fs/btrfs/ordered-data.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/slab.h>
   7 #include <linux/blkdev.h>
   8 #include <linux/writeback.h>
   9 #include <linux/sched/mm.h>
  10 #include "misc.h"
  11 #include "ctree.h"
  12 #include "transaction.h"
  13 #include "btrfs_inode.h"
  14 #include "extent_io.h"
  15 #include "disk-io.h"
  16 #include "compression.h"
  17 #include "delalloc-space.h"
  18 #include "qgroup.h"
  19
  20 static struct kmem_cache *btrfs_ordered_extent_cache;
  21
  22 static u64 entry_end(struct btrfs_ordered_extent *entry)
  23 {
  24         if (entry->file_offset + entry->num_bytes < entry->file_offset)
  25                 return (u64)-1;
  26         return entry->file_offset + entry->num_bytes;
  27 }
  28
  29 /* returns NULL if the insertion worked, or it returns the node it did find
  30  * in the tree
  31  */
  32 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
  33                                    struct rb_node *node)
  34 {
  35         struct rb_node **p = &root->rb_node;
  36         struct rb_node *parent = NULL;
  37         struct btrfs_ordered_extent *entry;
  38
  39         while (*p) {
  40                 parent = *p;
  41                 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
  42
  43                 if (file_offset < entry->file_offset)
  44                         p = &(*p)->rb_left;
  45                 else if (file_offset >= entry_end(entry))
  46                         p = &(*p)->rb_right;
  47                 else
  48                         return parent;
  49         }
  50
  51         rb_link_node(node, parent, p);
  52         rb_insert_color(node, root);
  53         return NULL;
  54 }
  55
  56 /*
  57  * look for a given offset in the tree, and if it can't be found return the
  58  * first lesser offset
  59  */
  60 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
  61                                      struct rb_node **prev_ret)
  62 {
  63         struct rb_node *n = root->rb_node;
  64         struct rb_node *prev = NULL;
  65         struct rb_node *test;
  66         struct btrfs_ordered_extent *entry;
  67         struct btrfs_ordered_extent *prev_entry = NULL;
  68
  69         while (n) {
  70                 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
  71                 prev = n;
  72                 prev_entry = entry;
  73
  74                 if (file_offset < entry->file_offset)
  75                         n = n->rb_left;
  76                 else if (file_offset >= entry_end(entry))
  77                         n = n->rb_right;
  78                 else
  79                         return n;
  80         }
  81         if (!prev_ret)
  82                 return NULL;
  83
  84         while (prev && file_offset >= entry_end(prev_entry)) {
  85                 test = rb_next(prev);
  86                 if (!test)
  87                         break;
  88                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
  89                                       rb_node);
  90                 if (file_offset < entry_end(prev_entry))
  91                         break;
  92
  93                 prev = test;
  94         }
  95         if (prev)
  96                 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
  97                                       rb_node);
  98         while (prev && file_offset < entry_end(prev_entry)) {
  99                 test = rb_prev(prev);
 100                 if (!test)
 101                         break;
 102                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
 103                                       rb_node);
 104                 prev = test;
 105         }
 106         *prev_ret = prev;
 107         return NULL;
 108 }
 109
 110 static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
 111                           u64 len)
 112 {
 113         if (file_offset + len <= entry->file_offset ||
 114             entry->file_offset + entry->num_bytes <= file_offset)
 115                 return 0;
 116         return 1;
 117 }
 118
 119 /*
 120  * look find the first ordered struct that has this offset, otherwise
 121  * the first one less than this offset
 122  */
 123 static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
 124                                           u64 file_offset)
 125 {
 126         struct rb_root *root = &tree->tree;
 127         struct rb_node *prev = NULL;
 128         struct rb_node *ret;
 129         struct btrfs_ordered_extent *entry;
 130
 131         if (tree->last) {
 132                 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
 133                                  rb_node);
 134                 if (in_range(file_offset, entry->file_offset, entry->num_bytes))
 135                         return tree->last;
 136         }
 137         ret = __tree_search(root, file_offset, &prev);
 138         if (!ret)
 139                 ret = prev;
 140         if (ret)
 141                 tree->last = ret;
 142         return ret;
 143 }
 144
 145 /*
 146  * Allocate and add a new ordered_extent into the per-inode tree.
 147  *
 148  * The tree is given a single reference on the ordered extent that was
 149  * inserted.
 150  */
 151 static int __btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
 152                                       u64 disk_bytenr, u64 num_bytes,
 153                                       u64 disk_num_bytes, int type, int dio,
 154                                       int compress_type)
 155 {
 156         struct btrfs_root *root = inode->root;
 157         struct btrfs_fs_info *fs_info = root->fs_info;
 158         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 159         struct rb_node *node;
 160         struct btrfs_ordered_extent *entry;
 161         int ret;
 162
 163         if (type == BTRFS_ORDERED_NOCOW || type == BTRFS_ORDERED_PREALLOC) {
 164                 /* For nocow write, we can release the qgroup rsv right now */
 165                 ret = btrfs_qgroup_free_data(inode, NULL, file_offset, num_bytes);
 166                 if (ret < 0)
 167                         return ret;
 168                 ret = 0;
 169         } else {
 170                 /*
 171                  * The ordered extent has reserved qgroup space, release now
 172                  * and pass the reserved number for qgroup_record to free.
