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 static int create_space_info(struct btrfs_fs_info *info, u64 flags) 185 { 186 187 struct btrfs_space_info *space_info; 188 int i; 189 int ret; 190 191 space_info = kzalloc(sizeof(*space_info), GFP_NOFS); 192 if (!space_info) 193 return -ENOMEM; 194 195 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) 196 INIT_LIST_HEAD(&space_info->block_groups[i]); 197 init_rwsem(&space_info->groups_sem); 198 spin_lock_init(&space_info->lock); 199 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK; 200 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE; 201 INIT_LIST_HEAD(&space_info->ro_bgs); 202 INIT_LIST_HEAD(&space_info->tickets); 203 INIT_LIST_HEAD(&space_info->priority_tickets); 204 space_info->clamp = 1; 205 206 ret = btrfs_sysfs_add_space_info_type(info, space_info); 207 if (ret) 208 return ret; 209 210 list_add(&space_info->list, &info->space_info); 211 if (flags & BTRFS_BLOCK_GROUP_DATA) 212 info->data_sinfo = space_info; 213 214 return ret; 215 } 216 217 int btrfs_init_space_info(struct btrfs_fs_info *fs_info) 218 { 219 struct btrfs_super_block *disk_super; 220 u64 features; 221 u64 flags; 222 int mixed = 0; 223 int ret; 224 225 disk_super = fs_info->super_copy; 226 if (!btrfs_super_root(disk_super)) 227 return -EINVAL; 228 229 features = btrfs_super_incompat_flags(disk_super); 230 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS) 231 mixed = 1; 232 233 flags = BTRFS_BLOCK_GROUP_SYSTEM; 234 ret = create_space_info(fs_info, flags); 235 if (ret) 236 goto out; 237 238 if (mixed) { 239 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA; 240 ret = create_space_info(fs_info, flags); 241 } else { 242 flags = BTRFS_BLOCK_GROUP_METADATA; 243 ret = create_space_info(fs_info, flags); 244 if (ret) 245 goto out; 246 247 flags = BTRFS_BLOCK_GROUP_DATA; 248 ret = create_space_info(fs_info, flags); 249 } 250 out: 251 return ret; 252 } 253 254 void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags, 255 u64 total_bytes, u64 bytes_used, 256 u64 bytes_readonly, u64 bytes_zone_unusable, 257 struct btrfs_space_info **space_info) 258 { 259 struct btrfs_space_info *found; 260 int factor; 261 262 factor = btrfs_bg_type_to_factor(flags); 263 264 found = btrfs_find_space_info(info, flags); 265 ASSERT(found); 266 spin_lock(&found->lock); 267 found->total_bytes += total_bytes; 268 found->disk_total += total_bytes * factor; 269 found->bytes_used += bytes_used; 270 found->disk_used += bytes_used * factor; 271 found->bytes_readonly += bytes_readonly; 272 found->bytes_zone_unusable += bytes_zone_unusable; 273 if (total_bytes > 0) 274 found->full = 0; 275 btrfs_try_granting_tickets(info, found); 276 spin_unlock(&found->lock); 277 *space_info = found; 278 } 279 280 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info, 281 u64 flags) 282 { 283 struct list_head *head = &info->space_info; 284 struct btrfs_space_info *found; 285 286 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK; 287 288 list_for_each_entry(found, head, list) { 289 if (found->flags & flags) 290 return found; 291 } 292 return NULL; 293 } 294 295 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info, 296 struct btrfs_space_info *space_info, 297 enum btrfs_reserve_flush_enum flush) 298 { 299 u64 profile; 300 u64 avail; 301 int factor; 302 303 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM) 304 profile = btrfs_system_alloc_profile(fs_info); 305 else 306 profile = btrfs_metadata_alloc_profile(fs_info); 307 308 avail = atomic64_read(&fs_info->free_chunk_space); 309 310 /* 311 * If we have dup, raid1 or raid10 then only half of the free 312 * space is actually usable. For raid56, the space info used 313 * doesn't include the parity drive, so we don't have to 314 * change the math 315 */ 316 factor = btrfs_bg_type_to_factor(profile); 317 avail = div_u64(avail, factor); 318 319 /* 320 * If we aren't flushing all things, let us overcommit up to 321 * 1/2th of the space. If we can flush, don't let us overcommit 322 * too much, let it overcommit up to 1/8 of the space. 323 */ 324 if (flush == BTRFS_RESERVE_FLUSH_ALL) 325 avail >>= 3; 326 else 327 avail >>= 1; 328 return avail; 329 } 330 331 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info, 332 struct btrfs_space_info *space_info, u64 bytes, 333 enum btrfs_reserve_flush_enum flush) 334 { 335 u64 avail; 336 u64 used; 337 338 /* Don't overcommit when in mixed mode */ 339 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA) 340 return 0; 341 342 used = btrfs_space_info_used(space_info, true); 343 avail = calc_available_free_space(fs_info, space_info, flush); 344 345 if (used + bytes < space_info->total_bytes + avail) 346 return 1; 347 return 0; 348 } 349 350 static void remove_ticket(struct btrfs_space_info *space_info, 351 struct reserve_ticket *ticket) 352 { 353 if (!list_empty(&ticket->list)) { 354 list_del_init(&ticket->list); 355 ASSERT(space_info->reclaim_size >= ticket->bytes); 356 space_info->reclaim_size -= ticket->bytes; 357 } 358 } 359 360 /* 361 * This is for space we already have accounted in space_info->bytes_may_use, so 362 * basically when we're returning space from block_rsv's. 363 */ 364 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info, 365 struct btrfs_space_info *space_info) 366 { 367 struct list_head *head; 368 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH; 369 370 lockdep_assert_held(&space_info->lock); 371 372 head = &space_info->priority_tickets; 373 again: 374 while (!list_empty(head)) { 375 struct reserve_ticket *ticket; 376 u64 used = btrfs_space_info_used(space_info, true); 377 378 ticket = list_first_entry(head, struct reserve_ticket, list); 379 380 /* Check and see if our