_put_root(root);
 750
 751                 spin_lock(&fs_info->ordered_root_lock);
 752                 if (nr != U64_MAX) {
 753                         nr -= done;
 754                 }
 755         }
 756         list_splice_tail(&splice, &fs_info->ordered_roots);
 757         spin_unlock(&fs_info->ordered_root_lock);
 758         mutex_unlock(&fs_info->ordered_operations_mutex);
 759 }
 760
 761 /*
 762  * Start IO and wait for a given ordered extent to finish.
 763  *
 764  * Wait on page writeback for all the pages in the extent and the IO completion
 765  * code to insert metadata into the btree corresponding to the extent.
 766  */
 767 void btrfs_start_ordered_extent(struct btrfs_ordered_extent *entry)
 768 {
 769         u64 start = entry->file_offset;
 770         u64 end = start + entry->num_bytes - 1;
 771         struct btrfs_inode *inode = BTRFS_I(entry->inode);
 772         bool freespace_inode;
 773
 774         trace_btrfs_ordered_extent_start(inode, entry);
 775
 776         /*
 777          * If this is a free space inode do not take the ordered extents lockdep
 778          * map.
 779          */
 780         freespace_inode = btrfs_is_free_space_inode(inode);
 781
 782         /*
 783          * pages in the range can be dirty, clean or writeback.  We
 784          * start IO on any dirty ones so the wait doesn't stall waiting
 785          * for the flusher thread to find them
 786          */
 787         if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
 788                 filemap_fdatawrite_range(inode->vfs_inode.i_mapping, start, end);
 789
 790         if (!freespace_inode)
 791                 btrfs_might_wait_for_event(inode->root->fs_info, btrfs_ordered_extent);
 792         wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE, &entry->flags));
 793 }
 794
 795 /*
 796  * Used to wait on ordered extents across a large range of bytes.
 797  */
 798 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
 799 {
 800         int ret = 0;
 801         int ret_wb = 0;
 802         u64 end;
 803         u64 orig_end;
 804         struct btrfs_ordered_extent *ordered;
 805
 806         if (start + len < start) {
 807                 orig_end = OFFSET_MAX;
 808         } else {
 809                 orig_end = start + len - 1;
 810                 if (orig_end > OFFSET_MAX)
 811                         orig_end = OFFSET_MAX;
 812         }
 813
 814         /* start IO across the range first to instantiate any delalloc
 815          * extents
 816          */
 817         ret = btrfs_fdatawrite_range(inode, start, orig_end);
 818         if (ret)
 819                 return ret;
 820
 821         /*
 822          * If we have a writeback error don't return immediately. Wait first
 823          * for any ordered extents that haven't completed yet. This is to make
 824          * sure no one can dirty the same page ranges and call writepages()
 825          * before the ordered extents complete - to avoid failures (-EEXIST)
 826          * when adding the new ordered extents to the ordered tree.
 827          */
 828         ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
 829
 830         end = orig_end;
 831         while (1) {
 832                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode), end);
 833                 if (!ordered)
 834                         break;
 835                 if (ordered->file_offset > orig_end) {
 836                         btrfs_put_ordered_extent(ordered);
 837                         break;
 838                 }
 839                 if (ordered->file_offset + ordered->num_bytes <= start) {
 840                         btrfs_put_ordered_extent(ordered);
 841                         break;
 842                 }
 843                 btrfs_start_ordered_extent(ordered);
 844                 end = ordered->file_offset;
 845                 /*
 846                  * If the ordered extent had an error save the error but don't
 847                  * exit without waiting first for all other ordered extents in
 848                  * the range to complete.