 173                  */
 174                 ret = btrfs_qgroup_release_data(inode, file_offset, num_bytes);
 175                 if (ret < 0)
 176                         return ret;
 177         }
 178         entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
 179         if (!entry)
 180                 return -ENOMEM;
 181
 182         entry->file_offset = file_offset;
 183         entry->disk_bytenr = disk_bytenr;
 184         entry->num_bytes = num_bytes;
 185         entry->disk_num_bytes = disk_num_bytes;
 186         entry->bytes_left = num_bytes;
 187         entry->inode = igrab(&inode->vfs_inode);
 188         entry->compress_type = compress_type;
 189         entry->truncated_len = (u64)-1;
 190         entry->qgroup_rsv = ret;
 191         entry->physical = (u64)-1;
 192         entry->disk = NULL;
 193         entry->partno = (u8)-1;
 194
 195         ASSERT(type == BTRFS_ORDERED_REGULAR ||
 196                type == BTRFS_ORDERED_NOCOW ||
 197                type == BTRFS_ORDERED_PREALLOC ||
 198                type == BTRFS_ORDERED_COMPRESSED);
 199         set_bit(type, &entry->flags);
 200
 201         percpu_counter_add_batch(&fs_info->ordered_bytes, num_bytes,
 202                                  fs_info->delalloc_batch);
 203
 204         if (dio)
 205                 set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
 206
 207         /* one ref for the tree */
 208         refcount_set(&entry->refs, 1);
 209         init_waitqueue_head(&entry->wait);
 210         INIT_LIST_HEAD(&entry->list);
 211         INIT_LIST_HEAD(&entry->log_list);
 212         INIT_LIST_HEAD(&entry->root_extent_list);
 213         INIT_LIST_HEAD(&entry->work_list);
 214         init_completion(&entry->completion);
 215
 216         trace_btrfs_ordered_extent_add(inode, entry);
 217
 218         spin_lock_irq(&tree->lock);
 219         node = tree_insert(&tree->tree, file_offset,
 220                            &entry->rb_node);
 221         if (node)
 222                 btrfs_panic(fs_info, -EEXIST,
 223                                 "inconsistency in ordered tree at offset %llu",
 224                                 file_offset);
 225         spin_unlock_irq(&tree->lock);
 226
 227         spin_lock(&root->ordered_extent_lock);
 228         list_add_tail(&entry->root_extent_list,
 229                       &root->ordered_extents);
 230         root->nr_ordered_extents++;
 231         if (root->nr_ordered_extents == 1) {
 232                 spin_lock(&fs_info->ordered_root_lock);
 233                 BUG_ON(!list_empty(&root->ordered_root));
 234                 list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
 235                 spin_unlock(&fs_info->ordered_root_lock);
 236         }
 237         spin_unlock(&root->ordered_extent_lock);
 238
 239         /*
 240          * We don't need the count_max_extents here, we can assume that all of
 241          * that work has been done at higher layers, so this is truly the
 242          * smallest the extent is going to get.
 243          */
 244         spin_lock(&inode->lock);
 245         btrfs_mod_outstanding_extents(inode, 1);
 246         spin_unlock(&inode->lock);
 247
 248         return 0;
 249 }
 250
 251 int btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
 252                              u64 disk_bytenr, u64 num_bytes, u64 disk_num_bytes,
 253                              int type)
 254 {
 255         ASSERT(type == BTRFS_ORDERED_REGULAR ||
 256                type == BTRFS_ORDERED_NOCOW ||
 257                type == BTRFS_ORDERED_PREALLOC);
 258         return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
 259                                           num_bytes, disk_num_bytes, type, 0,
 260                                           BTRFS_COMPRESS_NONE);
 261 }
 262
 263 int btrfs_add_ordered_extent_dio(struct btrfs_inode *inode, u64 file_offset,
 264                                  u64 disk_bytenr, u64 num_bytes,
 265                                  u64 disk_num_bytes, int type)
 266 {
 267         ASSERT(type == BTRFS_ORDERED_REGULAR ||
 268                type == BTRFS_ORDERED_NOCOW ||
 269                type == BTRFS_ORDERED_PREALLOC);
 270         return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
 271                                           num_bytes, disk_num_bytes, type, 1,
 272                                           BTRFS_COMPRESS_NONE);
 273 }
 274
 275 int btrfs_add_ordered_extent_compress(struct btrfs_inode *inode, u64 file_offset,
 276                                       u64 disk_bytenr, u64 num_bytes,
 277                                       u64 disk_num_bytes, int compress_type)
 278 {
 279         ASSERT(compress_type != BTRFS_COMPRESS_NONE);
 280         return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
 281                                           num_bytes, disk_num_bytes,
 282                                           BTRFS_ORDERED_COMPRESSED, 0,
 283                                           compress_type);
 284 }
 285
 286 /*
 287  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
 288  * when an ordered extent is finished.  If the list covers more than one
 289  * ordered extent, it is split across multiples.
