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