1 // SPDX-License-Identifier: GPL-2.0 2 3 #include "misc.h" 4 #include "ctree.h" 5 #include "space-info.h" 6 #include "sysfs.h" 7 #include "volumes.h" 8 #include "free-space-cache.h" 9 #include "ordered-data.h" 10 #include "transaction.h" 11 #include "block-group.h" 12 13 /* 14 * HOW DOES SPACE RESERVATION WORK 15 * 16 * If you want to know about delalloc specifically, there is a separate comment 17 * for that with the delalloc code. This comment is about how the whole system 18 * works generally. 19 * 20 * BASIC CONCEPTS 21 * 22 * 1) space_info. This is the ultimate arbiter of how much space we can use. 23 * There's a description of the bytes_ fields with the struct declaration, 24 * refer to that for specifics on each field. Suffice it to say that for 25 * reservations we care about total_bytes - SUM(space_info->bytes_) when 26 * determining if there is space to make an allocation. There is a space_info 27 * for METADATA, SYSTEM, and DATA areas. 28 * 29 * 2) block_rsv's. These are basically buckets for every different type of 30 * metadata reservation we have. You can see the comment in the block_rsv 31 * code on the rules for each type, but generally block_rsv->reserved is how 32 * much space is accounted for in space_info->bytes_may_use. 33 * 34 * 3) btrfs_calc*_size. These are the worst case calculations we used based 35 * on the number of items we will want to modify. We have one for changing 36 * items, and one for inserting new items. Generally we use these helpers to 37 * determine the size of the block reserves, and then use the actual bytes 38 * values to adjust the space_info counters. 39 * 40 * MAKING RESERVATIONS, THE NORMAL CASE 41 * 42 * We call into either btrfs_reserve_data_bytes() or 43 * btrfs_reserve_metadata_bytes(), depending on which we're looking for, with 44 * num_bytes we want to reserve. 45 * 46 * ->reserve 47 * space_info->bytes_may_reserve += num_bytes 48 * 49 * ->extent allocation 50 * Call btrfs_add_reserved_bytes() which does 51 * space_info->bytes_may_reserve -= num_bytes 52 * space_info->bytes_reserved += extent_bytes 53 * 54 * ->insert reference 55 * Call btrfs_update_block_group() which does 56 * space_info->bytes_reserved -= extent_bytes 57 * space_info->bytes_used += extent_bytes 58 * 59 * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority) 60 * 61 * Assume we are unable to simply make the reservation because we do not have 62 * enough space 63 * 64 * -> __reserve_bytes 65 * create a reserve_ticket with ->bytes set to our reservation, add it to 66 * the tail of space_info->tickets, kick async flush thread 67 * 68 * ->handle_reserve_ticket 69 * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set 70 * on the ticket. 71 * 72 * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space 73 * Flushes various things attempting to free up space. 74 * 75 * -> btrfs_try_granting_tickets() 76 * This is called by anything that either subtracts space from 77 * space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the 78 * space_info->total_bytes. This loops through the ->priority_tickets and 79 * then the ->tickets list checking to see if the reservation can be 80 * completed. If it can the space is added to space_info->bytes_may_use and 81 * the ticket is woken up. 82 * 83 * -> ticket wakeup 84 * Check if ->bytes == 0, if it does we got our reservation and we can carry 85 * on, if not return the appropriate error (ENOSPC, but can be EINTR if we 86 * were interrupted.) 87 * 88 * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY 89 * 90 * Same as the above, except we add ourselves to the 91 * space_info->priority_tickets, and we do not use ticket->wait, we simply 92 * call flush_space() ourselves for the states that are safe for us to call 93 * without deadlocking and hope for the best. 94 * 95 * THE FLUSHING STATES 96 * 97 * Generally speaking we will have two cases for each state, a "nice" state 98 * and a "ALL THE THINGS" state. In btrfs we delay a lot of work in order to 99 * reduce the locking over head on the various trees, and even to keep from 100 * doing any work at all in the case of delayed refs. Each of these delayed 101 * things however hold reservations, and so letting them run allows us to 102 * reclaim space so we can make new reservations. 103 * 104 * FLUSH_DELAYED_ITEMS 105 * Every inode has a delayed item to update the inode. Take a simple write 106 * for example, we would update the inode item at write time to update the 107 * mtime, and then again at finish_ordered_io() time in order to update the 108 * isize or bytes. We keep these delayed items to coalesce these operations 109 * into a single operation done on demand. These are an easy way to reclaim 110 * metadata space. 111 * 112 * FLUSH_DELALLOC 113 * Look at the delalloc comment to get an idea of how much space is reserved 114 * for delayed allocation. We can reclaim some of this space simply by 115 * running delalloc, but usually we need to wait for ordered extents to 116 * reclaim the bulk of this space. 117 * 118 * FLUSH_DELAYED_REFS 119 * We have a block reserve for the outstanding delayed refs space, and every 120 * delayed ref operation holds a reservation. Running these is a quick way 121 * to reclaim space, but we want to hold this until the end because COW can 122 * churn a lot and we can avoid making some extent tree modifications if we 123 * are able to delay for as long as possible. 124 * 125 * ALLOC_CHUNK 126 * We will skip this the first time through space reservation, because of 127 * overcommit and we don't want to have a lot of useless metadata space when 128 * our worst case reservations will likely never come true. 129 * 130 * RUN_DELAYED_IPUTS 131 * If we're freeing inodes we're likely freeing checksums, file extent 132 * items, and extent tree items. Loads of space could be freed up by these 133 * operations, however they won't be usable until the transaction commits. 134 * 135 * COMMIT_TRANS 136 * may_commit_transaction() is the ultimate arbiter on whether we commit the 137 * transaction or not. In order to avoid constantly churning we do all the 138 * above flushing first and then commit the transaction as the last resort. 