 290  */
 291 void btrfs_add_ordered_sum(struct btrfs_ordered_extent *entry,
 292                            struct btrfs_ordered_sum *sum)
 293 {
 294         struct btrfs_ordered_inode_tree *tree;
 295
 296         tree = &BTRFS_I(entry->inode)->ordered_tree;
 297         spin_lock_irq(&tree->lock);
 298         list_add_tail(&sum->list, &entry->list);
 299         spin_unlock_irq(&tree->lock);
 300 }
 301
 302 /*
 303  * Mark all ordered extents io inside the specified range finished.
 304  *
 305  * @page:        The invovled page for the opeartion.
 306  *               For uncompressed buffered IO, the page status also needs to be
 307  *               updated to indicate whether the pending ordered io is finished.
 308  *               Can be NULL for direct IO and compressed write.
 309  *               For these cases, callers are ensured they won't execute the
 310  *               endio function twice.
 311  * @finish_func: The function to be executed when all the IO of an ordered
 312  *               extent are finished.
 313  *
 314  * This function is called for endio, thus the range must have ordered
 315  * extent(s) coveri it.
 316  */
 317 void btrfs_mark_ordered_io_finished(struct btrfs_inode *inode,
 318                                 struct page *page, u64 file_offset,
 319                                 u64 num_bytes, btrfs_func_t finish_func,
 320                                 bool uptodate)
 321 {
 322         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 323         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 324         struct btrfs_workqueue *wq;
 325         struct rb_node *node;
 326         struct btrfs_ordered_extent *entry = NULL;
 327         unsigned long flags;
 328         u64 cur = file_offset;
 329
 330         if (btrfs_is_free_space_inode(inode))
 331                 wq = fs_info->endio_freespace_worker;
 332         else
 333                 wq = fs_info->endio_write_workers;
 334
 335         if (page)
 336                 ASSERT(page->mapping && page_offset(page) <= file_offset &&
 337                        file_offset + num_bytes <= page_offset(page) + PAGE_SIZE);
 338
 339         spin_lock_irqsave(&tree->lock, flags);
 340         while (cur < file_offset + num_bytes) {
 341                 u64 entry_end;
 342                 u64 end;
 343                 u32 len;
 344
 345                 node = tree_search(tree, cur);
 346                 /* No ordered extents at all */
 347                 if (!node)
 348                         break;
 349
 350                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 351                 entry_end = entry->file_offset + entry->num_bytes;
 352                 /*
 353                  * |<-- OE --->|  |
 354                  *                cur
 355                  * Go to next OE.
 356                  */
 357                 if (cur >= entry_end) {
 358                         node = rb_next(node);
 359                         /* No more ordered extents, exit */
 360                         if (!node)
 361                                 break;
 362                         entry = rb_entry(node, struct btrfs_ordered_extent,
 363                                          rb_node);
 364
 365                         /* Go to next ordered extent and continue */
 366                         cur = entry->file_offset;
 367                         continue;
 368                 }
 369                 /*
 370                  * |    |<--- OE --->|
 371                  * cur
 372                  * Go to the start of OE.
 373                  */
 374                 if (cur < entry->file_offset) {
 375                         cur = entry->file_offset;
 376                         continue;
 377                 }
 378
 379                 /*
 380                  * Now we are definitely inside one ordered extent.
 381                  *
 382                  * |<--- OE --->|
 383                  *      |
 384                  *      cur
 385                  */
 386                 end = min(entry->file_offset + entry->num_bytes,
 387                           file_offset + num_bytes) - 1;
 388                 ASSERT(end + 1 - cur < U32_MAX);
 389                 len = end + 1 - cur;
 390
 391                 if (page) {
 392                         /*
 393                          * Private2 bit indicates whether we still have pending
 394                          * io unfinished for the ordered extent.
 395                          *
 396                          * If there's no such bit, we need to skip to next range.
