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 #include "zoned.h" 13 14 /* 15 * HOW DOES SPACE RESERVATION WORK 16 * 17 * If you want to know about delalloc specifically, there is a separate comment 18 * for that with the delalloc code. This comment is about how the whole system 19 * works generally. 20 * 21 * BASIC CONCEPTS 22 * 23 * 1) space_info. This is the ultimate arbiter of how much space we can use. 24 * There's a description of the bytes_ fields with the struct declaration, 25 * refer to that for specifics on each field. Suffice it to say that for 26 * reservations we care about total_bytes - SUM(space_info->bytes_) when 27 * determining if there is space to make an allocation. There is a space_info 28 * for METADATA, SYSTEM, and DATA areas. 29 * 30 * 2) block_rsv's. These are basically buckets for every different type of 31 * metadata reservation we have. You can see the comment in the block_rsv 32 * code on the rules for each type, but generally block_rsv->reserved is how 33 * much space is accounted for in space_info->bytes_may_use. 34 * 35 * 3) btrfs_calc*_size. These are the worst case calculations we used based 36 * on the number of items we will want to modify. We have one for changing 37 * items, and one for inserting new items. Generally we use these helpers to 38 * determine the size of the block reserves, and then use the actual bytes 39 * values to adjust the space_info counters. 40 * 41 * MAKING RESERVATIONS, THE NORMAL CASE 42 * 43 * We call into either btrfs_reserve_data_bytes() or 44 * btrfs_reserve_metadata_bytes(), depending on which we're looking for, with 45 * num_bytes we want to reserve. 46 * 47 * ->reserve 48 * space_info->bytes_may_reserve += num_bytes 49 * 50 * ->extent allocation 51 * Call btrfs_add_reserved_bytes() which does 52 * space_info->bytes_may_reserve -= num_bytes 53 * space_info->bytes_reserved += extent_bytes 54 * 55 * ->insert reference 56 * Call btrfs_update_block_group() which does 57 * space_info->bytes_reserved -= extent_bytes 58 * space_info->bytes_used += extent_bytes 59 * 60 * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority) 61 * 62 * Assume we are unable to simply make the reservation because we do not have 63 * enough space 64 * 65 * -> __reserve_bytes 66 * create a reserve_ticket with ->bytes set to our reservation, add it to 67 * the tail of space_info->tickets, kick async flush thread 68 * 69 * ->handle_reserve_ticket 70 * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set 71 * on the ticket. 72 * 73 * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space 74 * Flushes various things attempting to free up space. 75 * 76 * -> btrfs_try_granting_tickets() 77 * This is called by anything that either subtracts space from 78 * space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the 79 * space_info->total_bytes. This loops through the ->priority_tickets and 80 * then the ->tickets list checking to see if the reservation can be 81 * completed. If it can the space is added to space_info->bytes_may_use and 82 * the ticket is woken up. 83 * 84 * -> ticket wakeup 85 * Check if ->bytes == 0, if it does we got our reservation and we can carry 86 * on, if not return the appropriate error (ENOSPC, but can be EINTR if we 87 * were interrupted.) 88 * 89 * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY 90 * 91 * Same as the above, except we add ourselves to the 92 * space_info->priority_tickets, and we do not use ticket->wait, we simply 93 * call flush_space() ourselves for the states that are safe for us to call 94 * without deadlocking and hope for the best. 95 * 96 * THE FLUSHING STATES 97 * 98 * Generally speaking we will have two cases for each state, a "nice" state 99 * and a "ALL THE THINGS" state. In btrfs we delay a lot of work in order to 100 * reduce the locking over head on the various trees, and even to keep from 101 * doing any work at all in the case of delayed refs. Each of these delayed 102 * things however hold reservations, and so letting them run allows us to 103 * reclaim space so we can make new reservations. 104 * 105 * FLUSH_DELAYED_ITEMS 106 * Every inode has a delayed item to update the inode. Take a simple write 107 * for example, we would update the inode item at write time to update the 108 * mtime, and then again at finish_ordered_io() time in order to update the 109 * isize or bytes. We keep these delayed items to coalesce these operations 110 * into a single operation done on demand. These are an easy way to reclaim 111 * metadata space. 112 * 113 * FLUSH_DELALLOC 114 * Look at the delalloc comment to get an idea of how much space is reserved 115 * for delayed allocation. We can reclaim some of this space simply by 116 * running delalloc, but usually we need to wait for ordered extents to 117 * reclaim the bulk of this space. 118 * 119 * FLUSH_DELAYED_REFS 120 * We have a block reserve for the outstanding delayed refs space, and every 121 * delayed ref operation holds a reservation. Running these is a quick way 122 * to reclaim space, but we want to hold this until the end because COW can 123 * churn a lot and we can avoid making some extent tree modifications if we 124 * are able to delay for as long as possible. 125 * 126 * ALLOC_CHUNK 127 * We will skip this the first time through space reservation, because of 128 * overcommit and we don't want to have a lot of useless metadata space when 129 * our worst case reservations will likely never come true. 130 * 131 * RUN_DELAYED_IPUTS 132 * If we're freeing inodes we're likely freeing checksums, file extent 133 * items, and extent tree items. Loads of space could be freed up by these 134 * operations, however they won't be usable until the transaction commits. 