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 21bd166ef1SJens Axboe void blk_mq_sched_free_hctx_data(struct request_queue *q, 22bd166ef1SJens Axboe void (*exit)(struct blk_mq_hw_ctx *)) 23bd166ef1SJens Axboe { 24bd166ef1SJens Axboe struct blk_mq_hw_ctx *hctx; 25bd166ef1SJens Axboe int i; 26bd166ef1SJens Axboe 27bd166ef1SJens Axboe queue_for_each_hw_ctx(q, hctx, i) { 28bd166ef1SJens Axboe if (exit && hctx->sched_data) 29bd166ef1SJens Axboe exit(hctx); 30bd166ef1SJens Axboe kfree(hctx->sched_data); 31bd166ef1SJens Axboe hctx->sched_data = NULL; 32bd166ef1SJens Axboe } 33bd166ef1SJens Axboe } 34bd166ef1SJens Axboe EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data); 35bd166ef1SJens Axboe 36e2b3fa5aSDamien Le Moal void blk_mq_sched_assign_ioc(struct request *rq) 37bd166ef1SJens Axboe { 3844e8c2bfSChristoph Hellwig struct request_queue *q = rq->q; 390c62bff1SJens Axboe struct io_context *ioc; 40bd166ef1SJens Axboe struct io_cq *icq; 41bd166ef1SJens Axboe 420c62bff1SJens Axboe /* 430c62bff1SJens Axboe * May not have an IO context if it's a passthrough request 440c62bff1SJens Axboe */ 450c62bff1SJens Axboe ioc = current->io_context; 460c62bff1SJens Axboe if (!ioc) 470c62bff1SJens Axboe return; 480c62bff1SJens Axboe 490d945c1fSChristoph Hellwig spin_lock_irq(&q->queue_lock); 50bd166ef1SJens Axboe icq = ioc_lookup_icq(ioc, q); 510d945c1fSChristoph Hellwig spin_unlock_irq(&q->queue_lock); 52bd166ef1SJens Axboe 53bd166ef1SJens Axboe if (!icq) { 54bd166ef1SJens Axboe icq = ioc_create_icq(ioc, q, GFP_ATOMIC); 55bd166ef1SJens Axboe if (!icq) 56bd166ef1SJens Axboe return; 57bd166ef1SJens Axboe } 58ea511e3cSChristoph Hellwig get_io_context(icq->ioc); 5944e8c2bfSChristoph Hellwig rq->elv.icq = icq; 60bd166ef1SJens Axboe } 61bd166ef1SJens Axboe 628e8320c9SJens Axboe /* 638e8320c9SJens Axboe * Mark a hardware queue as needing a restart. For shared queues, maintain 648e8320c9SJens Axboe * a count of how many hardware queues are marked for restart. 658e8320c9SJens Axboe */ 667211aef8SDamien Le Moal void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx) 678e8320c9SJens Axboe { 688e8320c9SJens Axboe if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) 698e8320c9SJens Axboe return; 708e8320c9SJens Axboe 718e8320c9SJens Axboe set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); 728e8320c9SJens Axboe } 737211aef8SDamien Le Moal EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx); 748e8320c9SJens Axboe 7597889f9aSMing Lei void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx) 768e8320c9SJens Axboe { 778e8320c9SJens Axboe if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) 7897889f9aSMing Lei return; 798e8320c9SJens Axboe clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); 808e8320c9SJens Axboe 81d7d8535fSMing Lei /* 82d7d8535fSMing Lei * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch) 83d7d8535fSMing Lei * in blk_mq_run_hw_queue(). Its pair is the barrier in 84d7d8535fSMing Lei * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART, 85d7d8535fSMing Lei * meantime new request added to hctx->dispatch is missed to check in 86d7d8535fSMing Lei * blk_mq_run_hw_queue(). 87d7d8535fSMing Lei */ 88d7d8535fSMing Lei smp_mb(); 89d7d8535fSMing Lei 9097889f9aSMing Lei blk_mq_run_hw_queue(hctx, true); 918e8320c9SJens Axboe } 928e8320c9SJens Axboe 936e6fcbc2SMing Lei static int sched_rq_cmp(void *priv, struct list_head *a, struct list_head *b) 946e6fcbc2SMing Lei { 956e6fcbc2SMing Lei struct request *rqa = container_of(a, struct request, queuelist); 966e6fcbc2SMing Lei struct request *rqb = container_of(b, struct request, queuelist); 976e6fcbc2SMing Lei 986e6fcbc2SMing Lei return rqa->mq_hctx > rqb->mq_hctx; 996e6fcbc2SMing Lei } 1006e6fcbc2SMing Lei 1016e6fcbc2SMing Lei static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list) 1026e6fcbc2SMing Lei { 1036e6fcbc2SMing Lei struct blk_mq_hw_ctx *hctx = 1046e6fcbc2SMing Lei list_first_entry(rq_list, struct request, queuelist)->mq_hctx; 1056e6fcbc2SMing Lei struct request *rq; 1066e6fcbc2SMing Lei LIST_HEAD(hctx_list); 1076e6fcbc2SMing Lei unsigned int count = 0; 1086e6fcbc2SMing