brintos / linux-shallow public Read only

1// SPDX-License-Identifier: GPL-2.02/*3 * Block multiqueue core code4 *5 * Copyright (C) 2013-2014 Jens Axboe6 * Copyright (C) 2013-2014 Christoph Hellwig7 */8#include <linux/kernel.h>9#include <linux/module.h>10#include <linux/backing-dev.h>11#include <linux/bio.h>12#include <linux/blkdev.h>13#include <linux/blk-integrity.h>14#include <linux/kmemleak.h>15#include <linux/mm.h>16#include <linux/init.h>17#include <linux/slab.h>18#include <linux/workqueue.h>19#include <linux/smp.h>20#include <linux/interrupt.h>21#include <linux/llist.h>22#include <linux/cpu.h>23#include <linux/cache.h>24#include <linux/sched/topology.h>25#include <linux/sched/signal.h>26#include <linux/delay.h>27#include <linux/crash_dump.h>28#include <linux/prefetch.h>29#include <linux/blk-crypto.h>30#include <linux/part_stat.h>31#include <linux/sched/isolation.h>32 33#include <trace/events/block.h>34 35#include <linux/t10-pi.h>36#include "blk.h"37#include "blk-mq.h"38#include "blk-mq-debugfs.h"39#include "blk-pm.h"40#include "blk-stat.h"41#include "blk-mq-sched.h"42#include "blk-rq-qos.h"43 44static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);45static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd);46 47static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);48static void blk_mq_request_bypass_insert(struct request *rq,49		blk_insert_t flags);50static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,51		struct list_head *list);52static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,53			 struct io_comp_batch *iob, unsigned int flags);54 55/*56 * Check if any of the ctx, dispatch list or elevator57 * have pending work in this hardware queue.58 */59static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)60{61	return !list_empty_careful(&hctx->dispatch) ||62		sbitmap_any_bit_set(&hctx->ctx_map) ||63			blk_mq_sched_has_work(hctx);64}65 66/*67 * Mark this ctx as having pending work in this hardware queue68 */69static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,70				     struct blk_mq_ctx *ctx)71{72	const int bit = ctx->index_hw[hctx->type];73 74	if (!sbitmap_test_bit(&hctx->ctx_map, bit))75		sbitmap_set_bit(&hctx->ctx_map, bit);76}77 78static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,79				      struct blk_mq_ctx *ctx)80{81	const int bit = ctx->index_hw[hctx->type];82 83	sbitmap_clear_bit(&hctx->ctx_map, bit);84}85 86struct mq_inflight {87	struct block_device *part;88	unsigned int inflight[2];89};90 91static bool blk_mq_check_inflight(struct request *rq, void *priv)92{93	struct mq_inflight *mi = priv;94 95	if (rq->part && blk_do_io_stat(rq) &&96	    (!bdev_is_partition(mi->part) || rq->part == mi->part) &&97	    blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)98		mi->inflight[rq_data_dir(rq)]++;99 100	return true;101}102 103unsigned int blk_mq_in_flight(struct request_queue *q,104		struct block_device *part)105{106	struct mq_inflight mi = { .part = part };107 108	blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);109 110	return mi.inflight[0] + mi.inflight[1];111}112 113void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,114		unsigned int inflight[2])115{116	struct mq_inflight mi = { .part = part };117 118	blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);119	inflight[0] = mi.inflight[0];120	inflight[1] = mi.inflight[1];121}122 123void blk_freeze_queue_start(struct request_queue *q)124{125	mutex_lock(&q->mq_freeze_lock);126	if (++q->mq_freeze_depth == 1) {127		percpu_ref_kill(&q->q_usage_counter);128		mutex_unlock(&q->mq_freeze_lock);129		if (queue_is_mq(q))130			blk_mq_run_hw_queues(q, false);131	} else {132		mutex_unlock(&q->mq_freeze_lock);133	}134}135EXPORT_SYMBOL_GPL(blk_freeze_queue_start);136 137void blk_mq_freeze_queue_wait(struct request_queue *q)138{139	wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));140}141EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);142 143int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,144				     unsigned long timeout)145{146	return wait_event_timeout(q->mq_freeze_wq,147					percpu_ref_is_zero(&q->q_usage_counter),148					timeout);149}150EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);151 152/*153 * Guarantee no request is in use, so we can change any data structure of154 * the queue afterward.155 */156void blk_freeze_queue(struct request_queue *q)157{158	/*159	 * In the !blk_mq case we are only calling this to kill the160	 * q_usage_counter, otherwise this increases the freeze depth161	 * and waits for it to return to zero.  For this reason there is162	 * no blk_unfreeze_queue(), and blk_freeze_queue() is not163	 * exported to drivers as the only user for unfreeze is blk_mq.164	 */165	blk_freeze_queue_start(q);166	blk_mq_freeze_queue_wait(q);167}168 169void blk_mq_freeze_queue(struct request_queue *q)170{171	/*172	 * ...just an alias to keep freeze and unfreeze actions balanced173	 * in the blk_mq_* namespace174	 */175	blk_freeze_queue(q);176}177EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);178 179void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)180{181	mutex_lock(&q->mq_freeze_lock);182	if (force_atomic)183		q->q_usage_counter.data->force_atomic = true;184	q->mq_freeze_depth--;185	WARN_ON_ONCE(q->mq_freeze_depth < 0);186	if (!q->mq_freeze_depth) {187		percpu_ref_resurrect(&q->q_usage_counter);188		wake_up_all(&q->mq_freeze_wq);189	}190	mutex_unlock(&q->mq_freeze_lock);191}192 193void blk_mq_unfreeze_queue(struct request_queue *q)194{195	__blk_mq_unfreeze_queue(q, false);196}197EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);198 199/*200 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the201 * mpt3sas driver such that this function can be removed.202 */203void blk_mq_quiesce_queue_nowait(struct request_queue *q)204{205	unsigned long flags;206 207	spin_lock_irqsave(&q->queue_lock, flags);208	if (!q->quiesce_depth++)209		blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);210	spin_unlock_irqrestore(&q->queue_lock, flags);211}212EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);213 214/**215 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done216 * @set: tag_set to wait on217 *218 * Note: it is driver's responsibility for making sure that quiesce has219 * been started on or more of the request_queues of the tag_set.  This220 * function only waits for the quiesce on those request_queues that had221 * the quiesce flag set using blk_mq_quiesce_queue_nowait.222 */223void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)224{225	if (set->flags & BLK_MQ_F_BLOCKING)226		synchronize_srcu(set->srcu);227	else228		synchronize_rcu();229}230EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);231 232/**233 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished234 * @q: request queue.235 *236 * Note: this function does not prevent that the struct request end_io()237 * callback function is invoked. Once this function is returned, we make238 * sure no dispatch can happen until the queue is unquiesced via239 * blk_mq_unquiesce_queue().240 */241void blk_mq_quiesce_queue(struct request_queue *q)242{243	blk_mq_quiesce_queue_nowait(q);244	/* nothing to wait for non-mq queues */245	if (queue_is_mq(q))246		blk_mq_wait_quiesce_done(q->tag_set);247}248EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);249 250/*251 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()252 * @q: request queue.253 *254 * This function recovers queue into the state before quiescing255 * which is done by blk_mq_quiesce_queue.256 */257void blk_mq_unquiesce_queue(struct request_queue *q)258{259	unsigned long flags;260	bool run_queue = false;261 262	spin_lock_irqsave(&q->queue_lock, flags);263	if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {264		;265	} else if (!--q->quiesce_depth) {266		blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);267		run_queue = true;268	}269	spin_unlock_irqrestore(&q->queue_lock, flags);270 271	/* dispatch requests which are inserted during quiescing */272	if (run_queue)273		blk_mq_run_hw_queues(q, true);274}275EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);276 277void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)278{279	struct request_queue *q;280 281	mutex_lock(&set->tag_list_lock);282	list_for_each_entry(q, &set->tag_list, tag_set_list) {283		if (!blk_queue_skip_tagset_quiesce(q))284			blk_mq_quiesce_queue_nowait(q);285	}286	blk_mq_wait_quiesce_done(set);287	mutex_unlock(&set->tag_list_lock);288}289EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);290 291void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)292{293	struct request_queue *q;294 295	mutex_lock(&set->tag_list_lock);296	list_for_each_entry(q, &set->tag_list, tag_set_list) {297		if (!blk_queue_skip_tagset_quiesce(q))298			blk_mq_unquiesce_queue(q);299	}300	mutex_unlock(&set->tag_list_lock);301}302EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);303 304void blk_mq_wake_waiters(struct request_queue *q)305{306	struct blk_mq_hw_ctx *hctx;307	unsigned long i;308 309	queue_for_each_hw_ctx(q, hctx, i)310		if (blk_mq_hw_queue_mapped(hctx))311			blk_mq_tag_wakeup_all(hctx->tags, true);312}313 314void blk_rq_init(struct request_queue *q, struct request *rq)315{316	memset(rq, 0, sizeof(*rq));317 318	INIT_LIST_HEAD(&rq->queuelist);319	rq->q = q;320	rq->__sector = (sector_t) -1;321	INIT_HLIST_NODE(&rq->hash);322	RB_CLEAR_NODE(&rq->rb_node);323	rq->tag = BLK_MQ_NO_TAG;324	rq->internal_tag = BLK_MQ_NO_TAG;325	rq->start_time_ns = blk_time_get_ns();326	rq->part = NULL;327	blk_crypto_rq_set_defaults(rq);328}329EXPORT_SYMBOL(blk_rq_init);330 331/* Set start and alloc time when the allocated request is actually used */332static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)333{334	if (blk_mq_need_time_stamp(rq))335		rq->start_time_ns = blk_time_get_ns();336	else337		rq->start_time_ns = 0;338 339#ifdef CONFIG_BLK_RQ_ALLOC_TIME340	if (blk_queue_rq_alloc_time(rq->q))341		rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;342	else343		rq->alloc_time_ns = 0;344#endif345}346 347static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,348		struct blk_mq_tags *tags, unsigned int tag)349{350	struct blk_mq_ctx *ctx = data->ctx;351	struct blk_mq_hw_ctx *hctx = data->hctx;352	struct request_queue *q = data->q;353	struct request *rq = tags->static_rqs[tag];354 355	rq->q = q;356	rq->mq_ctx = ctx;357	rq->mq_hctx = hctx;358	rq->cmd_flags = data->cmd_flags;359 360	if (data->flags & BLK_MQ_REQ_PM)361		data->rq_flags |= RQF_PM;362	if (blk_queue_io_stat(q))363		data->rq_flags |= RQF_IO_STAT;364	rq->rq_flags = data->rq_flags;365 366	if (data->rq_flags & RQF_SCHED_TAGS) {367		rq->tag = BLK_MQ_NO_TAG;368		rq->internal_tag = tag;369	} else {370		rq->tag = tag;371		rq->internal_tag = BLK_MQ_NO_TAG;372	}373	rq->timeout = 0;374 375	rq->part = NULL;376	rq->io_start_time_ns = 0;377	rq->stats_sectors = 0;378	rq->nr_phys_segments = 0;379	rq->nr_integrity_segments = 0;380	rq->end_io = NULL;381	rq->end_io_data = NULL;382 383	blk_crypto_rq_set_defaults(rq);384	INIT_LIST_HEAD(&rq->queuelist);385	/* tag was already set */386	WRITE_ONCE(rq->deadline, 0);387	req_ref_set(rq, 1);388 389	if (rq->rq_flags & RQF_USE_SCHED) {390		struct elevator_queue *e = data->q->elevator;391 392		INIT_HLIST_NODE(&rq->hash);393		RB_CLEAR_NODE(&rq->rb_node);394 395		if (e->type->ops.prepare_request)396			e->type->ops.prepare_request(rq);397	}398 399	return rq;400}401 402static inline struct request *403__blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)404{405	unsigned int tag, tag_offset;406	struct blk_mq_tags *tags;407	struct request *rq;408	unsigned long tag_mask;409	int i, nr = 0;410 411	tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);412	if (unlikely(!tag_mask))413		return NULL;414 415	tags = blk_mq_tags_from_data(data);416	for (i = 0; tag_mask; i++) {417		if (!(tag_mask & (1UL << i)))418			continue;419		tag = tag_offset + i;420		prefetch(tags->static_rqs[tag]);421		tag_mask &= ~(1UL << i);422		rq = blk_mq_rq_ctx_init(data, tags, tag);423		rq_list_add(data->cached_rq, rq);424		nr++;425	}426	if (!(data->rq_flags & RQF_SCHED_TAGS))427		blk_mq_add_active_requests(data->hctx, nr);428	/* caller already holds a reference, add for remainder */429	percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);430	data->nr_tags -= nr;431 432	return rq_list_pop(data->cached_rq);433}434 435static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)436{437	struct request_queue *q = data->q;438	u64 alloc_time_ns = 0;439	struct request *rq;440	unsigned int tag;441 442	/* alloc_time includes depth and tag waits */443	if (blk_queue_rq_alloc_time(q))444		alloc_time_ns = blk_time_get_ns();445 446	if (data->cmd_flags & REQ_NOWAIT)447		data->flags |= BLK_MQ_REQ_NOWAIT;448 449retry:450	data->ctx = blk_mq_get_ctx(q);451	data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);452 453	if (q->elevator) {454		/*455		 * All requests use scheduler tags when an I/O scheduler is456		 * enabled for the queue.457		 */458		data->rq_flags |= RQF_SCHED_TAGS;459 460		/*461		 * Flush/passthrough requests are special and go directly to the462		 * dispatch list.463		 */464		if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&465		    !blk_op_is_passthrough(data->cmd_flags)) {466			struct elevator_mq_ops *ops = &q->elevator->type->ops;467 468			WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);469 470			data->rq_flags |= RQF_USE_SCHED;471			if (ops->limit_depth)472				ops->limit_depth(data->cmd_flags, data);473		}474	} else {475		blk_mq_tag_busy(data->hctx);476	}477 478	if (data->flags & BLK_MQ_REQ_RESERVED)479		data->rq_flags |= RQF_RESV;480 481	/*482	 * Try batched alloc if we want more than 1 tag.483	 */484	if (data->nr_tags > 1) {485		rq = __blk_mq_alloc_requests_batch(data);486		if (rq) {487			blk_mq_rq_time_init(rq, alloc_time_ns);488			return rq;489		}490		data->nr_tags = 1;491	}492 493	/*494	 * Waiting allocations only fail because of an inactive hctx.  In that495	 * case just retry the hctx assignment and tag allocation as CPU hotplug496	 * should have migrated us to an online CPU by now.497	 */498	tag = blk_mq_get_tag(data);499	if (tag == BLK_MQ_NO_TAG) {500		if (data->flags & BLK_MQ_REQ_NOWAIT)501			return NULL;502		/*503		 * Give up the CPU and sleep for a random short time to504		 * ensure that thread using a realtime scheduling class505		 * are migrated off the CPU, and thus off the hctx that506		 * is going away.507		 */508		msleep(3);509		goto retry;510	}511 512	if (!(data->rq_flags & RQF_SCHED_TAGS))513		blk_mq_inc_active_requests(data->hctx);514	rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);515	blk_mq_rq_time_init(rq, alloc_time_ns);516	return rq;517}518 519static struct request *blk_mq_rq_cache_fill(struct request_queue *q,520					    struct blk_plug *plug,521					    blk_opf_t opf,522					    blk_mq_req_flags_t flags)523{524	struct blk_mq_alloc_data data = {525		.q		= q,526		.flags		= flags,527		.cmd_flags	= opf,528		.nr_tags	= plug->nr_ios,529		.cached_rq	= &plug->cached_rq,530	};531	struct request *rq;532 533	if (blk_queue_enter(q, flags))534		return NULL;535 536	plug->nr_ios = 1;537 538	rq = __blk_mq_alloc_requests(&data);539	if (unlikely(!rq))540		blk_queue_exit(q);541	return rq;542}543 544static struct request *blk_mq_alloc_cached_request(struct request_queue *q,545						   blk_opf_t opf,546						   blk_mq_req_flags_t flags)547{548	struct blk_plug *plug = current->plug;549	struct request *rq;550 551	if (!plug)552		return NULL;553 554	if (rq_list_empty(plug->cached_rq)) {555		if (plug->nr_ios == 1)556			return NULL;557		rq = blk_mq_rq_cache_fill(q, plug, opf, flags);558		if (!rq)559			return NULL;560	} else {561		rq = rq_list_peek(&plug->cached_rq);562		if (!rq || rq->q != q)563			return NULL;564 565		if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)566			return NULL;567		if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))568			return NULL;569 570		plug->cached_rq = rq_list_next(rq);571		blk_mq_rq_time_init(rq, 0);572	}573 574	rq->cmd_flags = opf;575	INIT_LIST_HEAD(&rq->queuelist);576	return rq;577}578 579struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,580		blk_mq_req_flags_t flags)581{582	struct request *rq;583 584	rq = blk_mq_alloc_cached_request(q, opf, flags);585	if (!rq) {586		struct blk_mq_alloc_data data = {587			.q		= q,588			.flags		= flags,589			.cmd_flags	= opf,590			.nr_tags	= 1,591		};592		int ret;593 594		ret = blk_queue_enter(q, flags);595		if (ret)596			return ERR_PTR(ret);597 598		rq = __blk_mq_alloc_requests(&data);599		if (!rq)600			goto out_queue_exit;601	}602	rq->__data_len = 0;603	rq->__sector = (sector_t) -1;604	rq->bio = rq->biotail = NULL;605	return rq;606out_queue_exit:607	blk_queue_exit(q);608	return ERR_PTR(-EWOULDBLOCK);609}610EXPORT_SYMBOL(blk_mq_alloc_request);611 612struct request *blk_mq_alloc_request_hctx(struct request_queue *q,613	blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)614{615	struct blk_mq_alloc_data data = {616		.q		= q,617		.flags		= flags,618		.cmd_flags	= opf,619		.nr_tags	= 1,620	};621	u64 alloc_time_ns = 0;622	struct request *rq;623	unsigned int cpu;624	unsigned int tag;625	int ret;626 627	/* alloc_time includes depth and tag waits */628	if (blk_queue_rq_alloc_time(q))629		alloc_time_ns = blk_time_get_ns();630 631	/*632	 * If the tag allocator sleeps we could get an allocation for a633	 * different hardware context.  