ticket can be satisfied now. */ 381 if ((used + ticket->bytes <= space_info->total_bytes) || 382 btrfs_can_overcommit(fs_info, space_info, ticket->bytes, 383 flush)) { 384 btrfs_space_info_update_bytes_may_use(fs_info, 385 space_info, 386 ticket->bytes); 387 remove_ticket(space_info, ticket); 388 ticket->bytes = 0; 389 space_info->tickets_id++; 390 wake_up(&ticket->wait); 391 } else { 392 break; 393 } 394 } 395 396 if (head == &space_info->priority_tickets) { 397 head = &space_info->tickets; 398 flush = BTRFS_RESERVE_FLUSH_ALL; 399 goto again; 400 } 401 } 402 403 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \ 404 do { \ 405 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \ 406 spin_lock(&__rsv->lock); \ 407 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \ 408 __rsv->size, __rsv->reserved); \ 409 spin_unlock(&__rsv->lock); \ 410 } while (0) 411 412 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 413 struct btrfs_space_info *info) 414 { 415 lockdep_assert_held(&info->lock); 416 417 btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull", 418 info->flags, 419 info->total_bytes - btrfs_space_info_used(info, true), 420 info->full ? "" : "not "); 421 btrfs_info(fs_info, 422 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu", 423 info->total_bytes, info->bytes_used, info->bytes_pinned, 424 info->bytes_reserved, info->bytes_may_use, 425 info->bytes_readonly, info->bytes_zone_unusable); 426 427 DUMP_BLOCK_RSV(fs_info, global_block_rsv); 428 DUMP_BLOCK_RSV(fs_info, trans_block_rsv); 429 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv); 430 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv); 431 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv); 432 433 } 434 435 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 436 struct btrfs_space_info *info, u64 bytes, 437 int dump_block_groups) 438 { 439 struct btrfs_block_group *cache; 440 int index = 0; 441 442 spin_lock(&info->lock); 443 __btrfs_dump_space_info(fs_info, info); 444 spin_unlock(&info->lock); 445 446 if (!dump_block_groups) 447 return; 448 449 down_read(&info->groups_sem); 450 again: 451 list_for_each_entry(cache, &info->block_groups[index], list) { 452 spin_lock(&cache->lock); 453 btrfs_info(fs_info, 454 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu zone_unusable %s", 455 cache->start, cache->length, cache->used, cache->pinned, 456 cache->reserved, cache->zone_unusable, 457 cache->ro ? "[readonly]" : ""); 458 spin_unlock(&cache->lock); 459 btrfs_dump_free_space(cache, bytes); 460 } 461 if (++index < BTRFS_NR_RAID_TYPES) 462 goto again; 463 up_read(&info->groups_sem); 464 } 465 466 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info, 467 u64 to_reclaim) 468 { 469 u64 bytes; 470 u64 nr; 471 472 bytes = btrfs_calc_insert_metadata_size(fs_info, 1); 473 nr = div64_u64(to_reclaim, bytes); 474 if (!nr) 475 nr = 1; 476 return nr; 477 } 478 479 #define EXTENT_SIZE_PER_ITEM SZ_256K 480 481 /* 482 * shrink metadata reservation for delalloc 483 */ 484 static void shrink_delalloc(struct btrfs_fs_info *fs_info, 485 struct btrfs_space_info *space_info, 486 u64 to_reclaim, bool wait_ordered, 487 bool for_preempt) 488 { 489 struct btrfs_trans_handle *trans; 490 u64 delalloc_bytes; 491 u64 ordered_bytes; 492 u64 items; 493 long time_left; 494 int loops; 495 496 delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes); 497 ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes); 498 if (delalloc_bytes == 0 && ordered_bytes == 0) 499 return; 500 501 /* Calc the number of the pages we need flush for space reservation */ 502 if (to_reclaim == U64_MAX) { 503 items = U64_MAX; 504 } else { 505 /* 506 * to_reclaim is set to however much metadata we need to 507 * reclaim, but reclaiming that much data doesn't really track 508 * exactly. What we really want to do is reclaim full inode's 509 * worth of reservations, however that's not available to us 510 * here. We will take a fraction of the delalloc bytes for our 511 * flushing loops and hope for the best. Delalloc will expand 512 * the amount we write to cover an entire dirty extent, which 513 * will reclaim the metadata reservation for that range. If 514 * it's not enough subsequent flush stages will be more 515 * aggressive. 516 */ 517 to_reclaim = max(to_reclaim, delalloc_bytes >> 3); 518 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2; 519 } 520 521 trans = (struct btrfs_trans_handle *)current->journal_info; 522 523 /* 524 * If we are doing more ordered than delalloc we need to just wait on 525 * ordered extents, otherwise we'll waste time trying to flush delalloc 526 * that likely won't give us the space back we need. 527 */ 528 if (ordered_bytes > delalloc_bytes && !for_preempt) 529 wait_ordered = true; 530 531 loops = 0; 532 while ((delalloc_bytes || ordered_bytes) && loops < 3) { 533 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT; 534 long nr_pages = min_t(u64, temp, LONG_MAX); 535 536 btrfs_start_delalloc_roots(fs_info, nr_pages, true); 537 538 loops++; 539 if (wait_ordered && !trans) { 540 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1); 541 } else { 542 time_left = schedule_timeout_killable(1); 543 if (time_left) 544 break; 545 } 546 547 /* 548 * If we are for preemption we just want a one-shot of delalloc 549 * flushing so we can stop flushing if we decide we don't need 550 * to anymore. 551 */ 552 if (for_preempt) 553 break; 554 555 spin_lock(&space_info->lock); 556 if (list_empty(&space_info->tickets) && 557 list_empty(&space_info->priority_tickets)) { 558 spin_unlock(&space_info->lock); 559 break; 560 } 561 spin_unlock(&space_info->lock); 562 563 delalloc_bytes = percpu_counter_sum_positive( 564 &fs_info->delalloc_bytes); 565 ordered_bytes = percpu_counter_sum_positive( 566 &fs_info->ordered_bytes); 567 } 568 } 569 570 /* 571 * Try to flush some data based on policy set by @state. This is only advisory 572 * and may fail for various reasons. The caller is supposed to examine the 573 * state of @space_info to detect the outcome. 