135 * 136 * COMMIT_TRANS 137 * This will commit the transaction. Historically we had a lot of logic 138 * surrounding whether or not we'd commit the transaction, but this waits born 139 * out of a pre-tickets era where we could end up committing the transaction 140 * thousands of times in a row without making progress. Now thanks to our 141 * ticketing system we know if we're not making progress and can error 142 * everybody out after a few commits rather than burning the disk hoping for 143 * a different answer. 144 * 145 * OVERCOMMIT 146 * 147 * Because we hold so many reservations for metadata we will allow you to 148 * reserve more space than is currently free in the currently allocate 149 * metadata space. This only happens with metadata, data does not allow 150 * overcommitting. 151 * 152 * You can see the current logic for when we allow overcommit in 153 * btrfs_can_overcommit(), but it only applies to unallocated space. If there 154 * is no unallocated space to be had, all reservations are kept within the 155 * free space in the allocated metadata chunks. 156 * 157 * Because of overcommitting, you generally want to use the 158 * btrfs_can_overcommit() logic for metadata allocations, as it does the right 159 * thing with or without extra unallocated space. 160 */ 161 162 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info, 163 bool may_use_included) 164 { 165 ASSERT(s_info); 166 return s_info->bytes_used + s_info->bytes_reserved + 167 s_info->bytes_pinned + s_info->bytes_readonly + 168 s_info->bytes_zone_unusable + 169 (may_use_included ? s_info->bytes_may_use : 0); 170 } 171 172 /* 173 * after adding space to the filesystem, we need to clear the full flags 174 * on all the space infos. 175 */ 176 void btrfs_clear_space_info_full(struct btrfs_fs_info *info) 177 { 178 struct list_head *head = &info->space_info; 179 struct btrfs_space_info *found; 180 181 list_for_each_entry(found, head, list) 182 found->full = 0; 183 } 184 185 /* 186 * Block groups with more than this value (percents) of unusable space will be 187 * scheduled for background reclaim. 188 */ 189 #define BTRFS_DEFAULT_ZONED_RECLAIM_THRESH (75) 190 191 /* 192 * Calculate chunk size depending on volume type (regular or zoned). 193 */ 194 static u64 calc_chunk_size(const struct btrfs_fs_info *fs_info, u64 flags) 195 { 196 if (btrfs_is_zoned(fs_info)) 197 return fs_info->zone_size; 198 199 ASSERT(flags & BTRFS_BLOCK_GROUP_TYPE_MASK); 200 201 if (flags & BTRFS_BLOCK_GROUP_DATA) 202 return BTRFS_MAX_DATA_CHUNK_SIZE; 203 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM) 204 return SZ_32M; 205 206 /* Handle BTRFS_BLOCK_GROUP_METADATA */ 207 if (fs_info->fs_devices->total_rw_bytes > 50ULL * SZ_1G) 208 return SZ_1G; 209 210 return SZ_256M; 211 } 212 213 /* 214 * Update default chunk size. 215 */ 216 void btrfs_update_space_info_chunk_size(struct btrfs_space_info *space_info, 217 u64 chunk_size) 218 { 219 WRITE_ONCE(space_info->chunk_size, chunk_size); 220 } 221 222 static int create_space_info(struct btrfs_fs_info *info, u64 flags) 223 { 224 225 struct btrfs_space_info *space_info; 226 int i; 227 int ret; 228 229 space_info = kzalloc(sizeof(*space_info), GFP_NOFS); 230 if (!space_info) 231 return -ENOMEM; 232 233 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) 234 INIT_LIST_HEAD(&space_info->block_groups[i]); 235 init_rwsem(&space_info->groups_sem); 236 spin_lock_init(&space_info->lock); 237 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK; 238 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE; 239 INIT_LIST_HEAD(&space_info->ro_bgs); 240 INIT_LIST_HEAD(&space_info->tickets); 241 INIT_LIST_HEAD(&space_info->priority_tickets); 242 space_info->clamp = 1; 243 btrfs_update_space_info_chunk_size(space_info, calc_chunk_size(info, flags)); 244 245 if (btrfs_is_zoned(info)) 246 space_info->bg_reclaim_threshold = BTRFS_DEFAULT_ZONED_RECLAIM_THRESH; 247 248 ret = btrfs_sysfs_add_space_info_type(info, space_info); 249 if (ret) 250 return ret; 251 252 list_add(&space_info->list, &info->space_info); 253 if (flags & BTRFS_BLOCK_GROUP_DATA) 254 info->data_sinfo = space_info; 255 256 return ret; 257 } 258 259 int btrfs_init_space_info(struct btrfs_fs_info *fs_info) 260 { 261 struct btrfs_super_block *disk_super; 262 u64 features; 263 u64 flags; 264 int mixed = 0; 265 int ret; 266 267 disk_super = fs_info->super_copy; 268 if (!btrfs_super_root(disk_super)) 269 return -EINVAL; 270 271 features = btrfs_super_incompat_flags(disk_super); 272 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS) 273 mixed = 1; 274 275 flags = BTRFS_BLOCK_GROUP_SYSTEM; 276 ret = create_space_info(fs_info, flags); 277 if (ret) 278 goto out; 279 280 if (mixed) { 281 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA; 282 ret = create_space_info(fs_info, flags); 283 } else { 284 flags = BTRFS_BLOCK_GROUP_METADATA; 285 ret = create_space_info(fs_info, flags); 286 if (ret) 287 goto out; 288 289 flags = BTRFS_BLOCK_GROUP_DATA; 290 ret = create_space_info(fs_info, flags); 291 } 292 out: 293 return ret; 294 } 295 296 void btrfs_add_bg_to_space_info(struct btrfs_fs_info *info, 297 struct btrfs_block_group *block_group, 298 struct btrfs_space_info **space_info) 299 { 300 struct btrfs_space_info *found; 301 int factor; 302 303 factor = btrfs_bg_type_to_factor(block_group->flags); 304 305 found = btrfs_find_space_info(info, block_group->flags); 306 ASSERT(found); 307 spin_lock(&found->lock); 308 found->total_bytes += block_group->length; 309 if (block_group->zone_is_active) 310 found->active_total_bytes += block_group->length; 311 found->disk_total += block_group->length * factor; 312 found->bytes_used += block_group->used; 313 found->disk_used += block_group->used * factor; 314 found->bytes_readonly += block_group->bytes_super; 315 found->bytes_zone_unusable += block_group->zone_unusable; 316 if (block_group->length > 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 /* 728 * For metadata space on zoned filesystem, reaching here means we 729 * don't have enough space left in active_total_bytes. Try to 730 * activate a block group first, because we may have inactive 731 * block group already allocated. 