No need to complicate the low level634	 * allocator for this for the rare use case of a command tied to635	 * a specific queue.636	 */637	if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||638	    WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))639		return ERR_PTR(-EINVAL);640 641	if (hctx_idx >= q->nr_hw_queues)642		return ERR_PTR(-EIO);643 644	ret = blk_queue_enter(q, flags);645	if (ret)646		return ERR_PTR(ret);647 648	/*649	 * Check if the hardware context is actually mapped to anything.650	 * If not tell the caller that it should skip this queue.651	 */652	ret = -EXDEV;653	data.hctx = xa_load(&q->hctx_table, hctx_idx);654	if (!blk_mq_hw_queue_mapped(data.hctx))655		goto out_queue_exit;656	cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);657	if (cpu >= nr_cpu_ids)658		goto out_queue_exit;659	data.ctx = __blk_mq_get_ctx(q, cpu);660 661	if (q->elevator)662		data.rq_flags |= RQF_SCHED_TAGS;663	else664		blk_mq_tag_busy(data.hctx);665 666	if (flags & BLK_MQ_REQ_RESERVED)667		data.rq_flags |= RQF_RESV;668 669	ret = -EWOULDBLOCK;670	tag = blk_mq_get_tag(&data);671	if (tag == BLK_MQ_NO_TAG)672		goto out_queue_exit;673	if (!(data.rq_flags & RQF_SCHED_TAGS))674		blk_mq_inc_active_requests(data.hctx);675	rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);676	blk_mq_rq_time_init(rq, alloc_time_ns);677	rq->__data_len = 0;678	rq->__sector = (sector_t) -1;679	rq->bio = rq->biotail = NULL;680	return rq;681 682out_queue_exit:683	blk_queue_exit(q);684	return ERR_PTR(ret);685}686EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);687 688static void blk_mq_finish_request(struct request *rq)689{690	struct request_queue *q = rq->q;691 692	blk_zone_finish_request(rq);693 694	if (rq->rq_flags & RQF_USE_SCHED) {695		q->elevator->type->ops.finish_request(rq);696		/*697		 * For postflush request that may need to be698		 * completed twice, we should clear this flag699		 * to avoid double finish_request() on the rq.700		 */701		rq->rq_flags &= ~RQF_USE_SCHED;702	}703}704 705static void __blk_mq_free_request(struct request *rq)706{707	struct request_queue *q = rq->q;708	struct blk_mq_ctx *ctx = rq->mq_ctx;709	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;710	const int sched_tag = rq->internal_tag;711 712	blk_crypto_free_request(rq);713	blk_pm_mark_last_busy(rq);714	rq->mq_hctx = NULL;715 716	if (rq->tag != BLK_MQ_NO_TAG) {717		blk_mq_dec_active_requests(hctx);718		blk_mq_put_tag(hctx->tags, ctx, rq->tag);719	}720	if (sched_tag != BLK_MQ_NO_TAG)721		blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);722	blk_mq_sched_restart(hctx);723	blk_queue_exit(q);724}725 726void blk_mq_free_request(struct request *rq)727{728	struct request_queue *q = rq->q;729 730	blk_mq_finish_request(rq);731 732	if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))733		laptop_io_completion(q->disk->bdi);734 735	rq_qos_done(q, rq);736 737	WRITE_ONCE(rq->state, MQ_RQ_IDLE);738	if (req_ref_put_and_test(rq))739		__blk_mq_free_request(rq);740}741EXPORT_SYMBOL_GPL(blk_mq_free_request);742 743void blk_mq_free_plug_rqs(struct blk_plug *plug)744{745	struct request *rq;746 747	while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)748		blk_mq_free_request(rq);749}750 751void blk_dump_rq_flags(struct request *rq, char *msg)752{753	printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,754		rq->q->disk ? rq->q->disk->disk_name : "?",755		(__force unsigned long long) rq->cmd_flags);756 757	printk(KERN_INFO "  sector %llu, nr/cnr %u/%u\n",758	       (unsigned long long)blk_rq_pos(rq),759	       blk_rq_sectors(rq), blk_rq_cur_sectors(rq));760	printk(KERN_INFO "  bio %p, biotail %p, len %u\n",761	       rq->bio, rq->biotail, blk_rq_bytes(rq));762}763EXPORT_SYMBOL(blk_dump_rq_flags);764 765static void blk_account_io_completion(struct request *req, unsigned int bytes)766{767	if (req->part && blk_do_io_stat(req)) {768		const int sgrp = op_stat_group(req_op(req));769 770		part_stat_lock();771		part_stat_add(req->part, sectors[sgrp], bytes >> 9);772		part_stat_unlock();773	}774}775 776static void blk_print_req_error(struct request *req, blk_status_t status)777{778	printk_ratelimited(KERN_ERR779		"%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "780		"phys_seg %u prio class %u\n",781		blk_status_to_str(status),782		req->q->disk ? req->q->disk->disk_name : "?",783		blk_rq_pos(req), (__force u32)req_op(req),784		blk_op_str(req_op(req)),785		(__force u32)(req->cmd_flags & ~REQ_OP_MASK),786		req->nr_phys_segments,787		IOPRIO_PRIO_CLASS(req->ioprio));788}789 790/*791 * Fully end IO on a request. Does not support partial completions, or792 * errors.793 */794static void blk_complete_request(struct request *req)795{796	const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;797	int total_bytes = blk_rq_bytes(req);798	struct bio *bio = req->bio;799 800	trace_block_rq_complete(req, BLK_STS_OK, total_bytes);801 802	if (!bio)803		return;804 805	if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)806		blk_integrity_complete(req, total_bytes);807 808	/*809	 * Upper layers may call blk_crypto_evict_key() anytime after the last810	 * bio_endio().  Therefore, the keyslot must be released before that.811	 */812	blk_crypto_rq_put_keyslot(req);813 814	blk_account_io_completion(req, total_bytes);815 816	do {817		struct bio *next = bio->bi_next;818 819		/* Completion has already been traced */820		bio_clear_flag(bio, BIO_TRACE_COMPLETION);821 822		blk_zone_update_request_bio(req, bio);823 824		if (!is_flush)825			bio_endio(bio);826		bio = next;827	} while (bio);828 829	/*830	 * Reset counters so that the request stacking driver831	 * can find how many bytes remain in the request832	 * later.833	 */834	if (!req->end_io) {835		req->bio = NULL;836		req->__data_len = 0;837	}838}839 840/**841 * blk_update_request - Complete multiple bytes without completing the request842 * @req:      the request being processed843 * @error:    block status code844 * @nr_bytes: number of bytes to complete for @req845 *846 * Description:847 *     Ends I/O on a number of bytes attached to @req, but doesn't complete848 *     the request structure even if @req doesn't have leftover.849 *     If @req has leftover, sets it up for the next range of segments.850 *851 *     Passing the result of blk_rq_bytes() as @nr_bytes guarantees852 *     %false return from this function.853 *854 * Note:855 *	The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function856 *      except in the consistency check at the end of this function.857 *858 * Return:859 *     %false - this request doesn't have any more data860 *     %true  - this request has more data861 **/862bool blk_update_request(struct request *req, blk_status_t error,863		unsigned int nr_bytes)864{865	bool is_flush = req->rq_flags & RQF_FLUSH_SEQ;866	bool quiet = req->rq_flags & RQF_QUIET;867	int total_bytes;868 869	trace_block_rq_complete(req, error, nr_bytes);870 871	if (!req->bio)872		return false;873 874	if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&875	    error == BLK_STS_OK)876		blk_integrity_complete(req, nr_bytes);877 878	/*879	 * Upper layers may call blk_crypto_evict_key() anytime after the last880	 * bio_endio().  Therefore, the keyslot must be released before that.881	 */882	if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))883		__blk_crypto_rq_put_keyslot(req);884 885	if (unlikely(error && !blk_rq_is_passthrough(req) && !quiet) &&886	    !test_bit(GD_DEAD, &req->q->disk->state)) {887		blk_print_req_error(req, error);888		trace_block_rq_error(req, error, nr_bytes);889	}890 891	blk_account_io_completion(req, nr_bytes);892 893	total_bytes = 0;894	while (req->bio) {895		struct bio *bio = req->bio;896		unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);897 898		if (unlikely(error))899			bio->bi_status = error;900 901		if (bio_bytes == bio->bi_iter.bi_size) {902			req->bio = bio->bi_next;903		} else if (bio_is_zone_append(bio) && error == BLK_STS_OK) {904			/*905			 * Partial zone append completions cannot be supported906			 * as the BIO fragments may end up not being written907			 * sequentially.908			 */909			bio->bi_status = BLK_STS_IOERR;910		}911 912		/* Completion has already been traced */913		bio_clear_flag(bio, BIO_TRACE_COMPLETION);914		if (unlikely(quiet))915			bio_set_flag(bio, BIO_QUIET);916 917		bio_advance(bio, bio_bytes);918 919		/* Don't actually finish bio if it's part of flush sequence */920		if (!bio->bi_iter.bi_size) {921			blk_zone_update_request_bio(req, bio);922			if (!is_flush)923				bio_endio(bio);924		}925 926		total_bytes += bio_bytes;927		nr_bytes -= bio_bytes;928 929		if (!nr_bytes)930			break;931	}932 933	/*934	 * completely done935	 */936	if (!req->bio) {937		/*938		 * Reset counters so that the request stacking driver939		 * can find how many bytes remain in the request940		 * later.941		 */942		req->__data_len = 0;943		return false;944	}945 946	req->__data_len -= total_bytes;947 948	/* update sector only for requests with clear definition of sector */949	if (!blk_rq_is_passthrough(req))950		req->__sector += total_bytes >> 9;951 952	/* mixed attributes always follow the first bio */953	if (req->rq_flags & RQF_MIXED_MERGE) {954		req->cmd_flags &= ~REQ_FAILFAST_MASK;955		req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;956	}957 958	if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {959		/*960		 * If total number of sectors is less than the first segment961		 * size, something has gone terribly wrong.962		 */963		if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {964			blk_dump_rq_flags(req, "request botched");965			req->__data_len = blk_rq_cur_bytes(req);966		}967 968		/* recalculate the number of segments */969		req->nr_phys_segments = blk_recalc_rq_segments(req);970	}971 972	return true;973}974EXPORT_SYMBOL_GPL(blk_update_request);975 976static inline void blk_account_io_done(struct request *req, u64 now)977{978	trace_block_io_done(req);979 980	/*981	 * Account IO completion.  flush_rq isn't accounted as a982	 * normal IO on queueing nor completion.  Accounting the983	 * containing request is enough.984	 */985	if (blk_do_io_stat(req) && req->part &&986	    !(req->rq_flags & RQF_FLUSH_SEQ)) {987		const int sgrp = op_stat_group(req_op(req));988 989		part_stat_lock();990		update_io_ticks(req->part, jiffies, true);991		part_stat_inc(req->part, ios[sgrp]);992		part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);993		part_stat_local_dec(req->part,994				    in_flight[op_is_write(req_op(req))]);995		part_stat_unlock();996	}997}998 999static inline void blk_account_io_start(struct request *req)1000{1001	trace_block_io_start(req);1002 1003	if (blk_do_io_stat(req)) {1004		/*1005		 * All non-passthrough requests are created from a bio with one1006		 * exception: when a flush command that is part of a flush sequence1007		 * generated by the state machine in blk-flush.c is cloned onto the1008		 * lower device by dm-multipath we can get here without a bio.1009		 */1010		if (req->bio)1011			req->part = req->bio->bi_bdev;1012		else1013			req->part = req->q->disk->part0;1014 1015		part_stat_lock();1016		update_io_ticks(req->part, jiffies, false);1017		part_stat_local_inc(req->part,1018				    in_flight[op_is_write(req_op(req))]);1019		part_stat_unlock();1020	}1021}1022 1023static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)1024{1025	if (rq->rq_flags & RQF_STATS)1026		blk_stat_add(rq, now);1027 1028	blk_mq_sched_completed_request(rq, now);1029	blk_account_io_done(rq, now);1030}1031 1032inline void __blk_mq_end_request(struct request *rq, blk_status_t error)1033{1034	if (blk_mq_need_time_stamp(rq))1035		__blk_mq_end_request_acct(rq, blk_time_get_ns());1036 1037	blk_mq_finish_request(rq);1038 1039	if (rq->end_io) {1040		rq_qos_done(rq->q, rq);1041		if (rq->end_io(rq, error) == RQ_END_IO_FREE)1042			blk_mq_free_request(rq);1043	} else {1044		blk_mq_free_request(rq);1045	}1046}1047EXPORT_SYMBOL(__blk_mq_end_request);1048 1049void blk_mq_end_request(struct request *rq, blk_status_t error)1050{1051	if (blk_update_request(rq, error, blk_rq_bytes(rq)))1052		BUG();1053	__blk_mq_end_request(rq, error);1054}1055EXPORT_SYMBOL(blk_mq_end_request);1056 1057#define TAG_COMP_BATCH		321058 1059static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,1060					  int *tag_array, int nr_tags)1061{1062	struct request_queue *q = hctx->queue;1063 1064	blk_mq_sub_active_requests(hctx, nr_tags);1065 1066	blk_mq_put_tags(hctx->tags, tag_array, nr_tags);1067	percpu_ref_put_many(&q->q_usage_counter, nr_tags);1068}1069 1070void blk_mq_end_request_batch(struct io_comp_batch *iob)1071{1072	int tags[TAG_COMP_BATCH], nr_tags = 0;1073	struct blk_mq_hw_ctx *cur_hctx = NULL;1074	struct request *rq;1075	u64 now = 0;1076 1077	if (iob->need_ts)1078		now = blk_time_get_ns();1079 1080	while ((rq = rq_list_pop(&iob->req_list)) != NULL) {1081		prefetch(rq->bio);1082		prefetch(rq->rq_next);1083 1084		blk_complete_request(rq);1085		if (iob->need_ts)1086			__blk_mq_end_request_acct(rq, now);1087 1088		blk_mq_finish_request(rq);1089 1090		rq_qos_done(rq->q, rq);1091 1092		/*1093		 * If end_io handler returns NONE, then it still has1094		 * ownership of the request.1095		 */1096		if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)1097			continue;1098 1099		WRITE_ONCE(rq->state, MQ_RQ_IDLE);1100		if (!req_ref_put_and_test(rq))1101			continue;1102 1103		blk_crypto_free_request(rq);1104		blk_pm_mark_last_busy(rq);1105 1106		if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {1107			if (cur_hctx)1108				blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);1109			nr_tags = 0;1110			cur_hctx = rq->mq_hctx;1111		}1112		tags[nr_tags++] = rq->tag;1113	}1114 1115	if (nr_tags)1116		blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);1117}1118EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);1119 1120static void blk_complete_reqs(struct llist_head *list)1121{1122	struct llist_node *entry = llist_reverse_order(llist_del_all(list));1123	struct request *rq, *next;1124 1125	llist_for_each_entry_safe(rq, next, entry, ipi_list)1126		rq->q->mq_ops->complete(rq);1127}1128 1129static __latent_entropy void blk_done_softirq(void)1130{1131	blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));1132}1133 1134static int blk_softirq_cpu_dead(unsigned int cpu)1135{1136	blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));1137	return 0;1138}1139 1140static void __blk_mq_complete_request_remote(void *data)1141{1142	__raise_softirq_irqoff(BLOCK_SOFTIRQ);1143}1144 1145static inline bool blk_mq_complete_need_ipi(struct request *rq)1146{1147	int cpu = raw_smp_processor_id();1148 1149	if (!IS_ENABLED(CONFIG_SMP) ||1150	    !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))1151		return false;1152	/*1153	 * With force threaded interrupts enabled, raising softirq from an SMP1154	 * function call will always result in waking the ksoftirqd thread.1155	 * This is probably worse than completing the request on a different1156	 * cache domain.1157	 */1158	if (force_irqthreads())1159		return false;1160 1161	/* same CPU or cache domain and capacity?  