139 * However we need to take into account things like pinned space that would 140 * be freed, plus any delayed work we may not have gotten rid of in the case 141 * of metadata. 142 * 143 * OVERCOMMIT 144 * 145 * Because we hold so many reservations for metadata we will allow you to 146 * reserve more space than is currently free in the currently allocate 147 * metadata space. This only happens with metadata, data does not allow 148 * overcommitting. 149 * 150 * You can see the current logic for when we allow overcommit in 151 * btrfs_can_overcommit(), but it only applies to unallocated space. If there 152 * is no unallocated space to be had, all reservations are kept within the 153 * free space in the allocated metadata chunks. 154 * 155 * Because of overcommitting, you generally want to use the 156 * btrfs_can_overcommit() logic for metadata allocations, as it does the right 157 * thing with or without extra unallocated space. 158 */ 159 160 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info, 161 bool may_use_included) 162 { 163 ASSERT(s_info); 164 return s_info->bytes_used + s_info->bytes_reserved + 165 s_info->bytes_pinned + s_info->bytes_readonly + 166 (may_use_included ? s_info->bytes_may_use : 0); 167 } 168 169 /* 170 * after adding space to the filesystem, we need to clear the full flags 171 * on all the space infos. 172 */ 173 void btrfs_clear_space_info_full(struct btrfs_fs_info *info) 174 { 175 struct list_head *head = &info->space_info; 176 struct btrfs_space_info *found; 177 178 rcu_read_lock(); 179 list_for_each_entry_rcu(found, head, list) 180 found->full = 0; 181 rcu_read_unlock(); 182 } 183 184 static int create_space_info(struct btrfs_fs_info *info, u64 flags) 185 { 186 187 struct btrfs_space_info *space_info; 188 int i; 189 int ret; 190 191 space_info = kzalloc(sizeof(*space_info), GFP_NOFS); 192 if (!space_info) 193 return -ENOMEM; 194 195 ret = percpu_counter_init(&space_info->total_bytes_pinned, 0, 196 GFP_KERNEL); 197 if (ret) { 198 kfree(space_info); 199 return ret; 200 } 201 202 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) 203 INIT_LIST_HEAD(&space_info->block_groups[i]); 204 init_rwsem(&space_info->groups_sem); 205 spin_lock_init(&space_info->lock); 206 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK; 207 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE; 208 INIT_LIST_HEAD(&space_info->ro_bgs); 209 INIT_LIST_HEAD(&space_info->tickets); 210 INIT_LIST_HEAD(&space_info->priority_tickets); 211 212 ret = btrfs_sysfs_add_space_info_type(info, space_info); 213 if (ret) 214 return ret; 215 216 list_add_rcu(&space_info->list, &info->space_info); 217 if (flags & BTRFS_BLOCK_GROUP_DATA) 218 info->data_sinfo = space_info; 219 220 return ret; 221 } 222 223 int btrfs_init_space_info(struct btrfs_fs_info *fs_info) 224 { 225 struct btrfs_super_block *disk_super; 226 u64 features; 227 u64 flags; 228 int mixed = 0; 229 int ret; 230 231 disk_super = fs_info->super_copy; 232 if (!btrfs_super_root(disk_super)) 233 return -EINVAL; 234 235 features = btrfs_super_incompat_flags(disk_super); 236 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS) 237 mixed = 1; 238 239 flags = BTRFS_BLOCK_GROUP_SYSTEM; 240 ret = create_space_info(fs_info, flags); 241 if (ret) 242 goto out; 243 244 if (mixed) { 245 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA; 246 ret = create_space_info(fs_info, flags); 247 } else { 248 flags = BTRFS_BLOCK_GROUP_METADATA; 249 ret = create_space_info(fs_info, flags); 250 if (ret) 251 goto out; 252 253 flags = BTRFS_BLOCK_GROUP_DATA; 254 ret = create_space_info(fs_info, flags); 255 } 256 out: 257 return ret; 258 } 259 260 void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags, 261 u64 total_bytes, u64 bytes_used, 262 u64 bytes_readonly, 263 struct btrfs_space_info **space_info) 264 { 265 struct btrfs_space_info *found; 266 int factor; 267 268 factor = btrfs_bg_type_to_factor(flags); 269 270 found = btrfs_find_space_info(info, flags); 271 ASSERT(found); 272 spin_lock(&found->lock); 273 found->total_bytes += total_bytes; 274 found->disk_total += total_bytes * factor; 275 found->bytes_used += bytes_used; 276 found->disk_used += bytes_used * factor; 277 found->bytes_readonly += bytes_readonly; 278 if (total_bytes > 0) 279 found->full = 0; 280 btrfs_try_granting_tickets(info, found); 281 spin_unlock(&found->lock); 282 *space_info = found; 283 } 284 285 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info, 286 u64 flags) 287 { 288 struct list_head *head = &info->space_info; 289 struct btrfs_space_info *found; 290 291 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK; 292 293 rcu_read_lock(); 294 list_for_each_entry_rcu(found, head, list) { 295 if (found->flags & flags) { 296 rcu_read_unlock(); 297 return found; 298 } 299 } 300 rcu_read_unlock(); 301 return NULL; 302 } 303 304 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info, 305 struct btrfs_space_info *space_info, 306 enum btrfs_reserve_flush_enum flush) 307 { 308 u64 profile; 309 u64 avail; 310 int factor; 311 312 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM) 313 profile = btrfs_system_alloc_profile(fs_info); 314 else 315 profile = btrfs_metadata_alloc_profile(fs_info); 316 317 avail = atomic64_read(&fs_info->free_chunk_space); 318 319 /* 320 * If we have dup, raid1 or raid10 then only half of the free 321 * space is actually usable. For raid56, the space info used 322 * doesn't include the parity drive, so we don't have to 323 * change the math 324 */ 325 factor = btrfs_bg_type_to_factor(profile); 326 avail = div_u64(avail, factor); 327 328 /* 329 * If we aren't flushing all things, let us overcommit up to 330 * 1/2th of the space. If we can flush, don't let us overcommit 331 * too much, let it overcommit up to 1/8 of the space. 