732 */ 733 ret = btrfs_zoned_activate_one_bg(fs_info, space_info, false); 734 if (ret < 0) 735 break; 736 else if (ret == 1) 737 break; 738 739 trans = btrfs_join_transaction(root); 740 if (IS_ERR(trans)) { 741 ret = PTR_ERR(trans); 742 break; 743 } 744 ret = btrfs_chunk_alloc(trans, 745 btrfs_get_alloc_profile(fs_info, space_info->flags), 746 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE : 747 CHUNK_ALLOC_FORCE); 748 btrfs_end_transaction(trans); 749 750 /* 751 * For metadata space on zoned filesystem, allocating a new chunk 752 * is not enough. We still need to activate the block * group. 753 * Active the newly allocated block group by (maybe) finishing 754 * a block group. 755 */ 756 if (ret == 1) { 757 ret = btrfs_zoned_activate_one_bg(fs_info, space_info, true); 758 /* 759 * Revert to the original ret regardless we could finish 760 * one block group or not. 761 */ 762 if (ret >= 0) 763 ret = 1; 764 } 765 766 if (ret > 0 || ret == -ENOSPC) 767 ret = 0; 768 break; 769 case RUN_DELAYED_IPUTS: 770 /* 771 * If we have pending delayed iputs then we could free up a 772 * bunch of pinned space, so make sure we run the iputs before 773 * we do our pinned bytes check below. 774 */ 775 btrfs_run_delayed_iputs(fs_info); 776 btrfs_wait_on_delayed_iputs(fs_info); 777 break; 778 case COMMIT_TRANS: 779 ASSERT(current->journal_info == NULL); 780 trans = btrfs_join_transaction(root); 781 if (IS_ERR(trans)) { 782 ret = PTR_ERR(trans); 783 break; 784 } 785 ret = btrfs_commit_transaction(trans); 786 break; 787 default: 788 ret = -ENOSPC; 789 break; 790 } 791 792 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state, 793 ret, for_preempt); 794 return; 795 } 796 797 static inline u64 798 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info, 799 struct btrfs_space_info *space_info) 800 { 801 u64 used; 802 u64 avail; 803 u64 total; 804 u64 to_reclaim = space_info->reclaim_size; 805 806 lockdep_assert_held(&space_info->lock); 807 808 avail = calc_available_free_space(fs_info, space_info, 809 BTRFS_RESERVE_FLUSH_ALL); 810 used = btrfs_space_info_used(space_info, true); 811 812 /* 813 * We may be flushing because suddenly we have less space than we had 814 * before, and now we're well over-committed based on our current free 815 * space. If that's the case add in our overage so we make sure to put 816 * appropriate pressure on the flushing state machine. 817 */ 818 total = writable_total_bytes(fs_info, space_info); 819 if (total + avail < used) 820 to_reclaim += used - (total + avail); 821 822 return to_reclaim; 823 } 824 825 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info, 826 struct btrfs_space_info *space_info) 827 { 828 u64 global_rsv_size = fs_info->global_block_rsv.reserved; 829 u64 ordered, delalloc; 830 u64 total = writable_total_bytes(fs_info, space_info); 831 u64 thresh; 832 u64 used; 833 834 thresh = div_factor_fine(total, 90); 835 836 lockdep_assert_held(&space_info->lock); 837 838 /* If we're just plain full then async reclaim just slows us down. */ 839 if ((space_info->bytes_used + space_info->bytes_reserved + 840 global_rsv_size) >= thresh) 841 return false; 842 843 used = space_info->bytes_may_use + space_info->bytes_pinned; 844 845 /* The total flushable belongs to the global rsv, don't flush. */ 846 if (global_rsv_size >= used) 847 return false; 848 849 /* 850 * 128MiB is 1/4 of the maximum global rsv size. If we have less than 851 * that devoted to other reservations then there's no sense in flushing, 852 * we don't have a lot of things that need flushing. 853 */ 854 if (used - global_rsv_size <= SZ_128M) 855 return false; 856 857 /* 858 * We have tickets queued, bail so we don't compete with the async 859 * flushers. 860 */ 861 if (space_info->reclaim_size) 862 return false; 863 864 /* 865 * If we have over half of the free space occupied by reservations or 866 * pinned then we want to start flushing. 867 * 868 * We do not do the traditional thing here, which is to say 869 * 870 * if (used >= ((total_bytes + avail) / 2)) 871 * return 1; 872 * 873 * because this doesn't quite work how we want. If we had more than 50% 874 * of the space_info used by bytes_used and we had 0 available we'd just 875 * constantly run the background flusher. Instead we want it to kick in 876 * if our reclaimable space exceeds our clamped free space. 877 * 878 * Our clamping range is 2^1 -> 2^8. Practically speaking that means 879 * the following: 880 * 881 * Amount of RAM Minimum threshold Maximum threshold 882 * 883 * 256GiB 1GiB 128GiB 884 * 128GiB 512MiB 64GiB 885 * 64GiB 256MiB 32GiB 886 * 32GiB 128MiB 16GiB 887 * 16GiB 64MiB 8GiB 888 * 889 * These are the range our thresholds will fall in, corresponding to how 890 * much delalloc we need for the background flusher to kick in. 891 */ 892 893 thresh = calc_available_free_space(fs_info, space_info, 894 BTRFS_RESERVE_FLUSH_ALL); 895 used = space_info->bytes_used + space_info->bytes_reserved + 896 space_info->bytes_readonly + global_rsv_size; 897 if (used < total) 898 thresh += total - used; 899 thresh >>= space_info->clamp; 900 901 used = space_info->bytes_pinned; 902 903 /* 904 * If we have more ordered bytes than delalloc bytes then we're either 905 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting 906 * around. Preemptive flushing is only useful in that it can free up 907 * space before tickets need to wait for things to finish. In the case 908 * of ordered extents, preemptively waiting on ordered extents gets us 909 * nothing, if our reservations are tied up in ordered extents we'll 910 * simply have to slow down writers by forcing them to wait on ordered 911 * extents. 