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 * This will commit the transaction. Historically we had a lot of logic 137 * surrounding whether or not we'd commit the transaction, but this waits born 138 * out of a pre-tickets era where we could end up committing the transaction 139 * thousands of times in a row without making progress. Now thanks to our 140 * ticketing system we know if we're not making progress and can error 141 * everybody out after a few commits rather than burning the disk hoping for 142 * a different answer. 143 * 144 * OVERCOMMIT 145 * 146 * Because we hold so many reservations for metadata we will allow you to 147 * reserve more space than is currently free in the currently allocate 148 * metadata space. This only happens with metadata, data does not allow 149 * overcommitting. 150 * 151 * You can see the current logic for when we allow overcommit in 152 * btrfs_can_overcommit(), but it only applies to unallocated space. If there 153 * is no unallocated space to be had, all reservations are kept within the 154 * free space in the allocated metadata chunks. 155 * 156 * Because of overcommitting, you generally want to use the 157 * btrfs_can_overcommit() logic for metadata allocations, as it does the right 158 * thing with or without extra unallocated space. 159 */ 160 161 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info, 162 bool may_use_included) 163 { 164 ASSERT(s_info); 165 return s_info->bytes_used + s_info->bytes_reserved + 166 s_info->bytes_pinned + s_info->bytes_readonly + 167 s_info->bytes_zone_unusable + 168 (may_use_included ? s_info->bytes_may_use : 0); 169 } 170 171 /* 172 * after adding space to the filesystem, we need to clear the full flags 173 * on all the space infos. 174 */ 175 void btrfs_clear_space_info_full(struct btrfs_fs_info *info) 176 { 177 struct list_head *head = &info->space_info; 178 struct btrfs_space_info *found; 179 180 list_for_each_entry(found, head, list) 181 found->full = 0; 182 } 183 184 /* 185 * Block groups with more than this value (percents) of unusable space will be 186 * scheduled for background reclaim. 187 */ 188 #define BTRFS_DEFAULT_ZONED_RECLAIM_THRESH (75) 189 190 /* 191 * Calculate chunk size depending on volume type (regular or zoned). 192 */ 193 static u64 calc_chunk_size(const struct btrfs_fs_info *fs_info, u64 flags) 194 { 195 if (btrfs_is_zoned(fs_info)) 196 return fs_info->zone_size; 197 198 ASSERT(flags & BTRFS_BLOCK_GROUP_TYPE_MASK); 199 200 if (flags & BTRFS_BLOCK_GROUP_DATA) 201 return SZ_1G; 202 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM) 203 return SZ_32M; 204 205 /* Handle BTRFS_BLOCK_GROUP_METADATA */ 206 if (fs_info->fs_devices->total_rw_bytes > 50ULL * SZ_1G) 207 return SZ_1G; 208 209 return SZ_256M; 210 } 211 212 /* 213 * Update default chunk size. 214 */ 215 void btrfs_update_space_info_chunk_size(struct btrfs_space_info *space_info, 216 u64 chunk_size) 217 { 218 WRITE_ONCE(space_info->chunk_size, chunk_size); 219 } 220 221 static int create_space_info(struct btrfs_fs_info *info, u64 flags) 222 { 223 224 struct btrfs_space_info *space_info; 225 int i; 226 int ret; 227 228 space_info = kzalloc(sizeof(*space_info), GFP_NOFS); 229 if (!space_info) 230 return -ENOMEM; 231 232 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) 233 INIT_LIST_HEAD(&space_info->block_groups[i]); 234 init_rwsem(&space_info->groups_sem); 235 spin_lock_init(&space_info->lock); 236 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK; 237 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE; 238 INIT_LIST_HEAD(&space_info->ro_bgs); 239 INIT_LIST_HEAD(&space_info->tickets); 240 INIT_LIST_HEAD(&space_info->priority_tickets); 241 space_info->clamp = 1; 242 btrfs_update_space_info_chunk_size(space_info, calc_chunk_size(info, flags)); 243 244 if (btrfs_is_zoned(info)) 245 space_info->bg_reclaim_threshold = BTRFS_DEFAULT_ZONED_RECLAIM_THRESH; 246 247 ret = btrfs_sysfs_add_space_info_type(info, space_info); 248 if (ret) 249 return ret; 250 251 list_add(&space_info->list, &info->space_info); 252 if (flags & BTRFS_BLOCK_GROUP_DATA) 253 info->data_sinfo = space_info; 254 255 return ret; 256 } 257 258 int btrfs_init_space_info(struct btrfs_fs_info *fs_info) 259 { 260 struct btrfs_super_block *disk_super; 261 u64 features; 262 u64 flags; 263 int mixed = 0; 264 int ret; 265 266 disk_super = fs_info->super_copy; 267 if (!btrfs_super_root(disk_super)) 268 return -EINVAL; 269 270 features = btrfs_super_incompat_flags(disk_super); 271 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS) 272 mixed = 1; 273 274 flags = BTRFS_BLOCK_GROUP_SYSTEM; 275 ret = create_space_info(fs_info, flags); 276 if (ret) 277 goto out; 278 279 if (mixed) { 280 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA; 281 ret = create_space_info(fs_info, flags); 282 } else { 283 flags = BTRFS_BLOCK_GROUP_METADATA; 284 ret = create_space_info(fs_info, flags); 285 if (ret) 286 goto out; 287 288 flags = BTRFS_BLOCK_GROUP_DATA; 289 ret = create_space_info(fs_info, flags); 290 } 291 out: 292 return ret; 293 } 294 295 void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags, 296 u64 total_bytes, u64 bytes_used, 297 u64 bytes_readonly, u64 bytes_zone_unusable, 298 bool active, struct btrfs_space_info **space_info) 299 { 300 struct btrfs_space_info *found; 301 int factor; 302 303 factor = btrfs_bg_type_to_factor(flags); 304 305 found = btrfs_find_space_info(info, flags); 306 ASSERT(found); 307 spin_lock(&found->lock); 308 found->total_bytes += total_bytes; 309 if (active) 310 found->active_total_bytes += total_bytes; 311 found->disk_total += total_bytes * factor; 312 found->bytes_used += bytes_used; 313 found->disk_used += bytes_used * factor; 314 found->bytes_readonly += bytes_readonly; 315 found->bytes_zone_unusable += bytes_zone_unusable; 316 if (total_bytes > 0) 317 found->full = 0; 318 btrfs_try_granting_tickets(info, found); 319 spin_unlock(&found->lock); 320 *space_info = found; 321 } 322 323 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info, 324 u64 flags) 325 { 326 struct list_head *head = &info->space_info; 327 struct btrfs_space_info *found; 328 329 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK; 330 331 list_for_each_entry(found, head, list) { 332 if (found->flags & flags) 333 return found; 334 } 335 return NULL; 336 } 337 338 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info, 339 struct btrfs_space_info *space_info, 340 enum btrfs_reserve_flush_enum flush) 341 { 342 u64 profile; 343 u64 avail; 344 int factor; 345 346 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM) 347 profile = btrfs_system_alloc_profile(fs_info); 348 else 349 profile = btrfs_metadata_alloc_profile(fs_info); 350 351 avail = atomic64_read(&fs_info->free_chunk_space); 352 353 /* 354 * If we have dup, raid1 or raid10 then only half of the free 355 * space is actually usable. For raid56, the space info used 356 * doesn't include the parity drive, so we don't have to 357 * change the math 358 */ 359 factor = btrfs_bg_type_to_factor(profile); 360 avail = div_u64(avail, factor); 361 362 /* 363 * If we aren't flushing all things, let us overcommit up to 364 * 1/2th of the space. If we can flush, don't let us overcommit 365 * too much, let it overcommit up to 1/8 of the space. 