Lei 1096e6fcbc2SMing Lei list_for_each_entry(rq, rq_list, queuelist) { 1106e6fcbc2SMing Lei if (rq->mq_hctx != hctx) { 1116e6fcbc2SMing Lei list_cut_before(&hctx_list, rq_list, &rq->queuelist); 1126e6fcbc2SMing Lei goto dispatch; 1136e6fcbc2SMing Lei } 1146e6fcbc2SMing Lei count++; 1156e6fcbc2SMing Lei } 1166e6fcbc2SMing Lei list_splice_tail_init(rq_list, &hctx_list); 1176e6fcbc2SMing Lei 1186e6fcbc2SMing Lei dispatch: 119106e71c5SBaolin Wang return blk_mq_dispatch_rq_list(hctx, &hctx_list, count); 1206e6fcbc2SMing Lei } 1216e6fcbc2SMing Lei 122a0823421SDouglas Anderson #define BLK_MQ_BUDGET_DELAY 3 /* ms units */ 123a0823421SDouglas Anderson 1241f460b63SMing Lei /* 1251f460b63SMing Lei * Only SCSI implements .get_budget and .put_budget, and SCSI restarts 1261f460b63SMing Lei * its queue by itself in its completion handler, so we don't need to 1271f460b63SMing Lei * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE. 12828d65729SSalman Qazi * 12928d65729SSalman Qazi * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to 13028d65729SSalman Qazi * be run again. This is necessary to avoid starving flushes. 1311f460b63SMing Lei */ 1326e6fcbc2SMing Lei static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) 133caf8eb0dSMing Lei { 134caf8eb0dSMing Lei struct request_queue *q = hctx->queue; 135caf8eb0dSMing Lei struct elevator_queue *e = q->elevator; 1366e6fcbc2SMing Lei bool multi_hctxs = false, run_queue = false; 1376e6fcbc2SMing Lei bool dispatched = false, busy = false; 1386e6fcbc2SMing Lei unsigned int max_dispatch; 139caf8eb0dSMing Lei LIST_HEAD(rq_list); 1406e6fcbc2SMing Lei int count = 0; 1416e6fcbc2SMing Lei 1426e6fcbc2SMing Lei if (hctx->dispatch_busy) 1436e6fcbc2SMing Lei max_dispatch = 1; 1446e6fcbc2SMing Lei else 1456e6fcbc2SMing Lei max_dispatch = hctx->queue->nr_requests; 146caf8eb0dSMing Lei 147445874e8SMing Lei do { 1486e6fcbc2SMing Lei struct request *rq; 1496e6fcbc2SMing Lei 150f9cd4bfeSJens Axboe if (e->type->ops.has_work && !e->type->ops.has_work(hctx)) 151caf8eb0dSMing Lei break; 152de148297SMing Lei 15328d65729SSalman Qazi if (!list_empty_careful(&hctx->dispatch)) { 1546e6fcbc2SMing Lei busy = true; 15528d65729SSalman Qazi break; 15628d65729SSalman Qazi } 15728d65729SSalman Qazi 15865c76369SMing Lei if (!blk_mq_get_dispatch_budget(q)) 1591f460b63SMing Lei break; 160de148297SMing Lei 161f9cd4bfeSJens Axboe rq = e->type->ops.dispatch_request(hctx); 162de148297SMing Lei if (!rq) { 16365c76369SMing Lei blk_mq_put_dispatch_budget(q); 164a0823421SDouglas Anderson /* 165a0823421SDouglas Anderson * We're releasing without dispatching. Holding the 166a0823421SDouglas Anderson * budget could have blocked any "hctx"s with the 167a0823421SDouglas Anderson * same queue and if we didn't dispatch then there's 168a0823421SDouglas Anderson * no guarantee anyone will kick the queue. Kick it 169a0823421SDouglas Anderson * ourselves. 170a0823421SDouglas Anderson */ 1716e6fcbc2SMing Lei run_queue = true; 172de148297SMing Lei break; 173caf8eb0dSMing Lei } 174caf8eb0dSMing Lei 175de148297SMing Lei /* 176de148297SMing Lei * Now this rq owns the budget which has to be released 177de148297SMing Lei * if this rq won't be queued to driver via .queue_rq() 178de148297SMing Lei * in blk_mq_dispatch_rq_list(). 179de148297SMing Lei */ 1806e6fcbc2SMing Lei list_add_tail(&rq->queuelist, &rq_list); 1816e6fcbc2SMing Lei if (rq->mq_hctx != hctx) 1826e6fcbc2SMing Lei multi_hctxs = true; 1836e6fcbc2SMing Lei } while (++count < max_dispatch); 1846e6fcbc2SMing Lei 1856e6fcbc2SMing Lei if (!count) { 1866e6fcbc2SMing Lei if (run_queue) 1876e6fcbc2SMing Lei blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); 1886e6fcbc2SMing Lei } else if (multi_hctxs) { 1896e6fcbc2SMing Lei /* 1906e6fcbc2SMing Lei * Requests from different hctx may be dequeued from some 1916e6fcbc2SMing Lei * schedulers, such as bfq and deadline. 