912 * 913 * In the case that ordered is larger than delalloc, only include the 914 * block reserves that we would actually be able to directly reclaim 915 * from. In this case if we're heavy on metadata operations this will 916 * clearly be heavy enough to warrant preemptive flushing. In the case 917 * of heavy DIO or ordered reservations, preemptive flushing will just 918 * waste time and cause us to slow down. 919 * 920 * We want to make sure we truly are maxed out on ordered however, so 921 * cut ordered in half, and if it's still higher than delalloc then we 922 * can keep flushing. This is to avoid the case where we start 923 * flushing, and now delalloc == ordered and we stop preemptively 924 * flushing when we could still have several gigs of delalloc to flush. 925 */ 926 ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1; 927 delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes); 928 if (ordered >= delalloc) 929 used += fs_info->delayed_refs_rsv.reserved + 930 fs_info->delayed_block_rsv.reserved; 931 else 932 used += space_info->bytes_may_use - global_rsv_size; 933 934 return (used >= thresh && !btrfs_fs_closing(fs_info) && 935 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state)); 936 } 937 938 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info, 939 struct btrfs_space_info *space_info, 940 struct reserve_ticket *ticket) 941 { 942 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 943 u64 min_bytes; 944 945 if (!ticket->steal) 946 return false; 947 948 if (global_rsv->space_info != space_info) 949 return false; 950 951 spin_lock(&global_rsv->lock); 952 min_bytes = div_factor(global_rsv->size, 1); 953 if (global_rsv->reserved < min_bytes + ticket->bytes) { 954 spin_unlock(&global_rsv->lock); 955 return false; 956 } 957 global_rsv->reserved -= ticket->bytes; 958 remove_ticket(space_info, ticket); 959 ticket->bytes = 0; 960 wake_up(&ticket->wait); 961 space_info->tickets_id++; 962 if (global_rsv->reserved < global_rsv->size) 963 global_rsv->full = 0; 964 spin_unlock(&global_rsv->lock); 965 966 return true; 967 } 968 969 /* 970 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets 971 * @fs_info - fs_info for this fs 972 * @space_info - the space info we were flushing 973 * 974 * We call this when we've exhausted our flushing ability and haven't made 975 * progress in satisfying tickets. The reservation code handles tickets in 976 * order, so if there is a large ticket first and then smaller ones we could 977 * very well satisfy the smaller tickets. This will attempt to wake up any 978 * tickets in the list to catch this case. 979 * 980 * This function returns true if it was able to make progress by clearing out 981 * other tickets, or if it stumbles across a ticket that was smaller than the 982 * first ticket. 983 */ 984 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info, 985 struct btrfs_space_info *space_info) 986 { 987 struct reserve_ticket *ticket; 988 u64 tickets_id = space_info->tickets_id; 989 const bool aborted = BTRFS_FS_ERROR(fs_info); 990 991 trace_btrfs_fail_all_tickets(fs_info, space_info); 992 993 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) { 994 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info"); 995 __btrfs_dump_space_info(fs_info, space_info); 996 } 997 998 while (!list_empty(&space_info->tickets) && 999 tickets_id == space_info->tickets_id) { 1000 ticket = list_first_entry(&space_info->tickets, 1001 struct reserve_ticket, list); 1002 1003 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket)) 1004 return true; 1005 1006 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1007 btrfs_info(fs_info, "failing ticket with %llu bytes", 1008 ticket->bytes); 1009 1010 remove_ticket(space_info, ticket); 1011 if (aborted) 1012 ticket->error = -EIO; 1013 else 1014 ticket->error = -ENOSPC; 1015 wake_up(&ticket->wait); 1016 1017 /* 1018 * We're just throwing tickets away, so more flushing may not 1019 * trip over btrfs_try_granting_tickets, so we need to call it 1020 * here to see if we can make progress with the next ticket in 1021 * the list. 1022 */ 1023 if (!aborted) 1024 btrfs_try_granting_tickets(fs_info, space_info); 1025 } 1026 return (tickets_id != space_info->tickets_id); 1027 } 1028 1029 /* 1030 * This is for normal flushers, we can wait all goddamned day if we want to. We 1031 * will loop and continuously try to flush as long as we are making progress. 1032 * We count progress as clearing off tickets each time we have to loop. 1033 */ 1034 static void btrfs_async_reclaim_metadata_space(struct work_struct *work) 1035 { 1036 struct btrfs_fs_info *fs_info; 1037 struct btrfs_space_info *space_info; 1038 u64 to_reclaim; 1039 enum btrfs_flush_state flush_state; 1040 int commit_cycles = 0; 1041 u64 last_tickets_id; 1042 1043 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work); 1044 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 1045 1046 spin_lock(&space_info->lock); 1047 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 1048 if (!to_reclaim) { 1049 space_info->flush = 0; 1050 spin_unlock(&space_info->lock); 1051 return; 1052 } 1053 last_tickets_id = space_info->tickets_id; 1054 spin_unlock(&space_info->lock); 1055 1056 flush_state = FLUSH_DELAYED_ITEMS_NR; 1057 do { 1058 flush_space(fs_info, space_info, to_reclaim, flush_state, false); 1059 spin_lock(&space_info->lock); 1060 if (list_empty(&space_info->tickets)) { 1061 space_info->flush = 0; 1062 spin_unlock(&space_info->lock); 1063 return; 1064 } 1065 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, 1066 space_info); 1067 if (last_tickets_id == space_info->tickets_id) { 1068 flush_state++; 1069 } else { 1070 last_tickets_id = space_info->tickets_id; 1071 flush_state = FLUSH_DELAYED_ITEMS_NR; 1072 if (commit_cycles) 1073 commit_cycles--; 1074 } 1075 1076 /* 1077 * We do not want to empty the system of delalloc unless we're 1078 * under heavy pressure, so allow one trip through the flushing 1079 * logic before we start doing a FLUSH_DELALLOC_FULL. 1080 */ 1081 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles) 1082 flush_state++; 1083 1084 /* 1085 * We don't want to force a chunk allocation until we've tried 1086 * pretty hard to reclaim space. Think of the case where we 1087 * freed up a bunch of space and so have a lot of pinned space 1088 * to reclaim. We would rather use that than possibly create a 1089 * underutilized metadata chunk. So if this is our first run 1090 * through the flushing state machine skip ALLOC_CHUNK_FORCE and 1091 * commit the transaction. If nothing has changed the next go 1092 * around then we can force a chunk allocation. 