332 */ 333 if (flush == BTRFS_RESERVE_FLUSH_ALL) 334 avail >>= 3; 335 else 336 avail >>= 1; 337 return avail; 338 } 339 340 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info, 341 struct btrfs_space_info *space_info, u64 bytes, 342 enum btrfs_reserve_flush_enum flush) 343 { 344 u64 avail; 345 u64 used; 346 347 /* Don't overcommit when in mixed mode */ 348 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA) 349 return 0; 350 351 used = btrfs_space_info_used(space_info, true); 352 avail = calc_available_free_space(fs_info, space_info, flush); 353 354 if (used + bytes < space_info->total_bytes + avail) 355 return 1; 356 return 0; 357 } 358 359 static void remove_ticket(struct btrfs_space_info *space_info, 360 struct reserve_ticket *ticket) 361 { 362 if (!list_empty(&ticket->list)) { 363 list_del_init(&ticket->list); 364 ASSERT(space_info->reclaim_size >= ticket->bytes); 365 space_info->reclaim_size -= ticket->bytes; 366 } 367 } 368 369 /* 370 * This is for space we already have accounted in space_info->bytes_may_use, so 371 * basically when we're returning space from block_rsv's. 372 */ 373 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info, 374 struct btrfs_space_info *space_info) 375 { 376 struct list_head *head; 377 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH; 378 379 lockdep_assert_held(&space_info->lock); 380 381 head = &space_info->priority_tickets; 382 again: 383 while (!list_empty(head)) { 384 struct reserve_ticket *ticket; 385 u64 used = btrfs_space_info_used(space_info, true); 386 387 ticket = list_first_entry(head, struct reserve_ticket, list); 388 389 /* Check and see if our ticket can be satisified now. */ 390 if ((used + ticket->bytes <= space_info->total_bytes) || 391 btrfs_can_overcommit(fs_info, space_info, ticket->bytes, 392 flush)) { 393 btrfs_space_info_update_bytes_may_use(fs_info, 394 space_info, 395 ticket->bytes); 396 remove_ticket(space_info, ticket); 397 ticket->bytes = 0; 398 space_info->tickets_id++; 399 wake_up(&ticket->wait); 400 } else { 401 break; 402 } 403 } 404 405 if (head == &space_info->priority_tickets) { 406 head = &space_info->tickets; 407 flush = BTRFS_RESERVE_FLUSH_ALL; 408 goto again; 409 } 410 } 411 412 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \ 413 do { \ 414 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \ 415 spin_lock(&__rsv->lock); \ 416 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \ 417 __rsv->size, __rsv->reserved); \ 418 spin_unlock(&__rsv->lock); \ 419 } while (0) 420 421 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 422 struct btrfs_space_info *info) 423 { 424 lockdep_assert_held(&info->lock); 425 426 btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull", 427 info->flags, 428 info->total_bytes - btrfs_space_info_used(info, true), 429 info->full ? "" : "not "); 430 btrfs_info(fs_info, 431 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu", 432 info->total_bytes, info->bytes_used, info->bytes_pinned, 433 info->bytes_reserved, info->bytes_may_use, 434 info->bytes_readonly); 435 436 DUMP_BLOCK_RSV(fs_info, global_block_rsv); 437 DUMP_BLOCK_RSV(fs_info, trans_block_rsv); 438 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv); 439 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv); 440 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv); 441 442 } 443 444 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 445 struct btrfs_space_info *info, u64 bytes, 446 int dump_block_groups) 447 { 448 struct btrfs_block_group *cache; 449 int index = 0; 450 451 spin_lock(&info->lock); 452 __btrfs_dump_space_info(fs_info, info); 453 spin_unlock(&info->lock); 454 455 if (!dump_block_groups) 456 return; 457 458 down_read(&info->groups_sem); 459 again: 460 list_for_each_entry(cache, &info->block_groups[index], list) { 461 spin_lock(&cache->lock); 462 btrfs_info(fs_info, 463 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s", 464 cache->start, cache->length, cache->used, cache->pinned, 465 cache->reserved, cache->ro ? "[readonly]" : ""); 466 spin_unlock(&cache->lock); 467 btrfs_dump_free_space(cache, bytes); 468 } 469 if (++index < BTRFS_NR_RAID_TYPES) 470 goto again; 471 up_read(&info->groups_sem); 472 } 473 474 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info, 475 u64 to_reclaim) 476 { 477 u64 bytes; 478 u64 nr; 479 480 bytes = btrfs_calc_insert_metadata_size(fs_info, 1); 481 nr = div64_u64(to_reclaim, bytes); 482 if (!nr) 483 nr = 1; 484 return nr; 485 } 486 487 #define EXTENT_SIZE_PER_ITEM SZ_256K 488 489 /* 490 * shrink metadata reservation for delalloc 491 */ 492 static void shrink_delalloc(struct btrfs_fs_info *fs_info, 493 struct btrfs_space_info *space_info, 494 u64 to_reclaim, bool wait_ordered) 495 { 496 struct btrfs_trans_handle *trans; 497 u64 delalloc_bytes; 498 u64 dio_bytes; 499 u64 items; 500 long time_left; 501 int loops; 502 503 /* Calc the number of the pages we need flush for space reservation */ 504 if (to_reclaim == U64_MAX) { 505 items = U64_MAX; 506 } else { 507 /* 508 * to_reclaim is set to however much metadata we need to 509 * reclaim, but reclaiming that much data doesn't really track 510 * exactly, so increase the amount to reclaim by 2x in order to 511 * make sure we're flushing enough delalloc to hopefully reclaim 512 * some metadata reservations. 513 */ 514 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2; 515 to_reclaim = items * EXTENT_SIZE_PER_ITEM; 516 } 517 518 trans = (struct btrfs_trans_handle *)current->journal_info; 519 520 delalloc_bytes = percpu_counter_sum_positive( 521 &fs_info->delalloc_bytes); 522 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes); 523 if (delalloc_bytes == 0 && dio_bytes == 0) { 524 if (trans) 525 return; 526 if (wait_ordered) 527 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1); 528 return; 529 } 530 531 /* 532 * If we are doing more ordered than delalloc we need to just wait on 533 * ordered extents, otherwise we'll waste time trying to flush delalloc 534 * that likely won't give us the space back we need. 535 */ 536 if (dio_bytes > delalloc_bytes) 537 wait_ordered = true; 538 539 loops = 0; 540 while ((delalloc_bytes || dio_bytes) && loops < 3) { 541 btrfs_start_delalloc_roots(fs_info, items); 542 543 loops++; 544 if (wait_ordered && !trans) { 545 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1); 546 } else { 547 time_left = schedule_timeout_killable(1); 548 if (time_left) 549 break; 550 } 551 552 spin_lock(&space_info->lock); 553 if (list_empty(&space_info->tickets) && 554 list_empty(&space_info->priority_tickets)) { 555 spin_unlock(&space_info->lock); 556 break; 557 } 558 spin_unlock(&space_info->lock); 559 560 delalloc_bytes = percpu_counter_sum_positive( 561 &fs_info->delalloc_bytes); 562 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes); 563 } 564 } 565 566 /** 567 * maybe_commit_transaction - possibly commit the transaction if its ok to 568 * @root - the root we're allocating for 569 * @bytes - the number of bytes we want to reserve 570 * @force - force the commit 571 * 572 * This will check to make sure that committing the transaction will actually 573 * get us somewhere and then commit the transaction if it does. Otherwise it 574 * will return -ENOSPC. 