13dcf60bcSChristoph Hellwig // SPDX-License-Identifier: GPL-2.0 2bd166ef1SJens Axboe /* 3bd166ef1SJens Axboe * blk-mq scheduling framework 4bd166ef1SJens Axboe * 5bd166ef1SJens Axboe * Copyright (C) 2016 Jens Axboe 6bd166ef1SJens Axboe */ 7bd166ef1SJens Axboe #include <linux/kernel.h> 8bd166ef1SJens Axboe #include <linux/module.h> 9bd166ef1SJens Axboe #include <linux/blk-mq.h> 106e6fcbc2SMing Lei #include <linux/list_sort.h> 11bd166ef1SJens Axboe 12bd166ef1SJens Axboe #include <trace/events/block.h> 13bd166ef1SJens Axboe 14bd166ef1SJens Axboe #include "blk.h" 15bd166ef1SJens Axboe #include "blk-mq.h" 16d332ce09SOmar Sandoval #include "blk-mq-debugfs.h" 17bd166ef1SJens Axboe #include "blk-mq-sched.h" 18bd166ef1SJens Axboe #include "blk-mq-tag.h" 19bd166ef1SJens Axboe #include "blk-wbt.h" 20bd166ef1SJens Axboe 21e2b3fa5aSDamien Le Moal void blk_mq_sched_assign_ioc(struct request *rq) 22bd166ef1SJens Axboe { 2344e8c2bfSChristoph Hellwig struct request_queue *q = rq->q; 240c62bff1SJens Axboe struct io_context *ioc; 25bd166ef1SJens Axboe struct io_cq *icq; 26bd166ef1SJens Axboe 270c62bff1SJens Axboe /* 280c62bff1SJens Axboe * May not have an IO context if it's a passthrough request 290c62bff1SJens Axboe */ 300c62bff1SJens Axboe ioc = current->io_context; 310c62bff1SJens Axboe if (!ioc) 320c62bff1SJens Axboe return; 330c62bff1SJens Axboe 340d945c1fSChristoph Hellwig spin_lock_irq(&q->queue_lock); 35bd166ef1SJens Axboe icq = ioc_lookup_icq(ioc, q); 360d945c1fSChristoph Hellwig spin_unlock_irq(&q->queue_lock); 37bd166ef1SJens Axboe 38bd166ef1SJens Axboe if (!icq) { 39bd166ef1SJens Axboe icq = ioc_create_icq(ioc, q, GFP_ATOMIC); 40bd166ef1SJens Axboe if (!icq) 41bd166ef1SJens Axboe return; 42bd166ef1SJens Axboe } 43ea511e3cSChristoph Hellwig get_io_context(icq->ioc); 4444e8c2bfSChristoph Hellwig rq->elv.icq = icq; 45bd166ef1SJens Axboe } 46bd166ef1SJens Axboe 478e8320c9SJens Axboe /* 488e8320c9SJens Axboe * Mark a hardware queue as needing a restart. For shared queues, maintain 498e8320c9SJens Axboe * a count of how many hardware queues are marked for restart. 508e8320c9SJens Axboe */ 517211aef8SDamien Le Moal void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx) 528e8320c9SJens Axboe { 538e8320c9SJens Axboe if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) 548e8320c9SJens Axboe return; 558e8320c9SJens Axboe 568e8320c9SJens Axboe set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); 578e8320c9SJens Axboe } 587211aef8SDamien Le Moal EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx); 598e8320c9SJens Axboe 6097889f9aSMing Lei void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx) 618e8320c9SJens Axboe { 628e8320c9SJens Axboe if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) 6397889f9aSMing Lei return; 648e8320c9SJens Axboe clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); 658e8320c9SJens Axboe 66d7d8535fSMing Lei /* 67d7d8535fSMing Lei * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch) 68d7d8535fSMing Lei * in blk_mq_run_hw_queue(). Its pair is the barrier in 69d7d8535fSMing Lei * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART, 70d7d8535fSMing Lei * meantime new request added to hctx->dispatch is missed to check in 71d7d8535fSMing Lei * blk_mq_run_hw_queue(). 72d7d8535fSMing Lei */ 73d7d8535fSMing Lei smp_mb(); 74d7d8535fSMing Lei 7597889f9aSMing Lei blk_mq_run_hw_queue(hctx, true); 768e8320c9SJens Axboe } 778e8320c9SJens Axboe 786e6fcbc2SMing Lei static int sched_rq_cmp(void *priv, struct list_head *a, struct list_head *b) 796e6fcbc2SMing Lei { 806e6fcbc2SMing Lei struct request *rqa = container_of(a, struct request, queuelist); 816e6fcbc2SMing Lei struct request *rqb = container_of(b, struct request, queuelist); 826e6fcbc2SMing Lei 836e6fcbc2SMing Lei return rqa->mq_hctx > rqb->mq_hctx; 846e6fcbc2SMing Lei } 856e6fcbc2SMing Lei 866e6fcbc2SMing Lei