1093 */ 1094 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles) 1095 flush_state++; 1096 1097 if (flush_state > COMMIT_TRANS) { 1098 commit_cycles++; 1099 if (commit_cycles > 2) { 1100 if (maybe_fail_all_tickets(fs_info, space_info)) { 1101 flush_state = FLUSH_DELAYED_ITEMS_NR; 1102 commit_cycles--; 1103 } else { 1104 space_info->flush = 0; 1105 } 1106 } else { 1107 flush_state = FLUSH_DELAYED_ITEMS_NR; 1108 } 1109 } 1110 spin_unlock(&space_info->lock); 1111 } while (flush_state <= COMMIT_TRANS); 1112 } 1113 1114 /* 1115 * This handles pre-flushing of metadata space before we get to the point that 1116 * we need to start blocking threads on tickets. The logic here is different 1117 * from the other flush paths because it doesn't rely on tickets to tell us how 1118 * much we need to flush, instead it attempts to keep us below the 80% full 1119 * watermark of space by flushing whichever reservation pool is currently the 1120 * largest. 1121 */ 1122 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work) 1123 { 1124 struct btrfs_fs_info *fs_info; 1125 struct btrfs_space_info *space_info; 1126 struct btrfs_block_rsv *delayed_block_rsv; 1127 struct btrfs_block_rsv *delayed_refs_rsv; 1128 struct btrfs_block_rsv *global_rsv; 1129 struct btrfs_block_rsv *trans_rsv; 1130 int loops = 0; 1131 1132 fs_info = container_of(work, struct btrfs_fs_info, 1133 preempt_reclaim_work); 1134 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 1135 delayed_block_rsv = &fs_info->delayed_block_rsv; 1136 delayed_refs_rsv = &fs_info->delayed_refs_rsv; 1137 global_rsv = &fs_info->global_block_rsv; 1138 trans_rsv = &fs_info->trans_block_rsv; 1139 1140 spin_lock(&space_info->lock); 1141 while (need_preemptive_reclaim(fs_info, space_info)) { 1142 enum btrfs_flush_state flush; 1143 u64 delalloc_size = 0; 1144 u64 to_reclaim, block_rsv_size; 1145 u64 global_rsv_size = global_rsv->reserved; 1146 1147 loops++; 1148 1149 /* 1150 * We don't have a precise counter for the metadata being 1151 * reserved for delalloc, so we'll approximate it by subtracting 1152 * out the block rsv's space from the bytes_may_use. If that 1153 * amount is higher than the individual reserves, then we can 1154 * assume it's tied up in delalloc reservations. 1155 */ 1156 block_rsv_size = global_rsv_size + 1157 delayed_block_rsv->reserved + 1158 delayed_refs_rsv->reserved + 1159 trans_rsv->reserved; 1160 if (block_rsv_size < space_info->bytes_may_use) 1161 delalloc_size = space_info->bytes_may_use - block_rsv_size; 1162 1163 /* 1164 * We don't want to include the global_rsv in our calculation, 1165 * because that's space we can't touch. Subtract it from the 1166 * block_rsv_size for the next checks. 1167 */ 1168 block_rsv_size -= global_rsv_size; 1169 1170 /* 1171 * We really want to avoid flushing delalloc too much, as it 1172 * could result in poor allocation patterns, so only flush it if 1173 * it's larger than the rest of the pools combined. 1174 */ 1175 if (delalloc_size > block_rsv_size) { 1176 to_reclaim = delalloc_size; 1177 flush = FLUSH_DELALLOC; 1178 } else if (space_info->bytes_pinned > 1179 (delayed_block_rsv->reserved + 1180 delayed_refs_rsv->reserved)) { 1181 to_reclaim = space_info->bytes_pinned; 1182 flush = COMMIT_TRANS; 1183 } else if (delayed_block_rsv->reserved > 1184 delayed_refs_rsv->reserved) { 1185 to_reclaim = delayed_block_rsv->reserved; 1186 flush = FLUSH_DELAYED_ITEMS_NR; 1187 } else { 1188 to_reclaim = delayed_refs_rsv->reserved; 1189 flush = FLUSH_DELAYED_REFS_NR; 1190 } 1191 1192 spin_unlock(&space_info->lock); 1193 1194 /* 1195 * We don't want to reclaim everything, just a portion, so scale 1196 * down the to_reclaim by 1/4. If it takes us down to 0, 1197 * reclaim 1 items worth. 1198 */ 1199 to_reclaim >>= 2; 1200 if (!to_reclaim) 1201 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1); 1202 flush_space(fs_info, space_info, to_reclaim, flush, true); 1203 cond_resched(); 1204 spin_lock(&space_info->lock); 1205 } 1206 1207 /* We only went through once, back off our clamping. */ 1208 if (loops == 1 && !space_info->reclaim_size) 1209 space_info->clamp = max(1, space_info->clamp - 1); 1210 trace_btrfs_done_preemptive_reclaim(fs_info, space_info); 1211 spin_unlock(&space_info->lock); 1212 } 1213 1214 /* 1215 * FLUSH_DELALLOC_WAIT: 1216 * Space is freed from flushing delalloc in one of two ways. 1217 * 1218 * 1) compression is on and we allocate less space than we reserved 1219 * 2) we are overwriting existing space 1220 * 1221 * For #1 that extra space is reclaimed as soon as the delalloc pages are 1222 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent 1223 * length to ->bytes_reserved, and subtracts the reserved space from 1224 * ->bytes_may_use. 1225 * 1226 * For #2 this is trickier. Once the ordered extent runs we will drop the 1227 * extent in the range we are overwriting, which creates a delayed ref for 1228 * that freed extent. This however is not reclaimed until the transaction 1229 * commits, thus the next stages. 