1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * blk-mq scheduling framework 4 * 5 * Copyright (C) 2016 Jens Axboe 6 */ 7 #include <linux/kernel.h> 8 #include <linux/module.h> 9 #include <linux/blk-mq.h> 10 #include <linux/list_sort.h> 11 12 #include <trace/events/block.h> 13 14 #include "blk.h" 15 #include "blk-mq.h" 16 #include "blk-mq-debugfs.h" 17 #include "blk-mq-sched.h" 18 #include "blk-mq-tag.h" 19 #include "blk-wbt.h" 20 21 void blk_mq_sched_assign_ioc(struct request *rq) 22 { 23 struct request_queue *q = rq->q; 24 struct io_context *ioc; 25 struct io_cq *icq; 26 27 /* 28 * May not have an IO context if it's a passthrough request 29 */ 30 ioc = current->io_context; 31 if (!ioc) 32 return; 33 34 spin_lock_irq(&q->queue_lock); 35 icq = ioc_lookup_icq(ioc, q); 36 spin_unlock_irq(&q->queue_lock); 37 38 if (!icq) { 39 icq = ioc_create_icq(ioc, q, GFP_ATOMIC); 40 if (!icq) 41 return; 42 } 43 get_io_context(icq->ioc); 44 rq->elv.icq = icq; 45 } 46 47 /* 48 * Mark a hardware queue as needing a restart. For shared queues, maintain 49 * a count of how many hardware queues are marked for restart. 50 */ 51 void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx) 52 { 53 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) 54 return; 55 56 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); 57 } 58 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx); 59 60 void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx) 61 { 62 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) 63 return; 64 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); 65 66 /* 67 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch) 68 * in blk_mq_run_hw_queue(). Its pair is the barrier in 69 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART, 70 * meantime new request added to hctx->dispatch is missed to check in 71 * blk_mq_run_hw_queue(). 72 */ 73 smp_mb(); 74 75 blk_mq_run_hw_queue(hctx, true); 76 } 77 78 static int sched_rq_cmp(void *priv, struct list_head *a, struct list_head *b) 79 { 80 struct request *rqa = container_of(a, struct request, queuelist); 81 struct request *rqb = container_of(b, struct request, queuelist); 82 83 return rqa->mq_hctx > rqb->mq_hctx; 84 } 85 86 static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list) 87 { 88 struct blk_mq_hw_ctx *hctx = 89 list_first_entry(rq_list, struct request, queuelist)->mq_hctx; 90 struct request *rq; 91 LIST_HEAD(hctx_list); 92 unsigned int count = 0; 93 94 list_for_each_entry(rq, rq_list, queuelist) { 95 if (rq->mq_hctx != hctx) { 96 list_cut_before(&hctx_list, rq_list, &rq->queuelist); 97 goto dispatch; 98 } 99 count++; 100 } 101 list_splice_tail_init(rq_list, &hctx_list); 102 103 dispatch: 104 return blk_mq_dispatch_rq_list(hctx, &hctx_list, count); 105 } 106 107 #define BLK_MQ_BUDGET_DELAY 3 /* ms units */ 108 109 /* 110 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts 111 * its queue by itself in its completion handler, so we don't need to 112 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE. 113 * 114 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to 115 * be run again. This is necessary to avoid starving flushes. 116 */ 117 static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) 118 { 119 struct request_queue *q = hctx->queue; 120 struct elevator_queue *e = q->elevator; 121 bool multi_hctxs = false, run_queue = false; 122 bool dispatched = false, busy = false; 123 unsigned int max_dispatch; 124 LIST_HEAD(rq_list); 125 int count = 0; 126 127 if (hctx->dispatch_busy) 128 max_dispatch = 1; 129 else 130 max_dispatch = hctx->queue->nr_requests; 131 132 do { 133 struct request *rq; 134 135 if (e->type->ops.has_work && !e->type->ops.has_work(hctx)) 136 break; 137 138 if (!list_empty_careful(&hctx->dispatch)) { 139 busy = true; 140 break; 141 } 142 143 if (!blk_mq_get_dispatch_budget(q)) 144 break; 145 146 rq = e->type->ops.dispatch_request(hctx); 147 if (!rq) { 148 blk_mq_put_dispatch_budget(q); 149 /* 150 * We're releasing without dispatching. Holding the 151 * budget could have blocked any "hctx"s with the 152 * same queue and if we didn't dispatch then there's 153 * no guarantee anyone will kick the queue. Kick it 154 * ourselves. 