static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list) 876e6fcbc2SMing Lei { 886e6fcbc2SMing Lei struct blk_mq_hw_ctx *hctx = 896e6fcbc2SMing Lei list_first_entry(rq_list, struct request, queuelist)->mq_hctx; 906e6fcbc2SMing Lei struct request *rq; 916e6fcbc2SMing Lei LIST_HEAD(hctx_list); 926e6fcbc2SMing Lei unsigned int count = 0; 936e6fcbc2SMing Lei 946e6fcbc2SMing Lei list_for_each_entry(rq, rq_list, queuelist) { 956e6fcbc2SMing Lei if (rq->mq_hctx != hctx) { 966e6fcbc2SMing Lei list_cut_before(&hctx_list, rq_list, &rq->queuelist); 976e6fcbc2SMing Lei goto dispatch; 986e6fcbc2SMing Lei } 996e6fcbc2SMing Lei count++; 1006e6fcbc2SMing Lei } 1016e6fcbc2SMing Lei list_splice_tail_init(rq_list, &hctx_list); 1026e6fcbc2SMing Lei 1036e6fcbc2SMing Lei dispatch: 104106e71c5SBaolin Wang return blk_mq_dispatch_rq_list(hctx, &hctx_list, count); 1056e6fcbc2SMing Lei } 1066e6fcbc2SMing Lei 107a0823421SDouglas Anderson #define BLK_MQ_BUDGET_DELAY 3 /* ms units */ 108a0823421SDouglas Anderson 1091f460b63SMing Lei /* 1101f460b63SMing Lei * Only SCSI implements .get_budget and .put_budget, and SCSI restarts 1111f460b63SMing Lei * its queue by itself in its completion handler, so we don't need to 1121f460b63SMing Lei * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE. 11328d65729SSalman Qazi * 11428d65729SSalman Qazi * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to 11528d65729SSalman Qazi * be run again. This is necessary to avoid starving flushes. 1161f460b63SMing Lei */ 1176e6fcbc2SMing Lei static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) 118caf8eb0dSMing Lei { 119caf8eb0dSMing Lei struct request_queue *q = hctx->queue; 120caf8eb0dSMing Lei struct elevator_queue *e = q->elevator; 1216e6fcbc2SMing Lei bool multi_hctxs = false, run_queue = false; 1226e6fcbc2SMing Lei bool dispatched = false, busy = false; 1236e6fcbc2SMing Lei unsigned int max_dispatch; 124caf8eb0dSMing Lei LIST_HEAD(rq_list); 1256e6fcbc2SMing Lei int count = 0; 1266e6fcbc2SMing Lei 1276e6fcbc2SMing Lei if (hctx->dispatch_busy) 1286e6fcbc2SMing Lei max_dispatch = 1; 1296e6fcbc2SMing Lei else 1306e6fcbc2SMing Lei max_dispatch = hctx->queue->nr_requests; 131caf8eb0dSMing Lei 132445874e8SMing Lei do { 1336e6fcbc2SMing Lei struct request *rq; 134*2a5a24aaSMing Lei int budget_token; 1356e6fcbc2SMing Lei 136f9cd4bfeSJens Axboe if (e->type->ops.has_work && !e->type->ops.has_work(hctx)) 137caf8eb0dSMing Lei break; 138de148297SMing Lei 13928d65729SSalman Qazi if (!list_empty_careful(&hctx->dispatch)) { 1406e6fcbc2SMing Lei busy = true; 14128d65729SSalman Qazi break; 14228d65729SSalman Qazi } 14328d65729SSalman Qazi 144*2a5a24aaSMing Lei budget_token = blk_mq_get_dispatch_budget(q); 145*2a5a24aaSMing Lei if (budget_token < 0) 1461f460b63SMing Lei break; 147de148297SMing Lei 148f9cd4bfeSJens Axboe rq = e->type->ops.dispatch_request(hctx); 149de148297SMing Lei if (!rq) { 150*2a5a24aaSMing Lei blk_mq_put_dispatch_budget(q, budget_token); 151a0823421SDouglas Anderson /* 152a0823421SDouglas Anderson * We're releasing without dispatching. Holding the 153a0823421SDouglas Anderson * budget could have blocked any "hctx"s with the 154a0823421SDouglas Anderson * same queue and if we didn't dispatch then there's 155a0823421SDouglas Anderson * no guarantee anyone will kick the queue. Kick it 156a0823421SDouglas Anderson * ourselves. 157a0823421SDouglas Anderson */ 1586e6fcbc2SMing Lei run_queue = true; 159de148297SMing Lei break; 160caf8eb0dSMing Lei } 161caf8eb0dSMing Lei 162*2a5a24aaSMing Lei blk_mq_set_rq_budget_token(rq, budget_token); 163*2a5a24aaSMing Lei 164de148297SMing Lei /* 165de148297SMing Lei * Now this rq owns the budget which has to be released 166de148297SMing Lei * if this rq won't be queued to driver via .queue_rq() 167de148297SMing Lei * in blk_mq_dispatch_rq_list(). 