575 */ 576 static int may_commit_transaction(struct btrfs_fs_info *fs_info, 577 struct btrfs_space_info *space_info) 578 { 579 struct reserve_ticket *ticket = NULL; 580 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv; 581 struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv; 582 struct btrfs_block_rsv *trans_rsv = &fs_info->trans_block_rsv; 583 struct btrfs_trans_handle *trans; 584 u64 reclaim_bytes = 0; 585 u64 bytes_needed = 0; 586 u64 cur_free_bytes = 0; 587 588 trans = (struct btrfs_trans_handle *)current->journal_info; 589 if (trans) 590 return -EAGAIN; 591 592 spin_lock(&space_info->lock); 593 cur_free_bytes = btrfs_space_info_used(space_info, true); 594 if (cur_free_bytes < space_info->total_bytes) 595 cur_free_bytes = space_info->total_bytes - cur_free_bytes; 596 else 597 cur_free_bytes = 0; 598 599 if (!list_empty(&space_info->priority_tickets)) 600 ticket = list_first_entry(&space_info->priority_tickets, 601 struct reserve_ticket, list); 602 else if (!list_empty(&space_info->tickets)) 603 ticket = list_first_entry(&space_info->tickets, 604 struct reserve_ticket, list); 605 if (ticket) 606 bytes_needed = ticket->bytes; 607 608 if (bytes_needed > cur_free_bytes) 609 bytes_needed -= cur_free_bytes; 610 else 611 bytes_needed = 0; 612 spin_unlock(&space_info->lock); 613 614 if (!bytes_needed) 615 return 0; 616 617 trans = btrfs_join_transaction(fs_info->extent_root); 618 if (IS_ERR(trans)) 619 return PTR_ERR(trans); 620 621 /* 622 * See if there is enough pinned space to make this reservation, or if 623 * we have block groups that are going to be freed, allowing us to 624 * possibly do a chunk allocation the next loop through. 625 */ 626 if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) || 627 __percpu_counter_compare(&space_info->total_bytes_pinned, 628 bytes_needed, 629 BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0) 630 goto commit; 631 632 /* 633 * See if there is some space in the delayed insertion reserve for this 634 * reservation. If the space_info's don't match (like for DATA or 635 * SYSTEM) then just go enospc, reclaiming this space won't recover any 636 * space to satisfy those reservations. 637 */ 638 if (space_info != delayed_rsv->space_info) 639 goto enospc; 640 641 spin_lock(&delayed_rsv->lock); 642 reclaim_bytes += delayed_rsv->reserved; 643 spin_unlock(&delayed_rsv->lock); 644 645 spin_lock(&delayed_refs_rsv->lock); 646 reclaim_bytes += delayed_refs_rsv->reserved; 647 spin_unlock(&delayed_refs_rsv->lock); 648 649 spin_lock(&trans_rsv->lock); 650 reclaim_bytes += trans_rsv->reserved; 651 spin_unlock(&trans_rsv->lock); 652 653 if (reclaim_bytes >= bytes_needed) 654 goto commit; 655 bytes_needed -= reclaim_bytes; 656 657 if (__percpu_counter_compare(&space_info->total_bytes_pinned, 658 bytes_needed, 659 BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0) 660 goto enospc; 661 662 commit: 663 return btrfs_commit_transaction(trans); 664 enospc: 665 btrfs_end_transaction(trans); 666 return -ENOSPC; 667 } 668 669 /* 670 * Try to flush some data based on policy set by @state. This is only advisory 671 * and may fail for various reasons. The caller is supposed to examine the 672 * state of @space_info to detect the outcome. 673 */ 674 static void flush_space(struct btrfs_fs_info *fs_info, 675 struct btrfs_space_info *space_info, u64 num_bytes, 676 int state) 677 { 678 struct btrfs_root *root = fs_info->extent_root; 679 struct btrfs_trans_handle *trans; 680 int nr; 681 int ret = 0; 682 683 switch (state) { 684 case FLUSH_DELAYED_ITEMS_NR: 685 case FLUSH_DELAYED_ITEMS: 686 if (state == FLUSH_DELAYED_ITEMS_NR) 687 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2; 688 else 689 nr = -1; 690 691 trans = btrfs_join_transaction(root); 692 if (IS_ERR(trans)) { 693 ret = PTR_ERR(trans); 694 break; 695 } 696 ret = btrfs_run_delayed_items_nr(trans, nr); 697 btrfs_end_transaction(trans); 698 break; 699 case FLUSH_DELALLOC: 700 case FLUSH_DELALLOC_WAIT: 701 shrink_delalloc(fs_info, space_info, num_bytes, 702 state == FLUSH_DELALLOC_WAIT); 703 break; 704 case FLUSH_DELAYED_REFS_NR: 705 case FLUSH_DELAYED_REFS: 706 trans = btrfs_join_transaction(root); 707 if (IS_ERR(trans)) { 708 ret = PTR_ERR(trans); 709 break; 710 } 711 if (state == FLUSH_DELAYED_REFS_NR) 712 nr = calc_reclaim_items_nr(fs_info, num_bytes); 713 else 714 nr = 0; 715 btrfs_run_delayed_refs(trans, nr); 716 btrfs_end_transaction(trans); 717 break; 718 case ALLOC_CHUNK: 719 case ALLOC_CHUNK_FORCE: 720 trans = btrfs_join_transaction(root); 721 if (IS_ERR(trans)) { 722 ret = PTR_ERR(trans); 723 break; 724 } 725 ret = btrfs_chunk_alloc(trans, 726 btrfs_get_alloc_profile(fs_info, space_info->flags), 727 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE : 728 CHUNK_ALLOC_FORCE); 729 btrfs_end_transaction(trans); 730 if (ret > 0 || ret == -ENOSPC) 731 ret = 0; 732 break; 733 case RUN_DELAYED_IPUTS: 734 /* 735 * If we have pending delayed iputs then we could free up a 736 * bunch of pinned space, so make sure we run the iputs before 737 * we do our pinned bytes check below. 738 */ 739 btrfs_run_delayed_iputs(fs_info); 740 btrfs_wait_on_delayed_iputs(fs_info); 741 break; 742 case COMMIT_TRANS: 743 ret = may_commit_transaction(fs_info, space_info); 744 break; 745 default: 746 ret = -ENOSPC; 747 break; 748 } 749 750 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state, 751 ret); 752 return; 753 } 754 755 static inline u64 756 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info, 757 struct btrfs_space_info *space_info) 758 { 759 u64 used; 760 u64 avail; 761 u64 expected; 762 u64 to_reclaim = space_info->reclaim_size; 763 764 lockdep_assert_held(&space_info->lock); 765 766 avail = calc_available_free_space(fs_info, space_info, 767 BTRFS_RESERVE_FLUSH_ALL); 768 used = btrfs_space_info_used(space_info, true); 769 770 /* 771 * We may be flushing because suddenly we have less space than we had 772 * before, and now we're well over-committed based on our current free 773 * space. If that's the case add in our overage so we make sure to put 774 * appropriate pressure on the flushing state machine. 