155 */ 156 run_queue = true; 157 break; 158 } 159 160 /* 161 * Now this rq owns the budget which has to be released 162 * if this rq won't be queued to driver via .queue_rq() 163 * in blk_mq_dispatch_rq_list(). 164 */ 165 list_add_tail(&rq->queuelist, &rq_list); 166 if (rq->mq_hctx != hctx) 167 multi_hctxs = true; 168 } while (++count < max_dispatch); 169 170 if (!count) { 171 if (run_queue) 172 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); 173 } else if (multi_hctxs) { 174 /* 175 * Requests from different hctx may be dequeued from some 176 * schedulers, such as bfq and deadline. 177 * 178 * Sort the requests in the list according to their hctx, 179 * dispatch batching requests from same hctx at a time. 180 */ 181 list_sort(NULL, &rq_list, sched_rq_cmp); 182 do { 183 dispatched |= blk_mq_dispatch_hctx_list(&rq_list); 184 } while (!list_empty(&rq_list)); 185 } else { 186 dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count); 187 } 188 189 if (busy) 190 return -EAGAIN; 191 return !!dispatched; 192 } 193 194 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) 195 { 196 int ret; 197 198 do { 199 ret = __blk_mq_do_dispatch_sched(hctx); 200 } while (ret == 1); 201 202 return ret; 203 } 204 205 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx, 206 struct blk_mq_ctx *ctx) 207 { 208 unsigned short idx = ctx->index_hw[hctx->type]; 209 210 if (++idx == hctx->nr_ctx) 211 idx = 0; 212 213 return hctx->ctxs[idx]; 214 } 215 216 /* 217 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts 218 * its queue by itself in its completion handler, so we don't need to 219 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE. 220 * 221 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to 222 * be run again. This is necessary to avoid starving flushes. 223 */ 224 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx) 225 { 226 struct request_queue *q = hctx->queue; 227 LIST_HEAD(rq_list); 228 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from); 229 int ret = 0; 230 struct request *rq; 231 232 do { 233 if (!list_empty_careful(&hctx->dispatch)) { 234 ret = -EAGAIN; 235 break; 236 } 237 238 if (!sbitmap_any_bit_set(&hctx->ctx_map)) 239 break; 240 241 if (!blk_mq_get_dispatch_budget(q)) 242 break; 243 244 rq = blk_mq_dequeue_from_ctx(hctx, ctx); 245 if (!rq) { 246 blk_mq_put_dispatch_budget(q); 247 /* 248 * We're releasing without dispatching. Holding the 249 * budget could have blocked any "hctx"s with the 250 * same queue and if we didn't dispatch then there's 251 * no guarantee anyone will kick the queue. Kick it 252 * ourselves. 253 */ 254 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); 255 break; 256 } 257 258 /* 259 * Now this rq owns the budget which has to be released 260 * if this rq won't be queued to driver via .queue_rq() 261 * in blk_mq_dispatch_rq_list(). 262 */ 263 list_add(&rq->queuelist, &rq_list); 264 265 /* round robin for fair dispatch */ 266 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx); 267 268 } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1)); 269 270 WRITE_ONCE(hctx->dispatch_from, ctx); 271 return ret; 272 } 273 274 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) 275 { 276 struct request_queue *q = hctx->queue; 277 struct elevator_queue *e = q->elevator; 278 const bool has_sched_dispatch = e && e->type->ops.dispatch_request; 279 int ret = 0; 280 LIST_HEAD(rq_list); 281 282 /* 283 * If we have previous entries on our dispatch list, grab them first for 284 * more fair dispatch. 