168de148297SMing Lei */ 1696e6fcbc2SMing Lei list_add_tail(&rq->queuelist, &rq_list); 1706e6fcbc2SMing Lei if (rq->mq_hctx != hctx) 1716e6fcbc2SMing Lei multi_hctxs = true; 1726e6fcbc2SMing Lei } while (++count < max_dispatch); 1736e6fcbc2SMing Lei 1746e6fcbc2SMing Lei if (!count) { 1756e6fcbc2SMing Lei if (run_queue) 1766e6fcbc2SMing Lei blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); 1776e6fcbc2SMing Lei } else if (multi_hctxs) { 1786e6fcbc2SMing Lei /* 1796e6fcbc2SMing Lei * Requests from different hctx may be dequeued from some 1806e6fcbc2SMing Lei * schedulers, such as bfq and deadline. 1816e6fcbc2SMing Lei * 1826e6fcbc2SMing Lei * Sort the requests in the list according to their hctx, 1836e6fcbc2SMing Lei * dispatch batching requests from same hctx at a time. 1846e6fcbc2SMing Lei */ 1856e6fcbc2SMing Lei list_sort(NULL, &rq_list, sched_rq_cmp); 1866e6fcbc2SMing Lei do { 1876e6fcbc2SMing Lei dispatched |= blk_mq_dispatch_hctx_list(&rq_list); 1886e6fcbc2SMing Lei } while (!list_empty(&rq_list)); 1896e6fcbc2SMing Lei } else { 1906e6fcbc2SMing Lei dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count); 1916e6fcbc2SMing Lei } 1926e6fcbc2SMing Lei 1936e6fcbc2SMing Lei if (busy) 1946e6fcbc2SMing Lei return -EAGAIN; 1956e6fcbc2SMing Lei return !!dispatched; 1966e6fcbc2SMing Lei } 1976e6fcbc2SMing Lei 1986e6fcbc2SMing Lei static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) 1996e6fcbc2SMing Lei { 2006e6fcbc2SMing Lei int ret; 2016e6fcbc2SMing Lei 2026e6fcbc2SMing Lei do { 2036e6fcbc2SMing Lei ret = __blk_mq_do_dispatch_sched(hctx); 2046e6fcbc2SMing Lei } while (ret == 1); 20528d65729SSalman Qazi 20628d65729SSalman Qazi return ret; 207de148297SMing Lei } 208de148297SMing Lei 209b347689fSMing Lei static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx, 210b347689fSMing Lei struct blk_mq_ctx *ctx) 211b347689fSMing Lei { 212f31967f0SJens Axboe unsigned short idx = ctx->index_hw[hctx->type]; 213b347689fSMing Lei 214b347689fSMing Lei if (++idx == hctx->nr_ctx) 215b347689fSMing Lei idx = 0; 216b347689fSMing Lei 217b347689fSMing Lei return hctx->ctxs[idx]; 218b347689fSMing Lei } 219b347689fSMing Lei 2201f460b63SMing Lei /* 2211f460b63SMing Lei * Only SCSI implements .get_budget and .put_budget, and SCSI restarts 2221f460b63SMing Lei * its queue by itself in its completion handler, so we don't need to 2231f460b63SMing Lei * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE. 22428d65729SSalman Qazi * 22528d65729SSalman Qazi * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to 226c4aecaa2SRandy Dunlap * be run again. This is necessary to avoid starving flushes. 2271f460b63SMing Lei */ 22828d65729SSalman Qazi static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx) 229b347689fSMing Lei { 230b347689fSMing Lei struct request_queue *q = hctx->queue; 231b347689fSMing Lei LIST_HEAD(rq_list); 232b347689fSMing Lei struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from); 23328d65729SSalman Qazi int ret = 0; 234b347689fSMing Lei struct request *rq; 235b347689fSMing Lei 236445874e8SMing Lei do { 237*2a5a24aaSMing Lei int budget_token; 238*2a5a24aaSMing Lei 23928d65729SSalman Qazi if (!list_empty_careful(&hctx->dispatch)) { 24028d65729SSalman Qazi ret = -EAGAIN; 24128d65729SSalman Qazi break; 24228d65729SSalman Qazi } 24328d65729SSalman Qazi 244b347689fSMing Lei if (!sbitmap_any_bit_set(&hctx->ctx_map)) 245b347689fSMing Lei break; 246b347689fSMing Lei 247*2a5a24aaSMing Lei budget_token = blk_mq_get_dispatch_budget(q); 248*2a5a24aaSMing Lei if (budget_token < 0) 2491f460b63SMing Lei break; 250b347689fSMing Lei 251b347689fSMing Lei rq = blk_mq_dequeue_from_ctx(hctx, ctx); 252b347689fSMing Lei if (!rq) { 253*2a5a24aaSMing Lei blk_mq_put_dispatch_budget(q, budget_token); 254a0823421SDouglas Anderson /* 255a0823421SDouglas Anderson * We're releasing without dispatching. Holding the 256a0823421SDouglas Anderson * budget could have blocked any "hctx"s with the 257a0823421SDouglas Anderson * same queue and if we didn't dispatch then there's 258a0823421SDouglas Anderson * no guarantee anyone will kick the queue. Kick it 259a0823421SDouglas Anderson * ourselves. 