1926e6fcbc2SMing Lei * 1936e6fcbc2SMing Lei * Sort the requests in the list according to their hctx, 1946e6fcbc2SMing Lei * dispatch batching requests from same hctx at a time. 1956e6fcbc2SMing Lei */ 1966e6fcbc2SMing Lei list_sort(NULL, &rq_list, sched_rq_cmp); 1976e6fcbc2SMing Lei do { 1986e6fcbc2SMing Lei dispatched |= blk_mq_dispatch_hctx_list(&rq_list); 1996e6fcbc2SMing Lei } while (!list_empty(&rq_list)); 2006e6fcbc2SMing Lei } else { 2016e6fcbc2SMing Lei dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count); 2026e6fcbc2SMing Lei } 2036e6fcbc2SMing Lei 2046e6fcbc2SMing Lei if (busy) 2056e6fcbc2SMing Lei return -EAGAIN; 2066e6fcbc2SMing Lei return !!dispatched; 2076e6fcbc2SMing Lei } 2086e6fcbc2SMing Lei 2096e6fcbc2SMing Lei static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) 2106e6fcbc2SMing Lei { 2116e6fcbc2SMing Lei int ret; 2126e6fcbc2SMing Lei 2136e6fcbc2SMing Lei do { 2146e6fcbc2SMing Lei ret = __blk_mq_do_dispatch_sched(hctx); 2156e6fcbc2SMing Lei } while (ret == 1); 21628d65729SSalman Qazi 21728d65729SSalman Qazi return ret; 218de148297SMing Lei } 219de148297SMing Lei 220b347689fSMing Lei static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx, 221b347689fSMing Lei struct blk_mq_ctx *ctx) 222b347689fSMing Lei { 223f31967f0SJens Axboe unsigned short idx = ctx->index_hw[hctx->type]; 224b347689fSMing Lei 225b347689fSMing Lei if (++idx == hctx->nr_ctx) 226b347689fSMing Lei idx = 0; 227b347689fSMing Lei 228b347689fSMing Lei return hctx->ctxs[idx]; 229b347689fSMing Lei } 230b347689fSMing Lei 2311f460b63SMing Lei /* 2321f460b63SMing Lei * Only SCSI implements .get_budget and .put_budget, and SCSI restarts 2331f460b63SMing Lei * its queue by itself in its completion handler, so we don't need to 2341f460b63SMing Lei * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE. 23528d65729SSalman Qazi * 23628d65729SSalman Qazi * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to 237c4aecaa2SRandy Dunlap * be run again. This is necessary to avoid starving flushes. 2381f460b63SMing Lei */ 23928d65729SSalman Qazi static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx) 240b347689fSMing Lei { 241b347689fSMing Lei struct request_queue *q = hctx->queue; 242b347689fSMing Lei LIST_HEAD(rq_list); 243b347689fSMing Lei struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from); 24428d65729SSalman Qazi int ret = 0; 245b347689fSMing Lei struct request *rq; 246b347689fSMing Lei 247445874e8SMing Lei do { 24828d65729SSalman Qazi if (!list_empty_careful(&hctx->dispatch)) { 24928d65729SSalman Qazi ret = -EAGAIN; 25028d65729SSalman Qazi break; 25128d65729SSalman Qazi } 25228d65729SSalman Qazi 253b347689fSMing Lei if (!sbitmap_any_bit_set(&hctx->ctx_map)) 254b347689fSMing Lei break; 255b347689fSMing Lei 25665c76369SMing Lei if (!blk_mq_get_dispatch_budget(q)) 2571f460b63SMing Lei break; 258b347689fSMing Lei 259b347689fSMing Lei rq = blk_mq_dequeue_from_ctx(hctx, ctx); 260b347689fSMing Lei if (!rq) { 26165c76369SMing Lei blk_mq_put_dispatch_budget(q); 262a0823421SDouglas Anderson /* 263a0823421SDouglas Anderson * We're releasing without dispatching. Holding the 264a0823421SDouglas Anderson * budget could have blocked any "hctx"s with the 265a0823421SDouglas Anderson * same queue and if we didn't dispatch then there's 266a0823421SDouglas Anderson * no guarantee anyone will kick the queue. Kick it 267a0823421SDouglas Anderson * ourselves. 268a0823421SDouglas Anderson */ 269a0823421SDouglas Anderson blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); 270b347689fSMing Lei break; 271b347689fSMing Lei } 272b347689fSMing Lei 273b347689fSMing Lei /* 274b347689fSMing Lei * Now this rq owns the budget which has to be released 275b347689fSMing Lei * if this rq won't be queued to driver via .queue_rq() 276b347689fSMing Lei * in blk_mq_dispatch_rq_list(). 277b347689fSMing Lei */ 278b347689fSMing Lei list_add(&rq->queuelist, &rq_list); 279b347689fSMing Lei 280b347689fSMing Lei /* round robin for fair dispatch */ 281b347689fSMing Lei ctx = blk_mq_next_ctx(hctx, rq->mq_ctx); 282b347689fSMing Lei 2831fd40b5eSMing Lei } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1)); 284b347689fSMing Lei 285b347689fSMing Lei WRITE_ONCE(hctx->dispatch_from, ctx); 28628d65729SSalman Qazi return ret; 287b347689fSMing Lei } 288b347689fSMing Lei 289e1b586f2SZheng Bin static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) 290bd166ef1SJens Axboe { 29181380ca1SOmar Sandoval struct request_queue *q = hctx->queue; 29281380ca1SOmar Sandoval struct elevator_queue *e = q->elevator; 293f9cd4bfeSJens Axboe const bool has_sched_dispatch = e && e->type->ops.dispatch_request; 29428d65729SSalman Qazi int ret = 0; 295bd166ef1SJens Axboe LIST_HEAD(rq_list); 296bd166ef1SJens Axboe 297bd166ef1SJens Axboe /* 298bd166ef1SJens Axboe * If we have previous entries on our dispatch list, grab them first for 299bd166ef1SJens Axboe * more fair dispatch. 