 849                  */
 850                 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
 851                         ret = -EIO;
 852                 btrfs_put_ordered_extent(ordered);
 853                 if (end == 0 || end == start)
 854                         break;
 855                 end--;
 856         }
 857         return ret_wb ? ret_wb : ret;
 858 }
 859
 860 /*
 861  * find an ordered extent corresponding to file_offset.  return NULL if
 862  * nothing is found, otherwise take a reference on the extent and return it
 863  */
 864 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct btrfs_inode *inode,
 865                                                          u64 file_offset)
 866 {
 867         struct btrfs_ordered_inode_tree *tree;
 868         struct rb_node *node;
 869         struct btrfs_ordered_extent *entry = NULL;
 870         unsigned long flags;
 871
 872         tree = &inode->ordered_tree;
 873         spin_lock_irqsave(&tree->lock, flags);
 874         node = tree_search(tree, file_offset);
 875         if (!node)
 876                 goto out;
 877
 878         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 879         if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
 880                 entry = NULL;
 881         if (entry) {
 882                 refcount_inc(&entry->refs);
 883                 trace_btrfs_ordered_extent_lookup(inode, entry);
 884         }
 885 out:
 886         spin_unlock_irqrestore(&tree->lock, flags);
 887         return entry;
 888 }
 889
 890 /* Since the DIO code tries to lock a wide area we need to look for any ordered
 891  * extents that exist in the range, rather than just the start of the range.
 892  */
 893 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
 894                 struct btrfs_inode *inode, u64 file_offset, u64 len)
 895 {
 896         struct btrfs_ordered_inode_tree *tree;
 897         struct rb_node *node;
 898         struct btrfs_ordered_extent *entry = NULL;
 899
 900         tree = &inode->ordered_tree;
 901         spin_lock_irq(&tree->lock);
 902         node = tree_search(tree, file_offset);
 903         if (!node) {
 904                 node = tree_search(tree, file_offset + len);
 905                 if (!node)
 906                         goto out;
 907         }
 908
 909         while (1) {
 910                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 911                 if (range_overlaps(entry, file_offset, len))
 912                         break;
 913
 914                 if (entry->file_offset >= file_offset + len) {
 915                         entry = NULL;
 916                         break;
 917                 }
 918                 entry = NULL;
 919                 node = rb_next(node);
 920                 if (!node)
 921                         break;
 922         }
 923 out:
 924         if (entry) {
 925                 refcount_inc(&entry->refs);
 926                 trace_btrfs_ordered_extent_lookup_range(inode, entry);
 927         }
 928         spin_unlock_irq(&tree->lock);
 929         return entry;
 930 }
 931
 932 /*
 933  * Adds all ordered extents to the given list. The list ends up sorted by the
 934  * file_offset of the ordered extents.
 935  */
 936 void btrfs_get_ordered_extents_for_logging(struct btrfs_inode *inode,
 937                                            struct list_head *list)
 938 {
 939         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 940         struct rb_node *n;
 941
 942         ASSERT(inode_is_locked(&inode->vfs_inode));
 943
 944         spin_lock_irq(&tree->lock);
 945         for (n = rb_first(&tree->tree); n; n = rb_next(n)) {
 946                 struct btrfs_ordered_extent *ordered;
 947
 948                 ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
 949
 950                 if (test_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
 951                         continue;
 952
 953                 ASSERT(list_empty(&ordered->log_list));
 954                 list_add_tail(&ordered->log_list, list);
 955                 refcount_inc(&ordered->refs);
 956                 trace_btrfs_ordered_extent_lookup_for_logging(inode, ordered);
 957         }
 958         spin_unlock_irq(&tree->lock);
 959 }
 960
 961 /*
 962  * lookup and return any extent before 'file_offset'.  NULL is returned
 963  * if none is found
 964  */
 965 struct btrfs_ordered_extent *
 966 btrfs_lookup_first_ordered_extent(struct btrfs_inode *inode, u64 file_offset)
 967 {
 968         struct btrfs_ordered_inode_tree *tree;
 969         struct rb_node *node;
 970         struct btrfs_ordered_extent *entry = NULL;
 971
 972         tree = &inode->ordered_tree;
 973         spin_lock_irq(&tree->lock);
 974         node = tree_search(tree, file_offset);
 975         if (!node)
 976                 goto out;
 977
 978         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 979         refcount_inc(&entry->refs);
 980         trace_btrfs_ordered_extent_lookup_first(inode, entry);
 981 out:
 982         spin_unlock_irq(&tree->lock);
 983         return entry;
 984 }
 985
 986 /*
 987  * Lookup the first ordered extent that overlaps the range
 988  * [@file_offset, @file_offset + @len).
 989  *
 990  * The difference between this and btrfs_lookup_first_ordered_extent() is
 991  * that this one won't return any ordered extent that does not overlap the range.
 992  * And the difference against btrfs_lookup_ordered_extent() is, this function
 993  * ensures the first ordered extent gets returned.