285 */ 286 if (!list_empty_careful(&hctx->dispatch)) { 287 spin_lock(&hctx->lock); 288 if (!list_empty(&hctx->dispatch)) 289 list_splice_init(&hctx->dispatch, &rq_list); 290 spin_unlock(&hctx->lock); 291 } 292 293 /* 294 * Only ask the scheduler for requests, if we didn't have residual 295 * requests from the dispatch list. This is to avoid the case where 296 * we only ever dispatch a fraction of the requests available because 297 * of low device queue depth. Once we pull requests out of the IO 298 * scheduler, we can no longer merge or sort them. So it's best to 299 * leave them there for as long as we can. Mark the hw queue as 300 * needing a restart in that case. 301 * 302 * We want to dispatch from the scheduler if there was nothing 303 * on the dispatch list or we were able to dispatch from the 304 * dispatch list. 305 */ 306 if (!list_empty(&rq_list)) { 307 blk_mq_sched_mark_restart_hctx(hctx); 308 if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) { 309 if (has_sched_dispatch) 310 ret = blk_mq_do_dispatch_sched(hctx); 311 else 312 ret = blk_mq_do_dispatch_ctx(hctx); 313 } 314 } else if (has_sched_dispatch) { 315 ret = blk_mq_do_dispatch_sched(hctx); 316 } else if (hctx->dispatch_busy) { 317 /* dequeue request one by one from sw queue if queue is busy */ 318 ret = blk_mq_do_dispatch_ctx(hctx); 319 } else { 320 blk_mq_flush_busy_ctxs(hctx, &rq_list); 321 blk_mq_dispatch_rq_list(hctx, &rq_list, 0); 322 } 323 324 return ret; 325 } 326 327 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) 328 { 329 struct request_queue *q = hctx->queue; 330 331 /* RCU or SRCU read lock is needed before checking quiesced flag */ 332 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) 333 return; 334 335 hctx->run++; 336 337 /* 338 * A return of -EAGAIN is an indication that hctx->dispatch is not 339 * empty and we must run again in order to avoid starving flushes. 340 */ 341 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) { 342 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) 343 blk_mq_run_hw_queue(hctx, true); 344 } 345 } 346 347 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio, 348 unsigned int nr_segs) 349 { 350 struct elevator_queue *e = q->elevator; 351 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); 352 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx); 353 bool ret = false; 354 enum hctx_type type; 355 356 if (e && e->type->ops.bio_merge) 357 return e->type->ops.bio_merge(hctx, bio, nr_segs); 358 359 type = hctx->type; 360 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) || 361 list_empty_careful(&ctx->rq_lists[type])) 362 return false; 363 364 /* default per sw-queue merge */ 365 spin_lock(&ctx->lock); 366 /* 367 * Reverse check our software queue for entries that we could 368 * potentially merge with. Currently includes a hand-wavy stop 369 * count of 8, to not spend too much time checking for merges. 370 */ 371 if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) { 372 ctx->rq_merged++; 373 ret = true; 374 } 375 376 spin_unlock(&ctx->lock); 377 378 return ret; 379 } 380 381 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq) 382 { 383 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq); 384 } 385 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge); 386 387 void blk_mq_sched_request_inserted(struct request *rq) 388 { 389 trace_block_rq_insert(rq); 390 } 391 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted); 392 393 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx, 394 bool has_sched, 395 struct request *rq) 396 { 397 /* 398 * dispatch flush and passthrough rq directly 399 * 400 * passthrough request has to be added to hctx->dispatch directly. 