574 */ 575 static void flush_space(struct btrfs_fs_info *fs_info, 576 struct btrfs_space_info *space_info, u64 num_bytes, 577 enum btrfs_flush_state state, bool for_preempt) 578 { 579 struct btrfs_root *root = fs_info->extent_root; 580 struct btrfs_trans_handle *trans; 581 int nr; 582 int ret = 0; 583 584 switch (state) { 585 case FLUSH_DELAYED_ITEMS_NR: 586 case FLUSH_DELAYED_ITEMS: 587 if (state == FLUSH_DELAYED_ITEMS_NR) 588 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2; 589 else 590 nr = -1; 591 592 trans = btrfs_join_transaction(root); 593 if (IS_ERR(trans)) { 594 ret = PTR_ERR(trans); 595 break; 596 } 597 ret = btrfs_run_delayed_items_nr(trans, nr); 598 btrfs_end_transaction(trans); 599 break; 600 case FLUSH_DELALLOC: 601 case FLUSH_DELALLOC_WAIT: 602 case FLUSH_DELALLOC_FULL: 603 if (state == FLUSH_DELALLOC_FULL) 604 num_bytes = U64_MAX; 605 shrink_delalloc(fs_info, space_info, num_bytes, 606 state != FLUSH_DELALLOC, for_preempt); 607 break; 608 case FLUSH_DELAYED_REFS_NR: 609 case FLUSH_DELAYED_REFS: 610 trans = btrfs_join_transaction(root); 611 if (IS_ERR(trans)) { 612 ret = PTR_ERR(trans); 613 break; 614 } 615 if (state == FLUSH_DELAYED_REFS_NR) 616 nr = calc_reclaim_items_nr(fs_info, num_bytes); 617 else 618 nr = 0; 619 btrfs_run_delayed_refs(trans, nr); 620 btrfs_end_transaction(trans); 621 break; 622 case ALLOC_CHUNK: 623 case ALLOC_CHUNK_FORCE: 624 trans = btrfs_join_transaction(root); 625 if (IS_ERR(trans)) { 626 ret = PTR_ERR(trans); 627 break; 628 } 629 ret = btrfs_chunk_alloc(trans, 630 btrfs_get_alloc_profile(fs_info, space_info->flags), 631 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE : 632 CHUNK_ALLOC_FORCE); 633 btrfs_end_transaction(trans); 634 if (ret > 0 || ret == -ENOSPC) 635 ret = 0; 636 break; 637 case RUN_DELAYED_IPUTS: 638 /* 639 * If we have pending delayed iputs then we could free up a 640 * bunch of pinned space, so make sure we run the iputs before 641 * we do our pinned bytes check below. 642 */ 643 btrfs_run_delayed_iputs(fs_info); 644 btrfs_wait_on_delayed_iputs(fs_info); 645 break; 646 case COMMIT_TRANS: 647 ASSERT(current->journal_info == NULL); 648 trans = btrfs_join_transaction(root); 649 if (IS_ERR(trans)) { 650 ret = PTR_ERR(trans); 651 break; 652 } 653 ret = btrfs_commit_transaction(trans); 654 break; 655 default: 656 ret = -ENOSPC; 657 break; 658 } 659 660 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state, 661 ret, for_preempt); 662 return; 663 } 664 665 static inline u64 666 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info, 667 struct btrfs_space_info *space_info) 668 { 669 u64 used; 670 u64 avail; 671 u64 to_reclaim = space_info->reclaim_size; 672 673 lockdep_assert_held(&space_info->lock); 674 675 avail = calc_available_free_space(fs_info, space_info, 676 BTRFS_RESERVE_FLUSH_ALL); 677 used = btrfs_space_info_used(space_info, true); 678 679 /* 680 * We may be flushing because suddenly we have less space than we had 681 * before, and now we're well over-committed based on our current free 682 * space. If that's the case add in our overage so we make sure to put 683 * appropriate pressure on the flushing state machine. 684 */ 685 if (space_info->total_bytes + avail < used) 686 to_reclaim += used - (space_info->total_bytes + avail); 687 688 return to_reclaim; 689 } 690 691 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info, 692 struct btrfs_space_info *space_info) 693 { 694 u64 global_rsv_size = fs_info->global_block_rsv.reserved; 695 u64 ordered, delalloc; 696 u64 thresh = div_factor_fine(space_info->total_bytes, 98); 697 u64 used; 698 699 /* If we're just plain full then async reclaim just slows us down. */ 700 if ((space_info->bytes_used + space_info->bytes_reserved + 701 global_rsv_size) >= thresh) 702 return false; 703 704 /* 705 * We have tickets queued, bail so we don't compete with the async 706 * flushers. 707 */ 708 if (space_info->reclaim_size) 709 return false; 710 711 /* 712 * If we have over half of the free space occupied by reservations or 713 * pinned then we want to start flushing. 714 * 715 * We do not do the traditional thing here, which is to say 716 * 717 * if (used >= ((total_bytes + avail) / 2)) 718 * return 1; 719 * 720 * because this doesn't quite work how we want. If we had more than 50% 721 * of the space_info used by bytes_used and we had 0 available we'd just 722 * constantly run the background flusher. Instead we want it to kick in 723 * if our reclaimable space exceeds our clamped free space. 724 * 725 * Our clamping range is 2^1 -> 2^8. Practically speaking that means 726 * the following: 727 * 728 * Amount of RAM Minimum threshold Maximum threshold 729 * 730 * 256GiB 1GiB 128GiB 731 * 128GiB 512MiB 64GiB 732 * 64GiB 256MiB 32GiB 733 * 32GiB 128MiB 16GiB 734 * 16GiB 64MiB 8GiB 735 * 736 * These are the range our thresholds will fall in, corresponding to how 737 * much delalloc we need for the background flusher to kick in. 738 */ 739 740 thresh = calc_available_free_space(fs_info, space_info, 741 BTRFS_RESERVE_FLUSH_ALL); 742 used = space_info->bytes_used + space_info->bytes_reserved + 743 space_info->bytes_readonly + global_rsv_size; 744 if (used < space_info->total_bytes) 745 thresh += space_info->total_bytes - used; 746 thresh >>= space_info->clamp; 747 748 used = space_info->bytes_pinned; 749 750 /* 751 * If we have more ordered bytes than delalloc bytes then we're either 752 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting 753 * around. Preemptive flushing is only useful in that it can free up 754 * space before tickets need to wait for things to finish. In the case 755 * of ordered extents, preemptively waiting on ordered extents gets us 756 * nothing, if our reservations are tied up in ordered extents we'll 757 * simply have to slow down writers by forcing them to wait on ordered 758 * extents. 