300bd166ef1SJens Axboe */ 301bd166ef1SJens Axboe if (!list_empty_careful(&hctx->dispatch)) { 302bd166ef1SJens Axboe spin_lock(&hctx->lock); 303bd166ef1SJens Axboe if (!list_empty(&hctx->dispatch)) 304bd166ef1SJens Axboe list_splice_init(&hctx->dispatch, &rq_list); 305bd166ef1SJens Axboe spin_unlock(&hctx->lock); 306bd166ef1SJens Axboe } 307bd166ef1SJens Axboe 308bd166ef1SJens Axboe /* 309bd166ef1SJens Axboe * Only ask the scheduler for requests, if we didn't have residual 310bd166ef1SJens Axboe * requests from the dispatch list. This is to avoid the case where 311bd166ef1SJens Axboe * we only ever dispatch a fraction of the requests available because 312bd166ef1SJens Axboe * of low device queue depth. Once we pull requests out of the IO 313bd166ef1SJens Axboe * scheduler, we can no longer merge or sort them. So it's best to 314bd166ef1SJens Axboe * leave them there for as long as we can. Mark the hw queue as 315bd166ef1SJens Axboe * needing a restart in that case. 316caf8eb0dSMing Lei * 3175e3d02bbSMing Lei * We want to dispatch from the scheduler if there was nothing 3185e3d02bbSMing Lei * on the dispatch list or we were able to dispatch from the 3195e3d02bbSMing Lei * dispatch list. 32064765a75SJens Axboe */ 321caf8eb0dSMing Lei if (!list_empty(&rq_list)) { 322caf8eb0dSMing Lei blk_mq_sched_mark_restart_hctx(hctx); 3231fd40b5eSMing Lei if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) { 324b347689fSMing Lei if (has_sched_dispatch) 32528d65729SSalman Qazi ret = blk_mq_do_dispatch_sched(hctx); 326b347689fSMing Lei else 32728d65729SSalman Qazi ret = blk_mq_do_dispatch_ctx(hctx); 328b347689fSMing Lei } 329caf8eb0dSMing Lei } else if (has_sched_dispatch) { 33028d65729SSalman Qazi ret = blk_mq_do_dispatch_sched(hctx); 3316e768717SMing Lei } else if (hctx->dispatch_busy) { 3326e768717SMing Lei /* dequeue request one by one from sw queue if queue is busy */ 33328d65729SSalman Qazi ret = blk_mq_do_dispatch_ctx(hctx); 334caf8eb0dSMing Lei } else { 335caf8eb0dSMing Lei blk_mq_flush_busy_ctxs(hctx, &rq_list); 3361fd40b5eSMing Lei blk_mq_dispatch_rq_list(hctx, &rq_list, 0); 337c13660a0SJens Axboe } 33828d65729SSalman Qazi 33928d65729SSalman Qazi return ret; 34028d65729SSalman Qazi } 34128d65729SSalman Qazi 34228d65729SSalman Qazi void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) 34328d65729SSalman Qazi { 34428d65729SSalman Qazi struct request_queue *q = hctx->queue; 34528d65729SSalman Qazi 34628d65729SSalman Qazi /* RCU or SRCU read lock is needed before checking quiesced flag */ 34728d65729SSalman Qazi if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) 34828d65729SSalman Qazi return; 34928d65729SSalman Qazi 35028d65729SSalman Qazi hctx->run++; 35128d65729SSalman Qazi 35228d65729SSalman Qazi /* 35328d65729SSalman Qazi * A return of -EAGAIN is an indication that hctx->dispatch is not 35428d65729SSalman Qazi * empty and we must run again in order to avoid starving flushes. 35528d65729SSalman Qazi */ 35628d65729SSalman Qazi if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) { 35728d65729SSalman Qazi if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) 35828d65729SSalman Qazi blk_mq_run_hw_queue(hctx, true); 35928d65729SSalman Qazi } 360bd166ef1SJens Axboe } 361bd166ef1SJens Axboe 362e4d750c9SJens Axboe bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio, 36314ccb66bSChristoph Hellwig unsigned int nr_segs, struct request **merged_request) 364bd166ef1SJens Axboe { 365bd166ef1SJens Axboe struct request *rq; 366bd166ef1SJens Axboe 36734fe7c05SChristoph Hellwig switch (elv_merge(q, &rq, bio)) { 36834fe7c05SChristoph Hellwig case ELEVATOR_BACK_MERGE: 369bd166ef1SJens Axboe if (!blk_mq_sched_allow_merge(q, rq, bio)) 370bd166ef1SJens Axboe return false; 3717d7ca7c5SBaolin Wang if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 37234fe7c05SChristoph Hellwig return false; 373e4d750c9SJens Axboe *merged_request = attempt_back_merge(q, rq); 374e4d750c9SJens Axboe if (!