775 */ 776 if (space_info->total_bytes + avail < used) 777 to_reclaim += used - (space_info->total_bytes + avail); 778 779 if (to_reclaim) 780 return to_reclaim; 781 782 to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M); 783 if (btrfs_can_overcommit(fs_info, space_info, to_reclaim, 784 BTRFS_RESERVE_FLUSH_ALL)) 785 return 0; 786 787 used = btrfs_space_info_used(space_info, true); 788 789 if (btrfs_can_overcommit(fs_info, space_info, SZ_1M, 790 BTRFS_RESERVE_FLUSH_ALL)) 791 expected = div_factor_fine(space_info->total_bytes, 95); 792 else 793 expected = div_factor_fine(space_info->total_bytes, 90); 794 795 if (used > expected) 796 to_reclaim = used - expected; 797 else 798 to_reclaim = 0; 799 to_reclaim = min(to_reclaim, space_info->bytes_may_use + 800 space_info->bytes_reserved); 801 return to_reclaim; 802 } 803 804 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info, 805 struct btrfs_space_info *space_info, 806 u64 used) 807 { 808 u64 thresh = div_factor_fine(space_info->total_bytes, 98); 809 810 /* If we're just plain full then async reclaim just slows us down. */ 811 if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh) 812 return 0; 813 814 if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info)) 815 return 0; 816 817 return (used >= thresh && !btrfs_fs_closing(fs_info) && 818 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state)); 819 } 820 821 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info, 822 struct btrfs_space_info *space_info, 823 struct reserve_ticket *ticket) 824 { 825 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 826 u64 min_bytes; 827 828 if (global_rsv->space_info != space_info) 829 return false; 830 831 spin_lock(&global_rsv->lock); 832 min_bytes = div_factor(global_rsv->size, 1); 833 if (global_rsv->reserved < min_bytes + ticket->bytes) { 834 spin_unlock(&global_rsv->lock); 835 return false; 836 } 837 global_rsv->reserved -= ticket->bytes; 838 remove_ticket(space_info, ticket); 839 ticket->bytes = 0; 840 wake_up(&ticket->wait); 841 space_info->tickets_id++; 842 if (global_rsv->reserved < global_rsv->size) 843 global_rsv->full = 0; 844 spin_unlock(&global_rsv->lock); 845 846 return true; 847 } 848 849 /* 850 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets 851 * @fs_info - fs_info for this fs 852 * @space_info - the space info we were flushing 853 * 854 * We call this when we've exhausted our flushing ability and haven't made 855 * progress in satisfying tickets. The reservation code handles tickets in 856 * order, so if there is a large ticket first and then smaller ones we could 857 * very well satisfy the smaller tickets. This will attempt to wake up any 858 * tickets in the list to catch this case. 859 * 860 * This function returns true if it was able to make progress by clearing out 861 * other tickets, or if it stumbles across a ticket that was smaller than the 862 * first ticket. 863 */ 864 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info, 865 struct btrfs_space_info *space_info) 866 { 867 struct reserve_ticket *ticket; 868 u64 tickets_id = space_info->tickets_id; 869 u64 first_ticket_bytes = 0; 870 871 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) { 872 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info"); 873 __btrfs_dump_space_info(fs_info, space_info); 874 } 875 876 while (!list_empty(&space_info->tickets) && 877 tickets_id == space_info->tickets_id) { 878 ticket = list_first_entry(&space_info->tickets, 879 struct reserve_ticket, list); 880 881 if (ticket->steal && 882 steal_from_global_rsv(fs_info, space_info, ticket)) 883 return true; 884 885 /* 886 * may_commit_transaction will avoid committing the transaction 887 * if it doesn't feel like the space reclaimed by the commit 888 * would result in the ticket succeeding. However if we have a 889 * smaller ticket in the queue it may be small enough to be 890 * satisified by committing the transaction, so if any 891 * subsequent ticket is smaller than the first ticket go ahead 892 * and send us back for another loop through the enospc flushing 893 * code. 894 */ 895 if (first_ticket_bytes == 0) 896 first_ticket_bytes = ticket->bytes; 897 else if (first_ticket_bytes > ticket->bytes) 898 return true; 899 900 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 901 btrfs_info(fs_info, "failing ticket with %llu bytes", 902 ticket->bytes); 903 904 remove_ticket(space_info, ticket); 905 ticket->error = -ENOSPC; 906 wake_up(&ticket->wait); 907 908 /* 909 * We're just throwing tickets away, so more flushing may not 910 * trip over btrfs_try_granting_tickets, so we need to call it 911 * here to see if we can make progress with the next ticket in 912 * the list. 913 */ 914 btrfs_try_granting_tickets(fs_info, space_info); 915 } 916 return (tickets_id != space_info->tickets_id); 917 } 918 919 /* 920 * This is for normal flushers, we can wait all goddamned day if we want to. We 921 * will loop and continuously try to flush as long as we are making progress. 922 * We count progress as clearing off tickets each time we have to loop. 923 */ 924 static void btrfs_async_reclaim_metadata_space(struct work_struct *work) 925 { 926 struct btrfs_fs_info *fs_info; 927 struct btrfs_space_info *space_info; 928 u64 to_reclaim; 929 int flush_state; 930 int commit_cycles = 0; 931 u64 last_tickets_id; 932 933 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work); 934 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 935 936 spin_lock(&space_info->lock); 937 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 938 if (!to_reclaim) { 939 space_info->flush = 0; 940 spin_unlock(&space_info->lock); 941 return; 942 } 943 last_tickets_id = space_info->tickets_id; 944 spin_unlock(&space_info->lock); 945 946 flush_state = FLUSH_DELAYED_ITEMS_NR; 947 do { 948 flush_space(fs_info, space_info, to_reclaim, flush_state); 949 spin_lock(&space_info->lock); 950 if (list_empty(&space_info->tickets)) { 951 space_info->flush = 0; 952 spin_unlock(&space_info->lock); 953 return; 954 } 955 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, 956 space_info); 957 if (last_tickets_id == space_info->tickets_id) { 958 flush_state++; 959 } else { 960 last_tickets_id = space_info->tickets_id; 961 flush_state = FLUSH_DELAYED_ITEMS_NR; 962 if (commit_cycles) 963 commit_cycles--; 964 } 965 966 /* 967 * We don't want to force a chunk allocation until we've tried 968 * pretty hard to reclaim space. Think of the case where we 969 * freed up a bunch of space and so have a lot of pinned space 970 * to reclaim. We would rather use that than possibly create a 971 * underutilized metadata chunk. So if this is our first run 972 * through the flushing state machine skip ALLOC_CHUNK_FORCE and 973 * commit the transaction. If nothing has changed the next go 974 * around then we can force a chunk allocation. 