366 */ 367 if (flush == BTRFS_RESERVE_FLUSH_ALL) 368 avail >>= 3; 369 else 370 avail >>= 1; 371 return avail; 372 } 373 374 static inline u64 writable_total_bytes(struct btrfs_fs_info *fs_info, 375 struct btrfs_space_info *space_info) 376 { 377 /* 378 * On regular filesystem, all total_bytes are always writable. On zoned 379 * filesystem, there may be a limitation imposed by max_active_zones. 380 * For metadata allocation, we cannot finish an existing active block 381 * group to avoid a deadlock. Thus, we need to consider only the active 382 * groups to be writable for metadata space. 383 */ 384 if (!btrfs_is_zoned(fs_info) || (space_info->flags & BTRFS_BLOCK_GROUP_DATA)) 385 return space_info->total_bytes; 386 387 return space_info->active_total_bytes; 388 } 389 390 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info, 391 struct btrfs_space_info *space_info, u64 bytes, 392 enum btrfs_reserve_flush_enum flush) 393 { 394 u64 avail; 395 u64 used; 396 397 /* Don't overcommit when in mixed mode */ 398 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA) 399 return 0; 400 401 used = btrfs_space_info_used(space_info, true); 402 if (btrfs_is_zoned(fs_info) && (space_info->flags & BTRFS_BLOCK_GROUP_METADATA)) 403 avail = 0; 404 else 405 avail = calc_available_free_space(fs_info, space_info, flush); 406 407 if (used + bytes < writable_total_bytes(fs_info, space_info) + avail) 408 return 1; 409 return 0; 410 } 411 412 static void remove_ticket(struct btrfs_space_info *space_info, 413 struct reserve_ticket *ticket) 414 { 415 if (!list_empty(&ticket->list)) { 416 list_del_init(&ticket->list); 417 ASSERT(space_info->reclaim_size >= ticket->bytes); 418 space_info->reclaim_size -= ticket->bytes; 419 } 420 } 421 422 /* 423 * This is for space we already have accounted in space_info->bytes_may_use, so 424 * basically when we're returning space from block_rsv's. 425 */ 426 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info, 427 struct btrfs_space_info *space_info) 428 { 429 struct list_head *head; 430 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH; 431 432 lockdep_assert_held(&space_info->lock); 433 434 head = &space_info->priority_tickets; 435 again: 436 while (!list_empty(head)) { 437 struct reserve_ticket *ticket; 438 u64 used = btrfs_space_info_used(space_info, true); 439 440 ticket = list_first_entry(head, struct reserve_ticket, list); 441 442 /* Check and see if our ticket can be satisfied now. */ 443 if ((used + ticket->bytes <= writable_total_bytes(fs_info, space_info)) || 444 btrfs_can_overcommit(fs_info, space_info, ticket->bytes, 445 flush)) { 446 btrfs_space_info_update_bytes_may_use(fs_info, 447 space_info, 448 ticket->bytes); 449 remove_ticket(space_info, ticket); 450 ticket->bytes = 0; 451 space_info->tickets_id++; 452 wake_up(&ticket->wait); 453 } else { 454 break; 455 } 456 } 457 458 if (head == &space_info->priority_tickets) { 459 head = &space_info->tickets; 460 flush = BTRFS_RESERVE_FLUSH_ALL; 461 goto again; 462 } 463 } 464 465 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \ 466 do { \ 467 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \ 468 spin_lock(&__rsv->lock); \ 469 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \ 470 __rsv->size, __rsv->reserved); \ 471 spin_unlock(&__rsv->lock); \ 472 } while (0) 473 474 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 475 struct btrfs_space_info *info) 476 { 477 lockdep_assert_held(&info->lock); 478 479 /* The free space could be negative in case of overcommit */ 480 btrfs_info(fs_info, "space_info %llu has %lld free, is %sfull", 481 info->flags, 482 (s64)(info->total_bytes - btrfs_space_info_used(info, true)), 483 info->full ? "" : "not "); 484 btrfs_info(fs_info, 485 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu", 486 info->total_bytes, info->bytes_used, info->bytes_pinned, 487 info->bytes_reserved, info->bytes_may_use, 488 info->bytes_readonly, info->bytes_zone_unusable); 489 490 DUMP_BLOCK_RSV(fs_info, global_block_rsv); 491 DUMP_BLOCK_RSV(fs_info, trans_block_rsv); 492 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv); 493 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv); 494 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv); 495 496 } 497 498 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 499 struct btrfs_space_info *info, u64 bytes, 500 int dump_block_groups) 501 { 502 struct btrfs_block_group *cache; 503 int index = 0; 504 505 spin_lock(&info->lock); 506 __btrfs_dump_space_info(fs_info, info); 507 spin_unlock(&info->lock); 508 509 if (!dump_block_groups) 510 return; 511 512 down_read(&info->groups_sem); 513 again: 514 list_for_each_entry(cache, &info->block_groups[index], list) { 515 spin_lock(&cache->lock); 516 btrfs_info(fs_info, 517 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu zone_unusable %s", 518 cache->start, cache->length, cache->used, cache->pinned, 519 cache->reserved, cache->zone_unusable, 520 cache->ro ? "[readonly]" : ""); 521 spin_unlock(&cache->lock); 522 btrfs_dump_free_space(cache, bytes); 523 } 524 if (++index < BTRFS_NR_RAID_TYPES) 525 goto again; 526 up_read(&info->groups_sem); 527 } 528 529 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info, 530 u64 to_reclaim) 531 { 532 u64 bytes; 533 u64 nr; 534 535 bytes = btrfs_calc_insert_metadata_size(fs_info, 1); 536 nr = div64_u64(to_reclaim, bytes); 537 if (!nr) 538 nr = 1; 539 return nr; 540 } 541 542 #define EXTENT_SIZE_PER_ITEM SZ_256K 543 544 /* 545 * shrink metadata reservation for delalloc 546 */ 547 static void shrink_delalloc(struct btrfs_fs_info *fs_info, 548 struct btrfs_space_info *space_info, 549 u64 to_reclaim, bool wait_ordered, 550 bool for_preempt) 551 { 552 struct btrfs_trans_handle *trans; 553 u64 delalloc_bytes; 554 u64 ordered_bytes; 555 u64 items; 556 long time_left; 557 int loops; 558 559 delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes); 560 ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes); 561 if (delalloc_bytes == 0 && ordered_bytes == 0) 562 return; 563 564 /* Calc the number of the pages we need flush for space reservation */ 565 if (to_reclaim == U64_MAX) { 566 items = U64_MAX; 567 } else { 568 /* 569 * to_reclaim is set to however much metadata we need to 570 * reclaim, but reclaiming that much data doesn't really track 571 * exactly. What we really want to do is reclaim full inode's 572 * worth of reservations, however that's not available to us 573 * here. We will take a fraction of the delalloc bytes for our 574 * flushing loops and hope for the best. Delalloc will expand 575 * the amount we write to cover an entire dirty extent, which 576 * will reclaim the metadata reservation for that range. If 577 * it's not enough subsequent flush stages will be more 578 * aggressive. 