Complete locally */1162	if (cpu == rq->mq_ctx->cpu ||1163	    (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&1164	     cpus_share_cache(cpu, rq->mq_ctx->cpu) &&1165	     cpus_equal_capacity(cpu, rq->mq_ctx->cpu)))1166		return false;1167 1168	/* don't try to IPI to an offline CPU */1169	return cpu_online(rq->mq_ctx->cpu);1170}1171 1172static void blk_mq_complete_send_ipi(struct request *rq)1173{1174	unsigned int cpu;1175 1176	cpu = rq->mq_ctx->cpu;1177	if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))1178		smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));1179}1180 1181static void blk_mq_raise_softirq(struct request *rq)1182{1183	struct llist_head *list;1184 1185	preempt_disable();1186	list = this_cpu_ptr(&blk_cpu_done);1187	if (llist_add(&rq->ipi_list, list))1188		raise_softirq(BLOCK_SOFTIRQ);1189	preempt_enable();1190}1191 1192bool blk_mq_complete_request_remote(struct request *rq)1193{1194	WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);1195 1196	/*1197	 * For request which hctx has only one ctx mapping,1198	 * or a polled request, always complete locally,1199	 * it's pointless to redirect the completion.1200	 */1201	if ((rq->mq_hctx->nr_ctx == 1 &&1202	     rq->mq_ctx->cpu == raw_smp_processor_id()) ||1203	     rq->cmd_flags & REQ_POLLED)1204		return false;1205 1206	if (blk_mq_complete_need_ipi(rq)) {1207		blk_mq_complete_send_ipi(rq);1208		return true;1209	}1210 1211	if (rq->q->nr_hw_queues == 1) {1212		blk_mq_raise_softirq(rq);1213		return true;1214	}1215	return false;1216}1217EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);1218 1219/**1220 * blk_mq_complete_request - end I/O on a request1221 * @rq:		the request being processed1222 *1223 * Description:1224 *	Complete a request by scheduling the ->complete_rq operation.1225 **/1226void blk_mq_complete_request(struct request *rq)1227{1228	if (!blk_mq_complete_request_remote(rq))1229		rq->q->mq_ops->complete(rq);1230}1231EXPORT_SYMBOL(blk_mq_complete_request);1232 1233/**1234 * blk_mq_start_request - Start processing a request1235 * @rq: Pointer to request to be started1236 *1237 * Function used by device drivers to notify the block layer that a request1238 * is going to be processed now, so blk layer can do proper initializations1239 * such as starting the timeout timer.1240 */1241void blk_mq_start_request(struct request *rq)1242{1243	struct request_queue *q = rq->q;1244 1245	trace_block_rq_issue(rq);1246 1247	if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) &&1248	    !blk_rq_is_passthrough(rq)) {1249		rq->io_start_time_ns = blk_time_get_ns();1250		rq->stats_sectors = blk_rq_sectors(rq);1251		rq->rq_flags |= RQF_STATS;1252		rq_qos_issue(q, rq);1253	}1254 1255	WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);1256 1257	blk_add_timer(rq);1258	WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);1259	rq->mq_hctx->tags->rqs[rq->tag] = rq;1260 1261	if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)1262		blk_integrity_prepare(rq);1263 1264	if (rq->bio && rq->bio->bi_opf & REQ_POLLED)1265	        WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);1266}1267EXPORT_SYMBOL(blk_mq_start_request);1268 1269/*1270 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple1271 * queues. This is important for md arrays to benefit from merging1272 * requests.1273 */1274static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)1275{1276	if (plug->multiple_queues)1277		return BLK_MAX_REQUEST_COUNT * 2;1278	return BLK_MAX_REQUEST_COUNT;1279}1280 1281static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)1282{1283	struct request *last = rq_list_peek(&plug->mq_list);1284 1285	if (!plug->rq_count) {1286		trace_block_plug(rq->q);1287	} else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||1288		   (!blk_queue_nomerges(rq->q) &&1289		    blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {1290		blk_mq_flush_plug_list(plug, false);1291		last = NULL;1292		trace_block_plug(rq->q);1293	}1294 1295	if (!plug->multiple_queues && last && last->q != rq->q)1296		plug->multiple_queues = true;1297	/*1298	 * Any request allocated from sched tags can't be issued to1299	 * ->queue_rqs() directly1300	 */1301	if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))1302		plug->has_elevator = true;1303	rq->rq_next = NULL;1304	rq_list_add(&plug->mq_list, rq);1305	plug->rq_count++;1306}1307 1308/**1309 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution1310 * @rq:		request to insert1311 * @at_head:    insert request at head or tail of queue1312 *1313 * Description:1314 *    Insert a fully prepared request at the back of the I/O scheduler queue1315 *    for execution.  Don't wait for completion.1316 *1317 * Note:1318 *    This function will invoke @done directly if the queue is dead.1319 */1320void blk_execute_rq_nowait(struct request *rq, bool at_head)1321{1322	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;1323 1324	WARN_ON(irqs_disabled());1325	WARN_ON(!blk_rq_is_passthrough(rq));1326 1327	blk_account_io_start(rq);1328 1329	if (current->plug && !at_head) {1330		blk_add_rq_to_plug(current->plug, rq);1331		return;1332	}1333 1334	blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);1335	blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);1336}1337EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);1338 1339struct blk_rq_wait {1340	struct completion done;1341	blk_status_t ret;1342};1343 1344static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)1345{1346	struct blk_rq_wait *wait = rq->end_io_data;1347 1348	wait->ret = ret;1349	complete(&wait->done);1350	return RQ_END_IO_NONE;1351}1352 1353bool blk_rq_is_poll(struct request *rq)1354{1355	if (!rq->mq_hctx)1356		return false;1357	if (rq->mq_hctx->type != HCTX_TYPE_POLL)1358		return false;1359	return true;1360}1361EXPORT_SYMBOL_GPL(blk_rq_is_poll);1362 1363static void blk_rq_poll_completion(struct request *rq, struct completion *wait)1364{1365	do {1366		blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);1367		cond_resched();1368	} while (!completion_done(wait));1369}1370 1371/**1372 * blk_execute_rq - insert a request into queue for execution1373 * @rq:		request to insert1374 * @at_head:    insert request at head or tail of queue1375 *1376 * Description:1377 *    Insert a fully prepared request at the back of the I/O scheduler queue1378 *    for execution and wait for completion.1379 * Return: The blk_status_t result provided to blk_mq_end_request().1380 */1381blk_status_t blk_execute_rq(struct request *rq, bool at_head)1382{1383	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;1384	struct blk_rq_wait wait = {1385		.done = COMPLETION_INITIALIZER_ONSTACK(wait.done),1386	};1387 1388	WARN_ON(irqs_disabled());1389	WARN_ON(!blk_rq_is_passthrough(rq));1390 1391	rq->end_io_data = &wait;1392	rq->end_io = blk_end_sync_rq;1393 1394	blk_account_io_start(rq);1395	blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);1396	blk_mq_run_hw_queue(hctx, false);1397 1398	if (blk_rq_is_poll(rq))1399		blk_rq_poll_completion(rq, &wait.done);1400	else1401		blk_wait_io(&wait.done);1402 1403	return wait.ret;1404}1405EXPORT_SYMBOL(blk_execute_rq);1406 1407static void __blk_mq_requeue_request(struct request *rq)1408{1409	struct request_queue *q = rq->q;1410 1411	blk_mq_put_driver_tag(rq);1412 1413	trace_block_rq_requeue(rq);1414	rq_qos_requeue(q, rq);1415 1416	if (blk_mq_request_started(rq)) {1417		WRITE_ONCE(rq->state, MQ_RQ_IDLE);1418		rq->rq_flags &= ~RQF_TIMED_OUT;1419	}1420}1421 1422void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)1423{1424	struct request_queue *q = rq->q;1425	unsigned long flags;1426 1427	__blk_mq_requeue_request(rq);1428 1429	/* this request will be re-inserted to io scheduler queue */1430	blk_mq_sched_requeue_request(rq);1431 1432	spin_lock_irqsave(&q->requeue_lock, flags);1433	list_add_tail(&rq->queuelist, &q->requeue_list);1434	spin_unlock_irqrestore(&q->requeue_lock, flags);1435 1436	if (kick_requeue_list)1437		blk_mq_kick_requeue_list(q);1438}1439EXPORT_SYMBOL(blk_mq_requeue_request);1440 1441static void blk_mq_requeue_work(struct work_struct *work)1442{1443	struct request_queue *q =1444		container_of(work, struct request_queue, requeue_work.work);1445	LIST_HEAD(rq_list);1446	LIST_HEAD(flush_list);1447	struct request *rq;1448 1449	spin_lock_irq(&q->requeue_lock);1450	list_splice_init(&q->requeue_list, &rq_list);1451	list_splice_init(&q->flush_list, &flush_list);1452	spin_unlock_irq(&q->requeue_lock);1453 1454	while (!list_empty(&rq_list)) {1455		rq = list_entry(rq_list.next, struct request, queuelist);1456		/*1457		 * If RQF_DONTPREP ist set, the request has been started by the1458		 * driver already and might have driver-specific data allocated1459		 * already.  Insert it into the hctx dispatch list to avoid1460		 * block layer merges for the request.1461		 */1462		if (rq->rq_flags & RQF_DONTPREP) {1463			list_del_init(&rq->queuelist);1464			blk_mq_request_bypass_insert(rq, 0);1465		} else {1466			list_del_init(&rq->queuelist);1467			blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);1468		}1469	}1470 1471	while (!list_empty(&flush_list)) {1472		rq = list_entry(flush_list.next, struct request, queuelist);1473		list_del_init(&rq->queuelist);1474		blk_mq_insert_request(rq, 0);1475	}1476 1477	blk_mq_run_hw_queues(q, false);1478}1479 1480void blk_mq_kick_requeue_list(struct request_queue *q)1481{1482	kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);1483}1484EXPORT_SYMBOL(blk_mq_kick_requeue_list);1485 1486void blk_mq_delay_kick_requeue_list(struct request_queue *q,1487				    unsigned long msecs)1488{1489	kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,1490				    msecs_to_jiffies(msecs));1491}1492EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);1493 1494static bool blk_is_flush_data_rq(struct request *rq)1495{1496	return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);1497}1498 1499static bool blk_mq_rq_inflight(struct request *rq, void *priv)1500{1501	/*1502	 * If we find a request that isn't idle we know the queue is busy1503	 * as it's checked in the iter.1504	 * Return false to stop the iteration.1505	 *1506	 * In case of queue quiesce, if one flush data request is completed,1507	 * don't count it as inflight given the flush sequence is suspended,1508	 * and the original flush data request is invisible to driver, just1509	 * like other pending requests because of quiesce1510	 */1511	if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&1512				blk_is_flush_data_rq(rq) &&1513				blk_mq_request_completed(rq))) {1514		bool *busy = priv;1515 1516		*busy = true;1517		return false;1518	}1519 1520	return true;1521}1522 1523bool blk_mq_queue_inflight(struct request_queue *q)1524{1525	bool busy = false;1526 1527	blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);1528	return busy;1529}1530EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);1531 1532static void blk_mq_rq_timed_out(struct request *req)1533{1534	req->rq_flags |= RQF_TIMED_OUT;1535	if (req->q->mq_ops->timeout) {1536		enum blk_eh_timer_return ret;1537 1538		ret = req->q->mq_ops->timeout(req);1539		if (ret == BLK_EH_DONE)1540			return;1541		WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);1542	}1543 1544	blk_add_timer(req);1545}1546 1547struct blk_expired_data {1548	bool has_timedout_rq;1549	unsigned long next;1550	unsigned long timeout_start;1551};1552 1553static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)1554{1555	unsigned long deadline;1556 1557	if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)1558		return false;1559	if (rq->rq_flags & RQF_TIMED_OUT)1560		return false;1561 1562	deadline = READ_ONCE(rq->deadline);1563	if (time_after_eq(expired->timeout_start, deadline))1564		return true;1565 1566	if (expired->next == 0)1567		expired->next = deadline;1568	else if (time_after(expired->next, deadline))1569		expired->next = deadline;1570	return false;1571}1572 1573void blk_mq_put_rq_ref(struct request *rq)1574{1575	if (is_flush_rq(rq)) {1576		if (rq->end_io(rq, 0) == RQ_END_IO_FREE)1577			blk_mq_free_request(rq);1578	} else if (req_ref_put_and_test(rq)) {1579		__blk_mq_free_request(rq);1580	}1581}1582 1583static bool blk_mq_check_expired(struct request *rq, void *priv)1584{1585	struct blk_expired_data *expired = priv;1586 1587	/*1588	 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot1589	 * be reallocated underneath the timeout handler's processing, then1590	 * the expire check is reliable. If the request is not expired, then1591	 * it was completed and reallocated as a new request after returning1592	 * from blk_mq_check_expired().1593	 */1594	if (blk_mq_req_expired(rq, expired)) {1595		expired->has_timedout_rq = true;1596		return false;1597	}1598	return true;1599}1600 1601static bool blk_mq_handle_expired(struct request *rq, void *priv)1602{1603	struct blk_expired_data *expired = priv;1604 1605	if (blk_mq_req_expired(rq, expired))1606		blk_mq_rq_timed_out(rq);1607	return true;1608}1609 1610static void blk_mq_timeout_work(struct work_struct *work)1611{1612	struct request_queue *q =1613		container_of(work, struct request_queue, timeout_work);1614	struct blk_expired_data expired = {1615		.timeout_start = jiffies,1616	};1617	struct blk_mq_hw_ctx *hctx;1618	unsigned long i;1619 1620	/* A deadlock might occur if a request is stuck requiring a1621	 * timeout at the same time a queue freeze is waiting1622	 * completion, since the timeout code would not be able to1623	 * acquire the queue reference here.1624	 *1625	 * That's why we don't use blk_queue_enter here; instead, we use1626	 * percpu_ref_tryget directly, because we need to be able to1627	 * obtain a reference even in the short window between the queue1628	 * starting to freeze, by dropping the first reference in1629	 * blk_freeze_queue_start, and the moment the last request is1630	 * consumed, marked by the instant q_usage_counter reaches1631	 * zero.1632	 */1633	if (!percpu_ref_tryget(&q->q_usage_counter))1634		return;1635 1636	/* check if there is any timed-out request */1637	blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);1638	if (expired.has_timedout_rq) {1639		/*1640		 * Before walking tags, we must ensure any submit started1641		 * before the current time has finished. Since the submit1642		 * uses srcu or rcu, wait for a synchronization point to1643		 * ensure all running submits have finished1644		 */1645		blk_mq_wait_quiesce_done(q->tag_set);1646 1647		expired.next = 0;1648		blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);1649	}1650 1651	if (expired.next != 0) {1652		mod_timer(&q->timeout, expired.next);1653	} else {1654		/*1655		 * Request timeouts are handled as a forward rolling timer. If1656		 * we end up here it means that no requests are pending and1657		 * also that no request has been pending for a while. Mark1658		 * each hctx as idle.1659		 */1660		queue_for_each_hw_ctx(q, hctx, i) {1661			/* the hctx may be unmapped, so check it here */1662			if (blk_mq_hw_queue_mapped(hctx))1663				blk_mq_tag_idle(hctx);1664		}1665	}1666	blk_queue_exit(q);1667}1668 1669struct flush_busy_ctx_data {1670	struct blk_mq_hw_ctx *hctx;1671	struct list_head *list;1672};1673 1674static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)1675{1676	struct flush_busy_ctx_data *flush_data = data;1677	struct blk_mq_hw_ctx *hctx = flush_data->hctx;1678	struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];1679	enum hctx_type type = hctx->type;1680 1681	spin_lock(&ctx->lock);1682	list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);1683	sbitmap_clear_bit(sb, bitnr);1684	spin_unlock(&ctx->lock);1685	return true;1686}1687 1688/*1689 * Process software queues that have been marked busy, splicing them1690 * to the for-dispatch1691 */1692void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)1693{1694	struct flush_busy_ctx_data data = {1695		.hctx = hctx,1696		.list = list,1697	};1698 1699	sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);1700}1701EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);1702 1703struct dispatch_rq_data {1704	struct blk_mq_hw_ctx *hctx;1705	struct request *rq;1706};1707 1708static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,1709		void *data)1710{1711	struct dispatch_rq_data *dispatch_data = data;1712	struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;1713	struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];1714	enum hctx_type type = hctx->type;1715 1716	spin_lock(&ctx->lock);1717	if (!list_empty(&ctx->rq_lists[type])) {1718		dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);1719		list_del_init(&dispatch_data->rq->queuelist);1720		if (list_empty(&ctx->rq_lists[type]))1721			sbitmap_clear_bit(sb, bitnr);1722	}1723	spin_unlock(&ctx->lock);1724 1725	return !dispatch_data->rq;1726}1727 1728struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,1729					struct blk_mq_ctx *start)1730{1731	unsigned off = start ? start->index_hw[hctx->type] : 0;1732	struct dispatch_rq_data data = {1733		.hctx = hctx,1734		.rq   = NULL,1735	};1736 1737	__sbitmap_for_each_set(&hctx->ctx_map, off,1738			       dispatch_rq_from_ctx, &data);1739 1740	return data.rq;1741}1742 1743bool __blk_mq_alloc_driver_tag(struct request *rq)1744{1745	struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;1746	unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;1747	int tag;1748 1749	blk_mq_tag_busy(rq->mq_hctx);1750 1751	if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {1752		bt = &rq->mq_hctx->tags->breserved_tags;1753		tag_offset = 0;1754	} else {1755		if (!hctx_may_queue(rq->mq_hctx, bt))1756			return false;1757	}1758 1759	tag = __sbitmap_queue_get(bt);1760	if (tag == BLK_MQ_NO_TAG)1761		return false;1762 1763	rq->tag = tag + tag_offset;1764	blk_mq_inc_active_requests(rq->mq_hctx);1765	return true;1766}1767 1768static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,1769				int flags, void *key)1770{1771	struct blk_mq_hw_ctx *hctx;1772 1773	hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);1774 1775	spin_lock(&hctx->dispatch_wait_lock);1776	if (!list_empty(&wait->entry)) {1777		struct sbitmap_queue *sbq;1778 1779		list_del_init(&wait->entry);1780		sbq = &hctx->tags->bitmap_tags;1781		atomic_dec(&sbq->ws_active);1782	}1783	spin_unlock(&hctx->dispatch_wait_lock);1784 1785	blk_mq_run_hw_queue(hctx, true);1786	return 1;1787}1788 1789/*1790 * Mark us waiting for a tag. For shared tags, this involves hooking us into1791 * the tag wakeups. For non-shared tags, we can simply mark us needing a1792 * restart. For both cases, take care to check the condition again after1793 * marking us as waiting.1794 */1795static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,1796				 struct request *rq)1797{1798	struct sbitmap_queue *sbq;1799	struct wait_queue_head *wq;1800	wait_queue_entry_t *wait;1801	bool ret;1802 1803	if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&1804	    !