759 * 760 * In the case that ordered is larger than delalloc, only include the 761 * block reserves that we would actually be able to directly reclaim 762 * from. In this case if we're heavy on metadata operations this will 763 * clearly be heavy enough to warrant preemptive flushing. In the case 764 * of heavy DIO or ordered reservations, preemptive flushing will just 765 * waste time and cause us to slow down. 766 * 767 * We want to make sure we truly are maxed out on ordered however, so 768 * cut ordered in half, and if it's still higher than delalloc then we 769 * can keep flushing. This is to avoid the case where we start 770 * flushing, and now delalloc == ordered and we stop preemptively 771 * flushing when we could still have several gigs of delalloc to flush. 772 */ 773 ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1; 774 delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes); 775 if (ordered >= delalloc) 776 used += fs_info->delayed_refs_rsv.reserved + 777 fs_info->delayed_block_rsv.reserved; 778 else 779 used += space_info->bytes_may_use - global_rsv_size; 780 781 return (used >= thresh && !btrfs_fs_closing(fs_info) && 782 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state)); 783 } 784 785 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info, 786 struct btrfs_space_info *space_info, 787 struct reserve_ticket *ticket) 788 { 789 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 790 u64 min_bytes; 791 792 if (global_rsv->space_info != space_info) 793 return false; 794 795 spin_lock(&global_rsv->lock); 796 min_bytes = div_factor(global_rsv->size, 1); 797 if (global_rsv->reserved < min_bytes + ticket->bytes) { 798 spin_unlock(&global_rsv->lock); 799 return false; 800 } 801 global_rsv->reserved -= ticket->bytes; 802 remove_ticket(space_info, ticket); 803 ticket->bytes = 0; 804 wake_up(&ticket->wait); 805 space_info->tickets_id++; 806 if (global_rsv->reserved < global_rsv->size) 807 global_rsv->full = 0; 808 spin_unlock(&global_rsv->lock); 809 810 return true; 811 } 812 813 /* 814 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets 815 * @fs_info - fs_info for this fs 816 * @space_info - the space info we were flushing 817 * 818 * We call this when we've exhausted our flushing ability and haven't made 819 * progress in satisfying tickets. The reservation code handles tickets in 820 * order, so if there is a large ticket first and then smaller ones we could 821 * very well satisfy the smaller tickets. This will attempt to wake up any 822 * tickets in the list to catch this case. 823 * 824 * This function returns true if it was able to make progress by clearing out 825 * other tickets, or if it stumbles across a ticket that was smaller than the 826 * first ticket. 827 */ 828 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info, 829 struct btrfs_space_info *space_info) 830 { 831 struct reserve_ticket *ticket; 832 u64 tickets_id = space_info->tickets_id; 833 834 trace_btrfs_fail_all_tickets(fs_info, space_info); 835 836 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) { 837 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info"); 838 __btrfs_dump_space_info(fs_info, space_info); 839 } 840 841 while (!list_empty(&space_info->tickets) && 842 tickets_id == space_info->tickets_id) { 843 ticket = list_first_entry(&space_info->tickets, 844 struct reserve_ticket, list); 845 846 if (ticket->steal && 847 steal_from_global_rsv(fs_info, space_info, ticket)) 848 return true; 849 850 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 851 btrfs_info(fs_info, "failing ticket with %llu bytes", 852 ticket->bytes); 853 854 remove_ticket(space_info, ticket); 855 ticket->error = -ENOSPC; 856 wake_up(&ticket->wait); 857 858 /* 859 * We're just throwing tickets away, so more flushing may not 860 * trip over btrfs_try_granting_tickets, so we need to call it 861 * here to see if we can make progress with the next ticket in 862 * the list. 863 */ 864 btrfs_try_granting_tickets(fs_info, space_info); 865 } 866 return (tickets_id != space_info->tickets_id); 867 } 868 869 /* 870 * This is for normal flushers, we can wait all goddamned day if we want to. We 871 * will loop and continuously try to flush as long as we are making progress. 872 * We count progress as clearing off tickets each time we have to loop. 873 */ 874 static void btrfs_async_reclaim_metadata_space(struct work_struct *work) 875 { 876 struct btrfs_fs_info *fs_info; 877 struct btrfs_space_info *space_info; 878 u64 to_reclaim; 879 enum btrfs_flush_state flush_state; 880 int commit_cycles = 0; 881 u64 last_tickets_id; 882 883 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work); 884 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 885 886 spin_lock(&space_info->lock); 887 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 888 if (!to_reclaim) { 889 space_info->flush = 0; 890 spin_unlock(&space_info->lock); 891 return; 892 } 893 last_tickets_id = space_info->tickets_id; 894 spin_unlock(&space_info->lock); 895 896 flush_state = FLUSH_DELAYED_ITEMS_NR; 897 do { 898 flush_space(fs_info, space_info, to_reclaim, flush_state, false); 899 spin_lock(&space_info->lock); 900 if (list_empty(&space_info->tickets)) { 901 space_info->flush = 0; 902 spin_unlock(&space_info->lock); 