579 */ 580 to_reclaim = max(to_reclaim, delalloc_bytes >> 3); 581 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2; 582 } 583 584 trans = current->journal_info; 585 586 /* 587 * If we are doing more ordered than delalloc we need to just wait on 588 * ordered extents, otherwise we'll waste time trying to flush delalloc 589 * that likely won't give us the space back we need. 590 */ 591 if (ordered_bytes > delalloc_bytes && !for_preempt) 592 wait_ordered = true; 593 594 loops = 0; 595 while ((delalloc_bytes || ordered_bytes) && loops < 3) { 596 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT; 597 long nr_pages = min_t(u64, temp, LONG_MAX); 598 int async_pages; 599 600 btrfs_start_delalloc_roots(fs_info, nr_pages, true); 601 602 /* 603 * We need to make sure any outstanding async pages are now 604 * processed before we continue. This is because things like 605 * sync_inode() try to be smart and skip writing if the inode is 606 * marked clean. We don't use filemap_fwrite for flushing 607 * because we want to control how many pages we write out at a 608 * time, thus this is the only safe way to make sure we've 609 * waited for outstanding compressed workers to have started 610 * their jobs and thus have ordered extents set up properly. 611 * 612 * This exists because we do not want to wait for each 613 * individual inode to finish its async work, we simply want to 614 * start the IO on everybody, and then come back here and wait 615 * for all of the async work to catch up. Once we're done with 616 * that we know we'll have ordered extents for everything and we 617 * can decide if we wait for that or not. 618 * 619 * If we choose to replace this in the future, make absolutely 620 * sure that the proper waiting is being done in the async case, 621 * as there have been bugs in that area before. 622 */ 623 async_pages = atomic_read(&fs_info->async_delalloc_pages); 624 if (!async_pages) 625 goto skip_async; 626 627 /* 628 * We don't want to wait forever, if we wrote less pages in this 629 * loop than we have outstanding, only wait for that number of 630 * pages, otherwise we can wait for all async pages to finish 631 * before continuing. 632 */ 633 if (async_pages > nr_pages) 634 async_pages -= nr_pages; 635 else 636 async_pages = 0; 637 wait_event(fs_info->async_submit_wait, 638 atomic_read(&fs_info->async_delalloc_pages) <= 639 async_pages); 640 skip_async: 641 loops++; 642 if (wait_ordered && !trans) { 643 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1); 644 } else { 645 time_left = schedule_timeout_killable(1); 646 if (time_left) 647 break; 648 } 649 650 /* 651 * If we are for preemption we just want a one-shot of delalloc 652 * flushing so we can stop flushing if we decide we don't need 653 * to anymore. 654 */ 655 if (for_preempt) 656 break; 657 658 spin_lock(&space_info->lock); 659 if (list_empty(&space_info->tickets) && 660 list_empty(&space_info->priority_tickets)) { 661 spin_unlock(&space_info->lock); 662 break; 663 } 664 spin_unlock(&space_info->lock); 665 666 delalloc_bytes = percpu_counter_sum_positive( 667 &fs_info->delalloc_bytes); 668 ordered_bytes = percpu_counter_sum_positive( 669 &fs_info->ordered_bytes); 670 } 671 } 672 673 /* 674 * Try to flush some data based on policy set by @state. This is only advisory 675 * and may fail for various reasons. The caller is supposed to examine the 676 * state of @space_info to detect the outcome. 677 */ 678 static void flush_space(struct btrfs_fs_info *fs_info, 679 struct btrfs_space_info *space_info, u64 num_bytes, 680 enum btrfs_flush_state state, bool for_preempt) 681 { 682 struct btrfs_root *root = fs_info->tree_root; 683 struct btrfs_trans_handle *trans; 684 int nr; 685 int ret = 0; 686 687 switch (state) { 688 case FLUSH_DELAYED_ITEMS_NR: 689 case FLUSH_DELAYED_ITEMS: 690 if (state == FLUSH_DELAYED_ITEMS_NR) 691 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2; 692 else 693 nr = -1; 694 695 trans = btrfs_join_transaction(root); 696 if (IS_ERR(trans)) { 697 ret = PTR_ERR(trans); 698 break; 699 } 700 ret = btrfs_run_delayed_items_nr(trans, nr); 701 btrfs_end_transaction(trans); 702 break; 703 case FLUSH_DELALLOC: 704 case FLUSH_DELALLOC_WAIT: 705 case FLUSH_DELALLOC_FULL: 706 if (state == FLUSH_DELALLOC_FULL) 707 num_bytes = U64_MAX; 708 shrink_delalloc(fs_info, space_info, num_bytes, 709 state != FLUSH_DELALLOC, for_preempt); 710 break; 711 case FLUSH_DELAYED_REFS_NR: 712 case FLUSH_DELAYED_REFS: 713 trans = btrfs_join_transaction(root); 714 if (IS_ERR(trans)) { 715 ret = PTR_ERR(trans); 716 break; 717 } 718 if (state == FLUSH_DELAYED_REFS_NR) 719 nr = calc_reclaim_items_nr(fs_info, num_bytes); 720 else 721 nr = 0; 722 btrfs_run_delayed_refs(trans, nr); 723 btrfs_end_transaction(trans); 724 break; 725 case ALLOC_CHUNK: 726 case ALLOC_CHUNK_FORCE: 727 trans = btrfs_join_transaction(root); 728 if (IS_ERR(trans)) { 729 ret = PTR_ERR(trans); 730 break; 731 } 732 ret = btrfs_chunk_alloc(trans, 733 btrfs_get_alloc_profile(fs_info, space_info->flags), 734 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE : 735 CHUNK_ALLOC_FORCE); 736 btrfs_end_transaction(trans); 737 if (ret > 0 || ret == -ENOSPC) 738 ret = 0; 739 break; 740 case RUN_DELAYED_IPUTS: 741 /* 742 * If we have pending delayed iputs then we could free up a 743 * bunch of pinned space, so make sure we run the iputs before 744 * we do our pinned bytes check below. 