260a0823421SDouglas Anderson */ 261a0823421SDouglas Anderson blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); 262b347689fSMing Lei break; 263b347689fSMing Lei } 264b347689fSMing Lei 265*2a5a24aaSMing Lei blk_mq_set_rq_budget_token(rq, budget_token); 266*2a5a24aaSMing Lei 267b347689fSMing Lei /* 268b347689fSMing Lei * Now this rq owns the budget which has to be released 269b347689fSMing Lei * if this rq won't be queued to driver via .queue_rq() 270b347689fSMing Lei * in blk_mq_dispatch_rq_list(). 271b347689fSMing Lei */ 272b347689fSMing Lei list_add(&rq->queuelist, &rq_list); 273b347689fSMing Lei 274b347689fSMing Lei /* round robin for fair dispatch */ 275b347689fSMing Lei ctx = blk_mq_next_ctx(hctx, rq->mq_ctx); 276b347689fSMing Lei 2771fd40b5eSMing Lei } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1)); 278b347689fSMing Lei 279b347689fSMing Lei WRITE_ONCE(hctx->dispatch_from, ctx); 28028d65729SSalman Qazi return ret; 281b347689fSMing Lei } 282b347689fSMing Lei 283e1b586f2SZheng Bin static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) 284bd166ef1SJens Axboe { 28581380ca1SOmar Sandoval struct request_queue *q = hctx->queue; 28681380ca1SOmar Sandoval struct elevator_queue *e = q->elevator; 287f9cd4bfeSJens Axboe const bool has_sched_dispatch = e && e->type->ops.dispatch_request; 28828d65729SSalman Qazi int ret = 0; 289bd166ef1SJens Axboe LIST_HEAD(rq_list); 290bd166ef1SJens Axboe 291bd166ef1SJens Axboe /* 292bd166ef1SJens Axboe * If we have previous entries on our dispatch list, grab them first for 293bd166ef1SJens Axboe * more fair dispatch. 294bd166ef1SJens Axboe */ 295bd166ef1SJens Axboe if (!list_empty_careful(&hctx->dispatch)) { 296bd166ef1SJens Axboe spin_lock(&hctx->lock); 297bd166ef1SJens Axboe if (!list_empty(&hctx->dispatch)) 298bd166ef1SJens Axboe list_splice_init(&hctx->dispatch, &rq_list); 299bd166ef1SJens Axboe spin_unlock(&hctx->lock); 300bd166ef1SJens Axboe } 301bd166ef1SJens Axboe 302bd166ef1SJens Axboe /* 303bd166ef1SJens Axboe * Only ask the scheduler for requests, if we didn't have residual 304bd166ef1SJens Axboe * requests from the dispatch list. This is to avoid the case where 305bd166ef1SJens Axboe * we only ever dispatch a fraction of the requests available because 306bd166ef1SJens Axboe * of low device queue depth. Once we pull requests out of the IO 307bd166ef1SJens Axboe * scheduler, we can no longer merge or sort them. So it's best to 308bd166ef1SJens Axboe * leave them there for as long as we can. Mark the hw queue as 309bd166ef1SJens Axboe * needing a restart in that case. 310caf8eb0dSMing Lei * 3115e3d02bbSMing Lei * We want to dispatch from the scheduler if there was nothing 3125e3d02bbSMing Lei * on the dispatch list or we were able to dispatch from the 3135e3d02bbSMing Lei * dispatch list. 31464765a75SJens Axboe */ 315caf8eb0dSMing Lei if (!list_empty(&rq_list)) { 316caf8eb0dSMing Lei blk_mq_sched_mark_restart_hctx(hctx); 3171fd40b5eSMing Lei if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) { 318b347689fSMing Lei if (has_sched_dispatch) 31928d65729SSalman Qazi ret = blk_mq_do_dispatch_sched(hctx); 320b347689fSMing Lei else 32128d65729SSalman Qazi ret = blk_mq_do_dispatch_ctx(hctx); 322b347689fSMing Lei } 323caf8eb0dSMing Lei } else if (has_sched_dispatch) { 32428d65729SSalman Qazi ret = blk_mq_do_dispatch_sched(hctx); 3256e768717SMing Lei } else if (hctx->dispatch_busy) { 3266e768717SMing Lei /* dequeue request one by one from sw queue if queue is busy */ 32728d65729SSalman Qazi ret = blk_mq_do_dispatch_ctx(hctx); 328caf8eb0dSMing Lei } else { 329caf8eb0dSMing Lei blk_mq_flush_busy_ctxs(hctx, &rq_list); 3301fd40b5eSMing Lei blk_mq_dispatch_rq_list(hctx, &rq_list, 0); 331c13660a0SJens Axboe } 33228d65729SSalman