(blk_mq_is_shared_tags(hctx->flags))) {1805		blk_mq_sched_mark_restart_hctx(hctx);1806 1807		/*1808		 * It's possible that a tag was freed in the window between the1809		 * allocation failure and adding the hardware queue to the wait1810		 * queue.1811		 *1812		 * Don't clear RESTART here, someone else could have set it.1813		 * At most this will cost an extra queue run.1814		 */1815		return blk_mq_get_driver_tag(rq);1816	}1817 1818	wait = &hctx->dispatch_wait;1819	if (!list_empty_careful(&wait->entry))1820		return false;1821 1822	if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))1823		sbq = &hctx->tags->breserved_tags;1824	else1825		sbq = &hctx->tags->bitmap_tags;1826	wq = &bt_wait_ptr(sbq, hctx)->wait;1827 1828	spin_lock_irq(&wq->lock);1829	spin_lock(&hctx->dispatch_wait_lock);1830	if (!list_empty(&wait->entry)) {1831		spin_unlock(&hctx->dispatch_wait_lock);1832		spin_unlock_irq(&wq->lock);1833		return false;1834	}1835 1836	atomic_inc(&sbq->ws_active);1837	wait->flags &= ~WQ_FLAG_EXCLUSIVE;1838	__add_wait_queue(wq, wait);1839 1840	/*1841	 * Add one explicit barrier since blk_mq_get_driver_tag() may1842	 * not imply barrier in case of failure.1843	 *1844	 * Order adding us to wait queue and allocating driver tag.1845	 *1846	 * The pair is the one implied in sbitmap_queue_wake_up() which1847	 * orders clearing sbitmap tag bits and waitqueue_active() in1848	 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless1849	 *1850	 * Otherwise, re-order of adding wait queue and getting driver tag1851	 * may cause __sbitmap_queue_wake_up() to wake up nothing because1852	 * the waitqueue_active() may not observe us in wait queue.1853	 */1854	smp_mb();1855 1856	/*1857	 * It's possible that a tag was freed in the window between the1858	 * allocation failure and adding the hardware queue to the wait1859	 * queue.1860	 */1861	ret = blk_mq_get_driver_tag(rq);1862	if (!ret) {1863		spin_unlock(&hctx->dispatch_wait_lock);1864		spin_unlock_irq(&wq->lock);1865		return false;1866	}1867 1868	/*1869	 * We got a tag, remove ourselves from the wait queue to ensure1870	 * someone else gets the wakeup.1871	 */1872	list_del_init(&wait->entry);1873	atomic_dec(&sbq->ws_active);1874	spin_unlock(&hctx->dispatch_wait_lock);1875	spin_unlock_irq(&wq->lock);1876 1877	return true;1878}1879 1880#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  81881#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  41882/*1883 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):1884 * - EWMA is one simple way to compute running average value1885 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially1886 * - take 4 as factor for avoiding to get too small(0) result, and this1887 *   factor doesn't matter because EWMA decreases exponentially1888 */1889static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)1890{1891	unsigned int ewma;1892 1893	ewma = hctx->dispatch_busy;1894 1895	if (!ewma && !busy)1896		return;1897 1898	ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;1899	if (busy)1900		ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;1901	ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;1902 1903	hctx->dispatch_busy = ewma;1904}1905 1906#define BLK_MQ_RESOURCE_DELAY	3		/* ms units */1907 1908static void blk_mq_handle_dev_resource(struct request *rq,1909				       struct list_head *list)1910{1911	list_add(&rq->queuelist, list);1912	__blk_mq_requeue_request(rq);1913}1914 1915enum prep_dispatch {1916	PREP_DISPATCH_OK,1917	PREP_DISPATCH_NO_TAG,1918	PREP_DISPATCH_NO_BUDGET,1919};1920 1921static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,1922						  bool need_budget)1923{1924	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;1925	int budget_token = -1;1926 1927	if (need_budget) {1928		budget_token = blk_mq_get_dispatch_budget(rq->q);1929		if (budget_token < 0) {1930			blk_mq_put_driver_tag(rq);1931			return PREP_DISPATCH_NO_BUDGET;1932		}1933		blk_mq_set_rq_budget_token(rq, budget_token);1934	}1935 1936	if (!blk_mq_get_driver_tag(rq)) {1937		/*1938		 * The initial allocation attempt failed, so we need to1939		 * rerun the hardware queue when a tag is freed. The1940		 * waitqueue takes care of that. If the queue is run1941		 * before we add this entry back on the dispatch list,1942		 * we'll re-run it below.1943		 */1944		if (!blk_mq_mark_tag_wait(hctx, rq)) {1945			/*1946			 * All budgets not got from this function will be put1947			 * together during handling partial dispatch1948			 */1949			if (need_budget)1950				blk_mq_put_dispatch_budget(rq->q, budget_token);1951			return PREP_DISPATCH_NO_TAG;1952		}1953	}1954 1955	return PREP_DISPATCH_OK;1956}1957 1958/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */1959static void blk_mq_release_budgets(struct request_queue *q,1960		struct list_head *list)1961{1962	struct request *rq;1963 1964	list_for_each_entry(rq, list, queuelist) {1965		int budget_token = blk_mq_get_rq_budget_token(rq);1966 1967		if (budget_token >= 0)1968			blk_mq_put_dispatch_budget(q, budget_token);1969	}1970}1971 1972/*1973 * blk_mq_commit_rqs will notify driver using bd->last that there is no1974 * more requests. (See comment in struct blk_mq_ops for commit_rqs for1975 * details)1976 * Attention, we should explicitly call this in unusual cases:1977 *  1) did not queue everything initially scheduled to queue1978 *  2) the last attempt to queue a request failed1979 */1980static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,1981			      bool from_schedule)1982{1983	if (hctx->queue->mq_ops->commit_rqs && queued) {1984		trace_block_unplug(hctx->queue, queued, !from_schedule);1985		hctx->queue->mq_ops->commit_rqs(hctx);1986	}1987}1988 1989/*1990 * Returns true if we did some work AND can potentially do more.1991 */1992bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,1993			     unsigned int nr_budgets)1994{1995	enum prep_dispatch prep;1996	struct request_queue *q = hctx->queue;1997	struct request *rq;1998	int queued;1999	blk_status_t ret = BLK_STS_OK;2000	bool needs_resource = false;2001 2002	if (list_empty(list))2003		return false;2004 2005	/*2006	 * Now process all the entries, sending them to the driver.2007	 */2008	queued = 0;2009	do {2010		struct blk_mq_queue_data bd;2011 2012		rq = list_first_entry(list, struct request, queuelist);2013 2014		WARN_ON_ONCE(hctx != rq->mq_hctx);2015		prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);2016		if (prep != PREP_DISPATCH_OK)2017			break;2018 2019		list_del_init(&rq->queuelist);2020 2021		bd.rq = rq;2022		bd.last = list_empty(list);2023 2024		/*2025		 * once the request is queued to lld, no need to cover the2026		 * budget any more2027		 */2028		if (nr_budgets)2029			nr_budgets--;2030		ret = q->mq_ops->queue_rq(hctx, &bd);2031		switch (ret) {2032		case BLK_STS_OK:2033			queued++;2034			break;2035		case BLK_STS_RESOURCE:2036			needs_resource = true;2037			fallthrough;2038		case BLK_STS_DEV_RESOURCE:2039			blk_mq_handle_dev_resource(rq, list);2040			goto out;2041		default:2042			blk_mq_end_request(rq, ret);2043		}2044	} while (!list_empty(list));2045out:2046	/* If we didn't flush the entire list, we could have told the driver2047	 * there was more coming, but that turned out to be a lie.2048	 */2049	if (!list_empty(list) || ret != BLK_STS_OK)2050		blk_mq_commit_rqs(hctx, queued, false);2051 2052	/*2053	 * Any items that need requeuing? Stuff them into hctx->dispatch,2054	 * that is where we will continue on next queue run.2055	 */2056	if (!list_empty(list)) {2057		bool needs_restart;2058		/* For non-shared tags, the RESTART check will suffice */2059		bool no_tag = prep == PREP_DISPATCH_NO_TAG &&2060			((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||2061			blk_mq_is_shared_tags(hctx->flags));2062 2063		if (nr_budgets)2064			blk_mq_release_budgets(q, list);2065 2066		spin_lock(&hctx->lock);2067		list_splice_tail_init(list, &hctx->dispatch);2068		spin_unlock(&hctx->lock);2069 2070		/*2071		 * Order adding requests to hctx->dispatch and checking2072		 * SCHED_RESTART flag. The pair of this smp_mb() is the one2073		 * in blk_mq_sched_restart(). Avoid restart code path to2074		 * miss the new added requests to hctx->dispatch, meantime2075		 * SCHED_RESTART is observed here.2076		 */2077		smp_mb();2078 2079		/*2080		 * If SCHED_RESTART was set by the caller of this function and2081		 * it is no longer set that means that it was cleared by another2082		 * thread and hence that a queue rerun is needed.2083		 *2084		 * If 'no_tag' is set, that means that we failed getting2085		 * a driver tag with an I/O scheduler attached. If our dispatch2086		 * waitqueue is no longer active, ensure that we run the queue2087		 * AFTER adding our entries back to the list.2088		 *2089		 * If no I/O scheduler has been configured it is possible that2090		 * the hardware queue got stopped and restarted before requests2091		 * were pushed back onto the dispatch list. Rerun the queue to2092		 * avoid starvation. Notes:2093		 * - blk_mq_run_hw_queue() checks whether or not a queue has2094		 *   been stopped before rerunning a queue.2095		 * - Some but not all block drivers stop a queue before2096		 *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq2097		 *   and dm-rq.2098		 *2099		 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART2100		 * bit is set, run queue after a delay to avoid IO stalls2101		 * that could otherwise occur if the queue is idle.  We'll do2102		 * similar if we couldn't get budget or couldn't lock a zone2103		 * and SCHED_RESTART is set.2104		 */2105		needs_restart = blk_mq_sched_needs_restart(hctx);2106		if (prep == PREP_DISPATCH_NO_BUDGET)2107			needs_resource = true;2108		if (!needs_restart ||2109		    (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))2110			blk_mq_run_hw_queue(hctx, true);2111		else if (needs_resource)2112			blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);2113 2114		blk_mq_update_dispatch_busy(hctx, true);2115		return false;2116	}2117 2118	blk_mq_update_dispatch_busy(hctx, false);2119	return true;2120}2121 2122static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)2123{2124	int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);2125 2126	if (cpu >= nr_cpu_ids)2127		cpu = cpumask_first(hctx->cpumask);2128	return cpu;2129}2130 2131/*2132 * ->next_cpu is always calculated from hctx->cpumask, so simply use2133 * it for speeding up the check2134 */2135static bool blk_mq_hctx_empty_cpumask(struct blk_mq_hw_ctx *hctx)2136{2137        return hctx->next_cpu >= nr_cpu_ids;2138}2139 2140/*2141 * It'd be great if the workqueue API had a way to pass2142 * in a mask and had some smarts for more clever placement.2143 * For now we just round-robin here, switching for every2144 * BLK_MQ_CPU_WORK_BATCH queued items.2145 */2146static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)2147{2148	bool tried = false;2149	int next_cpu = hctx->next_cpu;2150 2151	/* Switch to unbound if no allowable CPUs in this hctx */2152	if (hctx->queue->nr_hw_queues == 1 || blk_mq_hctx_empty_cpumask(hctx))2153		return WORK_CPU_UNBOUND;2154 2155	if (--hctx->next_cpu_batch <= 0) {2156select_cpu:2157		next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,2158				cpu_online_mask);2159		if (next_cpu >= nr_cpu_ids)2160			next_cpu = blk_mq_first_mapped_cpu(hctx);2161		hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;2162	}2163 2164	/*2165	 * Do unbound schedule if we can't find a online CPU for this hctx,2166	 * and it should only happen in the path of handling CPU DEAD.2167	 */2168	if (!cpu_online(next_cpu)) {2169		if (!tried) {2170			tried = true;2171			goto select_cpu;2172		}2173 2174		/*2175		 * Make sure to re-select CPU next time once after CPUs2176		 * in hctx->cpumask become online again.2177		 */2178		hctx->next_cpu = next_cpu;2179		hctx->next_cpu_batch = 1;2180		return WORK_CPU_UNBOUND;2181	}2182 2183	hctx->next_cpu = next_cpu;2184	return next_cpu;2185}2186 2187/**2188 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.2189 * @hctx: Pointer to the hardware queue to run.2190 * @msecs: Milliseconds of delay to wait before running the queue.2191 *2192 * Run a hardware queue asynchronously with a delay of @msecs.2193 */2194void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)2195{2196	if (unlikely(blk_mq_hctx_stopped(hctx)))2197		return;2198	kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,2199				    msecs_to_jiffies(msecs));2200}2201EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);2202 2203/**2204 * blk_mq_run_hw_queue - Start to run a hardware queue.2205 * @hctx: Pointer to the hardware queue to run.2206 * @async: If we want to run the queue asynchronously.2207 *2208 * Check if the request queue is not in a quiesced state and if there are2209 * pending requests to be sent. If this is true, run the queue to send requests2210 * to hardware.2211 */2212void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)2213{2214	bool need_run;2215 2216	/*2217	 * We can't run the queue inline with interrupts disabled.2218	 */2219	WARN_ON_ONCE(!async && in_interrupt());2220 2221	might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);2222 2223	/*2224	 * When queue is quiesced, we may be switching io scheduler, or2225	 * updating nr_hw_queues, or other things, and we can't run queue2226	 * any more, even __blk_mq_hctx_has_pending() can't be called safely.2227	 *2228	 * And queue will be rerun in blk_mq_unquiesce_queue() if it is2229	 * quiesced.2230	 */2231	__blk_mq_run_dispatch_ops(hctx->queue, false,2232		need_run = !blk_queue_quiesced(hctx->queue) &&2233		blk_mq_hctx_has_pending(hctx));2234 2235	if (!need_run)2236		return;2237 2238	if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {2239		blk_mq_delay_run_hw_queue(hctx, 0);2240		return;2241	}2242 2243	blk_mq_run_dispatch_ops(hctx->queue,2244				blk_mq_sched_dispatch_requests(hctx));2245}2246EXPORT_SYMBOL(blk_mq_run_hw_queue);2247 2248/*2249 * Return prefered queue to dispatch from (if any) for non-mq aware IO2250 * scheduler.2251 */2252static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)2253{2254	struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);2255	/*2256	 * If the IO scheduler does not respect hardware queues when2257	 * dispatching, we just don't bother with multiple HW queues and2258	 * dispatch from hctx for the current CPU since running multiple queues2259	 * just causes lock contention inside the scheduler and pointless cache2260	 * bouncing.2261	 */2262	struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];2263 2264	if (!blk_mq_hctx_stopped(hctx))2265		return hctx;2266	return NULL;2267}2268 2269/**2270 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.2271 * @q: Pointer to the request queue to run.2272 * @async: If we want to run the queue asynchronously.2273 */2274void blk_mq_run_hw_queues(struct request_queue *q, bool async)2275{2276	struct blk_mq_hw_ctx *hctx, *sq_hctx;2277	unsigned long i;2278 2279	sq_hctx = NULL;2280	if (blk_queue_sq_sched(q))2281		sq_hctx = blk_mq_get_sq_hctx(q);2282	queue_for_each_hw_ctx(q, hctx, i) {2283		if (blk_mq_hctx_stopped(hctx))2284			continue;2285		/*2286		 * Dispatch from this hctx either if there's no hctx preferred2287		 * by IO scheduler or if it has requests that bypass the2288		 * scheduler.2289		 */2290		if (!sq_hctx || sq_hctx == hctx ||2291		    !list_empty_careful(&hctx->dispatch))2292			blk_mq_run_hw_queue(hctx, async);2293	}2294}2295EXPORT_SYMBOL(blk_mq_run_hw_queues);2296 2297/**2298 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.2299 * @q: Pointer to the request queue to run.2300 * @msecs: Milliseconds of delay to wait before running the queues.2301 */2302void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)2303{2304	struct blk_mq_hw_ctx *hctx, *sq_hctx;2305	unsigned long i;2306 2307	sq_hctx = NULL;2308	if (blk_queue_sq_sched(q))2309		sq_hctx = blk_mq_get_sq_hctx(q);2310	queue_for_each_hw_ctx(q, hctx, i) {2311		if (blk_mq_hctx_stopped(hctx))2312			continue;2313		/*2314		 * If there is already a run_work pending, leave the2315		 * pending delay untouched. Otherwise, a hctx can stall2316		 * if another hctx is re-delaying the other's work2317		 * before the work executes.2318		 */2319		if (delayed_work_pending(&hctx->run_work))2320			continue;2321		/*2322		 * Dispatch from this hctx either if there's no hctx preferred2323		 * by IO scheduler or if it has requests that bypass the2324		 * scheduler.2325		 */2326		if (!sq_hctx || sq_hctx == hctx ||2327		    !list_empty_careful(&hctx->dispatch))2328			blk_mq_delay_run_hw_queue(hctx, msecs);2329	}2330}2331EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);2332 2333/*2334 * This function is often used for pausing .queue_rq() by driver when2335 * there isn't enough resource or some conditions aren't satisfied, and2336 * BLK_STS_RESOURCE is usually returned.2337 *2338 * We do not guarantee that dispatch can be drained or blocked2339 * after blk_mq_stop_hw_queue() returns. Please use2340 * blk_mq_quiesce_queue() for that requirement.2341 */2342void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)2343{2344	cancel_delayed_work(&hctx->run_work);2345 2346	set_bit(BLK_MQ_S_STOPPED, &hctx->state);2347}2348EXPORT_SYMBOL(blk_mq_stop_hw_queue);2349 2350/*2351 * This function is often used for pausing .queue_rq() by driver when2352 * there isn't enough resource or some conditions aren't satisfied, and2353 * BLK_STS_RESOURCE is usually returned.2354 *2355 * We do not guarantee that dispatch can be drained or blocked2356 * after blk_mq_stop_hw_queues() returns. Please use2357 * blk_mq_quiesce_queue() for that requirement.2358 */2359void blk_mq_stop_hw_queues(struct request_queue *q)2360{2361	struct blk_mq_hw_ctx *hctx;2362	unsigned long i;2363 2364	queue_for_each_hw_ctx(q, hctx, i)2365		blk_mq_stop_hw_queue(hctx);2366}2367EXPORT_SYMBOL(blk_mq_stop_hw_queues);2368 2369void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)2370{2371	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);2372 2373	blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);2374}2375EXPORT_SYMBOL(blk_mq_start_hw_queue);2376 2377void blk_mq_start_hw_queues(struct request_queue *q)2378{2379	struct blk_mq_hw_ctx *hctx;2380	unsigned long i;2381 2382	queue_for_each_hw_ctx(q, hctx, i)2383		blk_mq_start_hw_queue(hctx);2384}2385EXPORT_SYMBOL(blk_mq_start_hw_queues);2386 2387void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)2388{2389	if (!blk_mq_hctx_stopped(hctx))2390		return;2391 2392	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);2393	blk_mq_run_hw_queue(hctx, async);2394}2395EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);2396 2397void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)2398{2399	struct blk_mq_hw_ctx *hctx;2400	unsigned long i;2401 2402	queue_for_each_hw_ctx(q, hctx, i)2403		blk_mq_start_stopped_hw_queue(hctx, async ||2404					(hctx->flags & BLK_MQ_F_BLOCKING));2405}2406EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);2407 2408static void blk_mq_run_work_fn(struct work_struct *work)2409{2410	struct blk_mq_hw_ctx *hctx =2411		container_of(work, struct blk_mq_hw_ctx, run_work.work);2412 2413	blk_mq_run_dispatch_ops(hctx->queue,2414				blk_mq_sched_dispatch_requests(hctx));2415}2416 2417/**2418 * blk_mq_request_bypass_insert - Insert a request at dispatch list.2419 * @rq: Pointer to request to be inserted.2420 * @flags: BLK_MQ_INSERT_*2421 *2422 * Should only be used carefully, when the caller knows we want to2423 * bypass a potential IO scheduler on the target device.2424 */2425static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)2426{2427	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;2428 2429	spin_lock(&hctx->lock);2430	if (flags & BLK_MQ_INSERT_AT_HEAD)2431		list_add(&rq->queuelist, &hctx->dispatch);2432	else2433		list_add_tail(&rq->queuelist, &hctx->dispatch);2434	spin_unlock(&hctx->lock);2435}2436 2437static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,2438		struct blk_mq_ctx *ctx, struct list_head *list,2439		bool run_queue_async)2440{2441	struct request *rq;2442	enum hctx_type type = hctx->type;2443 2444	/*2445	 * Try to issue requests directly if the hw queue isn't busy to save an2446	 * extra enqueue & dequeue to the sw queue.2447	 */2448	if (!hctx->dispatch_busy && !run_queue_async) {2449		blk_mq_run_dispatch_ops(hctx->queue,2450			blk_mq_try_issue_list_directly(hctx, list));2451		if (list_empty(list))2452			goto out;2453	}2454 2455	/*2456	 * preemption doesn't flush plug list, so it's possible ctx->cpu is2457	 * offline now2458	 */2459	list_for_each_entry(rq, list, queuelist) {2460		BUG_ON(rq->mq_ctx != ctx);2461		trace_block_rq_insert(rq);2462		if (rq->cmd_flags & REQ_NOWAIT)2463			run_queue_async = true;2464	}2465 2466	spin_lock(&ctx->lock);2467	list_splice_tail_init(list, &ctx->rq_lists[type]);2468	blk_mq_hctx_mark_pending(hctx, ctx);2469	spin_unlock(&ctx->lock);2470out:2471	blk_mq_run_hw_queue(hctx, run_queue_async);2472}2473 2474static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)2475{2476	struct request_queue *q = rq->q;2477	struct blk_mq_ctx *ctx = rq->mq_ctx;2478	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;2479 2480	if (blk_rq_is_passthrough(rq)) {2481		/*2482		 * Passthrough request have to be added to hctx->dispatch2483		 * directly.  The device may be in a situation where it can't2484		 * handle FS request, and always returns BLK_STS_RESOURCE for2485		 * them, which gets them added to hctx->dispatch.2486		 *2487		 * If a passthrough request is required to unblock the queues,2488		 * and it is added to the scheduler queue, there is no chance to2489		 * dispatch it given we prioritize requests in hctx->dispatch.2490		 */2491		blk_mq_request_bypass_insert(rq, flags);2492	} else if (req_op(rq) == REQ_OP_FLUSH) {2493		/*2494		 * Firstly normal IO request is inserted to scheduler queue or2495		 * sw queue, meantime we add flush request to dispatch queue(2496		 * hctx->dispatch) directly and there is at most one in-flight2497		 * flush request for each hw queue, so it doesn't matter to add2498		 * flush request to tail or front of the dispatch queue.2499		 *2500		 * Secondly in case of NCQ, flush request belongs to non-NCQ2501		 * command, and queueing it will fail when there is any2502		 * in-flight normal IO request(NCQ command). When adding flush2503		 * rq to the front of hctx->dispatch, it is easier to introduce2504		 * extra time to flush rq's latency because of S_SCHED_RESTART2505		 * compared with adding to the tail of dispatch queue, then2506		 * chance of flush merge is increased, and less flush requests2507		 * will be issued to controller. It is observed that ~10% time2508		 * is saved in blktests block/004 on disk attached to AHCI/NCQ2509		 * drive when adding flush rq to the front of hctx->dispatch.2510		 *2511		 * Simply queue flush rq to the front of hctx->dispatch so that2512		 * intensive flush workloads can benefit in case of NCQ HW.2513		 */2514		blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);2515	} else if (q->elevator) {2516		LIST_HEAD(list);2517 2518		WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);2519 2520		list_add(&rq->queuelist, &list);2521		q->elevator->type->ops.insert_requests(hctx, &list, flags);2522	} else {2523		trace_block_rq_insert(rq);2524 2525		spin_lock(&ctx->lock);2526		if (flags & BLK_MQ_INSERT_AT_HEAD)2527			list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);2528		else2529			list_add_tail(&rq->queuelist,2530				      &ctx->rq_lists[hctx->type]);2531		blk_mq_hctx_mark_pending(hctx, ctx);2532		spin_unlock(&ctx->lock);2533	}2534}2535 2536static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,2537		unsigned int nr_segs)2538{2539	int err;2540 2541	if (bio->bi_opf & REQ_RAHEAD)2542		rq->cmd_flags |= REQ_FAILFAST_MASK;2543 2544	rq->__sector = bio->bi_iter.bi_sector;2545	rq->write_hint = bio->bi_write_hint;2546	blk_rq_bio_prep(rq, bio, nr_segs);2547	if (bio_integrity(bio))2548		rq->nr_integrity_segments = blk_rq_count_integrity_sg(rq->q,2549								      bio);2550 2551	/* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */2552	err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);2553	WARN_ON_ONCE(err);2554 2555	blk_account_io_start(rq);2556}2557 2558static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,2559					    struct request *rq, bool last)2560{2561	struct request_queue *q = rq->q;2562	struct blk_mq_queue_data bd = {2563		.rq = rq,2564		.last = last,2565	};2566	blk_status_t ret;2567 2568	/*2569	 * For OK queue, we are done. For error, caller may kill it.2570	 * Any other error (busy), just add it to our list as we2571	 * previously would have done.2572	 */2573	ret = q->mq_ops->queue_rq(hctx, &bd);2574	switch (ret) {2575	case BLK_STS_OK:2576		blk_mq_update_dispatch_busy(hctx, false);2577		break;2578	case BLK_STS_RESOURCE:2579	case BLK_STS_DEV_RESOURCE:2580		blk_mq_update_dispatch_busy(hctx, true);2581		__blk_mq_requeue_request(rq);2582		break;2583	default:2584		blk_mq_update_dispatch_busy(hctx, false);2585		break;2586	}2587 2588	return ret;2589}2590 2591static bool blk_mq_get_budget_and_tag(struct request *rq)2592{2593	int budget_token;2594 2595	budget_token = blk_mq_get_dispatch_budget(rq->q);2596	if (budget_token < 0)2597		return false;2598	blk_mq_set_rq_budget_token(rq, budget_token);2599	if (!blk_mq_get_driver_tag(rq)) {2600		blk_mq_put_dispatch_budget(rq->q, budget_token);2601		return false;2602	}2603	return true;2604}2605 2606/**2607 * blk_mq_try_issue_directly - Try to send a request directly to device driver.2608 * @hctx: Pointer of the associated hardware queue.2609 * @rq: Pointer to request to be sent.2610 *2611 * If the device has enough resources to accept a new request now, send the2612 * request directly to device driver. Else, insert at hctx->dispatch queue, so2613 * we can try send it another time in the future. Requests inserted at this2614 * queue have higher priority.2615 */2616static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,2617		struct request *rq)2618{2619	blk_status_t ret;2620 2621	if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {2622		blk_mq_insert_request(rq, 0);2623		return;2624	}2625 2626	if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {2627		blk_mq_insert_request(rq, 0);2628		blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);2629		return;2630	}2631 2632	ret = __blk_mq_issue_directly(hctx, rq, true);2633	switch (ret) {2634	case BLK_STS_OK:2635		break;2636	case BLK_STS_RESOURCE:2637	case BLK_STS_DEV_RESOURCE:2638		blk_mq_request_bypass_insert(rq, 0);2639		blk_mq_run_hw_queue(hctx, false);2640		break;2641	default:2642		blk_mq_end_request(rq, ret);2643		break;2644	}2645}2646 2647static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)2648{2649	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;2650 2651	if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {2652		blk_mq_insert_request(rq, 0);2653		return BLK_STS_OK;2654	}2655 2656	if (!blk_mq_get_budget_and_tag(rq))2657		return BLK_STS_RESOURCE;2658	return __blk_mq_issue_directly(hctx, rq, last);2659}2660 2661static void blk_mq_plug_issue_direct(struct blk_plug *plug)2662{2663	struct blk_mq_hw_ctx *hctx = NULL;2664	struct request *rq;2665	int queued = 0;2666	blk_status_t ret = BLK_STS_OK;2667 2668	while ((rq = rq_list_pop(&plug->mq_list))) {2669		bool last = rq_list_empty(plug->mq_list);2670 2671		if (hctx != rq->mq_hctx) {2672			if (hctx) {2673				blk_mq_commit_rqs(hctx, queued, false);2674				queued = 0;2675			}2676			hctx = rq->mq_hctx;2677		}2678 2679		ret = blk_mq_request_issue_directly(rq, last);2680		switch (ret) {2681		case BLK_STS_OK:2682			queued++;2683			break;2684		case BLK_STS_RESOURCE:2685		case BLK_STS_DEV_RESOURCE:2686			blk_mq_request_bypass_insert(rq, 0);2687			blk_mq_run_hw_queue(hctx, false);2688			goto out;2689		default:2690			blk_mq_end_request(rq, ret);2691			break;2692		}2693	}2694 2695out:2696	if (ret != BLK_STS_OK)2697		blk_mq_commit_rqs(hctx, queued, false);2698}2699 2700static void __blk_mq_flush_plug_list(struct request_queue *q,2701				     struct blk_plug *plug)2702{2703	if (blk_queue_quiesced(q))2704		return;2705	q->mq_ops->queue_rqs(&plug->mq_list);2706}2707 2708static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)2709{2710	struct blk_mq_hw_ctx *this_hctx = NULL;2711	struct blk_mq_ctx *this_ctx = NULL;2712	struct request *requeue_list = NULL;2713	struct request **requeue_lastp = &requeue_list;2714	unsigned int depth = 0;2715	bool is_passthrough = false;2716	LIST_HEAD(list);2717 2718	do {2719		struct request *rq = rq_list_pop(&plug->mq_list);2720 2721		if (!this_hctx) {2722			this_hctx = rq->mq_hctx;2723			this_ctx = rq->mq_ctx;2724			is_passthrough = blk_rq_is_passthrough(rq);2725		} else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||2726			   is_passthrough != blk_rq_is_passthrough(rq)) {2727			rq_list_add_tail(&requeue_lastp, rq);2728			continue;2729		}2730		list_add(&rq->queuelist, &list);2731		depth++;2732	} while (!rq_list_empty(plug->mq_list));2733 2734	plug->mq_list = requeue_list;2735	trace_block_unplug(this_hctx->queue, depth, !from_sched);2736 2737	percpu_ref_get(&this_hctx->queue->q_usage_counter);2738	/* passthrough requests should never be issued to the I/O scheduler */2739	if (is_passthrough) {2740		spin_lock(&this_hctx->lock);2741		list_splice_tail_init(&list, &this_hctx->dispatch);2742		spin_unlock(&this_hctx->lock);2743		blk_mq_run_hw_queue(this_hctx, from_sched);2744	} else if (this_hctx->queue->elevator) {2745		this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,2746				&list, 0);2747		blk_mq_run_hw_queue(this_hctx, from_sched);2748	} else {2749		blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);2750	}2751	percpu_ref_put(&this_hctx->queue->q_usage_counter);2752}2753 2754void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)2755{2756	struct request *rq;2757	unsigned int depth;2758 2759	/*2760	 * We may have been called recursively midway through handling2761	 * plug->mq_list via a schedule() in the driver's queue_rq() callback.2762	 * To avoid mq_list changing under our feet, clear rq_count early and2763	 * bail out specifically if rq_count is 0 rather than checking2764	 * whether the mq_list is empty.2765	 */2766	if (plug->rq_count == 0)2767		return;2768	depth = plug->rq_count;2769	plug->rq_count = 0;2770 2771	if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {2772		struct request_queue *q;2773 2774		rq = rq_list_peek(&plug->mq_list);2775		q = rq->q;2776		trace_block_unplug(q, depth, true);2777 2778		/*2779		 * Peek first request and see if we have a ->queue_rqs() hook.2780		 * If we do, we can dispatch the whole plug list in one go. We2781		 * already know at this point that all requests belong to the2782		 * same queue, caller must ensure that's the case.2783		 */2784		if (q->mq_ops->queue_rqs) {2785			blk_mq_run_dispatch_ops(q,2786				__blk_mq_flush_plug_list(q, plug));2787			if (rq_list_empty(plug->mq_list))2788				return;2789		}2790 2791		blk_mq_run_dispatch_ops(q,2792				blk_mq_plug_issue_direct(plug));2793		if (rq_list_empty(plug->mq_list))2794			return;2795	}2796 2797	do {2798		blk_mq_dispatch_plug_list(plug, from_schedule);2799	} while (!rq_list_empty(plug->mq_list));2800}2801 2802static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,2803		struct list_head *list)2804{2805	int queued = 0;2806	blk_status_t ret = BLK_STS_OK;2807 2808	while (!list_empty(list)) {2809		struct request *rq = list_first_entry(list, struct request,2810				queuelist);2811 2812		list_del_init(&rq->queuelist);2813		ret = blk_mq_request_issue_directly(rq, list_empty(list));2814		switch (ret) {2815		case BLK_STS_OK:2816			queued++;2817			break;2818		case BLK_STS_RESOURCE:2819		case BLK_STS_DEV_RESOURCE:2820			blk_mq_request_bypass_insert(rq, 0);2821			if (list_empty(list))2822				blk_mq_run_hw_queue(hctx, false);2823			goto out;2824		default:2825			blk_mq_end_request(rq, ret);2826			break;2827		}2828	}2829 2830out:2831	if (ret != BLK_STS_OK)2832		blk_mq_commit_rqs(hctx, queued, false);2833}2834 2835static bool blk_mq_attempt_bio_merge(struct request_queue *q,2836				     struct bio *bio, unsigned int nr_segs)2837{2838	if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {2839		if (blk_attempt_plug_merge(q, bio, nr_segs))2840			return true;2841		if (blk_mq_sched_bio_merge(q, bio, nr_segs))2842			return true;2843	}2844	return false;2845}2846 2847static struct request *blk_mq_get_new_requests(struct request_queue *q,2848					       struct blk_plug *plug,2849					       struct bio *bio,2850					       unsigned int nsegs)2851{2852	struct blk_mq_alloc_data data = {2853		.q		= q,2854		.nr_tags	= 1,2855		.cmd_flags	= bio->bi_opf,2856	};2857	struct request *rq;2858 2859	rq_qos_throttle(q, bio);2860 2861	if (plug) {2862		data.nr_tags = plug->nr_ios;2863		plug->nr_ios = 1;2864		data.cached_rq = &plug->cached_rq;2865	}2866 2867	rq = __blk_mq_alloc_requests(&data);2868	if (rq)2869		return rq;2870	rq_qos_cleanup(q, bio);2871	if (bio->bi_opf & REQ_NOWAIT)2872		bio_wouldblock_error(bio);2873	return NULL;2874}2875 2876/*2877 * Check if there is a suitable cached request and return it.2878 */2879static struct request *blk_mq_peek_cached_request(struct blk_plug *plug,2880		struct request_queue *q, blk_opf_t opf)2881{2882	enum hctx_type type = blk_mq_get_hctx_type(opf);2883	struct request *rq;2884 2885	if (!plug)2886		return NULL;2887	rq = rq_list_peek(&plug->cached_rq);2888	if (!rq || rq->q != q)2889		return NULL;2890	if (type != rq->mq_hctx->type &&2891	    (type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT))2892		return NULL;2893	if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))2894		return NULL;2895	return rq;2896}2897 2898static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug,2899		struct bio *bio)2900{2901	WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq);2902 2903	/*2904	 * If any qos ->throttle() end up blocking, we will have flushed the2905	 * plug and hence killed the cached_rq list as well. Pop this entry2906	 * before we throttle.2907	 */2908	plug->cached_rq = rq_list_next(rq);2909	rq_qos_throttle(rq->q, bio);2910 2911	blk_mq_rq_time_init(rq, 0);2912	rq->cmd_flags = bio->bi_opf;2913	INIT_LIST_HEAD(&rq->queuelist);2914}2915 2916static bool bio_unaligned(const struct bio *bio, struct request_queue *q)2917{2918	unsigned int bs_mask = queue_logical_block_size(q) - 1;2919 2920	/* .bi_sector of any zero sized bio need to be initialized */2921	if ((bio->bi_iter.bi_size & bs_mask) ||2922	    ((bio->bi_iter.bi_sector << SECTOR_SHIFT) & bs_mask))2923		return true;2924	return false;2925}2926 2927/**2928 * blk_mq_submit_bio - Create and send a request to block device.2929 * @bio: Bio pointer.2930 *2931 * Builds up a request structure from @q and @bio and send to the device. The2932 * request may not be queued directly to hardware if:2933 * * This request can be merged with another one2934 * * We want to place request at plug queue for possible future merging2935 * * There is an IO scheduler active at this queue2936 *2937 * It will not queue the request if there is an error with the bio, or at the2938 * request creation.2939 */2940void blk_mq_submit_bio(struct bio *bio)2941{2942	struct request_queue *q = bdev_get_queue(bio->bi_bdev);2943	struct blk_plug *plug = current->plug;2944	const int is_sync = op_is_sync(bio->bi_opf);2945	