 397                          */
 398                         if (!PagePrivate2(page)) {
 399                                 cur += len;
 400                                 continue;
 401                         }
 402                         ClearPagePrivate2(page);
 403                 }
 404
 405                 /* Now we're fine to update the accounting */
 406                 if (unlikely(len > entry->bytes_left)) {
 407                         WARN_ON(1);
 408                         btrfs_crit(fs_info,
 409 "bad ordered extent accounting, root=%llu ino=%llu OE offset=%llu OE len=%llu to_dec=%u left=%llu",
 410                                    inode->root->root_key.objectid,
 411                                    btrfs_ino(inode),
 412                                    entry->file_offset,
 413                                    entry->num_bytes,
 414                                    len, entry->bytes_left);
 415                         entry->bytes_left = 0;
 416                 } else {
 417                         entry->bytes_left -= len;
 418                 }
 419
 420                 if (!uptodate)
 421                         set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
 422
 423                 /*
 424                  * All the IO of the ordered extent is finished, we need to queue
 425                  * the finish_func to be executed.
 426                  */
 427                 if (entry->bytes_left == 0) {
 428                         set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 429                         cond_wake_up(&entry->wait);
 430                         refcount_inc(&entry->refs);
 431                         spin_unlock_irqrestore(&tree->lock, flags);
 432                         btrfs_init_work(&entry->work, finish_func, NULL, NULL);
 433                         btrfs_queue_work(wq, &entry->work);
 434                         spin_lock_irqsave(&tree->lock, flags);
 435                 }
 436                 cur += len;
 437         }
 438         spin_unlock_irqrestore(&tree->lock, flags);
 439 }
 440
 441 /*
 442  * Finish IO for one ordered extent across a given range.  The range can only
 443  * contain one ordered extent.
 444  *
 445  * @cached:      The cached ordered extent. If not NULL, we can skip the tree
 446  *               search and use the ordered extent directly.
 447  *               Will be also used to store the finished ordered extent.
 448  * @file_offset: File offset for the finished IO
 449  * @io_size:     Length of the finish IO range
 450  * @uptodate:    If the IO finishes without problem
 451  *
 452  * Return true if the ordered extent is finished in the range, and update
 453  * @cached.
 454  * Return false otherwise.
 455  *
 456  * NOTE: The range can NOT cross multiple ordered extents.
 457  * Thus caller should ensure the range doesn't cross ordered extents.
 458  */
 459 bool btrfs_dec_test_ordered_pending(struct btrfs_inode *inode,
 460                                     struct btrfs_ordered_extent **cached,
 461                                     u64 file_offset, u64 io_size, int uptodate)
 462 {
 463         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 464         struct rb_node *node;
 465         struct btrfs_ordered_extent *entry = NULL;
 466         unsigned long flags;
 467         bool finished = false;
 468
 469         spin_lock_irqsave(&tree->lock, flags);
 470         if (cached && *cached) {
 471                 entry = *cached;
 472                 goto have_entry;
 473         }
 474
 475         node = tree_search(tree, file_offset);
 476         if (!node)
 477                 goto out;
 478
 479         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 480 have_entry:
 481         if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
 482                 goto out;
 483
 484         if (io_size > entry->bytes_left)
 485                 btrfs_crit(inode->root->fs_info,
 486                            "bad ordered accounting left %llu size %llu",
 487                        entry->bytes_left, io_size);
 488
 489         entry->bytes_left -= io_size;
 490         if (!uptodate)
 491                 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
 492
 493         if (entry->bytes_left == 0) {
 494                 /*
 495                  * Ensure only one caller can set the flag and finished_ret
 496                  * accordingly
 497                  */
 498                 finished = !test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 499                 /* test_and_set_bit implies a barrier */
 500                 cond_wake_up_nomb(&entry->wait);
 501         }
 502 out:
 503         if (finished && cached && entry) {
 504                 *cached = entry;
 505                 refcount_inc(&entry->refs);
 506         }
 507         spin_unlock_irqrestore(&tree->lock, flags);
 508         return finished;
 509 }
 510
 511 /*
 512  * used to drop a reference on an ordered extent.  This will free
 513  * the extent if the last reference is dropped
 514  */
 515 void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
 516 {
 517         struct list_head *cur;
 518         struct btrfs_ordered_sum *sum;
 519
 520         trace_btrfs_ordered_extent_put(BTRFS_I(entry->inode), entry);
 521
 522         if (refcount_dec_and_test(&entry->refs)) {
 523                 ASSERT(list_empty(&entry->root_extent_list));
 524                 ASSERT(list_empty(&entry->log_list));
 525                 ASSERT(RB_EMPTY_NODE(&entry->rb_node));
 526                 if (entry->inode)
 527                         btrfs_add_delayed_iput(entry->inode);
 528                 while (!list_empty(&entry->list)) {
 529                         cur = entry->list.next;
 530                         sum = list_entry(cur, struct btrfs_ordered_sum, list);
 531                         list_del(&sum->list);
 532                         kvfree(sum);
 533                 }
 534                 kmem_cache_free(btrfs_ordered_extent_cache, entry);
 535         }
 536 }
 537
 538 /*
 539  * remove an ordered extent from the tree.  No references are dropped
 540  * and waiters are woken up.