Qazi 33328d65729SSalman Qazi return ret; 33428d65729SSalman Qazi } 33528d65729SSalman Qazi 33628d65729SSalman Qazi void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) 33728d65729SSalman Qazi { 33828d65729SSalman Qazi struct request_queue *q = hctx->queue; 33928d65729SSalman Qazi 34028d65729SSalman Qazi /* RCU or SRCU read lock is needed before checking quiesced flag */ 34128d65729SSalman Qazi if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) 34228d65729SSalman Qazi return; 34328d65729SSalman Qazi 34428d65729SSalman Qazi hctx->run++; 34528d65729SSalman Qazi 34628d65729SSalman Qazi /* 34728d65729SSalman Qazi * A return of -EAGAIN is an indication that hctx->dispatch is not 34828d65729SSalman Qazi * empty and we must run again in order to avoid starving flushes. 34928d65729SSalman Qazi */ 35028d65729SSalman Qazi if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) { 35128d65729SSalman Qazi if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) 35228d65729SSalman Qazi blk_mq_run_hw_queue(hctx, true); 35328d65729SSalman Qazi } 354bd166ef1SJens Axboe } 355bd166ef1SJens Axboe 35614ccb66bSChristoph Hellwig bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio, 35714ccb66bSChristoph Hellwig unsigned int nr_segs) 358bd166ef1SJens Axboe { 359bd166ef1SJens Axboe struct elevator_queue *e = q->elevator; 360bd166ef1SJens Axboe struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); 3618ccdf4a3SJianchao Wang struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx); 3629bddeb2aSMing Lei bool ret = false; 363c16d6b5aSMing Lei enum hctx_type type; 364bd166ef1SJens Axboe 365c05f4220SBart Van Assche if (e && e->type->ops.bio_merge) 36614ccb66bSChristoph Hellwig return e->type->ops.bio_merge(hctx, bio, nr_segs); 367bd166ef1SJens Axboe 368c16d6b5aSMing Lei type = hctx->type; 369cdfcef9eSBaolin Wang if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) || 370cdfcef9eSBaolin Wang list_empty_careful(&ctx->rq_lists[type])) 371cdfcef9eSBaolin Wang return false; 372cdfcef9eSBaolin Wang 3739bddeb2aSMing Lei /* default per sw-queue merge */ 3749bddeb2aSMing Lei spin_lock(&ctx->lock); 375cdfcef9eSBaolin Wang /* 376cdfcef9eSBaolin Wang * Reverse check our software queue for entries that we could 377cdfcef9eSBaolin Wang * potentially merge with. Currently includes a hand-wavy stop 378cdfcef9eSBaolin Wang * count of 8, to not spend too much time checking for merges. 379cdfcef9eSBaolin Wang */ 380cdfcef9eSBaolin Wang if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) { 381cdfcef9eSBaolin Wang ctx->rq_merged++; 382cdfcef9eSBaolin Wang ret = true; 3839bddeb2aSMing Lei } 3849bddeb2aSMing Lei 385cdfcef9eSBaolin Wang spin_unlock(&ctx->lock); 386cdfcef9eSBaolin Wang 3879bddeb2aSMing Lei return ret; 388bd166ef1SJens Axboe } 389bd166ef1SJens Axboe 390bd166ef1SJens Axboe bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq) 391bd166ef1SJens Axboe { 392bd166ef1SJens Axboe return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq); 393bd166ef1SJens Axboe } 394bd166ef1SJens Axboe EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge); 395bd166ef1SJens Axboe 3960cacba6cSOmar Sandoval static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx, 397a6a252e6SMing Lei bool has_sched, 3980cacba6cSOmar Sandoval struct request *rq) 399bd166ef1SJens Axboe { 40001e99aecSMing Lei /* 40101e99aecSMing Lei * dispatch flush and passthrough rq directly 40201e99aecSMing Lei * 40301e99aecSMing Lei * passthrough request has to be added to hctx->dispatch directly. 40401e99aecSMing Lei * For some reason, device may be in one situation which can't 40501e99aecSMing Lei * handle FS request, so STS_RESOURCE is always returned and the 40601e99aecSMing Lei * FS request will be added to hctx->dispatch. However passthrough 40701e99aecSMing Lei * request may be required at that time for fixing the problem. If 40801e99aecSMing Lei * passthrough request is added to scheduler queue, there isn't any 40901e99aecSMing Lei * chance to dispatch it given we prioritize requests in hctx->dispatch. 