975 */ 976 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles) 977 flush_state++; 978 979 if (flush_state > COMMIT_TRANS) { 980 commit_cycles++; 981 if (commit_cycles > 2) { 982 if (maybe_fail_all_tickets(fs_info, space_info)) { 983 flush_state = FLUSH_DELAYED_ITEMS_NR; 984 commit_cycles--; 985 } else { 986 space_info->flush = 0; 987 } 988 } else { 989 flush_state = FLUSH_DELAYED_ITEMS_NR; 990 } 991 } 992 spin_unlock(&space_info->lock); 993 } while (flush_state <= COMMIT_TRANS); 994 } 995 996 /* 997 * FLUSH_DELALLOC_WAIT: 998 * Space is freed from flushing delalloc in one of two ways. 999 * 1000 * 1) compression is on and we allocate less space than we reserved 1001 * 2) we are overwriting existing space 1002 * 1003 * For #1 that extra space is reclaimed as soon as the delalloc pages are 1004 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent 1005 * length to ->bytes_reserved, and subtracts the reserved space from 1006 * ->bytes_may_use. 1007 * 1008 * For #2 this is trickier. Once the ordered extent runs we will drop the 1009 * extent in the range we are overwriting, which creates a delayed ref for 1010 * that freed extent. This however is not reclaimed until the transaction 1011 * commits, thus the next stages. 1012 * 1013 * RUN_DELAYED_IPUTS 1014 * If we are freeing inodes, we want to make sure all delayed iputs have 1015 * completed, because they could have been on an inode with i_nlink == 0, and 1016 * thus have been truncated and freed up space. But again this space is not 1017 * immediately re-usable, it comes in the form of a delayed ref, which must be 1018 * run and then the transaction must be committed. 1019 * 1020 * FLUSH_DELAYED_REFS 1021 * The above two cases generate delayed refs that will affect 1022 * ->total_bytes_pinned. However this counter can be inconsistent with 1023 * reality if there are outstanding delayed refs. This is because we adjust 1024 * the counter based solely on the current set of delayed refs and disregard 1025 * any on-disk state which might include more refs. So for example, if we 1026 * have an extent with 2 references, but we only drop 1, we'll see that there 1027 * is a negative delayed ref count for the extent and assume that the space 1028 * will be freed, and thus increase ->total_bytes_pinned. 1029 * 1030 * Running the delayed refs gives us the actual real view of what will be 1031 * freed at the transaction commit time. This stage will not actually free 1032 * space for us, it just makes sure that may_commit_transaction() has all of 1033 * the information it needs to make the right decision. 1034 * 1035 * COMMIT_TRANS 1036 * This is where we reclaim all of the pinned space generated by the previous 1037 * two stages. We will not commit the transaction if we don't think we're 1038 * likely to satisfy our request, which means if our current free space + 1039 * total_bytes_pinned < reservation we will not commit. This is why the 1040 * previous states are actually important, to make sure we know for sure 1041 * whether committing the transaction will allow us to make progress. 1042 * 1043 * ALLOC_CHUNK_FORCE 1044 * For data we start with alloc chunk force, however we could have been full 1045 * before, and then the transaction commit could have freed new block groups, 1046 * so if we now have space to allocate do the force chunk allocation. 1047 */ 1048 static const enum btrfs_flush_state data_flush_states[] = { 1049 FLUSH_DELALLOC_WAIT, 1050 RUN_DELAYED_IPUTS, 1051 FLUSH_DELAYED_REFS, 1052 COMMIT_TRANS, 1053 ALLOC_CHUNK_FORCE, 1054 }; 1055 1056 static void btrfs_async_reclaim_data_space(struct work_struct *work) 1057 { 1058 struct btrfs_fs_info *fs_info; 1059 struct btrfs_space_info *space_info; 1060 u64 last_tickets_id; 1061 int flush_state = 0; 1062 1063 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work); 1064 space_info = fs_info->data_sinfo; 1065 1066 spin_lock(&space_info->lock); 1067 if (list_empty(&space_info->tickets)) { 1068 space_info->flush = 0; 1069 spin_unlock(&space_info->lock); 1070 return; 1071 } 1072 last_tickets_id = space_info->tickets_id; 1073 spin_unlock(&space_info->lock); 1074 1075 while (!space_info->full) { 1076 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE); 1077 spin_lock(&space_info->lock); 1078 if (list_empty(&space_info->tickets)) { 1079 space_info->flush = 0; 1080 spin_unlock(&space_info->lock); 1081 return; 1082 } 1083 last_tickets_id = space_info->tickets_id; 1084 spin_unlock(&space_info->lock); 1085 } 1086 1087 while (flush_state < ARRAY_SIZE(data_flush_states)) { 1088 flush_space(fs_info, space_info, U64_MAX, 1089 data_flush_states[flush_state]); 1090 spin_lock(&space_info->lock); 1091 if (list_empty(&space_info->tickets)) { 1092 space_info->flush = 0; 1093 spin_unlock(&space_info->lock); 1094 return; 1095 } 1096 1097 if (last_tickets_id == space_info->tickets_id) { 1098 flush_state++; 1099 } else { 1100 last_tickets_id = space_info->tickets_id; 1101 flush_state = 0; 1102 } 1103 1104 if (flush_state >= ARRAY_SIZE(data_flush_states)) { 1105 if (space_info->full) { 1106 if (maybe_fail_all_tickets(fs_info, space_info)) 1107 flush_state = 0; 1108 else 1109 space_info->flush = 0; 1110 } else { 1111 flush_state = 0; 1112 } 1113 } 1114 spin_unlock(&space_info->lock); 1115 } 1116 } 1117 1118 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info) 