*merged_request) 37534fe7c05SChristoph Hellwig elv_merged_request(q, rq, ELEVATOR_BACK_MERGE); 376bd166ef1SJens Axboe return true; 37734fe7c05SChristoph Hellwig case ELEVATOR_FRONT_MERGE: 378bd166ef1SJens Axboe if (!blk_mq_sched_allow_merge(q, rq, bio)) 379bd166ef1SJens Axboe return false; 3807d7ca7c5SBaolin Wang if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 38134fe7c05SChristoph Hellwig return false; 382e4d750c9SJens Axboe *merged_request = attempt_front_merge(q, rq); 383e4d750c9SJens Axboe if (!*merged_request) 38434fe7c05SChristoph Hellwig elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE); 385bd166ef1SJens Axboe return true; 386bea99a50SKeith Busch case ELEVATOR_DISCARD_MERGE: 3877d7ca7c5SBaolin Wang return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK; 38834fe7c05SChristoph Hellwig default: 389bd166ef1SJens Axboe return false; 390bd166ef1SJens Axboe } 39134fe7c05SChristoph Hellwig } 392bd166ef1SJens Axboe EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge); 393bd166ef1SJens Axboe 3949bddeb2aSMing Lei /* 3959c558734SJens Axboe * Reverse check our software queue for entries that we could potentially 3969c558734SJens Axboe * merge with. Currently includes a hand-wavy stop count of 8, to not spend 3979c558734SJens Axboe * too much time checking for merges. 3989c558734SJens Axboe */ 3999c558734SJens Axboe static bool blk_mq_attempt_merge(struct request_queue *q, 400c16d6b5aSMing Lei struct blk_mq_hw_ctx *hctx, 40114ccb66bSChristoph Hellwig struct blk_mq_ctx *ctx, struct bio *bio, 40214ccb66bSChristoph Hellwig unsigned int nr_segs) 4039c558734SJens Axboe { 404c16d6b5aSMing Lei enum hctx_type type = hctx->type; 405c16d6b5aSMing Lei 4069c558734SJens Axboe lockdep_assert_held(&ctx->lock); 4079c558734SJens Axboe 408bdc6a287SBaolin Wang if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) { 4099c558734SJens Axboe ctx->rq_merged++; 4109c558734SJens Axboe return true; 4119c558734SJens Axboe } 4129c558734SJens Axboe 4139c558734SJens Axboe return false; 4149c558734SJens Axboe } 4159bddeb2aSMing Lei 41614ccb66bSChristoph Hellwig bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio, 41714ccb66bSChristoph Hellwig unsigned int nr_segs) 418bd166ef1SJens Axboe { 419bd166ef1SJens Axboe struct elevator_queue *e = q->elevator; 420bd166ef1SJens Axboe struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); 4218ccdf4a3SJianchao Wang struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx); 4229bddeb2aSMing Lei bool ret = false; 423c16d6b5aSMing Lei enum hctx_type type; 424bd166ef1SJens Axboe 425c05f4220SBart Van Assche if (e && e->type->ops.bio_merge) 42614ccb66bSChristoph Hellwig return e->type->ops.bio_merge(hctx, bio, nr_segs); 427bd166ef1SJens Axboe 428c16d6b5aSMing Lei type = hctx->type; 429b04f50abSMing Lei if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) && 430c16d6b5aSMing Lei !list_empty_careful(&ctx->rq_lists[type])) { 4319bddeb2aSMing Lei /* default per sw-queue merge */ 4329bddeb2aSMing Lei spin_lock(&ctx->lock); 43314ccb66bSChristoph Hellwig ret = blk_mq_attempt_merge(q, hctx, ctx, bio, nr_segs); 4349bddeb2aSMing Lei spin_unlock(&ctx->lock); 4359bddeb2aSMing Lei } 4369bddeb2aSMing Lei 4379bddeb2aSMing Lei return ret; 438bd166ef1SJens Axboe } 439bd166ef1SJens Axboe 440bd166ef1SJens Axboe bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq) 441bd166ef1SJens Axboe { 442bd166ef1SJens Axboe return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq); 443bd166ef1SJens Axboe } 444bd166ef1SJens Axboe EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge); 445bd166ef1SJens Axboe 446bd166ef1SJens Axboe void blk_mq_sched_request_inserted(struct request *rq) 447bd166ef1SJens Axboe { 448bd166ef1SJens Axboe trace_block_rq_insert(rq->q, rq); 449bd166ef1SJens Axboe } 450bd166ef1SJens Axboe EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted); 451bd166ef1SJens Axboe 4520cacba6cSOmar Sandoval static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx, 453a6a252e6SMing Lei bool has_sched, 4540cacba6cSOmar Sandoval struct request *rq) 455bd166ef1SJens Axboe { 45601e99aecSMing Lei /* 45701e99aecSMing Lei * dispatch flush and passthrough rq directly 45801e99aecSMing Lei * 45901e99aecSMing Lei * passthrough request has to be added to hctx->dispatch directly. 