401 * For some reason, device may be in one situation which can't 402 * handle FS request, so STS_RESOURCE is always returned and the 403 * FS request will be added to hctx->dispatch. However passthrough 404 * request may be required at that time for fixing the problem. If 405 * passthrough request is added to scheduler queue, there isn't any 406 * chance to dispatch it given we prioritize requests in hctx->dispatch. 407 */ 408 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq)) 409 return true; 410 411 if (has_sched) 412 rq->rq_flags |= RQF_SORTED; 413 414 return false; 415 } 416 417 void blk_mq_sched_insert_request(struct request *rq, bool at_head, 418 bool run_queue, bool async) 419 { 420 struct request_queue *q = rq->q; 421 struct elevator_queue *e = q->elevator; 422 struct blk_mq_ctx *ctx = rq->mq_ctx; 423 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 424 425 WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG)); 426 427 if (blk_mq_sched_bypass_insert(hctx, !!e, rq)) { 428 /* 429 * Firstly normal IO request is inserted to scheduler queue or 430 * sw queue, meantime we add flush request to dispatch queue( 431 * hctx->dispatch) directly and there is at most one in-flight 432 * flush request for each hw queue, so it doesn't matter to add 433 * flush request to tail or front of the dispatch queue. 434 * 435 * Secondly in case of NCQ, flush request belongs to non-NCQ 436 * command, and queueing it will fail when there is any 437 * in-flight normal IO request(NCQ command). When adding flush 438 * rq to the front of hctx->dispatch, it is easier to introduce 439 * extra time to flush rq's latency because of S_SCHED_RESTART 440 * compared with adding to the tail of dispatch queue, then 441 * chance of flush merge is increased, and less flush requests 442 * will be issued to controller. It is observed that ~10% time 443 * is saved in blktests block/004 on disk attached to AHCI/NCQ 444 * drive when adding flush rq to the front of hctx->dispatch. 445 * 446 * Simply queue flush rq to the front of hctx->dispatch so that 447 * intensive flush workloads can benefit in case of NCQ HW. 448 */ 449 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head; 450 blk_mq_request_bypass_insert(rq, at_head, false); 451 goto run; 452 } 453 454 if (e && e->type->ops.insert_requests) { 455 LIST_HEAD(list); 456 457 list_add(&rq->queuelist, &list); 458 e->type->ops.insert_requests(hctx, &list, at_head); 459 } else { 460 spin_lock(&ctx->lock); 461 __blk_mq_insert_request(hctx, rq, at_head); 462 spin_unlock(&ctx->lock); 463 } 464 465 run: 466 if (run_queue) 467 blk_mq_run_hw_queue(hctx, async); 468 } 469 470 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx, 471 struct blk_mq_ctx *ctx, 472 struct list_head *list, bool run_queue_async) 473 { 474 struct elevator_queue *e; 475 struct request_queue *q = hctx->queue; 476 477 /* 478 * blk_mq_sched_insert_requests() is called from flush plug 479 * context only, and hold one usage counter to prevent queue 480 * from being released. 481 */ 482 percpu_ref_get(&q->q_usage_counter); 483 484 e = hctx->queue->elevator; 485 if (e && e->type->ops.insert_requests) 486 e->type->ops.insert_requests(hctx, list, false); 487 else { 488 /* 489 * try to issue requests directly if the hw queue isn't 490 * busy in case of 'none' scheduler, and this way may save 491 * us one extra enqueue & dequeue to sw queue. 