745 */ 746 btrfs_run_delayed_iputs(fs_info); 747 btrfs_wait_on_delayed_iputs(fs_info); 748 break; 749 case COMMIT_TRANS: 750 ASSERT(current->journal_info == NULL); 751 trans = btrfs_join_transaction(root); 752 if (IS_ERR(trans)) { 753 ret = PTR_ERR(trans); 754 break; 755 } 756 ret = btrfs_commit_transaction(trans); 757 break; 758 default: 759 ret = -ENOSPC; 760 break; 761 } 762 763 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state, 764 ret, for_preempt); 765 return; 766 } 767 768 static inline u64 769 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info, 770 struct btrfs_space_info *space_info) 771 { 772 u64 used; 773 u64 avail; 774 u64 total; 775 u64 to_reclaim = space_info->reclaim_size; 776 777 lockdep_assert_held(&space_info->lock); 778 779 avail = calc_available_free_space(fs_info, space_info, 780 BTRFS_RESERVE_FLUSH_ALL); 781 used = btrfs_space_info_used(space_info, true); 782 783 /* 784 * We may be flushing because suddenly we have less space than we had 785 * before, and now we're well over-committed based on our current free 786 * space. If that's the case add in our overage so we make sure to put 787 * appropriate pressure on the flushing state machine. 788 */ 789 total = writable_total_bytes(fs_info, space_info); 790 if (total + avail < used) 791 to_reclaim += used - (total + avail); 792 793 return to_reclaim; 794 } 795 796 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info, 797 struct btrfs_space_info *space_info) 798 { 799 u64 global_rsv_size = fs_info->global_block_rsv.reserved; 800 u64 ordered, delalloc; 801 u64 total = writable_total_bytes(fs_info, space_info); 802 u64 thresh; 803 u64 used; 804 805 thresh = div_factor_fine(total, 90); 806 807 lockdep_assert_held(&space_info->lock); 808 809 /* If we're just plain full then async reclaim just slows us down. */ 810 if ((space_info->bytes_used + space_info->bytes_reserved + 811 global_rsv_size) >= thresh) 812 return false; 813 814 used = space_info->bytes_may_use + space_info->bytes_pinned; 815 816 /* The total flushable belongs to the global rsv, don't flush. */ 817 if (global_rsv_size >= used) 818 return false; 819 820 /* 821 * 128MiB is 1/4 of the maximum global rsv size. If we have less than 822 * that devoted to other reservations then there's no sense in flushing, 823 * we don't have a lot of things that need flushing. 824 */ 825 if (used - global_rsv_size <= SZ_128M) 826 return false; 827 828 /* 829 * We have tickets queued, bail so we don't compete with the async 830 * flushers. 831 */ 832 if (space_info->reclaim_size) 833 return false; 834 835 /* 836 * If we have over half of the free space occupied by reservations or 837 * pinned then we want to start flushing. 838 * 839 * We do not do the traditional thing here, which is to say 840 * 841 * if (used >= ((total_bytes + avail) / 2)) 842 * return 1; 843 * 844 * because this doesn't quite work how we want. If we had more than 50% 845 * of the space_info used by bytes_used and we had 0 available we'd just 846 * constantly run the background flusher. Instead we want it to kick in 847 * if our reclaimable space exceeds our clamped free space. 848 * 849 * Our clamping range is 2^1 -> 2^8. Practically speaking that means 850 * the following: 851 * 852 * Amount of RAM Minimum threshold Maximum threshold 853 * 854 * 256GiB 1GiB 128GiB 855 * 128GiB 512MiB 64GiB 856 * 64GiB 256MiB 32GiB 857 * 32GiB 128MiB 16GiB 858 * 16GiB 64MiB 8GiB 859 * 860 * These are the range our thresholds will fall in, corresponding to how 861 * much delalloc we need for the background flusher to kick in. 862 */ 863 864 thresh = calc_available_free_space(fs_info, space_info, 865 BTRFS_RESERVE_FLUSH_ALL); 866 used = space_info->bytes_used + space_info->bytes_reserved + 867 space_info->bytes_readonly + global_rsv_size; 868 if (used < total) 869 thresh += total - used; 870 thresh >>= space_info->clamp; 871 872 used = space_info->bytes_pinned; 873 874 /* 875 * If we have more ordered bytes than delalloc bytes then we're either 876 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting 877 * around. Preemptive flushing is only useful in that it can free up 878 * space before tickets need to wait for things to finish. In the case 879 * of ordered extents, preemptively waiting on ordered extents gets us 880 * nothing, if our reservations are tied up in ordered extents we'll 881 * simply have to slow down writers by forcing them to wait on ordered 882 * extents. 883 * 884 * In the case that ordered is larger than delalloc, only include the 885 * block reserves that we would actually be able to directly reclaim 886 * from. In this case if we're heavy on metadata operations this will 887 * clearly be heavy enough to warrant preemptive flushing. In the case 888 * of heavy DIO or ordered reservations, preemptive flushing will just 889 * waste time and cause us to slow down. 890 * 891 * We want to make sure we truly are maxed out on ordered however, so 892 * cut ordered in half, and if it's still higher than delalloc then we 893 * can keep flushing. This is to avoid the case where we start 894 * flushing, and now delalloc == ordered and we stop preemptively 895 * flushing when we could still have several gigs of delalloc to flush. 896 */ 897 ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1; 898 delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes); 899 if (ordered >= delalloc) 900 used += fs_info->delayed_refs_rsv.reserved + 901 fs_info->delayed_block_rsv.reserved; 902 else 903 used += space_info->bytes_may_use - global_rsv_size; 904 905 return (used >= thresh && !btrfs_fs_closing(fs_info) && 906 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state)); 907 } 908 909 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info, 910 struct btrfs_space_info *space_info, 911 struct reserve_ticket *ticket) 912 { 913 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 914 u64 min_bytes; 915 916 if (!ticket->steal) 917 return false; 918 919 if (global_rsv->space_info != space_info) 920 return false; 921 922 spin_lock(&global_rsv->lock); 923 min_bytes = div_factor(global_rsv->size, 1); 924 if (global_rsv->reserved < min_bytes + ticket->bytes) { 925 spin_unlock(&global_rsv->lock); 926 return false; 927 } 928 global_rsv->reserved -= ticket->bytes; 929 remove_ticket(space_info, ticket); 930 ticket->bytes = 0; 931 wake_up(&ticket->wait); 932 space_info->tickets_id++; 933 if (global_rsv->reserved < global_rsv->size) 934 global_rsv->full = 0; 935 spin_unlock(&global_rsv->lock); 936 937 return true; 938 } 939 940 /* 941 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets 942 * @fs_info - fs_info for this fs 943 * @space_info - the space info we were flushing 944 * 945 * We call this when we've exhausted our flushing ability and haven't made 946 * progress in satisfying tickets. The reservation code handles tickets in 947 * order, so if there is a large ticket first and then smaller ones we could 948 * very well satisfy the smaller tickets. This will attempt to wake up any 949 * tickets in the list to catch this case. 950 * 951 * This function returns true if it was able to make progress by clearing out 952 * other tickets, or if it stumbles across a ticket that was smaller than the 953 * first ticket. 