41001e99aecSMing Lei */ 41101e99aecSMing Lei if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq)) 412bd166ef1SJens Axboe return true; 413bd166ef1SJens Axboe 414923218f6SMing Lei if (has_sched) 415a6a252e6SMing Lei rq->rq_flags |= RQF_SORTED; 416a6a252e6SMing Lei 417a6a252e6SMing Lei return false; 418a6a252e6SMing Lei } 419a6a252e6SMing Lei 420bd6737f1SJens Axboe void blk_mq_sched_insert_request(struct request *rq, bool at_head, 4219e97d295SMike Snitzer bool run_queue, bool async) 422bd6737f1SJens Axboe { 423bd6737f1SJens Axboe struct request_queue *q = rq->q; 424bd6737f1SJens Axboe struct elevator_queue *e = q->elevator; 425bd6737f1SJens Axboe struct blk_mq_ctx *ctx = rq->mq_ctx; 426ea4f995eSJens Axboe struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 427bd6737f1SJens Axboe 428e44a6a23SXianting Tian WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG)); 429923218f6SMing Lei 43001e99aecSMing Lei if (blk_mq_sched_bypass_insert(hctx, !!e, rq)) { 431cc3200eaSMing Lei /* 432cc3200eaSMing Lei * Firstly normal IO request is inserted to scheduler queue or 433cc3200eaSMing Lei * sw queue, meantime we add flush request to dispatch queue( 434cc3200eaSMing Lei * hctx->dispatch) directly and there is at most one in-flight 435cc3200eaSMing Lei * flush request for each hw queue, so it doesn't matter to add 436cc3200eaSMing Lei * flush request to tail or front of the dispatch queue. 437cc3200eaSMing Lei * 438cc3200eaSMing Lei * Secondly in case of NCQ, flush request belongs to non-NCQ 439cc3200eaSMing Lei * command, and queueing it will fail when there is any 440cc3200eaSMing Lei * in-flight normal IO request(NCQ command). When adding flush 441cc3200eaSMing Lei * rq to the front of hctx->dispatch, it is easier to introduce 442cc3200eaSMing Lei * extra time to flush rq's latency because of S_SCHED_RESTART 443cc3200eaSMing Lei * compared with adding to the tail of dispatch queue, then 444cc3200eaSMing Lei * chance of flush merge is increased, and less flush requests 445cc3200eaSMing Lei * will be issued to controller. It is observed that ~10% time 446cc3200eaSMing Lei * is saved in blktests block/004 on disk attached to AHCI/NCQ 447cc3200eaSMing Lei * drive when adding flush rq to the front of hctx->dispatch. 448cc3200eaSMing Lei * 449cc3200eaSMing Lei * Simply queue flush rq to the front of hctx->dispatch so that 450cc3200eaSMing Lei * intensive flush workloads can benefit in case of NCQ HW. 451cc3200eaSMing Lei */ 452cc3200eaSMing Lei at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head; 45301e99aecSMing Lei blk_mq_request_bypass_insert(rq, at_head, false); 4540cacba6cSOmar Sandoval goto run; 45501e99aecSMing Lei } 4560cacba6cSOmar Sandoval 457f9cd4bfeSJens Axboe if (e && e->type->ops.insert_requests) { 458bd6737f1SJens Axboe LIST_HEAD(list); 459bd6737f1SJens Axboe 460bd6737f1SJens Axboe list_add(&rq->queuelist, &list); 461f9cd4bfeSJens Axboe e->type->ops.insert_requests(hctx, &list, at_head); 462bd6737f1SJens Axboe } else { 463bd6737f1SJens Axboe spin_lock(&ctx->lock); 464bd6737f1SJens Axboe __blk_mq_insert_request(hctx, rq, at_head); 465bd6737f1SJens Axboe spin_unlock(&ctx->lock); 466bd6737f1SJens Axboe } 467bd6737f1SJens Axboe 4680cacba6cSOmar Sandoval run: 469bd6737f1SJens Axboe if (run_queue) 470bd6737f1SJens Axboe blk_mq_run_hw_queue(hctx, async); 471bd6737f1SJens Axboe } 472bd6737f1SJens Axboe 47367cae4c9SJens Axboe void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx, 474bd6737f1SJens Axboe struct blk_mq_ctx *ctx, 475bd6737f1SJens Axboe struct list_head *list, bool run_queue_async) 476bd6737f1SJens Axboe { 477f9afca4dSJens Axboe struct elevator_queue *e; 478e87eb301SMing Lei struct request_queue *q = hctx->queue; 479e87eb301SMing Lei 480e87eb301SMing Lei /* 481e87eb301SMing Lei * blk_mq_sched_insert_requests() is called from flush plug 482e87eb301SMing Lei * context only, and hold one usage counter to prevent queue 483e87eb301SMing Lei * from being released. 