 994  */
 995 struct btrfs_ordered_extent *btrfs_lookup_first_ordered_range(
 996                         struct btrfs_inode *inode, u64 file_offset, u64 len)
 997 {
 998         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 999         struct rb_node *node;
1000         struct rb_node *cur;
1001         struct rb_node *prev;
1002         struct rb_node *next;
1003         struct btrfs_ordered_extent *entry = NULL;
1004
1005         spin_lock_irq(&tree->lock);
1006         node = tree->tree.rb_node;
1007         /*
1008          * Here we don't want to use tree_search() which will use tree->last
1009          * and screw up the search order.
1010          * And __tree_search() can't return the adjacent ordered extents
1011          * either, thus here we do our own search.
1012          */
1013         while (node) {
1014                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
1015
1016                 if (file_offset < entry->file_offset) {
1017                         node = node->rb_left;
1018                 } else if (file_offset >= entry_end(entry)) {
1019                         node = node->rb_right;
1020                 } else {
1021                         /*
1022                          * Direct hit, got an ordered extent that starts at
1023                          * @file_offset
1024                          */
1025                         goto out;
1026                 }
1027         }
1028         if (!entry) {
1029                 /* Empty tree */
1030                 goto out;
1031         }
1032
1033         cur = &entry->rb_node;
1034         /* We got an entry around @file_offset, check adjacent entries */
1035         if (entry->file_offset < file_offset) {
1036                 prev = cur;
1037                 next = rb_next(cur);
1038         } else {
1039                 prev = rb_prev(cur);
1040                 next = cur;
1041         }
1042         if (prev) {
1043                 entry = rb_entry(prev, struct btrfs_ordered_extent, rb_node);
1044                 if (range_overlaps(entry, file_offset, len))
1045                         goto out;
1046         }
1047         if (next) {
1048                 entry = rb_entry(next, struct btrfs_ordered_extent, rb_node);
1049                 if (range_overlaps(entry, file_offset, len))
1050                         goto out;
1051         }
1052         /* No ordered extent in the range */
1053         entry = NULL;
1054 out:
1055         if (entry) {
1056                 refcount_inc(&entry->refs);
1057                 trace_btrfs_ordered_extent_lookup_first_range(inode, entry);
1058         }
1059
1060         spin_unlock_irq(&tree->lock);
1061         return entry;
1062 }
1063
1064 /*
1065  * Lock the passed range and ensures all pending ordered extents in it are run
1066  * to completion.
1067  *
1068  * @inode:        Inode whose ordered tree is to be searched
1069  * @start:        Beginning of range to flush
1070  * @end:          Last byte of range to lock
1071  * @cached_state: If passed, will return the extent state responsible for the
1072  *                locked range. It's the caller's responsibility to free the
1073  *                cached state.
1074  *
1075  * Always return with the given range locked, ensuring after it's called no
1076  * order extent can be pending.
1077  */
1078 void btrfs_lock_and_flush_ordered_range(struct btrfs_inode *inode, u64 start,
1079                                         u64 end,
1080                                         struct extent_state **cached_state)
1081 {
1082         struct btrfs_ordered_extent *ordered;
1083         struct extent_state *cache = NULL;
1084         struct extent_state **cachedp = &cache;
1085
1086         if (cached_state)
1087                 cachedp = cached_state;
1088
1089         while (1) {
1090                 lock_extent(&inode->io_tree, start, end, cachedp);
1091                 ordered = btrfs_lookup_ordered_range(inode, start,
1092                                                      end - start + 1);
1093                 if (!ordered) {
1094                         /*
1095                          * If no external cached_state has been passed then
1096                          * decrement the extra ref taken for cachedp since we
1097                          * aren't exposing it outside of this function
1098                          */
1099                         if (!cached_state)
1100                                 refcount_dec(&cache->refs);
1101                         break;
1102                 }
1103                 unlock_extent(&inode->io_tree, start, end, cachedp);
1104                 btrfs_start_ordered_extent(ordered);
1105                 btrfs_put_ordered_extent(ordered);
1106         }
1107 }
1108
1109 /*
1110  * Lock the passed range and ensure all pending ordered extents in it are run
1111  * to completion in nowait mode.
1112  *
1113  * Return true if btrfs_lock_ordered_range does not return any extents,
1114  * otherwise false.