954 */ 955 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info, 956 struct btrfs_space_info *space_info) 957 { 958 struct reserve_ticket *ticket; 959 u64 tickets_id = space_info->tickets_id; 960 const bool aborted = BTRFS_FS_ERROR(fs_info); 961 962 trace_btrfs_fail_all_tickets(fs_info, space_info); 963 964 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) { 965 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info"); 966 __btrfs_dump_space_info(fs_info, space_info); 967 } 968 969 while (!list_empty(&space_info->tickets) && 970 tickets_id == space_info->tickets_id) { 971 ticket = list_first_entry(&space_info->tickets, 972 struct reserve_ticket, list); 973 974 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket)) 975 return true; 976 977 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 978 btrfs_info(fs_info, "failing ticket with %llu bytes", 979 ticket->bytes); 980 981 remove_ticket(space_info, ticket); 982 if (aborted) 983 ticket->error = -EIO; 984 else 985 ticket->error = -ENOSPC; 986 wake_up(&ticket->wait); 987 988 /* 989 * We're just throwing tickets away, so more flushing may not 990 * trip over btrfs_try_granting_tickets, so we need to call it 991 * here to see if we can make progress with the next ticket in 992 * the list. 993 */ 994 if (!aborted) 995 btrfs_try_granting_tickets(fs_info, space_info); 996 } 997 return (tickets_id != space_info->tickets_id); 998 } 999 1000 /* 1001 * This is for normal flushers, we can wait all goddamned day if we want to. We 1002 * will loop and continuously try to flush as long as we are making progress. 1003 * We count progress as clearing off tickets each time we have to loop. 1004 */ 1005 static void btrfs_async_reclaim_metadata_space(struct work_struct *work) 1006 { 1007 struct btrfs_fs_info *fs_info; 1008 struct btrfs_space_info *space_info; 1009 u64 to_reclaim; 1010 enum btrfs_flush_state flush_state; 1011 int commit_cycles = 0; 1012 u64 last_tickets_id; 1013 1014 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work); 1015 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 1016 1017 spin_lock(&space_info->lock); 1018 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 1019 if (!to_reclaim) { 1020 space_info->flush = 0; 1021 spin_unlock(&space_info->lock); 1022 return; 1023 } 1024 last_tickets_id = space_info->tickets_id; 1025 spin_unlock(&space_info->lock); 1026 1027 flush_state = FLUSH_DELAYED_ITEMS_NR; 1028 do { 1029 flush_space(fs_info, space_info, to_reclaim, flush_state, false); 1030 spin_lock(&space_info->lock); 1031 if (list_empty(&space_info->tickets)) { 1032 space_info->flush = 0; 1033 spin_unlock(&space_info->lock); 1034 return; 1035 } 1036 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, 1037 space_info); 1038 if (last_tickets_id == space_info->tickets_id) { 1039 flush_state++; 1040 } else { 1041 last_tickets_id = space_info->tickets_id; 1042 flush_state = FLUSH_DELAYED_ITEMS_NR; 1043 if (commit_cycles) 1044 commit_cycles--; 1045 } 1046 1047 /* 1048 * We do not want to empty the system of delalloc unless we're 1049 * under heavy pressure, so allow one trip through the flushing 1050 * logic before we start doing a FLUSH_DELALLOC_FULL. 1051 */ 1052 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles) 1053 flush_state++; 1054 1055 /* 1056 * We don't want to force a chunk allocation until we've tried 1057 * pretty hard to reclaim space. Think of the case where we 1058 * freed up a bunch of space and so have a lot of pinned space 1059 * to reclaim. We would rather use that than possibly create a 1060 * underutilized metadata chunk. So if this is our first run 1061 * through the flushing state machine skip ALLOC_CHUNK_FORCE and 1062 * commit the transaction. If nothing has changed the next go 1063 * around then we can force a chunk allocation. 1064 */ 1065 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles) 1066 flush_state++; 1067 1068 if (flush_state > COMMIT_TRANS) { 1069 commit_cycles++; 1070 if (commit_cycles > 2) { 1071 if (maybe_fail_all_tickets(fs_info, space_info)) { 1072 flush_state = FLUSH_DELAYED_ITEMS_NR; 1073 commit_cycles--; 1074 } else { 1075 space_info->flush = 0; 1076 } 1077 } else { 1078 flush_state = FLUSH_DELAYED_ITEMS_NR; 1079 } 1080 } 1081 spin_unlock(&space_info->lock); 1082 } while (flush_state <= COMMIT_TRANS); 1083 } 1084 1085 /* 1086 * This handles pre-flushing of metadata space before we get to the point that 1087 * we need to start blocking threads on tickets. The logic here is different 1088 * from the other flush paths because it doesn't rely on tickets to tell us how 1089 * much we need to flush, instead it attempts to keep us below the 80% full 1090 * watermark of space by flushing whichever reservation pool is currently the 1091 * largest. 1092 */ 1093 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work) 1094 { 1095 struct btrfs_fs_info *fs_info; 1096 struct btrfs_space_info *space_info; 1097 struct btrfs_block_rsv *delayed_block_rsv; 1098 struct btrfs_block_rsv *delayed_refs_rsv; 1099 struct btrfs_block_rsv *global_rsv; 1100 struct btrfs_block_rsv *trans_rsv; 1101 int loops = 0; 1102 1103 fs_info = container_of(work, struct btrfs_fs_info, 1104 preempt_reclaim_work); 1105 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 1106 delayed_block_rsv = &fs_info->delayed_block_rsv; 1107 delayed_refs_rsv = &fs_info->delayed_refs_rsv; 1108 global_rsv = &fs_info->global_block_rsv; 1109 trans_rsv = &fs_info->trans_block_rsv; 1110 1111 spin_lock(&space_info->lock); 1112 while (need_preemptive_reclaim(fs_info, space_info)) { 1113 enum btrfs_flush_state flush; 1114 u64 delalloc_size = 0; 1115 u64 to_reclaim, block_rsv_size; 1116 u64 global_rsv_size = global_rsv->reserved; 1117 1118 loops++; 1119 1120 /* 1121 * We don't have a precise counter for the metadata being 1122 * reserved for delalloc, so we'll approximate it by subtracting 1123 * out the block rsv's space from the bytes_may_use. If that 1124 * amount is higher than the individual reserves, then we can 1125 * assume it's tied up in delalloc reservations. 1126 */ 1127 block_rsv_size = global_rsv_size + 1128 delayed_block_rsv->reserved + 1129 delayed_refs_rsv->reserved + 1130 trans_rsv->reserved; 1131 if (block_rsv_size < space_info->bytes_may_use) 1132 delalloc_size = space_info->bytes_may_use - block_rsv_size; 1133 1134 /* 1135 * We don't want to include the global_rsv in our calculation, 1136 * because that's space we can't touch. Subtract it from the 1137 * block_rsv_size for the next checks. 