 541  */
 542 void btrfs_remove_ordered_extent(struct btrfs_inode *btrfs_inode,
 543                                  struct btrfs_ordered_extent *entry)
 544 {
 545         struct btrfs_ordered_inode_tree *tree;
 546         struct btrfs_root *root = btrfs_inode->root;
 547         struct btrfs_fs_info *fs_info = root->fs_info;
 548         struct rb_node *node;
 549         bool pending;
 550
 551         /* This is paired with btrfs_add_ordered_extent. */
 552         spin_lock(&btrfs_inode->lock);
 553         btrfs_mod_outstanding_extents(btrfs_inode, -1);
 554         spin_unlock(&btrfs_inode->lock);
 555         if (root != fs_info->tree_root)
 556                 btrfs_delalloc_release_metadata(btrfs_inode, entry->num_bytes,
 557                                                 false);
 558
 559         percpu_counter_add_batch(&fs_info->ordered_bytes, -entry->num_bytes,
 560                                  fs_info->delalloc_batch);
 561
 562         tree = &btrfs_inode->ordered_tree;
 563         spin_lock_irq(&tree->lock);
 564         node = &entry->rb_node;
 565         rb_erase(node, &tree->tree);
 566         RB_CLEAR_NODE(node);
 567         if (tree->last == node)
 568                 tree->last = NULL;
 569         set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
 570         pending = test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags);
 571         spin_unlock_irq(&tree->lock);
 572
 573         /*
 574          * The current running transaction is waiting on us, we need to let it
 575          * know that we're complete and wake it up.
 576          */
 577         if (pending) {
 578                 struct btrfs_transaction *trans;
 579
 580                 /*
 581                  * The checks for trans are just a formality, it should be set,
 582                  * but if it isn't we don't want to deref/assert under the spin
 583                  * lock, so be nice and check if trans is set, but ASSERT() so
 584                  * if it isn't set a developer will notice.
 585                  */
 586                 spin_lock(&fs_info->trans_lock);
 587                 trans = fs_info->running_transaction;
 588                 if (trans)
 589                         refcount_inc(&trans->use_count);
 590                 spin_unlock(&fs_info->trans_lock);
 591
 592                 ASSERT(trans);
 593                 if (trans) {
 594                         if (atomic_dec_and_test(&trans->pending_ordered))
 595                                 wake_up(&trans->pending_wait);
 596                         btrfs_put_transaction(trans);
 597                 }
 598         }
 599
 600         spin_lock(&root->ordered_extent_lock);
 601         list_del_init(&entry->root_extent_list);
 602         root->nr_ordered_extents--;
 603
 604         trace_btrfs_ordered_extent_remove(btrfs_inode, entry);
 605
 606         if (!root->nr_ordered_extents) {
 607                 spin_lock(&fs_info->ordered_root_lock);
 608                 BUG_ON(list_empty(&root->ordered_root));
 609                 list_del_init(&root->ordered_root);
 610                 spin_unlock(&fs_info->ordered_root_lock);
 611         }
 612         spin_unlock(&root->ordered_extent_lock);
 613         wake_up(&entry->wait);
 614 }
 615
 616 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
 617 {
 618         struct btrfs_ordered_extent *ordered;
 619
 620         ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
 621         btrfs_start_ordered_extent(ordered, 1);
 622         complete(&ordered->completion);
 623 }
 624
 625 /*
 626  * wait for all the ordered extents in a root.  This is done when balancing
 627  * space between drives.
 628  */
 629 u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
 630                                const u64 range_start, const u64 range_len)
 631 {
 632         struct btrfs_fs_info *fs_info = root->fs_info;
 633         LIST_HEAD(splice);
 634         LIST_HEAD(skipped);
 635         LIST_HEAD(works);
 636         struct btrfs_ordered_extent *ordered, *next;
 637         u64 count = 0;
 638         const u64 range_end = range_start + range_len;
 639
 640         mutex_lock(&root->ordered_extent_mutex);
 641         spin_lock(&root->ordered_extent_lock);
 642         list_splice_init(&root->ordered_extents, &splice);
 643         while (!list_empty(&splice) && nr) {
 644                 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
 645                                            root_extent_list);
 646
 647                 if (range_end <= ordered->disk_bytenr ||
 648                     ordered->disk_bytenr + ordered->disk_num_bytes <= range_start) {
 649                         list_move_tail(&ordered->root_extent_list, &skipped);
 650                         cond_resched_lock(&root->ordered_extent_lock);
 651                         continue;
 652                 }
 653
 654                 list_move_tail(&ordered->root_extent_list,
 655                                &root->ordered_extents);
 656                 refcount_inc(&ordered->refs);
 657                 spin_unlock(&root->ordered_extent_lock);
 658
 659                 btrfs_init_work(&ordered->flush_work,
 660                                 btrfs_run_ordered_extent_work, NULL, NULL);
 661                 list_add_tail(&ordered->work_list, &works);
 662                 btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
 663
 664                 cond_resched();
 665                 spin_lock(&root->ordered_extent_lock);
 666                 if (nr != U64_MAX)
 667                         nr--;
 668                 count++;