492 */ 493 if (!hctx->dispatch_busy && !e && !run_queue_async) { 494 blk_mq_try_issue_list_directly(hctx, list); 495 if (list_empty(list)) 496 goto out; 497 } 498 blk_mq_insert_requests(hctx, ctx, list); 499 } 500 501 blk_mq_run_hw_queue(hctx, run_queue_async); 502 out: 503 percpu_ref_put(&q->q_usage_counter); 504 } 505 506 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set, 507 struct blk_mq_hw_ctx *hctx, 508 unsigned int hctx_idx) 509 { 510 unsigned int flags = set->flags & ~BLK_MQ_F_TAG_HCTX_SHARED; 511 512 if (hctx->sched_tags) { 513 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx); 514 blk_mq_free_rq_map(hctx->sched_tags, flags); 515 hctx->sched_tags = NULL; 516 } 517 } 518 519 static int blk_mq_sched_alloc_tags(struct request_queue *q, 520 struct blk_mq_hw_ctx *hctx, 521 unsigned int hctx_idx) 522 { 523 struct blk_mq_tag_set *set = q->tag_set; 524 /* Clear HCTX_SHARED so tags are init'ed */ 525 unsigned int flags = set->flags & ~BLK_MQ_F_TAG_HCTX_SHARED; 526 int ret; 527 528 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests, 529 set->reserved_tags, flags); 530 if (!hctx->sched_tags) 531 return -ENOMEM; 532 533 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests); 534 if (ret) 535 blk_mq_sched_free_tags(set, hctx, hctx_idx); 536 537 return ret; 538 } 539 540 /* called in queue's release handler, tagset has gone away */ 541 static void blk_mq_sched_tags_teardown(struct request_queue *q) 542 { 543 struct blk_mq_hw_ctx *hctx; 544 int i; 545 546 queue_for_each_hw_ctx(q, hctx, i) { 547 /* Clear HCTX_SHARED so tags are freed */ 548 unsigned int flags = hctx->flags & ~BLK_MQ_F_TAG_HCTX_SHARED; 549 550 if (hctx->sched_tags) { 551 blk_mq_free_rq_map(hctx->sched_tags, flags); 552 hctx->sched_tags = NULL; 553 } 554 } 555 } 556 557 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e) 558 { 559 struct blk_mq_hw_ctx *hctx; 560 struct elevator_queue *eq; 561 unsigned int i; 562 int ret; 563 564 if (!e) { 565 q->elevator = NULL; 566 q->nr_requests = q->tag_set->queue_depth; 567 return 0; 568 } 569 570 /* 571 * Default to double of smaller one between hw queue_depth and 128, 572 * since we don't split into sync/async like the old code did. 573 * Additionally, this is a per-hw queue depth. 574 */ 575 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth, 576 BLKDEV_MAX_RQ); 577 578 queue_for_each_hw_ctx(q, hctx, i) { 579 ret = blk_mq_sched_alloc_tags(q, hctx, i); 580 if (ret) 581 goto err; 582 } 583 584 ret = e->ops.init_sched(q, e); 585 if (ret) 586 goto err; 587 588 blk_mq_debugfs_register_sched(q); 589 590 queue_for_each_hw_ctx(q, hctx, i) { 591 if (e->ops.init_hctx) { 592 ret = e->ops.init_hctx(hctx, i); 593 if (ret) { 594 eq = q->elevator; 595 blk_mq_sched_free_requests(q); 596 blk_mq_exit_sched(q, eq); 597 kobject_put(&eq->kobj); 598 return ret; 599 } 600 } 601 blk_mq_debugfs_register_sched_hctx(q, hctx); 602 } 603 604 return 0; 605 606 err: 607 blk_mq_sched_free_requests(q); 608 blk_mq_sched_tags_teardown(q); 609 q->elevator = NULL; 610 return ret; 611 } 612 613 /* 614 * called in either blk_queue_cleanup or elevator_switch, tagset 615 * is required for freeing requests 616 */ 617 void blk_mq_sched_free_requests(struct request_queue *q) 618 { 619 struct blk_mq_hw_ctx *hctx; 620 int i; 621 622 queue_for_each_hw_ctx(q, hctx, i) { 623 if (hctx->sched_tags) 624 blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i); 625 } 626 } 627 628 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e) 629 { 630 struct blk_mq_hw_ctx *hctx; 631 unsigned int i; 632 633 queue_for_each_hw_ctx(q, hctx, i) { 634 blk_mq_debugfs_unregister_sched_hctx(hctx); 635 if (e->type->ops.exit_hctx && hctx->sched_data) { 636 e->type->ops.exit_hctx(hctx, i); 637 hctx->sched_data = NULL; 638 } 639 } 640 blk_mq_debugfs_unregister_sched(q); 641 if (e->type->ops.exit_sched) 642 e->type->ops.exit_sched(e); 643 blk_mq_sched_tags_teardown(q); 644 q->elevator = NULL; 645 } 646