46001e99aecSMing Lei * For some reason, device may be in one situation which can't 46101e99aecSMing Lei * handle FS request, so STS_RESOURCE is always returned and the 46201e99aecSMing Lei * FS request will be added to hctx->dispatch. However passthrough 46301e99aecSMing Lei * request may be required at that time for fixing the problem. If 46401e99aecSMing Lei * passthrough request is added to scheduler queue, there isn't any 46501e99aecSMing Lei * chance to dispatch it given we prioritize requests in hctx->dispatch. 46601e99aecSMing Lei */ 46701e99aecSMing Lei if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq)) 468bd166ef1SJens Axboe return true; 469bd166ef1SJens Axboe 470923218f6SMing Lei if (has_sched) 471a6a252e6SMing Lei rq->rq_flags |= RQF_SORTED; 472a6a252e6SMing Lei 473a6a252e6SMing Lei return false; 474a6a252e6SMing Lei } 475a6a252e6SMing Lei 476bd6737f1SJens Axboe void blk_mq_sched_insert_request(struct request *rq, bool at_head, 4779e97d295SMike Snitzer bool run_queue, bool async) 478bd6737f1SJens Axboe { 479bd6737f1SJens Axboe struct request_queue *q = rq->q; 480bd6737f1SJens Axboe struct elevator_queue *e = q->elevator; 481bd6737f1SJens Axboe struct blk_mq_ctx *ctx = rq->mq_ctx; 482ea4f995eSJens Axboe struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 483bd6737f1SJens Axboe 484a6a252e6SMing Lei /* flush rq in flush machinery need to be dispatched directly */ 485a6a252e6SMing Lei if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) { 486923218f6SMing Lei blk_insert_flush(rq); 487923218f6SMing Lei goto run; 488bd6737f1SJens Axboe } 489bd6737f1SJens Axboe 490923218f6SMing Lei WARN_ON(e && (rq->tag != -1)); 491923218f6SMing Lei 49201e99aecSMing Lei if (blk_mq_sched_bypass_insert(hctx, !!e, rq)) { 493cc3200eaSMing Lei /* 494cc3200eaSMing Lei * Firstly normal IO request is inserted to scheduler queue or 495cc3200eaSMing Lei * sw queue, meantime we add flush request to dispatch queue( 496cc3200eaSMing Lei * hctx->dispatch) directly and there is at most one in-flight 497cc3200eaSMing Lei * flush request for each hw queue, so it doesn't matter to add 498cc3200eaSMing Lei * flush request to tail or front of the dispatch queue. 499cc3200eaSMing Lei * 500cc3200eaSMing Lei * Secondly in case of NCQ, flush request belongs to non-NCQ 501cc3200eaSMing Lei * command, and queueing it will fail when there is any 502cc3200eaSMing Lei * in-flight normal IO request(NCQ command). When adding flush 503cc3200eaSMing Lei * rq to the front of hctx->dispatch, it is easier to introduce 504cc3200eaSMing Lei * extra time to flush rq's latency because of S_SCHED_RESTART 505cc3200eaSMing Lei * compared with adding to the tail of dispatch queue, then 506cc3200eaSMing Lei * chance of flush merge is increased, and less flush requests 507cc3200eaSMing Lei * will be issued to controller. It is observed that ~10% time 508cc3200eaSMing Lei * is saved in blktests block/004 on disk attached to AHCI/NCQ 509cc3200eaSMing Lei * drive when adding flush rq to the front of hctx->dispatch. 