struct blk_mq_hw_ctx *hctx;2946	unsigned int nr_segs;2947	struct request *rq;2948	blk_status_t ret;2949 2950	/*2951	 * If the plug has a cached request for this queue, try to use it.2952	 */2953	rq = blk_mq_peek_cached_request(plug, q, bio->bi_opf);2954 2955	/*2956	 * A BIO that was released from a zone write plug has already been2957	 * through the preparation in this function, already holds a reference2958	 * on the queue usage counter, and is the only write BIO in-flight for2959	 * the target zone. Go straight to preparing a request for it.2960	 */2961	if (bio_zone_write_plugging(bio)) {2962		nr_segs = bio->__bi_nr_segments;2963		if (rq)2964			blk_queue_exit(q);2965		goto new_request;2966	}2967 2968	bio = blk_queue_bounce(bio, q);2969 2970	/*2971	 * The cached request already holds a q_usage_counter reference and we2972	 * don't have to acquire a new one if we use it.2973	 */2974	if (!rq) {2975		if (unlikely(bio_queue_enter(bio)))2976			return;2977	}2978 2979	/*2980	 * Device reconfiguration may change logical block size, so alignment2981	 * check has to be done with queue usage counter held2982	 */2983	if (unlikely(bio_unaligned(bio, q))) {2984		bio_io_error(bio);2985		goto queue_exit;2986	}2987 2988	bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);2989	if (!bio)2990		goto queue_exit;2991 2992	if (!bio_integrity_prep(bio))2993		goto queue_exit;2994 2995	if (blk_mq_attempt_bio_merge(q, bio, nr_segs))2996		goto queue_exit;2997 2998	if (blk_queue_is_zoned(q) && blk_zone_plug_bio(bio, nr_segs))2999		goto queue_exit;3000 3001new_request:3002	if (!rq) {3003		rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);3004		if (unlikely(!rq))3005			goto queue_exit;3006	} else {3007		blk_mq_use_cached_rq(rq, plug, bio);3008	}3009 3010	trace_block_getrq(bio);3011 3012	rq_qos_track(q, rq, bio);3013 3014	blk_mq_bio_to_request(rq, bio, nr_segs);3015 3016	ret = blk_crypto_rq_get_keyslot(rq);3017	if (ret != BLK_STS_OK) {3018		bio->bi_status = ret;3019		bio_endio(bio);3020		blk_mq_free_request(rq);3021		return;3022	}3023 3024	if (bio_zone_write_plugging(bio))3025		blk_zone_write_plug_init_request(rq);3026 3027	if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))3028		return;3029 3030	if (plug) {3031		blk_add_rq_to_plug(plug, rq);3032		return;3033	}3034 3035	hctx = rq->mq_hctx;3036	if ((rq->rq_flags & RQF_USE_SCHED) ||3037	    (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {3038		blk_mq_insert_request(rq, 0);3039		blk_mq_run_hw_queue(hctx, true);3040	} else {3041		blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));3042	}3043	return;3044 3045queue_exit:3046	/*3047	 * Don't drop the queue reference if we were trying to use a cached3048	 * request and thus didn't acquire one.3049	 */3050	if (!rq)3051		blk_queue_exit(q);3052}3053 3054#ifdef CONFIG_BLK_MQ_STACKING3055/**3056 * blk_insert_cloned_request - Helper for stacking drivers to submit a request3057 * @rq: the request being queued3058 */3059blk_status_t blk_insert_cloned_request(struct request *rq)3060{3061	struct request_queue *q = rq->q;3062	unsigned int max_sectors = blk_queue_get_max_sectors(rq);3063	unsigned int max_segments = blk_rq_get_max_segments(rq);3064	blk_status_t ret;3065 3066	if (blk_rq_sectors(rq) > max_sectors) {3067		/*3068		 * SCSI device does not have a good way to return if3069		 * Write Same/Zero is actually supported. If a device rejects3070		 * a non-read/write command (discard, write same,etc.) the3071		 * low-level device driver will set the relevant queue limit to3072		 * 0 to prevent blk-lib from issuing more of the offending3073		 * operations. Commands queued prior to the queue limit being3074		 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O3075		 * errors being propagated to upper layers.3076		 */3077		if (max_sectors == 0)3078			return BLK_STS_NOTSUPP;3079 3080		printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",3081			__func__, blk_rq_sectors(rq), max_sectors);3082		return BLK_STS_IOERR;3083	}3084 3085	/*3086	 * The queue settings related to segment counting may differ from the3087	 * original queue.3088	 */3089	rq->nr_phys_segments = blk_recalc_rq_segments(rq);3090	if (rq->nr_phys_segments > max_segments) {3091		printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",3092			__func__, rq->nr_phys_segments, max_segments);3093		return BLK_STS_IOERR;3094	}3095 3096	if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))3097		return BLK_STS_IOERR;3098 3099	ret = blk_crypto_rq_get_keyslot(rq);3100	if (ret != BLK_STS_OK)3101		return ret;3102 3103	blk_account_io_start(rq);3104 3105	/*3106	 * Since we have a scheduler attached on the top device,3107	 * bypass a potential scheduler on the bottom device for3108	 * insert.3109	 */3110	blk_mq_run_dispatch_ops(q,3111			ret = blk_mq_request_issue_directly(rq, true));3112	if (ret)3113		blk_account_io_done(rq, blk_time_get_ns());3114	return ret;3115}3116EXPORT_SYMBOL_GPL(blk_insert_cloned_request);3117 3118/**3119 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request3120 * @rq: the clone request to be cleaned up3121 *3122 * Description:3123 *     Free all bios in @rq for a cloned request.3124 */3125void blk_rq_unprep_clone(struct request *rq)3126{3127	struct bio *bio;3128 3129	while ((bio = rq->bio) != NULL) {3130		rq->bio = bio->bi_next;3131 3132		bio_put(bio);3133	}3134}3135EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);3136 3137/**3138 * blk_rq_prep_clone - Helper function to setup clone request3139 * @rq: the request to be setup3140 * @rq_src: original request to be cloned3141 * @bs: bio_set that bios for clone are allocated from3142 * @gfp_mask: memory allocation mask for bio3143 * @bio_ctr: setup function to be called for each clone bio.3144 *           Returns %0 for success, non %0 for failure.3145 * @data: private data to be passed to @bio_ctr3146 *3147 * Description:3148 *     Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.3149 *     Also, pages which the original bios are pointing to are not copied3150 *     and the cloned bios just point same pages.3151 *     So cloned bios must be completed before original bios, which means3152 *     the caller must complete @rq before @rq_src.3153 */3154int blk_rq_prep_clone(struct request *rq, struct request *rq_src,3155		      struct bio_set *bs, gfp_t gfp_mask,3156		      int (*bio_ctr)(struct bio *, struct bio *, void *),3157		      void *data)3158{3159	struct bio *bio, *bio_src;3160 3161	if (!bs)3162		bs = &fs_bio_set;3163 3164	__rq_for_each_bio(bio_src, rq_src) {3165		bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,3166				      bs);3167		if (!bio)3168			goto free_and_out;3169 3170		if (bio_ctr && bio_ctr(bio, bio_src, data))3171			goto free_and_out;3172 3173		if (rq->bio) {3174			rq->biotail->bi_next = bio;3175			rq->biotail = bio;3176		} else {3177			rq->bio = rq->biotail = bio;3178		}3179		bio = NULL;3180	}3181 3182	/* Copy attributes of the original request to the clone request. */3183	rq->__sector = blk_rq_pos(rq_src);3184	rq->__data_len = blk_rq_bytes(rq_src);3185	if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {3186		rq->rq_flags |= RQF_SPECIAL_PAYLOAD;3187		rq->special_vec = rq_src->special_vec;3188	}3189	rq->nr_phys_segments = rq_src->nr_phys_segments;3190	rq->ioprio = rq_src->ioprio;3191	rq->write_hint = rq_src->write_hint;3192 3193	if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)3194		goto free_and_out;3195 3196	return 0;3197 3198free_and_out:3199	if (bio)3200		bio_put(bio);3201	blk_rq_unprep_clone(rq);3202 3203	return -ENOMEM;3204}3205EXPORT_SYMBOL_GPL(blk_rq_prep_clone);3206#endif /* CONFIG_BLK_MQ_STACKING */3207 3208/*3209 * Steal bios from a request and add them to a bio list.3210 * The request must not have been partially completed before.3211 */3212void blk_steal_bios(struct bio_list *list, struct request *rq)3213{3214	if (rq->bio) {3215		if (list->tail)3216			list->tail->bi_next = rq->bio;3217		else3218			list->head = rq->bio;3219		list->tail = rq->biotail;3220 3221		rq->bio = NULL;3222		rq->biotail = NULL;3223	}3224 3225	rq->__data_len = 0;3226}3227EXPORT_SYMBOL_GPL(blk_steal_bios);3228 3229static size_t order_to_size(unsigned int order)3230{3231	return (size_t)PAGE_SIZE << order;3232}3233 3234/* called before freeing request pool in @tags */3235static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,3236				    struct blk_mq_tags *tags)3237{3238	struct page *page;3239	unsigned long flags;3240 3241	/*3242	 * There is no need to clear mapping if driver tags is not initialized3243	 * or the mapping belongs to the driver tags.3244	 */3245	if (!drv_tags || drv_tags == tags)3246		return;3247 3248	list_for_each_entry(page, &tags->page_list, lru) {3249		unsigned long start = (unsigned long)page_address(page);3250		unsigned long end = start + order_to_size(page->private);3251		int i;3252 3253		for (i = 0; i < drv_tags->nr_tags; i++) {3254			struct request *rq = drv_tags->rqs[i];3255			unsigned long rq_addr = (unsigned long)rq;3256 3257			if (rq_addr >= start && rq_addr < end) {3258				WARN_ON_ONCE(req_ref_read(rq) != 0);3259				cmpxchg(&drv_tags->rqs[i], rq, NULL);3260			}3261		}3262	}3263 3264	/*3265	 * Wait until all pending iteration is done.3266	 *3267	 * Request reference is cleared and it is guaranteed to be observed3268	 * after the ->lock is released.3269	 */3270	spin_lock_irqsave(&drv_tags->lock, flags);3271	spin_unlock_irqrestore(&drv_tags->lock, flags);3272}3273 3274void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,3275		     unsigned int hctx_idx)3276{3277	struct blk_mq_tags *drv_tags;3278	struct page *page;3279 3280	if (list_empty(&tags->page_list))3281		return;3282 3283	if (blk_mq_is_shared_tags(set->flags))3284		drv_tags = set->shared_tags;3285	else3286		drv_tags = set->tags[hctx_idx];3287 3288	if (tags->static_rqs && set->ops->exit_request) {3289		int i;3290 3291		for (i = 0; i < tags->nr_tags; i++) {3292			struct request *rq = tags->static_rqs[i];3293 3294			if (!rq)3295				continue;3296			set->ops->exit_request(set, rq, hctx_idx);3297			tags->static_rqs[i] = NULL;3298		}3299	}3300 3301	blk_mq_clear_rq_mapping(drv_tags, tags);3302 3303	while (!list_empty(&tags->page_list)) {3304		page = list_first_entry(&tags->page_list, struct page, lru);3305		list_del_init(&page->lru);3306		/*3307		 * Remove kmemleak object previously allocated in3308		 * blk_mq_alloc_rqs().3309		 */3310		kmemleak_free(page_address(page));3311		__free_pages(page, page->private);3312	}3313}3314 3315void blk_mq_free_rq_map(struct blk_mq_tags *tags)3316{3317	kfree(tags->rqs);3318	tags->rqs = NULL;3319	kfree(tags->static_rqs);3320	tags->static_rqs = NULL;3321 3322	blk_mq_free_tags(tags);3323}3324 3325static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,3326		unsigned int hctx_idx)3327{3328	int i;3329 3330	for (i = 0; i < set->nr_maps; i++) {3331		unsigned int start = set->map[i].queue_offset;3332		unsigned int end = start + set->map[i].nr_queues;3333 3334		if (hctx_idx >= start && hctx_idx < end)3335			break;3336	}3337 3338	if (i >= set->nr_maps)3339		i = HCTX_TYPE_DEFAULT;3340 3341	return i;3342}3343 3344static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,3345		unsigned int hctx_idx)3346{3347	enum hctx_type type = hctx_idx_to_type(set, hctx_idx);3348 3349	return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);3350}3351 3352static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,3353					       unsigned int hctx_idx,3354					       unsigned int nr_tags,3355					       unsigned int reserved_tags)3356{3357	int node = blk_mq_get_hctx_node(set, hctx_idx);3358	struct blk_mq_tags *tags;3359 3360	if (node == NUMA_NO_NODE)3361		node = set->numa_node;3362 3363	tags = blk_mq_init_tags(nr_tags, reserved_tags, node,3364				BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));3365	if (!tags)3366		return NULL;3367 3368	tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),3369				 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,3370				 node);3371	if (!tags->rqs)3372		goto err_free_tags;3373 3374	tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),3375					GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,3376					node);3377	if (!tags->static_rqs)3378		goto err_free_rqs;3379 3380	return tags;3381 3382err_free_rqs:3383	kfree(tags->rqs);3384err_free_tags:3385	blk_mq_free_tags(tags);3386	return NULL;3387}3388 3389static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,3390			       unsigned int hctx_idx, int node)3391{3392	int ret;3393 3394	if (set->ops->init_request) {3395		ret = set->ops->init_request(set, rq, hctx_idx, node);3396		if (ret)3397			return ret;3398	}3399 3400	WRITE_ONCE(rq->state, MQ_RQ_IDLE);3401	return 0;3402}3403 3404static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,3405			    struct blk_mq_tags *tags,3406			    unsigned int hctx_idx, unsigned int depth)3407{3408	unsigned int i, j, entries_per_page, max_order = 4;3409	int node = blk_mq_get_hctx_node(set, hctx_idx);3410	size_t rq_size, left;3411 3412	if (node == NUMA_NO_NODE)3413		node = set->numa_node;3414 3415	INIT_LIST_HEAD(&tags->page_list);3416 3417	/*3418	 * rq_size is the size of the request plus driver payload, rounded3419	 * to the cacheline size3420	 */3421	rq_size = round_up(sizeof(struct request) + set->cmd_size,3422				cache_line_size());3423	left = rq_size * depth;3424 3425	for (i = 0; i < depth; ) {3426		int this_order = max_order;3427		struct page *page;3428		int to_do;3429		void *p;3430 3431		while (this_order && left < order_to_size(this_order - 1))3432			this_order--;3433 3434		do {3435			page = alloc_pages_node(node,3436				GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,3437				this_order);3438			if (page)3439				break;3440			if (!this_order--)3441				break;3442			if (order_to_size(this_order) < rq_size)3443				break;3444		} while (1);3445 3446		if (!page)3447			goto fail;3448 3449		page->private = this_order;3450		list_add_tail(&page->lru, &tags->page_list);3451 3452		p = page_address(page);3453		/*3454		 * Allow kmemleak to scan these pages as they contain pointers3455		 * to additional allocations like via ops->init_request().3456		 */3457		kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);3458		entries_per_page = order_to_size(this_order) / rq_size;3459		to_do = min(entries_per_page, depth - i);3460		left -= to_do * rq_size;3461		for (j = 0; j < to_do; j++) {3462			struct request *rq = p;3463 3464			tags->static_rqs[i] = rq;3465			if (blk_mq_init_request(set, rq, hctx_idx, node)) {3466				tags->static_rqs[i] = NULL;3467				goto fail;3468			}3469 3470			p += rq_size;3471			i++;3472		}3473	}3474	return 0;3475 3476fail:3477	blk_mq_free_rqs(set, tags, hctx_idx);3478	return -ENOMEM;3479}3480 3481struct rq_iter_data {3482	struct blk_mq_hw_ctx *hctx;3483	bool has_rq;3484};3485 3486static bool blk_mq_has_request(struct request *rq, void *data)3487{3488	struct rq_iter_data *iter_data = data;3489 3490	if (rq->mq_hctx != iter_data->hctx)3491		return true;3492	iter_data->has_rq = true;3493	return false;3494}3495 3496static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)3497{3498	struct blk_mq_tags *tags = hctx->sched_tags ?3499			hctx->sched_tags : hctx->tags;3500	struct rq_iter_data data = {3501		.hctx	= hctx,3502	};3503 3504	blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);3505	return data.has_rq;3506}3507 3508static bool blk_mq_hctx_has_online_cpu(struct blk_mq_hw_ctx *hctx,3509		unsigned int this_cpu)3510{3511	enum hctx_type type = hctx->type;3512	int cpu;3513 3514	/*3515	 * hctx->cpumask has to rule out isolated CPUs, but userspace still3516	 * might submit IOs on these isolated CPUs, so use the queue map to3517	 * check if all CPUs mapped to this hctx are offline3518	 */3519	for_each_online_cpu(cpu) {3520		struct blk_mq_hw_ctx *h = blk_mq_map_queue_type(hctx->queue,3521				type, cpu);3522 3523		if (h != hctx)3524			continue;3525 3526		/* this hctx has at least one online CPU */3527		if (this_cpu != cpu)3528			return true;3529	}3530 3531	return false;3532}3533 3534static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)3535{3536	struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,3537			struct blk_mq_hw_ctx, cpuhp_online);3538 3539	if (blk_mq_hctx_has_online_cpu(hctx, cpu))3540		return 0;3541 3542	/*3543	 * Prevent new request from being allocated on the current hctx.3544	 *3545	 * The smp_mb__after_atomic() Pairs with the implied barrier in3546	 * test_and_set_bit_lock in sbitmap_get().  Ensures the inactive flag is3547	 * seen once we return from the tag allocator.3548	 */3549	set_bit(BLK_MQ_S_INACTIVE, &hctx->state);3550	smp_mb__after_atomic();3551 3552	/*3553	 * Try to grab a reference to the queue and wait for any outstanding3554	 * requests.  