 669         }
 670         list_splice_tail(&skipped, &root->ordered_extents);
 671         list_splice_tail(&splice, &root->ordered_extents);
 672         spin_unlock(&root->ordered_extent_lock);
 673
 674         list_for_each_entry_safe(ordered, next, &works, work_list) {
 675                 list_del_init(&ordered->work_list);
 676                 wait_for_completion(&ordered->completion);
 677                 btrfs_put_ordered_extent(ordered);
 678                 cond_resched();
 679         }
 680         mutex_unlock(&root->ordered_extent_mutex);
 681
 682         return count;
 683 }
 684
 685 void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
 686                              const u64 range_start, const u64 range_len)
 687 {
 688         struct btrfs_root *root;
 689         struct list_head splice;
 690         u64 done;
 691
 692         INIT_LIST_HEAD(&splice);
 693
 694         mutex_lock(&fs_info->ordered_operations_mutex);
 695         spin_lock(&fs_info->ordered_root_lock);
 696         list_splice_init(&fs_info->ordered_roots, &splice);
 697         while (!list_empty(&splice) && nr) {
 698                 root = list_first_entry(&splice, struct btrfs_root,
 699                                         ordered_root);
 700                 root = btrfs_grab_root(root);
 701                 BUG_ON(!root);
 702                 list_move_tail(&root->ordered_root,
 703                                &fs_info->ordered_roots);
 704                 spin_unlock(&fs_info->ordered_root_lock);
 705
 706                 done = btrfs_wait_ordered_extents(root, nr,
 707                                                   range_start, range_len);
 708                 btrfs_put_root(root);
 709
 710                 spin_lock(&fs_info->ordered_root_lock);
 711                 if (nr != U64_MAX) {
 712                         nr -= done;
 713                 }
 714         }
 715         list_splice_tail(&splice, &fs_info->ordered_roots);
 716         spin_unlock(&fs_info->ordered_root_lock);
 717         mutex_unlock(&fs_info->ordered_operations_mutex);
 718 }
 719
 720 /*
 721  * Used to start IO or wait for a given ordered extent to finish.
 722  *
 723  * If wait is one, this effectively waits on page writeback for all the pages
 724  * in the extent, and it waits on the io completion code to insert
 725  * metadata into the btree corresponding to the extent
 726  */
 727 void btrfs_start_ordered_extent(struct btrfs_ordered_extent *entry, int wait)
 728 {
 729         u64 start = entry->file_offset;
 730         u64 end = start + entry->num_bytes - 1;
 731         struct btrfs_inode *inode = BTRFS_I(entry->inode);
 732
 733         trace_btrfs_ordered_extent_start(inode, entry);
 734
 735         /*
 736          * pages in the range can be dirty, clean or writeback.  We
 737          * start IO on any dirty ones so the wait doesn't stall waiting
 738          * for the flusher thread to find them
 739          */
 740         if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
 741                 filemap_fdatawrite_range(inode->vfs_inode.i_mapping, start, end);
 742         if (wait) {
 743                 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
 744                                                  &entry->flags));
 745         }
 746 }
 747
 748 /*
 749  * Used to wait on ordered extents across a large range of bytes.
 750  */
 751 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
 752 {
 753         int ret = 0;
 754         int ret_wb = 0;
 755         u64 end;
 756         u64 orig_end;
 757         struct btrfs_ordered_extent *ordered;
 758
 759         if (start + len < start) {
 760                 orig_end = INT_LIMIT(loff_t);
 761         } else {
 762                 orig_end = start + len - 1;
 763                 if (orig_end > INT_LIMIT(loff_t))
 764                         orig_end = INT_LIMIT(loff_t);
 765         }
 766
 767         /* start IO across the range first to instantiate any delalloc
 768          * extents
 769          */
 770         ret = btrfs_fdatawrite_range(inode, start, orig_end);
 771         if (ret)
 772                 return ret;
 773
 774         /*
 775          * If we have a writeback error don't return immediately. Wait first
 776          * for any ordered extents that haven't completed yet. This is to make
 777          * sure no one can dirty the same page ranges and call writepages()
 778          * before the ordered extents complete - to avoid failures (-EEXIST)
 779          * when adding the new ordered extents to the ordered tree.
 780          */
 781         ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
 782
 783         end = orig_end;
 784         while (1) {
 785                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode), end);
 786                 if (!ordered)
 787                         break;
 788                 if (ordered->file_offset > orig_end) {
 789                         btrfs_put_ordered_extent(ordered);
 790                         break;
 791                 }
 792                 if (ordered->file_offset + ordered->num_bytes <= start) {
 793                         btrfs_put_ordered_extent(ordered);
 794                         break;
 795                 }
 796                 btrfs_start_ordered_extent(ordered, 1);
 797                 end = ordered->file_offset;
 798                 /*
 799                  * If the ordered extent had an error save the error but don't
 800                  * exit without waiting first for all other ordered extents in
 801                  * the range to complete.