510cc3200eaSMing Lei * 511cc3200eaSMing Lei * Simply queue flush rq to the front of hctx->dispatch so that 512cc3200eaSMing Lei * intensive flush workloads can benefit in case of NCQ HW. 513cc3200eaSMing Lei */ 514cc3200eaSMing Lei at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head; 51501e99aecSMing Lei blk_mq_request_bypass_insert(rq, at_head, false); 5160cacba6cSOmar Sandoval goto run; 51701e99aecSMing Lei } 5180cacba6cSOmar Sandoval 519f9cd4bfeSJens Axboe if (e && e->type->ops.insert_requests) { 520bd6737f1SJens Axboe LIST_HEAD(list); 521bd6737f1SJens Axboe 522bd6737f1SJens Axboe list_add(&rq->queuelist, &list); 523f9cd4bfeSJens Axboe e->type->ops.insert_requests(hctx, &list, at_head); 524bd6737f1SJens Axboe } else { 525bd6737f1SJens Axboe spin_lock(&ctx->lock); 526bd6737f1SJens Axboe __blk_mq_insert_request(hctx, rq, at_head); 527bd6737f1SJens Axboe spin_unlock(&ctx->lock); 528bd6737f1SJens Axboe } 529bd6737f1SJens Axboe 5300cacba6cSOmar Sandoval run: 531bd6737f1SJens Axboe if (run_queue) 532bd6737f1SJens Axboe blk_mq_run_hw_queue(hctx, async); 533bd6737f1SJens Axboe } 534bd6737f1SJens Axboe 53567cae4c9SJens Axboe void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx, 536bd6737f1SJens Axboe struct blk_mq_ctx *ctx, 537bd6737f1SJens Axboe struct list_head *list, bool run_queue_async) 538bd6737f1SJens Axboe { 539f9afca4dSJens Axboe struct elevator_queue *e; 540e87eb301SMing Lei struct request_queue *q = hctx->queue; 541e87eb301SMing Lei 542e87eb301SMing Lei /* 543e87eb301SMing Lei * blk_mq_sched_insert_requests() is called from flush plug 544e87eb301SMing Lei * context only, and hold one usage counter to prevent queue 545e87eb301SMing Lei * from being released. 546e87eb301SMing Lei */ 547e87eb301SMing Lei percpu_ref_get(&q->q_usage_counter); 548f9afca4dSJens Axboe 549f9afca4dSJens Axboe e = hctx->queue->elevator; 550f9cd4bfeSJens Axboe if (e && e->type->ops.insert_requests) 551f9cd4bfeSJens Axboe e->type->ops.insert_requests(hctx, list, false); 5526ce3dd6eSMing Lei else { 5536ce3dd6eSMing Lei /* 5546ce3dd6eSMing Lei * try to issue requests directly if the hw queue isn't 5556ce3dd6eSMing Lei * busy in case of 'none' scheduler, and this way may save 5566ce3dd6eSMing Lei * us one extra enqueue & dequeue to sw queue. 5576ce3dd6eSMing Lei */ 558fd9c40f6SBart Van Assche if (!hctx->dispatch_busy && !e && !run_queue_async) { 5596ce3dd6eSMing Lei blk_mq_try_issue_list_directly(hctx, list); 560fd9c40f6SBart Van Assche if (list_empty(list)) 561e87eb301SMing Lei goto out; 562fd9c40f6SBart Van Assche } 563bd6737f1SJens Axboe blk_mq_insert_requests(hctx, ctx, list); 5646ce3dd6eSMing Lei } 565bd6737f1SJens Axboe 566bd6737f1SJens Axboe blk_mq_run_hw_queue(hctx, run_queue_async); 567e87eb301SMing Lei out: 568e87eb301SMing Lei percpu_ref_put(&q->q_usage_counter); 569bd6737f1SJens Axboe } 570bd6737f1SJens Axboe 571bd166ef1SJens Axboe static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set, 572bd166ef1SJens Axboe struct blk_mq_hw_ctx *hctx, 573bd166ef1SJens Axboe unsigned int hctx_idx) 574bd166ef1SJens Axboe { 575bd166ef1SJens Axboe if (hctx->sched_tags) { 576bd166ef1SJens Axboe blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx); 577bd166ef1SJens Axboe blk_mq_free_rq_map(hctx->sched_tags); 578bd166ef1SJens Axboe hctx->sched_tags = NULL; 579bd166ef1SJens Axboe } 580bd166ef1SJens Axboe } 581bd166ef1SJens Axboe 5826917ff0bSOmar Sandoval static int blk_mq_sched_alloc_tags(struct request_queue *q, 5836917ff0bSOmar Sandoval struct blk_mq_hw_ctx *hctx, 5846917ff0bSOmar Sandoval unsigned int hctx_idx) 585bd166ef1SJens Axboe { 586bd166ef1SJens Axboe struct blk_mq_tag_set *set = q->tag_set; 5876917ff0bSOmar Sandoval int ret; 588bd166ef1SJens Axboe 5896917ff0bSOmar Sandoval hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests, 5906917ff0bSOmar Sandoval set->reserved_tags); 591bd166ef1SJens Axboe if (!hctx->sched_tags) 5926917ff0bSOmar Sandoval return -ENOMEM; 5936917ff0bSOmar Sandoval 5946917ff0bSOmar Sandoval ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests); 5956917ff0bSOmar Sandoval if (ret) 5966917ff0bSOmar Sandoval blk_mq_sched_free_tags(set, hctx, hctx_idx); 597bd166ef1SJens Axboe 598bd166ef1SJens Axboe return ret; 599bd166ef1SJens Axboe } 600bd166ef1SJens Axboe 601c3e22192SMing Lei /* called in queue's release handler, tagset has gone away */ 60254d5329dSOmar Sandoval static void blk_mq_sched_tags_teardown(struct request_queue *q) 603bd166ef1SJens Axboe { 604bd166ef1SJens Axboe struct blk_mq_hw_ctx *hctx; 605bd166ef1SJens Axboe int i; 606bd166ef1SJens Axboe 607c3e22192SMing Lei queue_for_each_hw_ctx(q, hctx, i) { 608c3e22192SMing Lei if (hctx->sched_tags) { 609c3e22192SMing Lei blk_mq_free_rq_map(hctx->sched_tags); 