1138 */ 1139 block_rsv_size -= global_rsv_size; 1140 1141 /* 1142 * We really want to avoid flushing delalloc too much, as it 1143 * could result in poor allocation patterns, so only flush it if 1144 * it's larger than the rest of the pools combined. 1145 */ 1146 if (delalloc_size > block_rsv_size) { 1147 to_reclaim = delalloc_size; 1148 flush = FLUSH_DELALLOC; 1149 } else if (space_info->bytes_pinned > 1150 (delayed_block_rsv->reserved + 1151 delayed_refs_rsv->reserved)) { 1152 to_reclaim = space_info->bytes_pinned; 1153 flush = COMMIT_TRANS; 1154 } else if (delayed_block_rsv->reserved > 1155 delayed_refs_rsv->reserved) { 1156 to_reclaim = delayed_block_rsv->reserved; 1157 flush = FLUSH_DELAYED_ITEMS_NR; 1158 } else { 1159 to_reclaim = delayed_refs_rsv->reserved; 1160 flush = FLUSH_DELAYED_REFS_NR; 1161 } 1162 1163 spin_unlock(&space_info->lock); 1164 1165 /* 1166 * We don't want to reclaim everything, just a portion, so scale 1167 * down the to_reclaim by 1/4. If it takes us down to 0, 1168 * reclaim 1 items worth. 1169 */ 1170 to_reclaim >>= 2; 1171 if (!to_reclaim) 1172 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1); 1173 flush_space(fs_info, space_info, to_reclaim, flush, true); 1174 cond_resched(); 1175 spin_lock(&space_info->lock); 1176 } 1177 1178 /* We only went through once, back off our clamping. */ 1179 if (loops == 1 && !space_info->reclaim_size) 1180 space_info->clamp = max(1, space_info->clamp - 1); 1181 trace_btrfs_done_preemptive_reclaim(fs_info, space_info); 1182 spin_unlock(&space_info->lock); 1183 } 1184 1185 /* 1186 * FLUSH_DELALLOC_WAIT: 1187 * Space is freed from flushing delalloc in one of two ways. 1188 * 1189 * 1) compression is on and we allocate less space than we reserved 1190 * 2) we are overwriting existing space 1191 * 1192 * For #1 that extra space is reclaimed as soon as the delalloc pages are 1193 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent 1194 * length to ->bytes_reserved, and subtracts the reserved space from 1195 * ->bytes_may_use. 1196 * 1197 * For #2 this is trickier. Once the ordered extent runs we will drop the 1198 * extent in the range we are overwriting, which creates a delayed ref for 1199 * that freed extent. This however is not reclaimed until the transaction 1200 * commits, thus the next stages. 1201 * 1202 * RUN_DELAYED_IPUTS 1203 * If we are freeing inodes, we want to make sure all delayed iputs have 1204 * completed, because they could have been on an inode with i_nlink == 0, and 1205 * thus have been truncated and freed up space. But again this space is not 1206 * immediately re-usable, it comes in the form of a delayed ref, which must be 1207 * run and then the transaction must be committed. 1208 * 1209 * COMMIT_TRANS 1210 * This is where we reclaim all of the pinned space generated by running the 1211 * iputs 1212 * 1213 * ALLOC_CHUNK_FORCE 1214 * For data we start with alloc chunk force, however we could have been full 1215 * before, and then the transaction commit could have freed new block groups, 1216 * so if we now have space to allocate do the force chunk allocation. 1217 */ 1218 static const enum btrfs_flush_state data_flush_states[] = { 1219 FLUSH_DELALLOC_FULL, 1220 RUN_DELAYED_IPUTS, 1221 COMMIT_TRANS, 1222 ALLOC_CHUNK_FORCE, 1223 }; 1224 1225 static void btrfs_async_reclaim_data_space(struct work_struct *work) 1226 { 1227 struct btrfs_fs_info *fs_info; 1228 struct btrfs_space_info *space_info; 1229 u64 last_tickets_id; 1230 enum btrfs_flush_state flush_state = 0; 1231 1232 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work); 1233 space_info = fs_info->data_sinfo; 1234 1235 spin_lock(&space_info->lock); 1236 if (list_empty(&space_info->tickets)) { 1237 space_info->flush = 0; 1238 spin_unlock(&space_info->lock); 1239 return; 1240 } 1241 last_tickets_id = space_info->tickets_id; 1242 spin_unlock(&space_info->lock); 1243 1244 while (!space_info->full) { 1245 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); 1246 spin_lock(&space_info->lock); 1247 if (list_empty(&space_info->tickets)) { 1248 space_info->flush = 0; 1249 spin_unlock(&space_info->lock); 1250 return; 1251 } 1252 1253 /* Something happened, fail everything and bail. */ 1254 if (BTRFS_FS_ERROR(fs_info)) 1255 goto aborted_fs; 1256 last_tickets_id = space_info->tickets_id; 1257 spin_unlock(&space_info->lock); 1258 } 1259 1260 while (flush_state < ARRAY_SIZE(data_flush_states)) { 1261 flush_space(fs_info, space_info, U64_MAX, 1262 data_flush_states[flush_state], false); 1263 spin_lock(&space_info->lock); 1264 if (list_empty(&space_info->tickets)) { 1265 space_info->flush = 0; 1266 spin_unlock(&space_info->lock); 1267 return; 1268 } 1269 1270 if (last_tickets_id == space_info->tickets_id) { 1271 flush_state++; 1272 } else { 1273 last_tickets_id = space_info->tickets_id; 1274 flush_state = 0; 1275 } 1276 1277 if (flush_state >= ARRAY_SIZE(data_flush_states)) { 1278 if (space_info->full) { 1279 if (maybe_fail_all_tickets(fs_info, space_info)) 1280 flush_state = 0; 1281 else 1282 space_info->flush = 0; 1283 } else { 1284 flush_state = 0; 1285 } 1286 1287 /* Something happened, fail everything and bail. */ 1288 if (BTRFS_FS_ERROR(fs_info)) 1289 goto aborted_fs; 1290 1291 } 1292 spin_unlock(&space_info->lock); 1293 } 1294 return; 1295 1296 aborted_fs: 1297 maybe_fail_all_tickets(fs_info, space_info); 1298 space_info->flush = 0; 1299 spin_unlock(&space_info->lock); 1300 } 1301 1302 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info) 1303 { 1304 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space); 1305 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space); 1306 INIT_WORK(&fs_info->preempt_reclaim_work, 1307 btrfs_preempt_reclaim_metadata_space); 1308 } 1309 1310 static const enum btrfs_flush_state priority_flush_states[] = { 1311 FLUSH_DELAYED_ITEMS_NR, 1312 FLUSH_DELAYED_ITEMS, 1313 ALLOC_CHUNK, 1314 }; 1315 1316 static const enum btrfs_flush_state evict_flush_states[] = { 1317 FLUSH_DELAYED_ITEMS_NR, 1318 FLUSH_DELAYED_ITEMS, 1319 FLUSH_DELAYED_REFS_NR, 1320 FLUSH_DELAYED_REFS, 1321 FLUSH_DELALLOC, 1322 FLUSH_DELALLOC_WAIT, 1323 FLUSH_DELALLOC_FULL, 1324 ALLOC_CHUNK, 1325 COMMIT_TRANS, 1326 }; 1327 1328 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info, 1329 struct btrfs_space_info *space_info, 1330 struct reserve_ticket *ticket, 1331 const enum btrfs_flush_state *states, 1332 int states_nr) 1333 { 1334 u64 to_reclaim; 1335 int flush_state = 0; 1336 1337 spin_lock(&space_info->lock); 1338 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 1339 /* 1340 * This is the priority reclaim path, so to_reclaim could be >0 still 1341 * because we may have only satisfied the priority tickets and still 1342 * left non priority tickets on the list. We would then have 1343 * to_reclaim but ->bytes == 0. 