 802                  */
 803                 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
 804                         ret = -EIO;
 805                 btrfs_put_ordered_extent(ordered);
 806                 if (end == 0 || end == start)
 807                         break;
 808                 end--;
 809         }
 810         return ret_wb ? ret_wb : ret;
 811 }
 812
 813 /*
 814  * find an ordered extent corresponding to file_offset.  return NULL if
 815  * nothing is found, otherwise take a reference on the extent and return it
 816  */
 817 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct btrfs_inode *inode,
 818                                                          u64 file_offset)
 819 {
 820         struct btrfs_ordered_inode_tree *tree;
 821         struct rb_node *node;
 822         struct btrfs_ordered_extent *entry = NULL;
 823         unsigned long flags;
 824
 825         tree = &inode->ordered_tree;
 826         spin_lock_irqsave(&tree->lock, flags);
 827         node = tree_search(tree, file_offset);
 828         if (!node)
 829                 goto out;
 830
 831         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 832         if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
 833                 entry = NULL;
 834         if (entry)
 835                 refcount_inc(&entry->refs);
 836 out:
 837         spin_unlock_irqrestore(&tree->lock, flags);
 838         return entry;
 839 }
 840
 841 /* Since the DIO code tries to lock a wide area we need to look for any ordered
 842  * extents that exist in the range, rather than just the start of the range.
 843  */
 844 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
 845                 struct btrfs_inode *inode, u64 file_offset, u64 len)
 846 {
 847         struct btrfs_ordered_inode_tree *tree;
 848         struct rb_node *node;
 849         struct btrfs_ordered_extent *entry = NULL;
 850
 851         tree = &inode->ordered_tree;
 852         spin_lock_irq(&tree->lock);
 853         node = tree_search(tree, file_offset);
 854         if (!node) {
 855                 node = tree_search(tree, file_offset + len);
 856                 if (!node)
 857                         goto out;
 858         }
 859
 860         while (1) {
 861                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 862                 if (range_overlaps(entry, file_offset, len))
 863                         break;
 864
 865                 if (entry->file_offset >= file_offset + len) {
 866                         entry = NULL;
 867                         break;
 868                 }
 869                 entry = NULL;
 870                 node = rb_next(node);
 871                 if (!node)
 872                         break;
 873         }
 874 out:
 875         if (entry)
 876                 refcount_inc(&entry->refs);
 877         spin_unlock_irq(&tree->lock);
 878         return entry;
 879 }
 880
 881 /*
 882  * Adds all ordered extents to the given list. The list ends up sorted by the
 883  * file_offset of the ordered extents.
 884  */
 885 void btrfs_get_ordered_extents_for_logging(struct btrfs_inode *inode,
 886                                            struct list_head *list)
 887 {
 888         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 889         struct rb_node *n;
 890
 891         ASSERT(inode_is_locked(&inode->vfs_inode));
 892
 893         spin_lock_irq(&tree->lock);
 894         for (n = rb_first(&tree->tree); n; n = rb_next(n)) {
 895                 struct btrfs_ordered_extent *ordered;
 896
 897                 ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
 898
 899                 if (test_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
 900                         continue;
 901
 902                 ASSERT(list_empty(&ordered->log_list));
 903                 list_add_tail(&ordered->log_list, list);
 904                 refcount_inc(&ordered->refs);
 905         }
 906         spin_unlock_irq(&tree->lock);
 907 }
 908
 909 /*
 910  * lookup and return any extent before 'file_offset'.  NULL is returned
 911  * if none is found
 912  */
 913 struct btrfs_ordered_extent *
 914 btrfs_lookup_first_ordered_extent(struct btrfs_inode *inode, u64 file_offset)
 915 {
 916         struct btrfs_ordered_inode_tree *tree;
 917         struct rb_node *node;
 918         struct btrfs_ordered_extent *entry = NULL;
 919
 920         tree = &inode->ordered_tree;
 921         spin_lock_irq(&tree->lock);
 922         node = tree_search(tree, file_offset);
 923         if (!node)
 924                 goto out;
 925
 926         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 927         refcount_inc(&entry->refs);
 928 out:
 929         spin_unlock_irq(&tree->lock);
 930         return entry;
 931 }
 932
 933 /*
 934  * btrfs_flush_ordered_range - Lock the passed range and ensures all pending
 935  * ordered extents in it are run to completion.
 936  *
 937  * @inode:        Inode whose ordered tree is to be searched
 938  * @start:        Beginning of range to flush
 939  * @end:          Last byte of range to lock
 940  * @cached_state: If passed, will return the extent state responsible for the
 941  * locked range. It's the caller's responsibility to free the cached state.
 942  *
 943  * This function always returns with the given range locked, ensuring after it's
 944  * called no order extent can be pending.