903 return; 904 } 905 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, 906 space_info); 907 if (last_tickets_id == space_info->tickets_id) { 908 flush_state++; 909 } else { 910 last_tickets_id = space_info->tickets_id; 911 flush_state = FLUSH_DELAYED_ITEMS_NR; 912 if (commit_cycles) 913 commit_cycles--; 914 } 915 916 /* 917 * We do not want to empty the system of delalloc unless we're 918 * under heavy pressure, so allow one trip through the flushing 919 * logic before we start doing a FLUSH_DELALLOC_FULL. 920 */ 921 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles) 922 flush_state++; 923 924 /* 925 * We don't want to force a chunk allocation until we've tried 926 * pretty hard to reclaim space. Think of the case where we 927 * freed up a bunch of space and so have a lot of pinned space 928 * to reclaim. We would rather use that than possibly create a 929 * underutilized metadata chunk. So if this is our first run 930 * through the flushing state machine skip ALLOC_CHUNK_FORCE and 931 * commit the transaction. If nothing has changed the next go 932 * around then we can force a chunk allocation. 933 */ 934 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles) 935 flush_state++; 936 937 if (flush_state > COMMIT_TRANS) { 938 commit_cycles++; 939 if (commit_cycles > 2) { 940 if (maybe_fail_all_tickets(fs_info, space_info)) { 941 flush_state = FLUSH_DELAYED_ITEMS_NR; 942 commit_cycles--; 943 } else { 944 space_info->flush = 0; 945 } 946 } else { 947 flush_state = FLUSH_DELAYED_ITEMS_NR; 948 } 949 } 950 spin_unlock(&space_info->lock); 951 } while (flush_state <= COMMIT_TRANS); 952 } 953 954 /* 955 * This handles pre-flushing of metadata space before we get to the point that 956 * we need to start blocking threads on tickets. The logic here is different 957 * from the other flush paths because it doesn't rely on tickets to tell us how 958 * much we need to flush, instead it attempts to keep us below the 80% full 959 * watermark of space by flushing whichever reservation pool is currently the 960 * largest. 961 */ 962 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work) 963 { 964 struct btrfs_fs_info *fs_info; 965 struct btrfs_space_info *space_info; 966 struct btrfs_block_rsv *delayed_block_rsv; 967 struct btrfs_block_rsv *delayed_refs_rsv; 968 struct btrfs_block_rsv *global_rsv; 969 struct btrfs_block_rsv *trans_rsv; 970 int loops = 0; 971 972 fs_info = container_of(work, struct btrfs_fs_info, 973 preempt_reclaim_work); 974 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 975 delayed_block_rsv = &fs_info->delayed_block_rsv; 976 delayed_refs_rsv = &fs_info->delayed_refs_rsv; 977 global_rsv = &fs_info->global_block_rsv; 978 trans_rsv = &fs_info->trans_block_rsv; 979 980 spin_lock(&space_info->lock); 981 while (need_preemptive_reclaim(fs_info, space_info)) { 982 enum btrfs_flush_state flush; 983 u64 delalloc_size = 0; 984 u64 to_reclaim, block_rsv_size; 985 u64 global_rsv_size = global_rsv->reserved; 986 987 loops++; 988 989 /* 990 * We don't have a precise counter for the metadata being 991 * reserved for delalloc, so we'll approximate it by subtracting 992 * out the block rsv's space from the bytes_may_use. If that 993 * amount is higher than the individual reserves, then we can 994 * assume it's tied up in delalloc reservations. 995 */ 996 block_rsv_size = global_rsv_size + 997 delayed_block_rsv->reserved + 998 delayed_refs_rsv->reserved + 999 trans_rsv->reserved; 1000 if (block_rsv_size < space_info->bytes_may_use) 1001 delalloc_size = space_info->bytes_may_use - block_rsv_size; 1002 spin_unlock(&space_info->lock); 1003 1004 /* 1005 * We don't want to include the global_rsv in our calculation, 1006 * because that's space we can't touch. Subtract it from the 1007 * block_rsv_size for the next checks. 1008 */ 1009 block_rsv_size -= global_rsv_size; 1010 1011 /* 1012 * We really want to avoid flushing delalloc too much, as it 1013 * could result in poor allocation patterns, so only flush it if 1014 * it's larger than the rest of the pools combined. 1015 */ 1016 if (delalloc_size > block_rsv_size) { 1017 to_reclaim = delalloc_size; 1018 flush = FLUSH_DELALLOC; 1019 } else if (space_info->bytes_pinned > 1020 (delayed_block_rsv->reserved + 1021 delayed_refs_rsv->reserved)) { 1022 to_reclaim = space_info->bytes_pinned; 1023 flush = COMMIT_TRANS; 1024 } else if (delayed_block_rsv->reserved > 1025 delayed_refs_rsv->reserved) { 1026 to_reclaim = delayed_block_rsv->reserved; 1027 flush = FLUSH_DELAYED_ITEMS_NR; 1028 } else { 1029 to_reclaim = delayed_refs_rsv->reserved; 1030 flush = FLUSH_DELAYED_REFS_NR; 1031 } 1032 1033 /* 1034 * We don't want to reclaim everything, just a portion, so scale 1035 * down the to_reclaim by 1/4. If it takes us down to 0, 1036 * reclaim 1 items worth. 1037 */ 1038 to_reclaim >>= 2; 1039 if (!to_reclaim) 1040 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1); 1041 flush_space(fs_info, space_info, to_reclaim, flush, true); 1042 cond_resched(); 1043 spin_lock(&space_info->lock); 1044 } 1045 1046 /* We only went through once, back off our clamping. */ 1047 if (loops == 1 && !space_info->reclaim_size) 1048 space_info->clamp = max(1, space_info->clamp - 1); 1049 trace_btrfs_done_preemptive_reclaim(fs_info, space_info); 1050 spin_unlock(&space_info->lock); 1051 } 1052 1053 /* 1054 * FLUSH_DELALLOC_WAIT: 1055 * Space is freed from flushing delalloc in one of two ways. 1056 * 1057 * 1) compression is on and we allocate less space than we reserved 1058 * 2) we are overwriting existing space 1059 * 1060 * For #1 that extra space is reclaimed as soon as the delalloc pages are 1061 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent 1062 * length to ->bytes_reserved, and subtracts the reserved space from 1063 * ->bytes_may_use. 1064 * 1065 * For #2 this is trickier. Once the ordered extent runs we will drop the 1066 * extent in the range we are overwriting, which creates a delayed ref for 1067 * that freed extent. This however is not reclaimed until the transaction 1068 * commits, thus the next stages. 