1119 { 1120 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space); 1121 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space); 1122 } 1123 1124 static const enum btrfs_flush_state priority_flush_states[] = { 1125 FLUSH_DELAYED_ITEMS_NR, 1126 FLUSH_DELAYED_ITEMS, 1127 ALLOC_CHUNK, 1128 }; 1129 1130 static const enum btrfs_flush_state evict_flush_states[] = { 1131 FLUSH_DELAYED_ITEMS_NR, 1132 FLUSH_DELAYED_ITEMS, 1133 FLUSH_DELAYED_REFS_NR, 1134 FLUSH_DELAYED_REFS, 1135 FLUSH_DELALLOC, 1136 FLUSH_DELALLOC_WAIT, 1137 ALLOC_CHUNK, 1138 COMMIT_TRANS, 1139 }; 1140 1141 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info, 1142 struct btrfs_space_info *space_info, 1143 struct reserve_ticket *ticket, 1144 const enum btrfs_flush_state *states, 1145 int states_nr) 1146 { 1147 u64 to_reclaim; 1148 int flush_state; 1149 1150 spin_lock(&space_info->lock); 1151 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 1152 if (!to_reclaim) { 1153 spin_unlock(&space_info->lock); 1154 return; 1155 } 1156 spin_unlock(&space_info->lock); 1157 1158 flush_state = 0; 1159 do { 1160 flush_space(fs_info, space_info, to_reclaim, states[flush_state]); 1161 flush_state++; 1162 spin_lock(&space_info->lock); 1163 if (ticket->bytes == 0) { 1164 spin_unlock(&space_info->lock); 1165 return; 1166 } 1167 spin_unlock(&space_info->lock); 1168 } while (flush_state < states_nr); 1169 } 1170 1171 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info, 1172 struct btrfs_space_info *space_info, 1173 struct reserve_ticket *ticket) 1174 { 1175 while (!space_info->full) { 1176 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE); 1177 spin_lock(&space_info->lock); 1178 if (ticket->bytes == 0) { 1179 spin_unlock(&space_info->lock); 1180 return; 1181 } 1182 spin_unlock(&space_info->lock); 1183 } 1184 } 1185 1186 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info, 1187 struct btrfs_space_info *space_info, 1188 struct reserve_ticket *ticket) 1189 1190 { 1191 DEFINE_WAIT(wait); 1192 int ret = 0; 1193 1194 spin_lock(&space_info->lock); 1195 while (ticket->bytes > 0 && ticket->error == 0) { 1196 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE); 1197 if (ret) { 1198 /* 1199 * Delete us from the list. After we unlock the space 1200 * info, we don't want the async reclaim job to reserve 1201 * space for this ticket. If that would happen, then the 1202 * ticket's task would not known that space was reserved 1203 * despite getting an error, resulting in a space leak 1204 * (bytes_may_use counter of our space_info). 1205 */ 1206 remove_ticket(space_info, ticket); 1207 ticket->error = -EINTR; 1208 break; 1209 } 1210 spin_unlock(&space_info->lock); 1211 1212 schedule(); 1213 1214 finish_wait(&ticket->wait, &wait); 1215 spin_lock(&space_info->lock); 1216 } 1217 spin_unlock(&space_info->lock); 1218 } 1219 1220 /** 1221 * handle_reserve_ticket - do the appropriate flushing and waiting for a ticket 1222 * @fs_info - the fs 1223 * @space_info - the space_info for the reservation 1224 * @ticket - the ticket for the reservation 1225 * @flush - how much we can flush 1226 * 1227 * This does the work of figuring out how to flush for the ticket, waiting for 1228 * the reservation, and returning the appropriate error if there is one. 1229 */ 1230 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info, 1231 struct btrfs_space_info *space_info, 1232 struct reserve_ticket *ticket, 1233 enum btrfs_reserve_flush_enum flush) 1234 { 1235 int ret; 1236 1237 switch (flush) { 1238 case BTRFS_RESERVE_FLUSH_DATA: 1239 case BTRFS_RESERVE_FLUSH_ALL: 1240 case BTRFS_RESERVE_FLUSH_ALL_STEAL: 1241 wait_reserve_ticket(fs_info, space_info, ticket); 1242 break; 1243 case BTRFS_RESERVE_FLUSH_LIMIT: 1244 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1245 priority_flush_states, 1246 ARRAY_SIZE(priority_flush_states)); 1247 break; 1248 case BTRFS_RESERVE_FLUSH_EVICT: 1249 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1250 evict_flush_states, 1251 ARRAY_SIZE(evict_flush_states)); 1252 break; 1253 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE: 1254 priority_reclaim_data_space(fs_info, space_info, ticket); 1255 break; 1256 default: 1257 ASSERT(0); 1258 break; 1259 } 1260 1261 spin_lock(&space_info->lock); 1262 ret = ticket->error; 1263 if (ticket->bytes || ticket->error) { 1264 /* 1265 * We were a priority ticket, so we need to delete ourselves 1266 * from the list. Because we could have other priority tickets 1267 * behind us that require less space, run 1268 * btrfs_try_granting_tickets() to see if their reservations can 1269 * now be made. 1270 */ 1271 if (!list_empty(&ticket->list)) { 1272 remove_ticket(space_info, ticket); 1273 btrfs_try_granting_tickets(fs_info, space_info); 1274 } 1275 1276 if (!ret) 1277 ret = -ENOSPC; 1278 } 1279 spin_unlock(&space_info->lock); 1280 ASSERT(list_empty(&ticket->list)); 1281 /* 1282 * Check that we can't have an error set if the reservation succeeded, 1283 * as that would confuse tasks and lead them to error out without 1284 * releasing reserved space (if an error happens the expectation is that 1285 * space wasn't reserved at all). 1286 */ 1287 ASSERT(!(ticket->bytes == 0 && ticket->error)); 1288 return ret; 1289 } 1290 1291 /* 1292 * This returns true if this flush state will go through the ordinary flushing 1293 * code. 1294 */ 1295 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush) 1296 { 1297 return (flush == BTRFS_RESERVE_FLUSH_ALL) || 1298 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL); 1299 } 1300 1301 /** 1302 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space 1303 * @root - the root we're allocating for 1304 * @space_info - the space info we want to allocate from 1305 * @orig_bytes - the number of bytes we want 1306 * @flush - whether or not we can flush to make our reservation 1307 * 1308 * This will reserve orig_bytes number of bytes from the space info associated 1309 * with the block_rsv. If there is not enough space it will make an attempt to 1310 * flush out space to make room. It will do this by flushing delalloc if 1311 * possible or committing the transaction. If flush is 0 then no attempts to 1312 * regain reservations will be made and this will fail if there is not enough 1313 * space already. 