610c3e22192SMing Lei hctx->sched_tags = NULL; 611c3e22192SMing Lei } 612c3e22192SMing Lei } 613bd166ef1SJens Axboe } 614d3484991SJens Axboe 6156917ff0bSOmar Sandoval int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e) 6166917ff0bSOmar Sandoval { 6176917ff0bSOmar Sandoval struct blk_mq_hw_ctx *hctx; 618ee056f98SOmar Sandoval struct elevator_queue *eq; 6196917ff0bSOmar Sandoval unsigned int i; 6206917ff0bSOmar Sandoval int ret; 6216917ff0bSOmar Sandoval 6226917ff0bSOmar Sandoval if (!e) { 6236917ff0bSOmar Sandoval q->elevator = NULL; 62432a50fabSMing Lei q->nr_requests = q->tag_set->queue_depth; 6256917ff0bSOmar Sandoval return 0; 6266917ff0bSOmar Sandoval } 6276917ff0bSOmar Sandoval 6286917ff0bSOmar Sandoval /* 62932825c45SMing Lei * Default to double of smaller one between hw queue_depth and 128, 63032825c45SMing Lei * since we don't split into sync/async like the old code did. 63132825c45SMing Lei * Additionally, this is a per-hw queue depth. 6326917ff0bSOmar Sandoval */ 63332825c45SMing Lei q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth, 63432825c45SMing Lei BLKDEV_MAX_RQ); 6356917ff0bSOmar Sandoval 6366917ff0bSOmar Sandoval queue_for_each_hw_ctx(q, hctx, i) { 6376917ff0bSOmar Sandoval ret = blk_mq_sched_alloc_tags(q, hctx, i); 6386917ff0bSOmar Sandoval if (ret) 6396917ff0bSOmar Sandoval goto err; 6406917ff0bSOmar Sandoval } 6416917ff0bSOmar Sandoval 642f9cd4bfeSJens Axboe ret = e->ops.init_sched(q, e); 6436917ff0bSOmar Sandoval if (ret) 6446917ff0bSOmar Sandoval goto err; 6456917ff0bSOmar Sandoval 646d332ce09SOmar Sandoval blk_mq_debugfs_register_sched(q); 647d332ce09SOmar Sandoval 648ee056f98SOmar Sandoval queue_for_each_hw_ctx(q, hctx, i) { 649f9cd4bfeSJens Axboe if (e->ops.init_hctx) { 650f9cd4bfeSJens Axboe ret = e->ops.init_hctx(hctx, i); 651ee056f98SOmar Sandoval if (ret) { 652ee056f98SOmar Sandoval eq = q->elevator; 653c3e22192SMing Lei blk_mq_sched_free_requests(q); 654ee056f98SOmar Sandoval blk_mq_exit_sched(q, eq); 655ee056f98SOmar Sandoval kobject_put(&eq->kobj); 656ee056f98SOmar Sandoval return ret; 657ee056f98SOmar Sandoval } 658ee056f98SOmar Sandoval } 659d332ce09SOmar Sandoval blk_mq_debugfs_register_sched_hctx(q, hctx); 660ee056f98SOmar Sandoval } 661ee056f98SOmar Sandoval 6626917ff0bSOmar Sandoval return 0; 6636917ff0bSOmar Sandoval 6646917ff0bSOmar Sandoval err: 665c3e22192SMing Lei blk_mq_sched_free_requests(q); 66654d5329dSOmar Sandoval blk_mq_sched_tags_teardown(q); 66754d5329dSOmar Sandoval q->elevator = NULL; 6686917ff0bSOmar Sandoval return ret; 6696917ff0bSOmar Sandoval } 6706917ff0bSOmar Sandoval 671c3e22192SMing Lei /* 672c3e22192SMing Lei * called in either blk_queue_cleanup or elevator_switch, tagset 673c3e22192SMing Lei * is required for freeing requests 674c3e22192SMing Lei */ 675c3e22192SMing Lei void blk_mq_sched_free_requests(struct request_queue *q) 676c3e22192SMing Lei { 677c3e22192SMing Lei struct blk_mq_hw_ctx *hctx; 678c3e22192SMing Lei int i; 679c3e22192SMing Lei 680c3e22192SMing Lei queue_for_each_hw_ctx(q, hctx, i) { 681c3e22192SMing Lei if (hctx->sched_tags) 682c3e22192SMing Lei blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i); 683c3e22192SMing Lei } 684c3e22192SMing Lei } 685c3e22192SMing Lei 68654d5329dSOmar Sandoval void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e) 68754d5329dSOmar Sandoval { 688ee056f98SOmar Sandoval struct blk_mq_hw_ctx *hctx; 689ee056f98SOmar Sandoval unsigned int i; 690ee056f98SOmar Sandoval 691ee056f98SOmar Sandoval queue_for_each_hw_ctx(q, hctx, i) { 692d332ce09SOmar Sandoval blk_mq_debugfs_unregister_sched_hctx(hctx); 693f9cd4bfeSJens Axboe if (e->type->ops.exit_hctx && hctx->sched_data) { 694f9cd4bfeSJens Axboe e->type->ops.exit_hctx(hctx, i); 695ee056f98SOmar Sandoval hctx->sched_data = NULL; 696ee056f98SOmar Sandoval } 697ee056f98SOmar Sandoval } 698d332ce09SOmar Sandoval blk_mq_debugfs_unregister_sched(q); 699f9cd4bfeSJens Axboe if (e->type->ops.exit_sched) 700f9cd4bfeSJens Axboe e->type->ops.exit_sched(e); 70154d5329dSOmar Sandoval blk_mq_sched_tags_teardown(q); 70254d5329dSOmar Sandoval q->elevator = NULL; 70354d5329dSOmar Sandoval } 704