1069 * 1070 * RUN_DELAYED_IPUTS 1071 * If we are freeing inodes, we want to make sure all delayed iputs have 1072 * completed, because they could have been on an inode with i_nlink == 0, and 1073 * thus have been truncated and freed up space. But again this space is not 1074 * immediately re-usable, it comes in the form of a delayed ref, which must be 1075 * run and then the transaction must be committed. 1076 * 1077 * COMMIT_TRANS 1078 * This is where we reclaim all of the pinned space generated by running the 1079 * iputs 1080 * 1081 * ALLOC_CHUNK_FORCE 1082 * For data we start with alloc chunk force, however we could have been full 1083 * before, and then the transaction commit could have freed new block groups, 1084 * so if we now have space to allocate do the force chunk allocation. 1085 */ 1086 static const enum btrfs_flush_state data_flush_states[] = { 1087 FLUSH_DELALLOC_FULL, 1088 RUN_DELAYED_IPUTS, 1089 COMMIT_TRANS, 1090 ALLOC_CHUNK_FORCE, 1091 }; 1092 1093 static void btrfs_async_reclaim_data_space(struct work_struct *work) 1094 { 1095 struct btrfs_fs_info *fs_info; 1096 struct btrfs_space_info *space_info; 1097 u64 last_tickets_id; 1098 enum btrfs_flush_state flush_state = 0; 1099 1100 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work); 1101 space_info = fs_info->data_sinfo; 1102 1103 spin_lock(&space_info->lock); 1104 if (list_empty(&space_info->tickets)) { 1105 space_info->flush = 0; 1106 spin_unlock(&space_info->lock); 1107 return; 1108 } 1109 last_tickets_id = space_info->tickets_id; 1110 spin_unlock(&space_info->lock); 1111 1112 while (!space_info->full) { 1113 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); 1114 spin_lock(&space_info->lock); 1115 if (list_empty(&space_info->tickets)) { 1116 space_info->flush = 0; 1117 spin_unlock(&space_info->lock); 1118 return; 1119 } 1120 last_tickets_id = space_info->tickets_id; 1121 spin_unlock(&space_info->lock); 1122 } 1123 1124 while (flush_state < ARRAY_SIZE(data_flush_states)) { 1125 flush_space(fs_info, space_info, U64_MAX, 1126 data_flush_states[flush_state], false); 1127 spin_lock(&space_info->lock); 1128 if (list_empty(&space_info->tickets)) { 1129 space_info->flush = 0; 1130 spin_unlock(&space_info->lock); 1131 return; 1132 } 1133 1134 if (last_tickets_id == space_info->tickets_id) { 1135 flush_state++; 1136 } else { 1137 last_tickets_id = space_info->tickets_id; 1138 flush_state = 0; 1139 } 1140 1141 if (flush_state >= ARRAY_SIZE(data_flush_states)) { 1142 if (space_info->full) { 1143 if (maybe_fail_all_tickets(fs_info, space_info)) 1144 flush_state = 0; 1145 else 1146 space_info->flush = 0; 1147 } else { 1148 flush_state = 0; 1149 } 1150 } 1151 spin_unlock(&space_info->lock); 1152 } 1153 } 1154 1155 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info) 1156 { 1157 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space); 1158 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space); 1159 INIT_WORK(&fs_info->preempt_reclaim_work, 1160 btrfs_preempt_reclaim_metadata_space); 1161 } 1162 1163 static const enum btrfs_flush_state priority_flush_states[] = { 1164 FLUSH_DELAYED_ITEMS_NR, 1165 FLUSH_DELAYED_ITEMS, 1166 ALLOC_CHUNK, 1167 }; 1168 1169 static const enum btrfs_flush_state evict_flush_states[] = { 1170 FLUSH_DELAYED_ITEMS_NR, 1171 FLUSH_DELAYED_ITEMS, 1172 FLUSH_DELAYED_REFS_NR, 1173 FLUSH_DELAYED_REFS, 1174 FLUSH_DELALLOC, 1175 FLUSH_DELALLOC_WAIT, 1176 FLUSH_DELALLOC_FULL, 1177 ALLOC_CHUNK, 1178 COMMIT_TRANS, 1179 }; 1180 1181 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info, 1182 struct btrfs_space_info *space_info, 1183 struct reserve_ticket *ticket, 1184 const enum btrfs_flush_state *states, 1185 int states_nr) 1186 { 1187 u64 to_reclaim; 1188 int flush_state; 1189 1190 spin_lock(&space_info->lock); 1191 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 1192 if (!to_reclaim) { 1193 spin_unlock(&space_info->lock); 1194 return; 1195 } 1196 spin_unlock(&space_info->lock); 1197 1198 flush_state = 0; 1199 do { 1200 flush_space(fs_info, space_info, to_reclaim, states[flush_state], 1201 false); 1202 flush_state++; 1203 spin_lock(&space_info->lock); 1204 if (ticket->bytes == 0) { 1205 spin_unlock(&space_info->lock); 1206 return; 1207 } 1208 spin_unlock(&space_info->lock); 1209 } while (flush_state < states_nr); 1210 } 1211 1212 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info, 1213 struct btrfs_space_info *space_info, 1214 struct reserve_ticket *ticket) 1215 { 1216 while (!space_info->full) { 1217 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); 1218 spin_lock(&space_info->lock); 1219 if (ticket->bytes == 0) { 1220 spin_unlock(&space_info->lock); 1221 return; 1222 } 1223 spin_unlock(&space_info->lock); 1224 } 1225 } 1226 1227 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info, 1228 struct btrfs_space_info *space_info, 1229 struct reserve_ticket *ticket) 1230 1231 { 1232 DEFINE_WAIT(wait); 1233 int ret = 0; 1234 1235 spin_lock(&space_info->lock); 1236 while (ticket->bytes > 0 && ticket->error == 0) { 1237 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE); 1238 if (ret) { 1239 /* 1240 * Delete us from the list. After we unlock the space 1241 * info, we don't want the async reclaim job to reserve 1242 * space for this ticket. If that would happen, then the 1243 * ticket's task would not known that space was reserved 1244 * despite getting an error, resulting in a space leak 1245 * (bytes_may_use counter of our space_info). 