 945  */
 946 void btrfs_lock_and_flush_ordered_range(struct btrfs_inode *inode, u64 start,
 947                                         u64 end,
 948                                         struct extent_state **cached_state)
 949 {
 950         struct btrfs_ordered_extent *ordered;
 951         struct extent_state *cache = NULL;
 952         struct extent_state **cachedp = &cache;
 953
 954         if (cached_state)
 955                 cachedp = cached_state;
 956
 957         while (1) {
 958                 lock_extent_bits(&inode->io_tree, start, end, cachedp);
 959                 ordered = btrfs_lookup_ordered_range(inode, start,
 960                                                      end - start + 1);
 961                 if (!ordered) {
 962                         /*
 963                          * If no external cached_state has been passed then
 964                          * decrement the extra ref taken for cachedp since we
 965                          * aren't exposing it outside of this function
 966                          */
 967                         if (!cached_state)
 968                                 refcount_dec(&cache->refs);
 969                         break;
 970                 }
 971                 unlock_extent_cached(&inode->io_tree, start, end, cachedp);
 972                 btrfs_start_ordered_extent(ordered, 1);
 973                 btrfs_put_ordered_extent(ordered);
 974         }
 975 }
 976
 977 static int clone_ordered_extent(struct btrfs_ordered_extent *ordered, u64 pos,
 978                                 u64 len)
 979 {
 980         struct inode *inode = ordered->inode;
 981         u64 file_offset = ordered->file_offset + pos;
 982         u64 disk_bytenr = ordered->disk_bytenr + pos;
 983         u64 num_bytes = len;
 984         u64 disk_num_bytes = len;
 985         int type;
 986         unsigned long flags_masked = ordered->flags & ~(1 << BTRFS_ORDERED_DIRECT);
 987         int compress_type = ordered->compress_type;
 988         unsigned long weight;
 989         int ret;
 990
 991         weight = hweight_long(flags_masked);
 992         WARN_ON_ONCE(weight > 1);
 993         if (!weight)
 994                 type = 0;
 995         else
 996                 type = __ffs(flags_masked);
 997
 998         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered->flags)) {
 999                 WARN_ON_ONCE(1);
1000                 ret = btrfs_add_ordered_extent_compress(BTRFS_I(inode),
1001                                 file_offset, disk_bytenr, num_bytes,
1002                                 disk_num_bytes, compress_type);
1003         } else if (test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags)) {
1004                 ret = btrfs_add_ordered_extent_dio(BTRFS_I(inode), file_offset,
1005                                 disk_bytenr, num_bytes, disk_num_bytes, type);
1006         } else {
1007                 ret = btrfs_add_ordered_extent(BTRFS_I(inode), file_offset,
1008                                 disk_bytenr, num_bytes, disk_num_bytes, type);
1009         }
1010
1011         return ret;
1012 }
1013
1014 int btrfs_split_ordered_extent(struct btrfs_ordered_extent *ordered, u64 pre,
1015                                 u64 post)
1016 {
1017         struct inode *inode = ordered->inode;
1018         struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
1019         struct rb_node *node;
1020         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1021         int ret = 0;
1022
1023         spin_lock_irq(&tree->lock);
1024         /* Remove from tree once */
1025         node = &ordered->rb_node;
1026         rb_erase(node, &tree->tree);
1027         RB_CLEAR_NODE(node);
1028         if (tree->last == node)
1029                 tree->last = NULL;
1030
1031         ordered->file_offset += pre;
1032         ordered->disk_bytenr += pre;
1033         ordered->num_bytes -= (pre + post);
1034         ordered->disk_num_bytes -= (pre + post);
1035         ordered->bytes_left -= (pre + post);
1036
1037         /* Re-insert the node */
1038         node = tree_insert(&tree->tree, ordered->file_offset, &ordered->rb_node);
1039         if (node)
1040                 btrfs_panic(fs_info, -EEXIST,
1041                         "zoned: inconsistency in ordered tree at offset %llu",
1042                             ordered->file_offset);
1043
1044         spin_unlock_irq(&tree->lock);
1045
1046         if (pre)
1047                 ret = clone_ordered_extent(ordered, 0, pre);
1048         if (ret == 0 && post)
1049                 ret = clone_ordered_extent(ordered, pre + ordered->disk_num_bytes,
1050                                            post);
1051
1052         return ret;
1053 }
1054
1055 int __init ordered_data_init(void)
1056 {
1057         btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1058                                      sizeof(struct btrfs_ordered_extent), 0,
1059                                      SLAB_MEM_SPREAD,
1060                                      NULL);
1061         if (!btrfs_ordered_extent_cache)
1062                 return -ENOMEM;
1063
1064         return 0;
1065 }
1066
1067 void __cold ordered_data_exit(void)
1068 {
1069         kmem_cache_destroy(btrfs_ordered_extent_cache);
1070 }