1230 * 1231 * RUN_DELAYED_IPUTS 1232 * If we are freeing inodes, we want to make sure all delayed iputs have 1233 * completed, because they could have been on an inode with i_nlink == 0, and 1234 * thus have been truncated and freed up space. But again this space is not 1235 * immediately re-usable, it comes in the form of a delayed ref, which must be 1236 * run and then the transaction must be committed. 1237 * 1238 * COMMIT_TRANS 1239 * This is where we reclaim all of the pinned space generated by running the 1240 * iputs 1241 * 1242 * ALLOC_CHUNK_FORCE 1243 * For data we start with alloc chunk force, however we could have been full 1244 * before, and then the transaction commit could have freed new block groups, 1245 * so if we now have space to allocate do the force chunk allocation. 1246 */ 1247 static const enum btrfs_flush_state data_flush_states[] = { 1248 FLUSH_DELALLOC_FULL, 1249 RUN_DELAYED_IPUTS, 1250 COMMIT_TRANS, 1251 ALLOC_CHUNK_FORCE, 1252 }; 1253 1254 static void btrfs_async_reclaim_data_space(struct work_struct *work) 1255 { 1256 struct btrfs_fs_info *fs_info; 1257 struct btrfs_space_info *space_info; 1258 u64 last_tickets_id; 1259 enum btrfs_flush_state flush_state = 0; 1260 1261 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work); 1262 space_info = fs_info->data_sinfo; 1263 1264 spin_lock(&space_info->lock); 1265 if (list_empty(&space_info->tickets)) { 1266 space_info->flush = 0; 1267 spin_unlock(&space_info->lock); 1268 return; 1269 } 1270 last_tickets_id = space_info->tickets_id; 1271 spin_unlock(&space_info->lock); 1272 1273 while (!space_info->full) { 1274 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); 1275 spin_lock(&space_info->lock); 1276 if (list_empty(&space_info->tickets)) { 1277 space_info->flush = 0; 1278 spin_unlock(&space_info->lock); 1279 return; 1280 } 1281 1282 /* Something happened, fail everything and bail. */ 1283 if (BTRFS_FS_ERROR(fs_info)) 1284 goto aborted_fs; 1285 last_tickets_id = space_info->tickets_id; 1286 spin_unlock(&space_info->lock); 1287 } 1288 1289 while (flush_state < ARRAY_SIZE(data_flush_states)) { 1290 flush_space(fs_info, space_info, U64_MAX, 1291 data_flush_states[flush_state], false); 1292 spin_lock(&space_info->lock); 1293 if (list_empty(&space_info->tickets)) { 1294 space_info->flush = 0; 1295 spin_unlock(&space_info->lock); 1296 return; 1297 } 1298 1299 if (last_tickets_id == space_info->tickets_id) { 1300 flush_state++; 1301 } else { 1302 last_tickets_id = space_info->tickets_id; 1303 flush_state = 0; 1304 } 1305 1306 if (flush_state >= ARRAY_SIZE(data_flush_states)) { 1307 if (space_info->full) { 1308 if (maybe_fail_all_tickets(fs_info, space_info)) 1309 flush_state = 0; 1310 else 1311 space_info->flush = 0; 1312 } else { 1313 flush_state = 0; 1314 } 1315 1316 /* Something happened, fail everything and bail. */ 1317 if (BTRFS_FS_ERROR(fs_info)) 1318 goto aborted_fs; 1319 1320 } 1321 spin_unlock(&space_info->lock); 1322 } 1323 return; 1324 1325 aborted_fs: 1326 maybe_fail_all_tickets(fs_info, space_info); 1327 space_info->flush = 0; 1328 spin_unlock(&space_info->lock); 1329 } 1330 1331 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info) 1332 { 1333 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space); 1334 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space); 1335 INIT_WORK(&fs_info->preempt_reclaim_work, 1336 btrfs_preempt_reclaim_metadata_space); 1337 } 1338 1339 static const enum btrfs_flush_state priority_flush_states[] = { 1340 FLUSH_DELAYED_ITEMS_NR, 1341 FLUSH_DELAYED_ITEMS, 1342 ALLOC_CHUNK, 1343 }; 1344 1345 static const enum btrfs_flush_state evict_flush_states[] = { 1346 FLUSH_DELAYED_ITEMS_NR, 1347 FLUSH_DELAYED_ITEMS, 1348 FLUSH_DELAYED_REFS_NR, 1349 FLUSH_DELAYED_REFS, 1350 FLUSH_DELALLOC, 1351 FLUSH_DELALLOC_WAIT, 1352 FLUSH_DELALLOC_FULL, 1353 ALLOC_CHUNK, 1354 COMMIT_TRANS, 1355 }; 1356 1357 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info, 1358 struct btrfs_space_info *space_info, 1359 struct reserve_ticket *ticket, 1360 const enum btrfs_flush_state *states, 1361 int states_nr) 1362 { 1363 u64 to_reclaim; 1364 int flush_state = 0; 1365 1366 spin_lock(&space_info->lock); 1367 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 1368 /* 1369 * This is the priority reclaim path, so to_reclaim could be >0 still 1370 * because we may have only satisfied the priority tickets and still 1371 * left non priority tickets on the list. We would then have 1372 * to_reclaim but ->bytes == 0. 1373 */ 1374 if (ticket->bytes == 0) { 1375 spin_unlock(&space_info->lock); 1376 return; 1377 } 1378 1379 while (flush_state < states_nr) { 1380 spin_unlock(&space_info->lock); 1381 flush_space(fs_info, space_info, to_reclaim, states[flush_state], 1382 false); 1383 flush_state++; 1384 spin_lock(&space_info->lock); 1385 if (ticket->bytes == 0) { 1386 spin_unlock(&space_info->lock); 1387 return; 1388 } 1389 } 1390 1391 /* Attempt to steal from the global rsv if we can. */ 1392 if (!steal_from_global_rsv(fs_info, space_info, ticket)) { 1393 ticket->error = -ENOSPC; 1394 remove_ticket(space_info, ticket); 1395 } 1396 1397 /* 1398 * We must run try_granting_tickets here because we could be a large 1399 * ticket in front of a smaller ticket that can now be satisfied with 1400 * the available space. 