1246 */ 1247 remove_ticket(space_info, ticket); 1248 ticket->error = -EINTR; 1249 break; 1250 } 1251 spin_unlock(&space_info->lock); 1252 1253 schedule(); 1254 1255 finish_wait(&ticket->wait, &wait); 1256 spin_lock(&space_info->lock); 1257 } 1258 spin_unlock(&space_info->lock); 1259 } 1260 1261 /** 1262 * Do the appropriate flushing and waiting for a ticket 1263 * 1264 * @fs_info: the filesystem 1265 * @space_info: space info for the reservation 1266 * @ticket: ticket for the reservation 1267 * @start_ns: timestamp when the reservation started 1268 * @orig_bytes: amount of bytes originally reserved 1269 * @flush: how much we can flush 1270 * 1271 * This does the work of figuring out how to flush for the ticket, waiting for 1272 * the reservation, and returning the appropriate error if there is one. 1273 */ 1274 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info, 1275 struct btrfs_space_info *space_info, 1276 struct reserve_ticket *ticket, 1277 u64 start_ns, u64 orig_bytes, 1278 enum btrfs_reserve_flush_enum flush) 1279 { 1280 int ret; 1281 1282 switch (flush) { 1283 case BTRFS_RESERVE_FLUSH_DATA: 1284 case BTRFS_RESERVE_FLUSH_ALL: 1285 case BTRFS_RESERVE_FLUSH_ALL_STEAL: 1286 wait_reserve_ticket(fs_info, space_info, ticket); 1287 break; 1288 case BTRFS_RESERVE_FLUSH_LIMIT: 1289 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1290 priority_flush_states, 1291 ARRAY_SIZE(priority_flush_states)); 1292 break; 1293 case BTRFS_RESERVE_FLUSH_EVICT: 1294 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1295 evict_flush_states, 1296 ARRAY_SIZE(evict_flush_states)); 1297 break; 1298 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE: 1299 priority_reclaim_data_space(fs_info, space_info, ticket); 1300 break; 1301 default: 1302 ASSERT(0); 1303 break; 1304 } 1305 1306 spin_lock(&space_info->lock); 1307 ret = ticket->error; 1308 if (ticket->bytes || ticket->error) { 1309 /* 1310 * We were a priority ticket, so we need to delete ourselves 1311 * from the list. Because we could have other priority tickets 1312 * behind us that require less space, run 1313 * btrfs_try_granting_tickets() to see if their reservations can 1314 * now be made. 1315 */ 1316 if (!list_empty(&ticket->list)) { 1317 remove_ticket(space_info, ticket); 1318 btrfs_try_granting_tickets(fs_info, space_info); 1319 } 1320 1321 if (!ret) 1322 ret = -ENOSPC; 1323 } 1324 spin_unlock(&space_info->lock); 1325 ASSERT(list_empty(&ticket->list)); 1326 /* 1327 * Check that we can't have an error set if the reservation succeeded, 1328 * as that would confuse tasks and lead them to error out without 1329 * releasing reserved space (if an error happens the expectation is that 1330 * space wasn't reserved at all). 1331 */ 1332 ASSERT(!(ticket->bytes == 0 && ticket->error)); 1333 trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes, 1334 start_ns, flush, ticket->error); 1335 return ret; 1336 } 1337 1338 /* 1339 * This returns true if this flush state will go through the ordinary flushing 1340 * code. 1341 */ 1342 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush) 1343 { 1344 return (flush == BTRFS_RESERVE_FLUSH_ALL) || 1345 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL); 1346 } 1347 1348 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info, 1349 struct btrfs_space_info *space_info) 1350 { 1351 u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes); 1352 u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes); 1353 1354 /* 1355 * If we're heavy on ordered operations then clamping won't help us. We 1356 * need to clamp specifically to keep up with dirty'ing buffered 1357 * writers, because there's not a 1:1 correlation of writing delalloc 1358 * and freeing space, like there is with flushing delayed refs or 1359 * delayed nodes. If we're already more ordered than delalloc then 1360 * we're keeping up, otherwise we aren't and should probably clamp. 1361 */ 1362 if (ordered < delalloc) 1363 space_info->clamp = min(space_info->clamp + 1, 8); 1364 } 1365 1366 /** 1367 * Try to reserve bytes from the block_rsv's space 1368 * 1369 * @fs_info: the filesystem 1370 * @space_info: space info we want to allocate from 1371 * @orig_bytes: number of bytes we want 1372 * @flush: whether or not we can flush to make our reservation 1373 * 1374 * This will reserve orig_bytes number of bytes from the space info associated 1375 * with the block_rsv. If there is not enough space it will make an attempt to 1376 * flush out space to make room. It will do this by flushing delalloc if 1377 * possible or committing the transaction. If flush is 0 then no attempts to 1378 * regain reservations will be made and this will fail if there is not enough 1379 * space already. 1380 */ 1381 static int __reserve_bytes(struct btrfs_fs_info *fs_info, 1382 struct btrfs_space_info *space_info, u64 orig_bytes, 1383 enum btrfs_reserve_flush_enum flush) 1384 { 1385 struct work_struct *async_work; 1386 struct reserve_ticket ticket; 1387 u64 start_ns = 0; 1388 u64 used; 1389 int ret = 0; 1390 bool pending_tickets; 1391 1392 ASSERT(orig_bytes); 1393 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL); 1394 1395 if (flush == BTRFS_RESERVE_FLUSH_DATA) 1396 async_work = &fs_info->async_data_reclaim_work; 1397 else 1398 async_work = &fs_info->async_reclaim_work; 1399 1400 spin_lock(&space_info->lock); 1401 ret = -ENOSPC; 1402 used = btrfs_space_info_used(space_info, true); 1403 1404 /* 1405 * We don't want NO_FLUSH allocations to jump everybody, they can 1406 * generally handle ENOSPC in a different way, so treat them the same as 1407 * normal flushers when it comes to skipping pending tickets. 