484e87eb301SMing Lei */ 485e87eb301SMing Lei percpu_ref_get(&q->q_usage_counter); 486f9afca4dSJens Axboe 487f9afca4dSJens Axboe e = hctx->queue->elevator; 488f9cd4bfeSJens Axboe if (e && e->type->ops.insert_requests) 489f9cd4bfeSJens Axboe e->type->ops.insert_requests(hctx, list, false); 4906ce3dd6eSMing Lei else { 4916ce3dd6eSMing Lei /* 4926ce3dd6eSMing Lei * try to issue requests directly if the hw queue isn't 4936ce3dd6eSMing Lei * busy in case of 'none' scheduler, and this way may save 4946ce3dd6eSMing Lei * us one extra enqueue & dequeue to sw queue. 4956ce3dd6eSMing Lei */ 496fd9c40f6SBart Van Assche if (!hctx->dispatch_busy && !e && !run_queue_async) { 4976ce3dd6eSMing Lei blk_mq_try_issue_list_directly(hctx, list); 498fd9c40f6SBart Van Assche if (list_empty(list)) 499e87eb301SMing Lei goto out; 500fd9c40f6SBart Van Assche } 501bd6737f1SJens Axboe blk_mq_insert_requests(hctx, ctx, list); 5026ce3dd6eSMing Lei } 503bd6737f1SJens Axboe 504bd6737f1SJens Axboe blk_mq_run_hw_queue(hctx, run_queue_async); 505e87eb301SMing Lei out: 506e87eb301SMing Lei percpu_ref_put(&q->q_usage_counter); 507bd6737f1SJens Axboe } 508bd6737f1SJens Axboe 509bd166ef1SJens Axboe static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set, 510bd166ef1SJens Axboe struct blk_mq_hw_ctx *hctx, 511bd166ef1SJens Axboe unsigned int hctx_idx) 512bd166ef1SJens Axboe { 51332bc15afSJohn Garry unsigned int flags = set->flags & ~BLK_MQ_F_TAG_HCTX_SHARED; 5141c0706a7SJohn Garry 515bd166ef1SJens Axboe if (hctx->sched_tags) { 516bd166ef1SJens Axboe blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx); 5171c0706a7SJohn Garry blk_mq_free_rq_map(hctx->sched_tags, flags); 518bd166ef1SJens Axboe hctx->sched_tags = NULL; 519bd166ef1SJens Axboe } 520bd166ef1SJens Axboe } 521bd166ef1SJens Axboe 5226917ff0bSOmar Sandoval static int blk_mq_sched_alloc_tags(struct request_queue *q, 5236917ff0bSOmar Sandoval struct blk_mq_hw_ctx *hctx, 5246917ff0bSOmar Sandoval unsigned int hctx_idx) 525bd166ef1SJens Axboe { 526bd166ef1SJens Axboe struct blk_mq_tag_set *set = q->tag_set; 52732bc15afSJohn Garry /* Clear HCTX_SHARED so tags are init'ed */ 52832bc15afSJohn Garry unsigned int flags = set->flags & ~BLK_MQ_F_TAG_HCTX_SHARED; 5296917ff0bSOmar Sandoval int ret; 530bd166ef1SJens Axboe 5316917ff0bSOmar Sandoval hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests, 5321c0706a7SJohn Garry set->reserved_tags, flags); 533bd166ef1SJens Axboe if (!hctx->sched_tags) 5346917ff0bSOmar Sandoval return -ENOMEM; 5356917ff0bSOmar Sandoval 5366917ff0bSOmar Sandoval ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests); 5376917ff0bSOmar Sandoval if (ret) 5386917ff0bSOmar Sandoval blk_mq_sched_free_tags(set, hctx, hctx_idx); 539bd166ef1SJens Axboe 540bd166ef1SJens Axboe return ret; 541bd166ef1SJens Axboe } 542bd166ef1SJens Axboe 543c3e22192SMing Lei /* called in queue's release handler, tagset has gone away */ 54454d5329dSOmar Sandoval static void blk_mq_sched_tags_teardown(struct request_queue *q) 545bd166ef1SJens Axboe { 546bd166ef1SJens Axboe struct blk_mq_hw_ctx *hctx; 547bd166ef1SJens Axboe int i; 548bd166ef1SJens Axboe 549c3e22192SMing Lei queue_for_each_hw_ctx(q, hctx, i) { 55032bc15afSJohn Garry /* Clear HCTX_SHARED so tags are freed */ 55132bc15afSJohn Garry unsigned int flags = hctx->flags & ~BLK_MQ_F_TAG_HCTX_SHARED; 5521c0706a7SJohn Garry 553c3e22192SMing Lei if (hctx->sched_tags) { 5541c0706a7SJohn Garry blk_mq_free_rq_map(hctx->sched_tags, flags); 555c3e22192SMing Lei hctx->sched_tags = NULL; 556c3e22192SMing Lei } 557c3e22192SMing Lei } 558bd166ef1SJens Axboe } 559d3484991SJens Axboe 5606917ff0bSOmar Sandoval int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e) 