1344 */ 1345 if (ticket->bytes == 0) { 1346 spin_unlock(&space_info->lock); 1347 return; 1348 } 1349 1350 while (flush_state < states_nr) { 1351 spin_unlock(&space_info->lock); 1352 flush_space(fs_info, space_info, to_reclaim, states[flush_state], 1353 false); 1354 flush_state++; 1355 spin_lock(&space_info->lock); 1356 if (ticket->bytes == 0) { 1357 spin_unlock(&space_info->lock); 1358 return; 1359 } 1360 } 1361 1362 /* Attempt to steal from the global rsv if we can. */ 1363 if (!steal_from_global_rsv(fs_info, space_info, ticket)) { 1364 ticket->error = -ENOSPC; 1365 remove_ticket(space_info, ticket); 1366 } 1367 1368 /* 1369 * We must run try_granting_tickets here because we could be a large 1370 * ticket in front of a smaller ticket that can now be satisfied with 1371 * the available space. 1372 */ 1373 btrfs_try_granting_tickets(fs_info, space_info); 1374 spin_unlock(&space_info->lock); 1375 } 1376 1377 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info, 1378 struct btrfs_space_info *space_info, 1379 struct reserve_ticket *ticket) 1380 { 1381 spin_lock(&space_info->lock); 1382 1383 /* We could have been granted before we got here. */ 1384 if (ticket->bytes == 0) { 1385 spin_unlock(&space_info->lock); 1386 return; 1387 } 1388 1389 while (!space_info->full) { 1390 spin_unlock(&space_info->lock); 1391 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); 1392 spin_lock(&space_info->lock); 1393 if (ticket->bytes == 0) { 1394 spin_unlock(&space_info->lock); 1395 return; 1396 } 1397 } 1398 1399 ticket->error = -ENOSPC; 1400 remove_ticket(space_info, ticket); 1401 btrfs_try_granting_tickets(fs_info, space_info); 1402 spin_unlock(&space_info->lock); 1403 } 1404 1405 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info, 1406 struct btrfs_space_info *space_info, 1407 struct reserve_ticket *ticket) 1408 1409 { 1410 DEFINE_WAIT(wait); 1411 int ret = 0; 1412 1413 spin_lock(&space_info->lock); 1414 while (ticket->bytes > 0 && ticket->error == 0) { 1415 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE); 1416 if (ret) { 1417 /* 1418 * Delete us from the list. After we unlock the space 1419 * info, we don't want the async reclaim job to reserve 1420 * space for this ticket. If that would happen, then the 1421 * ticket's task would not known that space was reserved 1422 * despite getting an error, resulting in a space leak 1423 * (bytes_may_use counter of our space_info). 1424 */ 1425 remove_ticket(space_info, ticket); 1426 ticket->error = -EINTR; 1427 break; 1428 } 1429 spin_unlock(&space_info->lock); 1430 1431 schedule(); 1432 1433 finish_wait(&ticket->wait, &wait); 1434 spin_lock(&space_info->lock); 1435 } 1436 spin_unlock(&space_info->lock); 1437 } 1438 1439 /** 1440 * Do the appropriate flushing and waiting for a ticket 1441 * 1442 * @fs_info: the filesystem 1443 * @space_info: space info for the reservation 1444 * @ticket: ticket for the reservation 1445 * @start_ns: timestamp when the reservation started 1446 * @orig_bytes: amount of bytes originally reserved 1447 * @flush: how much we can flush 1448 * 1449 * This does the work of figuring out how to flush for the ticket, waiting for 1450 * the reservation, and returning the appropriate error if there is one. 1451 */ 1452 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info, 1453 struct btrfs_space_info *space_info, 1454 struct reserve_ticket *ticket, 1455 u64 start_ns, u64 orig_bytes, 1456 enum btrfs_reserve_flush_enum flush) 1457 { 1458 int ret; 1459 1460 switch (flush) { 1461 case BTRFS_RESERVE_FLUSH_DATA: 1462 case BTRFS_RESERVE_FLUSH_ALL: 1463 case BTRFS_RESERVE_FLUSH_ALL_STEAL: 1464 wait_reserve_ticket(fs_info, space_info, ticket); 1465 break; 1466 case BTRFS_RESERVE_FLUSH_LIMIT: 1467 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1468 priority_flush_states, 1469 ARRAY_SIZE(priority_flush_states)); 1470 break; 1471 case BTRFS_RESERVE_FLUSH_EVICT: 1472 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1473 evict_flush_states, 1474 ARRAY_SIZE(evict_flush_states)); 1475 break; 1476 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE: 1477 priority_reclaim_data_space(fs_info, space_info, ticket); 1478 break; 1479 default: 1480 ASSERT(0); 1481 break; 1482 } 1483 1484 ret = ticket->error; 1485 ASSERT(list_empty(&ticket->list)); 1486 /* 1487 * Check that we can't have an error set if the reservation succeeded, 1488 * as that would confuse tasks and lead them to error out without 1489 * releasing reserved space (if an error happens the expectation is that 1490 * space wasn't reserved at all). 1491 */ 1492 ASSERT(!(ticket->bytes == 0 && ticket->error)); 1493 trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes, 1494 start_ns, flush, ticket->error); 1495 return ret; 1496 } 1497 1498 /* 1499 * This returns true if this flush state will go through the ordinary flushing 1500 * code. 1501 */ 1502 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush) 1503 { 1504 return (flush == BTRFS_RESERVE_FLUSH_ALL) || 1505 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL); 1506 } 1507 1508 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info, 1509 struct btrfs_space_info *space_info) 1510 { 1511 u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes); 1512 u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes); 1513 1514 /* 1515 * If we're heavy on ordered operations then clamping won't help us. We 1516 * need to clamp specifically to keep up with dirty'ing buffered 1517 * writers, because there's not a 1:1 correlation of writing delalloc 1518 * and freeing space, like there is with flushing delayed refs or 1519 * delayed nodes. If we're already more ordered than delalloc then 1520 * we're keeping up, otherwise we aren't and should probably clamp. 1521 */ 1522 if (ordered < delalloc) 1523 space_info->clamp = min(space_info->clamp + 1, 8); 1524 } 1525 1526 static inline bool can_steal(enum btrfs_reserve_flush_enum flush) 1527 { 1528 return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || 1529 flush == BTRFS_RESERVE_FLUSH_EVICT); 1530 } 1531 1532 /** 1533 * Try to reserve bytes from the block_rsv's space 1534 * 1535 * @fs_info: the filesystem 1536 * @space_info: space info we want to allocate from 1537 * @orig_bytes: number of bytes we want 1538 * @flush: whether or not we can flush to make our reservation 1539 * 1540 * This will reserve orig_bytes number of bytes from the space info associated 1541 * with the block_rsv. If there is not enough space it will make an attempt to 1542 * flush out space to make room. It will do this by flushing delalloc if 1543 * possible or committing the transaction. If flush is 0 then no attempts to 1544 * regain reservations will be made and this will fail if there is not enough 1545 * space already. 