1408 */ 1409 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH)) 1410 pending_tickets = !list_empty(&space_info->tickets) || 1411 !list_empty(&space_info->priority_tickets); 1412 else 1413 pending_tickets = !list_empty(&space_info->priority_tickets); 1414 1415 /* 1416 * Carry on if we have enough space (short-circuit) OR call 1417 * can_overcommit() to ensure we can overcommit to continue. 1418 */ 1419 if (!pending_tickets && 1420 ((used + orig_bytes <= space_info->total_bytes) || 1421 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) { 1422 btrfs_space_info_update_bytes_may_use(fs_info, space_info, 1423 orig_bytes); 1424 ret = 0; 1425 } 1426 1427 /* 1428 * If we couldn't make a reservation then setup our reservation ticket 1429 * and kick the async worker if it's not already running. 1430 * 1431 * If we are a priority flusher then we just need to add our ticket to 1432 * the list and we will do our own flushing further down. 1433 */ 1434 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) { 1435 ticket.bytes = orig_bytes; 1436 ticket.error = 0; 1437 space_info->reclaim_size += ticket.bytes; 1438 init_waitqueue_head(&ticket.wait); 1439 ticket.steal = (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL); 1440 if (trace_btrfs_reserve_ticket_enabled()) 1441 start_ns = ktime_get_ns(); 1442 1443 if (flush == BTRFS_RESERVE_FLUSH_ALL || 1444 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || 1445 flush == BTRFS_RESERVE_FLUSH_DATA) { 1446 list_add_tail(&ticket.list, &space_info->tickets); 1447 if (!space_info->flush) { 1448 /* 1449 * We were forced to add a reserve ticket, so 1450 * our preemptive flushing is unable to keep 1451 * up. Clamp down on the threshold for the 1452 * preemptive flushing in order to keep up with 1453 * the workload. 1454 */ 1455 maybe_clamp_preempt(fs_info, space_info); 1456 1457 space_info->flush = 1; 1458 trace_btrfs_trigger_flush(fs_info, 1459 space_info->flags, 1460 orig_bytes, flush, 1461 "enospc"); 1462 queue_work(system_unbound_wq, async_work); 1463 } 1464 } else { 1465 list_add_tail(&ticket.list, 1466 &space_info->priority_tickets); 1467 } 1468 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) { 1469 used += orig_bytes; 1470 /* 1471 * We will do the space reservation dance during log replay, 1472 * which means we won't have fs_info->fs_root set, so don't do 1473 * the async reclaim as we will panic. 1474 */ 1475 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) && 1476 !work_busy(&fs_info->preempt_reclaim_work) && 1477 need_preemptive_reclaim(fs_info, space_info)) { 1478 trace_btrfs_trigger_flush(fs_info, space_info->flags, 1479 orig_bytes, flush, "preempt"); 1480 queue_work(system_unbound_wq, 1481 &fs_info->preempt_reclaim_work); 1482 } 1483 } 1484 spin_unlock(&space_info->lock); 1485 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH) 1486 return ret; 1487 1488 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns, 1489 orig_bytes, flush); 1490 } 1491 1492 /** 1493 * Trye to reserve metadata bytes from the block_rsv's space 1494 * 1495 * @root: the root we're allocating for 1496 * @block_rsv: block_rsv we're allocating for 1497 * @orig_bytes: number of bytes we want 1498 * @flush: whether or not we can flush to make our reservation 1499 * 1500 * This will reserve orig_bytes number of bytes from the space info associated 1501 * with the block_rsv. If there is not enough space it will make an attempt to 1502 * flush out space to make room. It will do this by flushing delalloc if 1503 * possible or committing the transaction. If flush is 0 then no attempts to 1504 * regain reservations will be made and this will fail if there is not enough 1505 * space already. 1506 */ 1507 int btrfs_reserve_metadata_bytes(struct btrfs_root *root, 1508 struct btrfs_block_rsv *block_rsv, 1509 u64 orig_bytes, 1510 enum btrfs_reserve_flush_enum flush) 1511 { 1512 struct btrfs_fs_info *fs_info = root->fs_info; 1513 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 1514 int ret; 1515 1516 ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush); 1517 if (ret == -ENOSPC && 1518 unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) { 1519 if (block_rsv != global_rsv && 1520 !btrfs_block_rsv_use_bytes(global_rsv, orig_bytes)) 1521 ret = 0; 1522 } 1523 if (ret == -ENOSPC) { 1524 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1525 block_rsv->space_info->flags, 1526 orig_bytes, 1); 1527 1528 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1529 btrfs_dump_space_info(fs_info, block_rsv->space_info, 1530 orig_bytes, 0); 1531 } 1532 return ret; 1533 } 1534 1535 /** 1536 * Try to reserve data bytes for an allocation 1537 * 1538 * @fs_info: the filesystem 1539 * @bytes: number of bytes we need 1540 * @flush: how we are allowed to flush 1541 * 1542 * This will reserve bytes from the data space info. If there is not enough 1543 * space then we will attempt to flush space as specified by flush. 1544 */ 1545 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes, 1546 enum btrfs_reserve_flush_enum flush) 1547 { 1548 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo; 1549 int ret; 1550 1551 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA || 1552 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE); 1553 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA); 1554 1555 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush); 1556 if (ret == -ENOSPC) { 1557 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1558 data_sinfo->flags, bytes, 1); 1559 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1560 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0); 1561 } 1562 return ret; 1563 } 1564