1401 */ 1402 btrfs_try_granting_tickets(fs_info, space_info); 1403 spin_unlock(&space_info->lock); 1404 } 1405 1406 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info, 1407 struct btrfs_space_info *space_info, 1408 struct reserve_ticket *ticket) 1409 { 1410 spin_lock(&space_info->lock); 1411 1412 /* We could have been granted before we got here. */ 1413 if (ticket->bytes == 0) { 1414 spin_unlock(&space_info->lock); 1415 return; 1416 } 1417 1418 while (!space_info->full) { 1419 spin_unlock(&space_info->lock); 1420 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); 1421 spin_lock(&space_info->lock); 1422 if (ticket->bytes == 0) { 1423 spin_unlock(&space_info->lock); 1424 return; 1425 } 1426 } 1427 1428 ticket->error = -ENOSPC; 1429 remove_ticket(space_info, ticket); 1430 btrfs_try_granting_tickets(fs_info, space_info); 1431 spin_unlock(&space_info->lock); 1432 } 1433 1434 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info, 1435 struct btrfs_space_info *space_info, 1436 struct reserve_ticket *ticket) 1437 1438 { 1439 DEFINE_WAIT(wait); 1440 int ret = 0; 1441 1442 spin_lock(&space_info->lock); 1443 while (ticket->bytes > 0 && ticket->error == 0) { 1444 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE); 1445 if (ret) { 1446 /* 1447 * Delete us from the list. After we unlock the space 1448 * info, we don't want the async reclaim job to reserve 1449 * space for this ticket. If that would happen, then the 1450 * ticket's task would not known that space was reserved 1451 * despite getting an error, resulting in a space leak 1452 * (bytes_may_use counter of our space_info). 1453 */ 1454 remove_ticket(space_info, ticket); 1455 ticket->error = -EINTR; 1456 break; 1457 } 1458 spin_unlock(&space_info->lock); 1459 1460 schedule(); 1461 1462 finish_wait(&ticket->wait, &wait); 1463 spin_lock(&space_info->lock); 1464 } 1465 spin_unlock(&space_info->lock); 1466 } 1467 1468 /** 1469 * Do the appropriate flushing and waiting for a ticket 1470 * 1471 * @fs_info: the filesystem 1472 * @space_info: space info for the reservation 1473 * @ticket: ticket for the reservation 1474 * @start_ns: timestamp when the reservation started 1475 * @orig_bytes: amount of bytes originally reserved 1476 * @flush: how much we can flush 1477 * 1478 * This does the work of figuring out how to flush for the ticket, waiting for 1479 * the reservation, and returning the appropriate error if there is one. 1480 */ 1481 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info, 1482 struct btrfs_space_info *space_info, 1483 struct reserve_ticket *ticket, 1484 u64 start_ns, u64 orig_bytes, 1485 enum btrfs_reserve_flush_enum flush) 1486 { 1487 int ret; 1488 1489 switch (flush) { 1490 case BTRFS_RESERVE_FLUSH_DATA: 1491 case BTRFS_RESERVE_FLUSH_ALL: 1492 case BTRFS_RESERVE_FLUSH_ALL_STEAL: 1493 wait_reserve_ticket(fs_info, space_info, ticket); 1494 break; 1495 case BTRFS_RESERVE_FLUSH_LIMIT: 1496 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1497 priority_flush_states, 1498 ARRAY_SIZE(priority_flush_states)); 1499 break; 1500 case BTRFS_RESERVE_FLUSH_EVICT: 1501 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1502 evict_flush_states, 1503 ARRAY_SIZE(evict_flush_states)); 1504 break; 1505 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE: 1506 priority_reclaim_data_space(fs_info, space_info, ticket); 1507 break; 1508 default: 1509 ASSERT(0); 1510 break; 1511 } 1512 1513 ret = ticket->error; 1514 ASSERT(list_empty(&ticket->list)); 1515 /* 1516 * Check that we can't have an error set if the reservation succeeded, 1517 * as that would confuse tasks and lead them to error out without 1518 * releasing reserved space (if an error happens the expectation is that 1519 * space wasn't reserved at all). 1520 */ 1521 ASSERT(!(ticket->bytes == 0 && ticket->error)); 1522 trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes, 1523 start_ns, flush, ticket->error); 1524 return ret; 1525 } 1526 1527 /* 1528 * This returns true if this flush state will go through the ordinary flushing 1529 * code. 1530 */ 1531 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush) 1532 { 1533 return (flush == BTRFS_RESERVE_FLUSH_ALL) || 1534 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL); 1535 } 1536 1537 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info, 1538 struct btrfs_space_info *space_info) 1539 { 1540 u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes); 1541 u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes); 1542 1543 /* 1544 * If we're heavy on ordered operations then clamping won't help us. We 1545 * need to clamp specifically to keep up with dirty'ing buffered 1546 * writers, because there's not a 1:1 correlation of writing delalloc 1547 * and freeing space, like there is with flushing delayed refs or 1548 * delayed nodes. If we're already more ordered than delalloc then 1549 * we're keeping up, otherwise we aren't and should probably clamp. 1550 */ 1551 if (ordered < delalloc) 1552 space_info->clamp = min(space_info->clamp + 1, 8); 1553 } 1554 1555 static inline bool can_steal(enum btrfs_reserve_flush_enum flush) 1556 { 1557 return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || 1558 flush == BTRFS_RESERVE_FLUSH_EVICT); 1559 } 1560 1561 /** 1562 * Try to reserve bytes from the block_rsv's space 1563 * 1564 * @fs_info: the filesystem 1565 * @space_info: space info we want to allocate from 1566 * @orig_bytes: number of bytes we want 1567 * @flush: whether or not we can flush to make our reservation 1568 * 1569 * This will reserve orig_bytes number of bytes from the space info associated 1570 * with the block_rsv. If there is not enough space it will make an attempt to 1571 * flush out space to make room. It will do this by flushing delalloc if 1572 * possible or committing the transaction. If flush is 0 then no attempts to 1573 * regain reservations will be made and this will fail if there is not enough 1574 * space already. 