1314 */ 1315 static int __reserve_bytes(struct btrfs_fs_info *fs_info, 1316 struct btrfs_space_info *space_info, u64 orig_bytes, 1317 enum btrfs_reserve_flush_enum flush) 1318 { 1319 struct work_struct *async_work; 1320 struct reserve_ticket ticket; 1321 u64 used; 1322 int ret = 0; 1323 bool pending_tickets; 1324 1325 ASSERT(orig_bytes); 1326 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL); 1327 1328 if (flush == BTRFS_RESERVE_FLUSH_DATA) 1329 async_work = &fs_info->async_data_reclaim_work; 1330 else 1331 async_work = &fs_info->async_reclaim_work; 1332 1333 spin_lock(&space_info->lock); 1334 ret = -ENOSPC; 1335 used = btrfs_space_info_used(space_info, true); 1336 1337 /* 1338 * We don't want NO_FLUSH allocations to jump everybody, they can 1339 * generally handle ENOSPC in a different way, so treat them the same as 1340 * normal flushers when it comes to skipping pending tickets. 1341 */ 1342 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH)) 1343 pending_tickets = !list_empty(&space_info->tickets) || 1344 !list_empty(&space_info->priority_tickets); 1345 else 1346 pending_tickets = !list_empty(&space_info->priority_tickets); 1347 1348 /* 1349 * Carry on if we have enough space (short-circuit) OR call 1350 * can_overcommit() to ensure we can overcommit to continue. 1351 */ 1352 if (!pending_tickets && 1353 ((used + orig_bytes <= space_info->total_bytes) || 1354 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) { 1355 btrfs_space_info_update_bytes_may_use(fs_info, space_info, 1356 orig_bytes); 1357 ret = 0; 1358 } 1359 1360 /* 1361 * If we couldn't make a reservation then setup our reservation ticket 1362 * and kick the async worker if it's not already running. 1363 * 1364 * If we are a priority flusher then we just need to add our ticket to 1365 * the list and we will do our own flushing further down. 1366 */ 1367 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) { 1368 ticket.bytes = orig_bytes; 1369 ticket.error = 0; 1370 space_info->reclaim_size += ticket.bytes; 1371 init_waitqueue_head(&ticket.wait); 1372 ticket.steal = (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL); 1373 if (flush == BTRFS_RESERVE_FLUSH_ALL || 1374 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || 1375 flush == BTRFS_RESERVE_FLUSH_DATA) { 1376 list_add_tail(&ticket.list, &space_info->tickets); 1377 if (!space_info->flush) { 1378 space_info->flush = 1; 1379 trace_btrfs_trigger_flush(fs_info, 1380 space_info->flags, 1381 orig_bytes, flush, 1382 "enospc"); 1383 queue_work(system_unbound_wq, async_work); 1384 } 1385 } else { 1386 list_add_tail(&ticket.list, 1387 &space_info->priority_tickets); 1388 } 1389 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) { 1390 used += orig_bytes; 1391 /* 1392 * We will do the space reservation dance during log replay, 1393 * which means we won't have fs_info->fs_root set, so don't do 1394 * the async reclaim as we will panic. 1395 */ 1396 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) && 1397 need_do_async_reclaim(fs_info, space_info, used) && 1398 !work_busy(&fs_info->async_reclaim_work)) { 1399 trace_btrfs_trigger_flush(fs_info, space_info->flags, 1400 orig_bytes, flush, "preempt"); 1401 queue_work(system_unbound_wq, 1402 &fs_info->async_reclaim_work); 1403 } 1404 } 1405 spin_unlock(&space_info->lock); 1406 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH) 1407 return ret; 1408 1409 return handle_reserve_ticket(fs_info, space_info, &ticket, flush); 1410 } 1411 1412 /** 1413 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space 1414 * @root - the root we're allocating for 1415 * @block_rsv - the block_rsv we're allocating for 1416 * @orig_bytes - the number of bytes we want 1417 * @flush - whether or not we can flush to make our reservation 1418 * 1419 * This will reserve orig_bytes number of bytes from the space info associated 1420 * with the block_rsv. If there is not enough space it will make an attempt to 1421 * flush out space to make room. It will do this by flushing delalloc if 1422 * possible or committing the transaction. If flush is 0 then no attempts to 1423 * regain reservations will be made and this will fail if there is not enough 1424 * space already. 1425 */ 1426 int btrfs_reserve_metadata_bytes(struct btrfs_root *root, 1427 struct btrfs_block_rsv *block_rsv, 1428 u64 orig_bytes, 1429 enum btrfs_reserve_flush_enum flush) 1430 { 1431 struct btrfs_fs_info *fs_info = root->fs_info; 1432 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 1433 int ret; 1434 1435 ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush); 1436 if (ret == -ENOSPC && 1437 unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) { 1438 if (block_rsv != global_rsv && 1439 !btrfs_block_rsv_use_bytes(global_rsv, orig_bytes)) 1440 ret = 0; 1441 } 1442 if (ret == -ENOSPC) { 1443 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1444 block_rsv->space_info->flags, 1445 orig_bytes, 1); 1446 1447 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1448 btrfs_dump_space_info(fs_info, block_rsv->space_info, 1449 orig_bytes, 0); 1450 } 1451 return ret; 1452 } 1453 1454 /** 1455 * btrfs_reserve_data_bytes - try to reserve data bytes for an allocation 1456 * @fs_info - the filesystem 1457 * @bytes - the number of bytes we need 1458 * @flush - how we are allowed to flush 1459 * 1460 * This will reserve bytes from the data space info. If there is not enough 1461 * space then we will attempt to flush space as specified by flush. 1462 */ 1463 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes, 1464 enum btrfs_reserve_flush_enum flush) 1465 { 1466 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo; 1467 int ret; 1468 1469 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA || 1470 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE); 1471 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA); 1472 1473 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush); 1474 if (ret == -ENOSPC) { 1475 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1476 data_sinfo->flags, bytes, 1); 1477 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1478 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0); 1479 } 1480 return ret; 1481 } 1482