1115  */
1116 bool btrfs_try_lock_ordered_range(struct btrfs_inode *inode, u64 start, u64 end,
1117                                   struct extent_state **cached_state)
1118 {
1119         struct btrfs_ordered_extent *ordered;
1120
1121         if (!try_lock_extent(&inode->io_tree, start, end, cached_state))
1122                 return false;
1123
1124         ordered = btrfs_lookup_ordered_range(inode, start, end - start + 1);
1125         if (!ordered)
1126                 return true;
1127
1128         btrfs_put_ordered_extent(ordered);
1129         unlock_extent(&inode->io_tree, start, end, cached_state);
1130
1131         return false;
1132 }
1133
1134 /* Split out a new ordered extent for this first @len bytes of @ordered. */
1135 struct btrfs_ordered_extent *btrfs_split_ordered_extent(
1136                         struct btrfs_ordered_extent *ordered, u64 len)
1137 {
1138         struct inode *inode = ordered->inode;
1139         struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
1140         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1141         u64 file_offset = ordered->file_offset;
1142         u64 disk_bytenr = ordered->disk_bytenr;
1143         unsigned long flags = ordered->flags & BTRFS_ORDERED_TYPE_FLAGS;
1144         struct rb_node *node;
1145
1146         trace_btrfs_ordered_extent_split(BTRFS_I(inode), ordered);
1147
1148         ASSERT(!(flags & (1U << BTRFS_ORDERED_COMPRESSED)));
1149
1150         /*
1151          * The entire bio must be covered by the ordered extent, but we can't
1152          * reduce the original extent to a zero length either.
1153          */
1154         if (WARN_ON_ONCE(len >= ordered->num_bytes))
1155                 return ERR_PTR(-EINVAL);
1156         /* We cannot split once ordered extent is past end_bio. */
1157         if (WARN_ON_ONCE(ordered->bytes_left != ordered->disk_num_bytes))
1158                 return ERR_PTR(-EINVAL);
1159         /* We cannot split a compressed ordered extent. */
1160         if (WARN_ON_ONCE(ordered->disk_num_bytes != ordered->num_bytes))
1161                 return ERR_PTR(-EINVAL);
1162         /* Checksum list should be empty. */
1163         if (WARN_ON_ONCE(!list_empty(&ordered->list)))
1164                 return ERR_PTR(-EINVAL);
1165
1166         spin_lock_irq(&tree->lock);
1167         /* Remove from tree once */
1168         node = &ordered->rb_node;
1169         rb_erase(node, &tree->tree);
1170         RB_CLEAR_NODE(node);
1171         if (tree->last == node)
1172                 tree->last = NULL;
1173
1174         ordered->file_offset += len;
1175         ordered->disk_bytenr += len;
1176         ordered->num_bytes -= len;
1177         ordered->disk_num_bytes -= len;
1178         ordered->bytes_left -= len;
1179
1180         /* Re-insert the node */
1181         node = tree_insert(&tree->tree, ordered->file_offset, &ordered->rb_node);
1182         if (node)
1183                 btrfs_panic(fs_info, -EEXIST,
1184                         "zoned: inconsistency in ordered tree at offset %llu",
1185                             ordered->file_offset);
1186
1187         spin_unlock_irq(&tree->lock);
1188
1189         /*
1190          * The splitting extent is already counted and will be added again in
1191          * btrfs_alloc_ordered_extent(). Subtract len to avoid double counting.
1192          */
1193         percpu_counter_add_batch(&fs_info->ordered_bytes, -len, fs_info->delalloc_batch);
1194
1195         return btrfs_alloc_ordered_extent(BTRFS_I(inode), file_offset, len, len,
1196                                           disk_bytenr, len, 0, flags,
1197                                           ordered->compress_type);
1198 }
1199
1200 int __init ordered_data_init(void)
1201 {
1202         btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1203                                      sizeof(struct btrfs_ordered_extent), 0,
1204                                      SLAB_MEM_SPREAD,
1205                                      NULL);
1206         if (!btrfs_ordered_extent_cache)
1207                 return -ENOMEM;
1208
1209         return 0;
1210 }
1211
1212 void __cold ordered_data_exit(void)
1213 {
1214         kmem_cache_destroy(btrfs_ordered_extent_cache);
1215 }