5616917ff0bSOmar Sandoval { 5626917ff0bSOmar Sandoval struct blk_mq_hw_ctx *hctx; 563ee056f98SOmar Sandoval struct elevator_queue *eq; 5646917ff0bSOmar Sandoval unsigned int i; 5656917ff0bSOmar Sandoval int ret; 5666917ff0bSOmar Sandoval 5676917ff0bSOmar Sandoval if (!e) { 5686917ff0bSOmar Sandoval q->elevator = NULL; 56932a50fabSMing Lei q->nr_requests = q->tag_set->queue_depth; 5706917ff0bSOmar Sandoval return 0; 5716917ff0bSOmar Sandoval } 5726917ff0bSOmar Sandoval 5736917ff0bSOmar Sandoval /* 57432825c45SMing Lei * Default to double of smaller one between hw queue_depth and 128, 57532825c45SMing Lei * since we don't split into sync/async like the old code did. 57632825c45SMing Lei * Additionally, this is a per-hw queue depth. 5776917ff0bSOmar Sandoval */ 57832825c45SMing Lei q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth, 57932825c45SMing Lei BLKDEV_MAX_RQ); 5806917ff0bSOmar Sandoval 5816917ff0bSOmar Sandoval queue_for_each_hw_ctx(q, hctx, i) { 5826917ff0bSOmar Sandoval ret = blk_mq_sched_alloc_tags(q, hctx, i); 5836917ff0bSOmar Sandoval if (ret) 5846917ff0bSOmar Sandoval goto err; 5856917ff0bSOmar Sandoval } 5866917ff0bSOmar Sandoval 587f9cd4bfeSJens Axboe ret = e->ops.init_sched(q, e); 5886917ff0bSOmar Sandoval if (ret) 5896917ff0bSOmar Sandoval goto err; 5906917ff0bSOmar Sandoval 591d332ce09SOmar Sandoval blk_mq_debugfs_register_sched(q); 592d332ce09SOmar Sandoval 593ee056f98SOmar Sandoval queue_for_each_hw_ctx(q, hctx, i) { 594f9cd4bfeSJens Axboe if (e->ops.init_hctx) { 595f9cd4bfeSJens Axboe ret = e->ops.init_hctx(hctx, i); 596ee056f98SOmar Sandoval if (ret) { 597ee056f98SOmar Sandoval eq = q->elevator; 598c3e22192SMing Lei blk_mq_sched_free_requests(q); 599ee056f98SOmar Sandoval blk_mq_exit_sched(q, eq); 600ee056f98SOmar Sandoval kobject_put(&eq->kobj); 601ee056f98SOmar Sandoval return ret; 602ee056f98SOmar Sandoval } 603ee056f98SOmar Sandoval } 604d332ce09SOmar Sandoval blk_mq_debugfs_register_sched_hctx(q, hctx); 605ee056f98SOmar Sandoval } 606ee056f98SOmar Sandoval 6076917ff0bSOmar Sandoval return 0; 6086917ff0bSOmar Sandoval 6096917ff0bSOmar Sandoval err: 610c3e22192SMing Lei blk_mq_sched_free_requests(q); 61154d5329dSOmar Sandoval blk_mq_sched_tags_teardown(q); 61254d5329dSOmar Sandoval q->elevator = NULL; 6136917ff0bSOmar Sandoval return ret; 6146917ff0bSOmar Sandoval } 6156917ff0bSOmar Sandoval 616c3e22192SMing Lei /* 617c3e22192SMing Lei * called in either blk_queue_cleanup or elevator_switch, tagset 618c3e22192SMing Lei * is required for freeing requests 619c3e22192SMing Lei */ 620c3e22192SMing Lei void blk_mq_sched_free_requests(struct request_queue *q) 621c3e22192SMing Lei { 622c3e22192SMing Lei struct blk_mq_hw_ctx *hctx; 623c3e22192SMing Lei int i; 624c3e22192SMing Lei 625c3e22192SMing Lei queue_for_each_hw_ctx(q, hctx, i) { 626c3e22192SMing Lei if (hctx->sched_tags) 627c3e22192SMing Lei blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i); 628c3e22192SMing Lei } 629c3e22192SMing Lei } 630c3e22192SMing Lei 63154d5329dSOmar Sandoval void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e) 63254d5329dSOmar Sandoval { 633ee056f98SOmar Sandoval struct blk_mq_hw_ctx *hctx; 634ee056f98SOmar Sandoval unsigned int i; 635ee056f98SOmar Sandoval 636ee056f98SOmar Sandoval queue_for_each_hw_ctx(q, hctx, i) { 637d332ce09SOmar Sandoval blk_mq_debugfs_unregister_sched_hctx(hctx); 638f9cd4bfeSJens Axboe if (e->type->ops.exit_hctx && hctx->sched_data) { 639f9cd4bfeSJens Axboe e->type->ops.exit_hctx(hctx, i); 640ee056f98SOmar Sandoval hctx->sched_data = NULL; 641ee056f98SOmar Sandoval } 642ee056f98SOmar Sandoval } 643d332ce09SOmar Sandoval blk_mq_debugfs_unregister_sched(q); 644f9cd4bfeSJens Axboe if (e->type->ops.exit_sched) 645f9cd4bfeSJens Axboe e->type->ops.exit_sched(e); 64654d5329dSOmar Sandoval blk_mq_sched_tags_teardown(q); 64754d5329dSOmar Sandoval q->elevator = NULL; 64854d5329dSOmar Sandoval } 649