1546 */ 1547 static int __reserve_bytes(struct btrfs_fs_info *fs_info, 1548 struct btrfs_space_info *space_info, u64 orig_bytes, 1549 enum btrfs_reserve_flush_enum flush) 1550 { 1551 struct work_struct *async_work; 1552 struct reserve_ticket ticket; 1553 u64 start_ns = 0; 1554 u64 used; 1555 int ret = 0; 1556 bool pending_tickets; 1557 1558 ASSERT(orig_bytes); 1559 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL); 1560 1561 if (flush == BTRFS_RESERVE_FLUSH_DATA) 1562 async_work = &fs_info->async_data_reclaim_work; 1563 else 1564 async_work = &fs_info->async_reclaim_work; 1565 1566 spin_lock(&space_info->lock); 1567 ret = -ENOSPC; 1568 used = btrfs_space_info_used(space_info, true); 1569 1570 /* 1571 * We don't want NO_FLUSH allocations to jump everybody, they can 1572 * generally handle ENOSPC in a different way, so treat them the same as 1573 * normal flushers when it comes to skipping pending tickets. 1574 */ 1575 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH)) 1576 pending_tickets = !list_empty(&space_info->tickets) || 1577 !list_empty(&space_info->priority_tickets); 1578 else 1579 pending_tickets = !list_empty(&space_info->priority_tickets); 1580 1581 /* 1582 * Carry on if we have enough space (short-circuit) OR call 1583 * can_overcommit() to ensure we can overcommit to continue. 1584 */ 1585 if (!pending_tickets && 1586 ((used + orig_bytes <= writable_total_bytes(fs_info, space_info)) || 1587 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) { 1588 btrfs_space_info_update_bytes_may_use(fs_info, space_info, 1589 orig_bytes); 1590 ret = 0; 1591 } 1592 1593 /* 1594 * If we couldn't make a reservation then setup our reservation ticket 1595 * and kick the async worker if it's not already running. 1596 * 1597 * If we are a priority flusher then we just need to add our ticket to 1598 * the list and we will do our own flushing further down. 1599 */ 1600 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) { 1601 ticket.bytes = orig_bytes; 1602 ticket.error = 0; 1603 space_info->reclaim_size += ticket.bytes; 1604 init_waitqueue_head(&ticket.wait); 1605 ticket.steal = can_steal(flush); 1606 if (trace_btrfs_reserve_ticket_enabled()) 1607 start_ns = ktime_get_ns(); 1608 1609 if (flush == BTRFS_RESERVE_FLUSH_ALL || 1610 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || 1611 flush == BTRFS_RESERVE_FLUSH_DATA) { 1612 list_add_tail(&ticket.list, &space_info->tickets); 1613 if (!space_info->flush) { 1614 /* 1615 * We were forced to add a reserve ticket, so 1616 * our preemptive flushing is unable to keep 1617 * up. Clamp down on the threshold for the 1618 * preemptive flushing in order to keep up with 1619 * the workload. 1620 */ 1621 maybe_clamp_preempt(fs_info, space_info); 1622 1623 space_info->flush = 1; 1624 trace_btrfs_trigger_flush(fs_info, 1625 space_info->flags, 1626 orig_bytes, flush, 1627 "enospc"); 1628 queue_work(system_unbound_wq, async_work); 1629 } 1630 } else { 1631 list_add_tail(&ticket.list, 1632 &space_info->priority_tickets); 1633 } 1634 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) { 1635 used += orig_bytes; 1636 /* 1637 * We will do the space reservation dance during log replay, 1638 * which means we won't have fs_info->fs_root set, so don't do 1639 * the async reclaim as we will panic. 1640 */ 1641 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) && 1642 !work_busy(&fs_info->preempt_reclaim_work) && 1643 need_preemptive_reclaim(fs_info, space_info)) { 1644 trace_btrfs_trigger_flush(fs_info, space_info->flags, 1645 orig_bytes, flush, "preempt"); 1646 queue_work(system_unbound_wq, 1647 &fs_info->preempt_reclaim_work); 1648 } 1649 } 1650 spin_unlock(&space_info->lock); 1651 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH) 1652 return ret; 1653 1654 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns, 1655 orig_bytes, flush); 1656 } 1657 1658 /** 1659 * Trye to reserve metadata bytes from the block_rsv's space 1660 * 1661 * @fs_info: the filesystem 1662 * @block_rsv: block_rsv we're allocating for 1663 * @orig_bytes: number of bytes we want 1664 * @flush: whether or not we can flush to make our reservation 1665 * 1666 * This will reserve orig_bytes number of bytes from the space info associated 1667 * with the block_rsv. If there is not enough space it will make an attempt to 1668 * flush out space to make room. It will do this by flushing delalloc if 1669 * possible or committing the transaction. If flush is 0 then no attempts to 1670 * regain reservations will be made and this will fail if there is not enough 1671 * space already. 1672 */ 1673 int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info, 1674 struct btrfs_block_rsv *block_rsv, 1675 u64 orig_bytes, 1676 enum btrfs_reserve_flush_enum flush) 1677 { 1678 int ret; 1679 1680 ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush); 1681 if (ret == -ENOSPC) { 1682 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1683 block_rsv->space_info->flags, 1684 orig_bytes, 1); 1685 1686 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1687 btrfs_dump_space_info(fs_info, block_rsv->space_info, 1688 orig_bytes, 0); 1689 } 1690 return ret; 1691 } 1692 1693 /** 1694 * Try to reserve data bytes for an allocation 1695 * 1696 * @fs_info: the filesystem 1697 * @bytes: number of bytes we need 1698 * @flush: how we are allowed to flush 1699 * 1700 * This will reserve bytes from the data space info. If there is not enough 1701 * space then we will attempt to flush space as specified by flush. 1702 */ 1703 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes, 1704 enum btrfs_reserve_flush_enum flush) 1705 { 1706 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo; 1707 int ret; 1708 1709 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA || 1710 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE); 1711 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA); 1712 1713 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush); 1714 if (ret == -ENOSPC) { 1715 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1716 data_sinfo->flags, bytes, 1); 1717 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1718 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0); 1719 } 1720 return ret; 1721 } 1722