1575 */ 1576 static int __reserve_bytes(struct btrfs_fs_info *fs_info, 1577 struct btrfs_space_info *space_info, u64 orig_bytes, 1578 enum btrfs_reserve_flush_enum flush) 1579 { 1580 struct work_struct *async_work; 1581 struct reserve_ticket ticket; 1582 u64 start_ns = 0; 1583 u64 used; 1584 int ret = 0; 1585 bool pending_tickets; 1586 1587 ASSERT(orig_bytes); 1588 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL); 1589 1590 if (flush == BTRFS_RESERVE_FLUSH_DATA) 1591 async_work = &fs_info->async_data_reclaim_work; 1592 else 1593 async_work = &fs_info->async_reclaim_work; 1594 1595 spin_lock(&space_info->lock); 1596 ret = -ENOSPC; 1597 used = btrfs_space_info_used(space_info, true); 1598 1599 /* 1600 * We don't want NO_FLUSH allocations to jump everybody, they can 1601 * generally handle ENOSPC in a different way, so treat them the same as 1602 * normal flushers when it comes to skipping pending tickets. 1603 */ 1604 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH)) 1605 pending_tickets = !list_empty(&space_info->tickets) || 1606 !list_empty(&space_info->priority_tickets); 1607 else 1608 pending_tickets = !list_empty(&space_info->priority_tickets); 1609 1610 /* 1611 * Carry on if we have enough space (short-circuit) OR call 1612 * can_overcommit() to ensure we can overcommit to continue. 1613 */ 1614 if (!pending_tickets && 1615 ((used + orig_bytes <= writable_total_bytes(fs_info, space_info)) || 1616 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) { 1617 btrfs_space_info_update_bytes_may_use(fs_info, space_info, 1618 orig_bytes); 1619 ret = 0; 1620 } 1621 1622 /* 1623 * If we couldn't make a reservation then setup our reservation ticket 1624 * and kick the async worker if it's not already running. 1625 * 1626 * If we are a priority flusher then we just need to add our ticket to 1627 * the list and we will do our own flushing further down. 1628 */ 1629 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) { 1630 ticket.bytes = orig_bytes; 1631 ticket.error = 0; 1632 space_info->reclaim_size += ticket.bytes; 1633 init_waitqueue_head(&ticket.wait); 1634 ticket.steal = can_steal(flush); 1635 if (trace_btrfs_reserve_ticket_enabled()) 1636 start_ns = ktime_get_ns(); 1637 1638 if (flush == BTRFS_RESERVE_FLUSH_ALL || 1639 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || 1640 flush == BTRFS_RESERVE_FLUSH_DATA) { 1641 list_add_tail(&ticket.list, &space_info->tickets); 1642 if (!space_info->flush) { 1643 /* 1644 * We were forced to add a reserve ticket, so 1645 * our preemptive flushing is unable to keep 1646 * up. Clamp down on the threshold for the 1647 * preemptive flushing in order to keep up with 1648 * the workload. 1649 */ 1650 maybe_clamp_preempt(fs_info, space_info); 1651 1652 space_info->flush = 1; 1653 trace_btrfs_trigger_flush(fs_info, 1654 space_info->flags, 1655 orig_bytes, flush, 1656 "enospc"); 1657 queue_work(system_unbound_wq, async_work); 1658 } 1659 } else { 1660 list_add_tail(&ticket.list, 1661 &space_info->priority_tickets); 1662 } 1663 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) { 1664 used += orig_bytes; 1665 /* 1666 * We will do the space reservation dance during log replay, 1667 * which means we won't have fs_info->fs_root set, so don't do 1668 * the async reclaim as we will panic. 1669 */ 1670 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) && 1671 !work_busy(&fs_info->preempt_reclaim_work) && 1672 need_preemptive_reclaim(fs_info, space_info)) { 1673 trace_btrfs_trigger_flush(fs_info, space_info->flags, 1674 orig_bytes, flush, "preempt"); 1675 queue_work(system_unbound_wq, 1676 &fs_info->preempt_reclaim_work); 1677 } 1678 } 1679 spin_unlock(&space_info->lock); 1680 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH) 1681 return ret; 1682 1683 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns, 1684 orig_bytes, flush); 1685 } 1686 1687 /** 1688 * Trye to reserve metadata bytes from the block_rsv's space 1689 * 1690 * @fs_info: the filesystem 1691 * @block_rsv: block_rsv we're allocating for 1692 * @orig_bytes: number of bytes we want 1693 * @flush: whether or not we can flush to make our reservation 1694 * 1695 * This will reserve orig_bytes number of bytes from the space info associated 1696 * with the block_rsv. If there is not enough space it will make an attempt to 1697 * flush out space to make room. It will do this by flushing delalloc if 1698 * possible or committing the transaction. If flush is 0 then no attempts to 1699 * regain reservations will be made and this will fail if there is not enough 1700 * space already. 1701 */ 1702 int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info, 1703 struct btrfs_block_rsv *block_rsv, 1704 u64 orig_bytes, 1705 enum btrfs_reserve_flush_enum flush) 1706 { 1707 int ret; 1708 1709 ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush); 1710 if (ret == -ENOSPC) { 1711 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1712 block_rsv->space_info->flags, 1713 orig_bytes, 1); 1714 1715 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1716 btrfs_dump_space_info(fs_info, block_rsv->space_info, 1717 orig_bytes, 0); 1718 } 1719 return ret; 1720 } 1721 1722 /** 1723 * Try to reserve data bytes for an allocation 1724 * 1725 * @fs_info: the filesystem 1726 * @bytes: number of bytes we need 1727 * @flush: how we are allowed to flush 1728 * 1729 * This will reserve bytes from the data space info. If there is not enough 1730 * space then we will attempt to flush space as specified by flush. 1731 */ 1732 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes, 1733 enum btrfs_reserve_flush_enum flush) 1734 { 1735 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo; 1736 int ret; 1737 1738 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA || 1739 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE); 1740 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA); 1741 1742 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush); 1743 if (ret == -ENOSPC) { 1744 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1745 data_sinfo->flags, bytes, 1); 1746 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1747 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0); 1748 } 1749 return ret; 1750 } 1751