If we could not grab a reference the queue has been3555	 * frozen and there are no requests.3556	 */3557	if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {3558		while (blk_mq_hctx_has_requests(hctx))3559			msleep(5);3560		percpu_ref_put(&hctx->queue->q_usage_counter);3561	}3562 3563	return 0;3564}3565 3566/*3567 * Check if one CPU is mapped to the specified hctx3568 *3569 * Isolated CPUs have been ruled out from hctx->cpumask, which is supposed3570 * to be used for scheduling kworker only. For other usage, please call this3571 * helper for checking if one CPU belongs to the specified hctx3572 */3573static bool blk_mq_cpu_mapped_to_hctx(unsigned int cpu,3574		const struct blk_mq_hw_ctx *hctx)3575{3576	struct blk_mq_hw_ctx *mapped_hctx = blk_mq_map_queue_type(hctx->queue,3577			hctx->type, cpu);3578 3579	return mapped_hctx == hctx;3580}3581 3582static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)3583{3584	struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,3585			struct blk_mq_hw_ctx, cpuhp_online);3586 3587	if (blk_mq_cpu_mapped_to_hctx(cpu, hctx))3588		clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);3589	return 0;3590}3591 3592/*3593 * 'cpu' is going away. splice any existing rq_list entries from this3594 * software queue to the hw queue dispatch list, and ensure that it3595 * gets run.3596 */3597static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)3598{3599	struct blk_mq_hw_ctx *hctx;3600	struct blk_mq_ctx *ctx;3601	LIST_HEAD(tmp);3602	enum hctx_type type;3603 3604	hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);3605	if (!blk_mq_cpu_mapped_to_hctx(cpu, hctx))3606		return 0;3607 3608	ctx = __blk_mq_get_ctx(hctx->queue, cpu);3609	type = hctx->type;3610 3611	spin_lock(&ctx->lock);3612	if (!list_empty(&ctx->rq_lists[type])) {3613		list_splice_init(&ctx->rq_lists[type], &tmp);3614		blk_mq_hctx_clear_pending(hctx, ctx);3615	}3616	spin_unlock(&ctx->lock);3617 3618	if (list_empty(&tmp))3619		return 0;3620 3621	spin_lock(&hctx->lock);3622	list_splice_tail_init(&tmp, &hctx->dispatch);3623	spin_unlock(&hctx->lock);3624 3625	blk_mq_run_hw_queue(hctx, true);3626	return 0;3627}3628 3629static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)3630{3631	if (!(hctx->flags & BLK_MQ_F_STACKING))3632		cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,3633						    &hctx->cpuhp_online);3634	cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,3635					    &hctx->cpuhp_dead);3636}3637 3638/*3639 * Before freeing hw queue, clearing the flush request reference in3640 * tags->rqs[] for avoiding potential UAF.3641 */3642static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,3643		unsigned int queue_depth, struct request *flush_rq)3644{3645	int i;3646	unsigned long flags;3647 3648	/* The hw queue may not be mapped yet */3649	if (!tags)3650		return;3651 3652	WARN_ON_ONCE(req_ref_read(flush_rq) != 0);3653 3654	for (i = 0; i < queue_depth; i++)3655		cmpxchg(&tags->rqs[i], flush_rq, NULL);3656 3657	/*3658	 * Wait until all pending iteration is done.3659	 *3660	 * Request reference is cleared and it is guaranteed to be observed3661	 * after the ->lock is released.3662	 */3663	spin_lock_irqsave(&tags->lock, flags);3664	spin_unlock_irqrestore(&tags->lock, flags);3665}3666 3667/* hctx->ctxs will be freed in queue's release handler */3668static void blk_mq_exit_hctx(struct request_queue *q,3669		struct blk_mq_tag_set *set,3670		struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)3671{3672	struct request *flush_rq = hctx->fq->flush_rq;3673 3674	if (blk_mq_hw_queue_mapped(hctx))3675		blk_mq_tag_idle(hctx);3676 3677	if (blk_queue_init_done(q))3678		blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],3679				set->queue_depth, flush_rq);3680	if (set->ops->exit_request)3681		set->ops->exit_request(set, flush_rq, hctx_idx);3682 3683	if (set->ops->exit_hctx)3684		set->ops->exit_hctx(hctx, hctx_idx);3685 3686	blk_mq_remove_cpuhp(hctx);3687 3688	xa_erase(&q->hctx_table, hctx_idx);3689 3690	spin_lock(&q->unused_hctx_lock);3691	list_add(&hctx->hctx_list, &q->unused_hctx_list);3692	spin_unlock(&q->unused_hctx_lock);3693}3694 3695static void blk_mq_exit_hw_queues(struct request_queue *q,3696		struct blk_mq_tag_set *set, int nr_queue)3697{3698	struct blk_mq_hw_ctx *hctx;3699	unsigned long i;3700 3701	queue_for_each_hw_ctx(q, hctx, i) {3702		if (i == nr_queue)3703			break;3704		blk_mq_exit_hctx(q, set, hctx, i);3705	}3706}3707 3708static int blk_mq_init_hctx(struct request_queue *q,3709		struct blk_mq_tag_set *set,3710		struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)3711{3712	hctx->queue_num = hctx_idx;3713 3714	if (!(hctx->flags & BLK_MQ_F_STACKING))3715		cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,3716				&hctx->cpuhp_online);3717	cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);3718 3719	hctx->tags = set->tags[hctx_idx];3720 3721	if (set->ops->init_hctx &&3722	    set->ops->init_hctx(hctx, set->driver_data, hctx_idx))3723		goto unregister_cpu_notifier;3724 3725	if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,3726				hctx->numa_node))3727		goto exit_hctx;3728 3729	if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))3730		goto exit_flush_rq;3731 3732	return 0;3733 3734 exit_flush_rq:3735	if (set->ops->exit_request)3736		set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);3737 exit_hctx:3738	if (set->ops->exit_hctx)3739		set->ops->exit_hctx(hctx, hctx_idx);3740 unregister_cpu_notifier:3741	blk_mq_remove_cpuhp(hctx);3742	return -1;3743}3744 3745static struct blk_mq_hw_ctx *3746blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,3747		int node)3748{3749	struct blk_mq_hw_ctx *hctx;3750	gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;3751 3752	hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);3753	if (!hctx)3754		goto fail_alloc_hctx;3755 3756	if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))3757		goto free_hctx;3758 3759	atomic_set(&hctx->nr_active, 0);3760	if (node == NUMA_NO_NODE)3761		node = set->numa_node;3762	hctx->numa_node = node;3763 3764	INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);3765	spin_lock_init(&hctx->lock);3766	INIT_LIST_HEAD(&hctx->dispatch);3767	hctx->queue = q;3768	hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;3769 3770	INIT_LIST_HEAD(&hctx->hctx_list);3771 3772	/*3773	 * Allocate space for all possible cpus to avoid allocation at3774	 * runtime3775	 */3776	hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),3777			gfp, node);3778	if (!hctx->ctxs)3779		goto free_cpumask;3780 3781	if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),3782				gfp, node, false, false))3783		goto free_ctxs;3784	hctx->nr_ctx = 0;3785 3786	spin_lock_init(&hctx->dispatch_wait_lock);3787	init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);3788	INIT_LIST_HEAD(&hctx->dispatch_wait.entry);3789 3790	hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);3791	if (!hctx->fq)3792		goto free_bitmap;3793 3794	blk_mq_hctx_kobj_init(hctx);3795 3796	return hctx;3797 3798 free_bitmap:3799	sbitmap_free(&hctx->ctx_map);3800 free_ctxs:3801	kfree(hctx->ctxs);3802 free_cpumask:3803	free_cpumask_var(hctx->cpumask);3804 free_hctx:3805	kfree(hctx);3806 fail_alloc_hctx:3807	return NULL;3808}3809 3810static void blk_mq_init_cpu_queues(struct request_queue *q,3811				   unsigned int nr_hw_queues)3812{3813	struct blk_mq_tag_set *set = q->tag_set;3814	unsigned int i, j;3815 3816	for_each_possible_cpu(i) {3817		struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);3818		struct blk_mq_hw_ctx *hctx;3819		int k;3820 3821		__ctx->cpu = i;3822		spin_lock_init(&__ctx->lock);3823		for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)3824			INIT_LIST_HEAD(&__ctx->rq_lists[k]);3825 3826		__ctx->queue = q;3827 3828		/*3829		 * Set local node, IFF we have more than one hw queue. If3830		 * not, we remain on the home node of the device3831		 */3832		for (j = 0; j < set->nr_maps; j++) {3833			hctx = blk_mq_map_queue_type(q, j, i);3834			if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)3835				hctx->numa_node = cpu_to_node(i);3836		}3837	}3838}3839 3840struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,3841					     unsigned int hctx_idx,3842					     unsigned int depth)3843{3844	struct blk_mq_tags *tags;3845	int ret;3846 3847	tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);3848	if (!tags)3849		return NULL;3850 3851	ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);3852	if (ret) {3853		blk_mq_free_rq_map(tags);3854		return NULL;3855	}3856 3857	return tags;3858}3859 3860static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,3861				       int hctx_idx)3862{3863	if (blk_mq_is_shared_tags(set->flags)) {3864		set->tags[hctx_idx] = set->shared_tags;3865 3866		return true;3867	}3868 3869	set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,3870						       set->queue_depth);3871 3872	return set->tags[hctx_idx];3873}3874 3875void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,3876			     struct blk_mq_tags *tags,3877			     unsigned int hctx_idx)3878{3879	if (tags) {3880		blk_mq_free_rqs(set, tags, hctx_idx);3881		blk_mq_free_rq_map(tags);3882	}3883}3884 3885static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,3886				      unsigned int hctx_idx)3887{3888	if (!blk_mq_is_shared_tags(set->flags))3889		blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);3890 3891	set->tags[hctx_idx] = NULL;3892}3893 3894static void blk_mq_map_swqueue(struct request_queue *q)3895{3896	unsigned int j, hctx_idx;3897	unsigned long i;3898	struct blk_mq_hw_ctx *hctx;3899	struct blk_mq_ctx *ctx;3900	struct blk_mq_tag_set *set = q->tag_set;3901 3902	queue_for_each_hw_ctx(q, hctx, i) {3903		cpumask_clear(hctx->cpumask);3904		hctx->nr_ctx = 0;3905		hctx->dispatch_from = NULL;3906	}3907 3908	/*3909	 * Map software to hardware queues.3910	 *3911	 * If the cpu isn't present, the cpu is mapped to first hctx.3912	 */3913	for_each_possible_cpu(i) {3914 3915		ctx = per_cpu_ptr(q->queue_ctx, i);3916		for (j = 0; j < set->nr_maps; j++) {3917			if (!set->map[j].nr_queues) {3918				ctx->hctxs[j] = blk_mq_map_queue_type(q,3919						HCTX_TYPE_DEFAULT, i);3920				continue;3921			}3922			hctx_idx = set->map[j].mq_map[i];3923			/* unmapped hw queue can be remapped after CPU topo changed */3924			if (!set->tags[hctx_idx] &&3925			    !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {3926				/*3927				 * If tags initialization fail for some hctx,3928				 * that hctx won't be brought online.  In this3929				 * case, remap the current ctx to hctx[0] which3930				 * is guaranteed to always have tags allocated3931				 */3932				set->map[j].mq_map[i] = 0;3933			}3934 3935			hctx = blk_mq_map_queue_type(q, j, i);3936			ctx->hctxs[j] = hctx;3937			/*3938			 * If the CPU is already set in the mask, then we've3939			 * mapped this one already. This can happen if3940			 * devices share queues across queue maps.3941			 */3942			if (cpumask_test_cpu(i, hctx->cpumask))3943				continue;3944 3945			cpumask_set_cpu(i, hctx->cpumask);3946			hctx->type = j;3947			ctx->index_hw[hctx->type] = hctx->nr_ctx;3948			hctx->ctxs[hctx->nr_ctx++] = ctx;3949 3950			/*3951			 * If the nr_ctx type overflows, we have exceeded the3952			 * amount of sw queues we can support.3953			 */3954			BUG_ON(!hctx->nr_ctx);3955		}3956 3957		for (; j < HCTX_MAX_TYPES; j++)3958			ctx->hctxs[j] = blk_mq_map_queue_type(q,3959					HCTX_TYPE_DEFAULT, i);3960	}3961 3962	queue_for_each_hw_ctx(q, hctx, i) {3963		int cpu;3964 3965		/*3966		 * If no software queues are mapped to this hardware queue,3967		 * disable it and free the request entries.3968		 */3969		if (!hctx->nr_ctx) {3970			/* Never unmap queue 0.  We need it as a3971			 * fallback in case of a new remap fails3972			 * allocation3973			 */3974			if (i)3975				__blk_mq_free_map_and_rqs(set, i);3976 3977			hctx->tags = NULL;3978			continue;3979		}3980 3981		hctx->tags = set->tags[i];3982		WARN_ON(!hctx->tags);3983 3984		/*3985		 * Set the map size to the number of mapped software queues.3986		 * This is more accurate and more efficient than looping3987		 * over all possibly mapped software queues.3988		 */3989		sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);3990 3991		/*3992		 * Rule out isolated CPUs from hctx->cpumask to avoid3993		 * running block kworker on isolated CPUs3994		 */3995		for_each_cpu(cpu, hctx->cpumask) {3996			if (cpu_is_isolated(cpu))3997				cpumask_clear_cpu(cpu, hctx->cpumask);3998		}3999 4000		/*4001		 * Initialize batch roundrobin counts4002		 */4003		hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);4004		hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;4005	}4006}4007 4008/*4009 * Caller needs to ensure that we're either frozen/quiesced, or that4010 * the queue isn't live yet.4011 */4012static void queue_set_hctx_shared(struct request_queue *q, bool shared)4013{4014	struct blk_mq_hw_ctx *hctx;4015	unsigned long i;4016 4017	queue_for_each_hw_ctx(q, hctx, i) {4018		if (shared) {4019			hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;4020		} else {4021			blk_mq_tag_idle(hctx);4022			hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;4023		}4024	}4025}4026 4027static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,4028					 bool shared)4029{4030	struct request_queue *q;4031 4032	lockdep_assert_held(&set->tag_list_lock);4033 4034	list_for_each_entry(q, &set->tag_list, tag_set_list) {4035		blk_mq_freeze_queue(q);4036		queue_set_hctx_shared(q, shared);4037		blk_mq_unfreeze_queue(q);4038	}4039}4040 4041static void blk_mq_del_queue_tag_set(struct request_queue *q)4042{4043	struct blk_mq_tag_set *set = q->tag_set;4044 4045	mutex_lock(&set->tag_list_lock);4046	list_del(&q->tag_set_list);4047	if (list_is_singular(&set->tag_list)) {4048		/* just transitioned to unshared */4049		set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;4050		/* update existing queue */4051		blk_mq_update_tag_set_shared(set, false);4052	}4053	mutex_unlock(&set->tag_list_lock);4054	INIT_LIST_HEAD(&q->tag_set_list);4055}4056 4057static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,4058				     struct request_queue *q)4059{4060	mutex_lock(&set->tag_list_lock);4061 4062	/*4063	 * Check to see if we're transitioning to shared (from 1 to 2 queues).4064	 */4065	if (!list_empty(&set->tag_list) &&4066	    !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {4067		set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;4068		/* update existing queue */4069		blk_mq_update_tag_set_shared(set, true);4070	}4071	if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)4072		queue_set_hctx_shared(q, true);4073	list_add_tail(&q->tag_set_list, &set->tag_list);4074 4075	mutex_unlock(&set->tag_list_lock);4076}4077 4078/* All allocations will be freed in release handler of q->mq_kobj */4079static int blk_mq_alloc_ctxs(struct request_queue *q)4080{4081	struct blk_mq_ctxs *ctxs;4082	int cpu;4083 4084	ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);4085	if (!ctxs)4086		return -ENOMEM;4087 4088	ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);4089	if (!ctxs->queue_ctx)4090		goto fail;4091 4092	for_each_possible_cpu(cpu) {4093		struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);4094		ctx->ctxs = ctxs;4095	}4096 4097	q->mq_kobj = &ctxs->kobj;4098	q->queue_ctx = ctxs->queue_ctx;4099 4100	return 0;4101 fail:4102	kfree(ctxs);4103	return -ENOMEM;4104}4105 4106/*4107 * It is the actual release handler for mq, but we do it from4108 * request queue's release handler for avoiding use-after-free4109 * and headache because q->mq_kobj shouldn't have been introduced,4110 * but we can't group ctx/kctx kobj without it.4111 */4112void blk_mq_release(struct request_queue *q)4113{4114	struct blk_mq_hw_ctx *hctx, *next;4115	unsigned long i;4116 4117	queue_for_each_hw_ctx(q, hctx, i)4118		WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));4119 4120	/* all hctx are in .unused_hctx_list now */4121	list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {4122		list_del_init(&hctx->hctx_list);4123		kobject_put(&hctx->kobj);4124	}4125 4126	xa_destroy(&q->hctx_table);4127 4128	/*4129	 * release .mq_kobj and sw queue's kobject now because4130	 * both share lifetime with request queue.4131	 */4132	blk_mq_sysfs_deinit(q);4133}4134 4135static bool blk_mq_can_poll(struct blk_mq_tag_set *set)4136{4137	return set->nr_maps > HCTX_TYPE_POLL &&4138		set->map[HCTX_TYPE_POLL].nr_queues;4139}4140 4141struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set,4142		struct queue_limits *lim, void *queuedata)4143{4144	struct queue_limits default_lim = { };4145	struct request_queue *q;4146	int ret;4147 4148	if (!lim)4149		lim = &default_lim;4150	lim->features |= BLK_FEAT_IO_STAT | BLK_FEAT_NOWAIT;4151	if (blk_mq_can_poll(set))4152		lim->features |= BLK_FEAT_POLL;4153 4154	q = blk_alloc_queue(lim, set->numa_node);4155	if (IS_ERR(q))4156		return q;4157	q->queuedata = queuedata;4158	ret = blk_mq_init_allocated_queue(set, q);4159	if (ret) {4160		blk_put_queue(q);4161		return ERR_PTR(ret);4162	}4163	return q;4164}4165EXPORT_SYMBOL(blk_mq_alloc_queue);4166 4167/**4168 * blk_mq_destroy_queue - shutdown a request queue4169 * @q: request queue to shutdown4170 *4171 * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future4172 * requests will be failed with -ENODEV. The caller is responsible for dropping4173 * the reference from blk_mq_alloc_queue() by calling blk_put_queue().4174 *4175 * Context: can sleep4176 */4177void blk_mq_destroy_queue(struct request_queue *q)4178{4179	WARN_ON_ONCE(!queue_is_mq(q));4180	WARN_ON_ONCE(blk_queue_registered(q));4181 4182	might_sleep();4183 4184	blk_queue_flag_set(QUEUE_FLAG_DYING, q);4185	blk_queue_start_drain(q);4186	blk_mq_freeze_queue_wait(q);4187 4188	blk_sync_queue(q);4189	blk_mq_cancel_work_sync(q);4190	blk_mq_exit_queue(q);4191}4192EXPORT_SYMBOL(blk_mq_destroy_queue);4193 4194struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set,4195		struct queue_limits *lim, void *queuedata,4196		struct lock_class_key *lkclass)4197{4198	struct request_queue *q;4199	struct gendisk *disk;4200 4201	q = blk_mq_alloc_queue(set, lim, queuedata);4202	if (IS_ERR(q))4203		return ERR_CAST(q);4204 4205	disk = __alloc_disk_node(q, set->numa_node, lkclass);4206	if (!disk) {4207		blk_mq_destroy_queue(q);4208		blk_put_queue(q);4209		return ERR_PTR(-ENOMEM);4210	}4211	set_bit(GD_OWNS_QUEUE, &disk->state);4212	return disk;4213}4214EXPORT_SYMBOL(__blk_mq_alloc_disk);4215 4216struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,4217		struct lock_class_key *lkclass)4218{4219	struct gendisk *disk;4220 4221	if (!blk_get_queue(q))4222		return NULL;4223	disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);4224	if (!disk)4225		blk_put_queue(q);4226	return disk;4227}4228EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);4229 4230static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(4231		struct blk_mq_tag_set *set, struct request_queue *q,4232		int hctx_idx, int node)4233{4234	struct blk_mq_hw_ctx *hctx = NULL, *tmp;4235 4236	/* reuse dead hctx first */4237	spin_lock(&q->unused_hctx_lock);4238	list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {4239		if (tmp->numa_node == node) {4240			hctx = tmp;4241			break;4242		}4243	}4244	if (hctx)4245		list_del_init(&hctx->hctx_list);4246	spin_unlock(&q->unused_hctx_lock);4247 4248	if (!hctx)4249		hctx = blk_mq_alloc_hctx(q, set, node);4250	if (!hctx)4251		goto fail;4252 4253	if (blk_mq_init_hctx(q, set, hctx, hctx_idx))4254		goto free_hctx;4255 4256	return hctx;4257 4258 free_hctx:4259	kobject_put(&hctx->kobj);4260 fail:4261	return NULL;4262}4263 4264static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,4265						struct request_queue *q)4266{4267	struct blk_mq_hw_ctx *hctx;4268	unsigned long i, j;4269 4270	/* protect against switching io scheduler  */4271	mutex_lock(&q->sysfs_lock);4272	for (i = 0; i < set->nr_hw_queues; i++) {4273		int old_node;4274		int node = blk_mq_get_hctx_node(set, i);4275		struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);4276 4277		if (old_hctx) {4278			old_node = old_hctx->numa_node;4279			blk_mq_exit_hctx(q, set, old_hctx, i);4280		}4281 4282		if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {4283			if (!old_hctx)4284				break;4285			pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",4286					node, old_node);4287			hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);4288			WARN_ON_ONCE(!hctx);4289		}4290	}4291	/*4292	 * Increasing nr_hw_queues fails. Free the newly allocated4293	 * hctxs and keep the previous q->nr_hw_queues.4294	 */4295	if (i != set->nr_hw_queues) {4296		j = q->nr_hw_queues;4297	} else {4298		j = i;4299		q->nr_hw_queues = set->nr_hw_queues;4300	}4301 4302	xa_for_each_start(&q->hctx_table, j, hctx, j)4303		blk_mq_exit_hctx(q, set, hctx, j);4304	mutex_unlock(&q->sysfs_lock);4305}4306 4307int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,4308		struct request_queue *q)4309{4310	/* mark the queue as mq asap */4311	q->mq_ops = set->ops;4312 4313	/*4314	 * ->tag_set has to be setup before initialize hctx, which cpuphp4315	 * handler needs it for checking queue mapping4316	 */4317	q->tag_set = set;4318 4319	if (blk_mq_alloc_ctxs(q))4320		goto err_exit;4321 4322	/* init q->mq_kobj and sw queues' kobjects */4323	blk_mq_sysfs_init(q);4324 4325	INIT_LIST_HEAD(&q->unused_hctx_list);4326	spin_lock_init(&q->unused_hctx_lock);4327 4328	xa_init(&q->hctx_table);4329 4330	blk_mq_realloc_hw_ctxs(set, q);4331	if (!q->nr_hw_queues)4332		goto err_hctxs;4333 4334	INIT_WORK(&q->timeout_work, blk_mq_timeout_work);4335	blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);4336 4337	q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;4338 4339	INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);4340	INIT_LIST_HEAD(&q->flush_list);4341	INIT_LIST_HEAD(&q->requeue_list);4342	spin_lock_init(&q->requeue_lock);4343 4344	q->nr_requests = set->queue_depth;4345 4346	blk_mq_init_cpu_queues(q, set->nr_hw_queues);4347	blk_mq_add_queue_tag_set(set, q);4348	blk_mq_map_swqueue(q);4349	return 0;4350 4351err_hctxs:4352	blk_mq_release(q);4353err_exit:4354	q->mq_ops = NULL;4355	return -ENOMEM;4356}4357EXPORT_SYMBOL(blk_mq_init_allocated_queue);4358 4359/* tags can _not_ be used after returning from blk_mq_exit_queue */4360void blk_mq_exit_queue(struct request_queue *q)4361{4362	struct blk_mq_tag_set *set = q->tag_set;4363 4364	/* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */4365	blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);4366	/* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */4367	blk_mq_del_queue_tag_set(q);4368}4369 4370static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)4371{4372	int i;4373 4374	if (blk_mq_is_shared_tags(set->flags)) {4375		set->shared_tags = blk_mq_alloc_map_and_rqs(set,4376						BLK_MQ_NO_HCTX_IDX,4377						set->queue_depth);4378		if (!set->shared_tags)4379			return -ENOMEM;4380	}4381 4382	for (i = 0; i < set->nr_hw_queues; i++) {4383		if (!__blk_mq_alloc_map_and_rqs(set, i))4384			goto out_unwind;4385		cond_resched();4386	}4387 4388	return 0;4389 4390out_unwind:4391	while (--i >= 0)4392		__blk_mq_free_map_and_rqs(set, i);4393 4394	if (blk_mq_is_shared_tags(set->flags)) {4395		blk_mq_free_map_and_rqs(set, set->shared_tags,4396					BLK_MQ_NO_HCTX_IDX);4397	}4398 4399	return -ENOMEM;4400}4401 4402/*4403 * Allocate the request maps associated with this tag_set. Note that this4404 * may reduce the depth asked for, if memory is tight. set->queue_depth4405 * will be updated to reflect the allocated depth.4406 */4407static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)4408{4409	unsigned int depth;4410	int err;4411 4412	depth = set->queue_depth;4413	do {4414		err = __blk_mq_alloc_rq_maps(set);4415		if (!err)4416			break;4417 4418		set->queue_depth >>= 1;4419		if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {4420			err = -ENOMEM;4421			break;4422		}4423	} while (set->queue_depth);4424 4425	if (!set->queue_depth || err) {4426		pr_err("blk-mq: failed to allocate request map\n");4427		return -ENOMEM;4428	}4429 4430	if (depth != set->queue_depth)4431		pr_info("blk-mq: reduced tag depth (%u -> %u)\n",4432						depth, set->queue_depth);4433 4434	return 0;4435}4436 4437static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)4438{4439	/*4440	 * blk_mq_map_queues() and multiple .map_queues() implementations4441	 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the4442	 * number of hardware queues.4443	 */4444	if (set->nr_maps == 1)4445		set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;4446 4447	if (set->ops->map_queues) {4448		int i;4449 4450		/*4451		 * transport .map_queues is usually done in the following4452		 * way:4453		 *4454		 * for (queue = 0; queue < set->nr_hw_queues; queue++) {4455		 * 	mask = get_cpu_mask(queue)4456		 * 	for_each_cpu(cpu, mask)4457		 * 		set->map[x].mq_map[cpu] = queue;4458		 * }4459		 *4460		 * When we need to remap, the table has to be cleared for4461		 * killing stale mapping since one CPU may not be mapped4462		 * to any hw queue.4463		 */4464		for (i = 0; i < set->nr_maps; i++)4465			blk_mq_clear_mq_map(&set->map[i]);4466 4467		set->ops->map_queues(set);4468	} else {4469		BUG_ON(set->nr_maps > 1);4470		blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);4471	}4472}4473 4474static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,4475				       int new_nr_hw_queues)4476{4477	struct blk_mq_tags **new_tags;4478	int i;4479 4480	if (set->nr_hw_queues >= new_nr_hw_queues)4481		goto done;4482 4483	new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),4484				GFP_KERNEL, set->numa_node);4485	if (!new_tags)4486		return -ENOMEM;4487 4488	if (set->tags)4489		memcpy(new_tags, set->tags, set->nr_hw_queues *4490		       sizeof(*set->tags));4491	kfree(set->tags);4492	set->tags = new_tags;4493 4494	for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {4495		if (!__blk_mq_alloc_map_and_rqs(set, i)) {4496			while (--i >= set->nr_hw_queues)4497				__blk_mq_free_map_and_rqs(set, i);4498			return -ENOMEM;4499		}4500		cond_resched();4501	}4502 4503done:4504	set->nr_hw_queues = new_nr_hw_queues;4505	return 0;4506}4507 4508/*4509 * Alloc a tag set to be associated with one or more request queues.4510 * May fail with EINVAL for various error conditions. May adjust the4511 * requested depth down, if it's too large. In that case, the set4512 * value will be stored in set->queue_depth.4513 */4514int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)4515{4516	int i, ret;4517 4518	BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);4519 4520	if (!set->nr_hw_queues)4521		return -EINVAL;4522	if (!set->queue_depth)4523		return -EINVAL;4524	if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)4525		return -EINVAL;4526 4527	if (!set->ops->queue_rq)4528		return -EINVAL;4529 4530	if (!set->ops->get_budget ^ !set->ops->put_budget)4531		return -EINVAL;4532 4533	if (set->queue_depth > BLK_MQ_MAX_DEPTH) {4534		pr_info("blk-mq: reduced tag depth to %u\n",4535			BLK_MQ_MAX_DEPTH);4536		set->queue_depth = BLK_MQ_MAX_DEPTH;4537	}4538 4539	if (!set->nr_maps)4540		set->nr_maps = 1;4541	else if (set->nr_maps > HCTX_MAX_TYPES)4542		return -EINVAL;4543 4544	/*4545	 * If a crashdump is active, then we are potentially in a very4546	 * memory constrained environment. Limit us to  64 tags to prevent4547	 * using too much memory.4548	 */4549	if (is_kdump_kernel())4550		set->queue_depth = min(64U, set->queue_depth);4551 4552	/*4553	 * There is no use for more h/w queues than cpus if we just have4554	 * a single map4555	 */4556	if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)4557		set->nr_hw_queues = nr_cpu_ids;4558 4559	if (set->flags & BLK_MQ_F_BLOCKING) {4560		set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);4561		if (!set->srcu)4562			return -ENOMEM;4563		ret = init_srcu_struct(set->srcu);4564		if (ret)4565			goto out_free_srcu;4566	}4567 4568	ret = -ENOMEM;4569	set->tags = kcalloc_node(set->nr_hw_queues,4570				 sizeof(struct blk_mq_tags *), GFP_KERNEL,4571				 set->numa_node);4572	if (!set->tags)4573		goto out_cleanup_srcu;4574 4575	for (i = 0; i < set->nr_maps; i++) {4576		set->map[i].mq_map = kcalloc_node(nr_cpu_ids,4577						  sizeof(set->map[i].mq_map[0]),4578						  GFP_KERNEL, set->numa_node);4579		if (!set->map[i].mq_map)4580			goto out_free_mq_map;4581		set->map[i].nr_queues = set->nr_hw_queues;4582	}4583 4584	blk_mq_update_queue_map(set);4585 4586	ret = blk_mq_alloc_set_map_and_rqs(set);4587	if (ret)4588		goto out_free_mq_map;4589 4590	mutex_init(&set->tag_list_lock);4591	INIT_LIST_HEAD(&set->tag_list);4592 4593	return 0;4594 4595out_free_mq_map:4596	for (i = 0; i < set->nr_maps; i++) {4597		kfree(set->map[i].mq_map);4598		set->map[i].mq_map = NULL;4599	}4600	kfree(set->tags);4601	set->tags = NULL;4602out_cleanup_srcu:4603	if (set->flags & BLK_MQ_F_BLOCKING)4604		cleanup_srcu_struct(set->srcu);4605out_free_srcu:4606	if (set->flags & BLK_MQ_F_BLOCKING)4607		kfree(set->srcu);4608	return ret;4609}4610EXPORT_SYMBOL(blk_mq_alloc_tag_set);4611 4612/* allocate and initialize a tagset for a simple single-queue device */4613int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,4614		const struct blk_mq_ops *ops, unsigned int queue_depth,4615		unsigned int set_flags)4616{4617	memset(set, 0, sizeof(*set));4618	set->ops = ops;4619	set->nr_hw_queues = 1;4620	set->nr_maps = 1;4621	set->queue_depth = queue_depth;4622	set->numa_node = NUMA_NO_NODE;4623	set->flags = set_flags;4624	return blk_mq_alloc_tag_set(set);4625}4626EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);4627 4628void blk_mq_free_tag_set(struct blk_mq_tag_set *set)4629{4630	int i, j;4631 4632	for (i = 0; i < set->nr_hw_queues; i++)4633		__blk_mq_free_map_and_rqs(set, i);4634 4635	if (blk_mq_is_shared_tags(set->flags)) {4636		blk_mq_free_map_and_rqs(set, set->shared_tags,4637					BLK_MQ_NO_HCTX_IDX);4638	}4639 4640	for (j = 0; j < set->nr_maps; j++) {4641		kfree(set->map[j].mq_map);4642		set->map[j].mq_map = NULL;4643	}4644 4645	kfree(set->tags);4646	set->tags = NULL;4647	if (set->flags & BLK_MQ_F_BLOCKING) {4648		cleanup_srcu_struct(set->srcu);4649		kfree(set->srcu);4650	}4651}4652EXPORT_SYMBOL(blk_mq_free_tag_set);4653 4654int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)4655{4656	struct blk_mq_tag_set *set = q->tag_set;4657	struct blk_mq_hw_ctx *hctx;4658	int ret;4659	unsigned long i;4660 4661	if (WARN_ON_ONCE(!q->mq_freeze_depth))4662		return -EINVAL;4663 4664	if (!set)4665		return -EINVAL;4666 4667	if (q->nr_requests == nr)4668		return 0;4669 4670	blk_mq_quiesce_queue(q);4671 4672	ret = 0;4673	queue_for_each_hw_ctx(q, hctx, i) {4674		if (!hctx->tags)4675			continue;4676		/*4677		 * If we're using an MQ scheduler, just update the scheduler4678		 * queue depth. This is similar to what the old code would do.4679		 */4680		if (hctx->sched_tags) {4681			ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,4682						      nr, true);4683		} else {4684			ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,4685						      false);4686		}4687		if (ret)4688			break;4689		if (q->elevator && q->elevator->type->ops.depth_updated)4690			q->elevator->type->ops.depth_updated(hctx);4691	}4692	if (!ret) {4693		q->nr_requests = nr;4694		if (blk_mq_is_shared_tags(set->flags)) {4695			if (q->elevator)4696				blk_mq_tag_update_sched_shared_tags(q);4697			else4698				blk_mq_tag_resize_shared_tags(set, nr);4699		}4700	}4701 4702	blk_mq_unquiesce_queue(q);4703 4704	return ret;4705}4706 4707/*4708 * request_queue and elevator_type pair.4709 * It is just used by __blk_mq_update_nr_hw_queues to cache4710 * the elevator_type associated with a request_queue.4711 */4712struct blk_mq_qe_pair {4713	struct list_head node;4714	struct request_queue *q;4715	struct elevator_type *type;4716};4717 4718/*4719 * Cache the elevator_type in qe pair list and switch the4720 * io scheduler to 'none'4721 */4722static bool blk_mq_elv_switch_none(struct list_head *head,4723		struct request_queue *q)4724{4725	struct blk_mq_qe_pair *qe;4726 4727	qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);4728	if (!qe)4729		return false;4730 4731	/* q->elevator needs protection from ->sysfs_lock */4732	mutex_lock(&q->sysfs_lock);4733 4734	/* the check has to be done with holding sysfs_lock */4735	if (!q->elevator) {4736		kfree(qe);4737		goto unlock;4738	}4739 4740	INIT_LIST_HEAD(&qe->node);4741	qe->q = q;4742	qe->type = q->elevator->type;4743	/* keep a reference to the elevator module as we'll switch back */4744	__elevator_get(qe->type);4745	list_add(&qe->node, head);4746	elevator_disable(q);4747unlock:4748	mutex_unlock(&q->sysfs_lock);4749 4750	return true;4751}4752 4753static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,4754						struct request_queue *q)4755{4756	struct blk_mq_qe_pair *qe;4757 4758	list_for_each_entry(qe, head, node)4759		if (qe->q == q)4760			return qe;4761 4762	return NULL;4763}4764 4765static void blk_mq_elv_switch_back(struct list_head *head,4766				  struct request_queue *q)4767{4768	struct blk_mq_qe_pair *qe;4769	struct elevator_type *t;4770 4771	qe = blk_lookup_qe_pair(head, q);4772	if (!qe)4773		return;4774	t = qe->type;4775	list_del(&qe->node);4776	kfree(qe);4777 4778	mutex_lock(&q->sysfs_lock);4779	elevator_switch(q, t);4780	/* drop the reference acquired in blk_mq_elv_switch_none */4781	elevator_put(t);4782	mutex_unlock(&q->sysfs_lock);4783}4784 4785static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,4786							int nr_hw_queues)4787{4788	struct request_queue *q;4789	LIST_HEAD(head);4790	int prev_nr_hw_queues = set->nr_hw_queues;4791	int i;4792 4793	lockdep_assert_held(&set->tag_list_lock);4794 4795	if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)4796		nr_hw_queues = nr_cpu_ids;4797	if (nr_hw_queues < 1)4798		return;4799	if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)4800		return;4801 4802	list_for_each_entry(q, &set->tag_list, tag_set_list)4803		blk_mq_freeze_queue(q);4804	/*4805	 * Switch IO scheduler to 'none', cleaning up the data associated4806	 * with the previous scheduler. We will switch back once we are done4807	 * updating the new sw to hw queue mappings.4808	 */4809	list_for_each_entry(q, &set->tag_list, tag_set_list)4810		if (!blk_mq_elv_switch_none(&head, q))4811			goto switch_back;4812 4813	list_for_each_entry(q, &set->tag_list, tag_set_list) {4814		blk_mq_debugfs_unregister_hctxs(q);4815		blk_mq_sysfs_unregister_hctxs(q);4816	}4817 4818	if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)4819		goto reregister;4820 4821fallback:4822	blk_mq_update_queue_map(set);4823	list_for_each_entry(q, &set->tag_list, tag_set_list) {4824		struct queue_limits lim;4825 4826		blk_mq_realloc_hw_ctxs(set, q);4827 4828		if (q->nr_hw_queues != set->nr_hw_queues) {4829			int i = prev_nr_hw_queues;4830 4831			pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",4832					nr_hw_queues, prev_nr_hw_queues);4833			for (; i < set->nr_hw_queues; i++)4834				__blk_mq_free_map_and_rqs(set, i);4835 4836			set->nr_hw_queues = prev_nr_hw_queues;4837			goto fallback;4838		}4839		lim = queue_limits_start_update(q);4840		if (blk_mq_can_poll(set))4841			lim.features |= BLK_FEAT_POLL;4842		else4843			lim.features &= ~BLK_FEAT_POLL;4844		if (queue_limits_commit_update(q, &lim) < 0)4845			pr_warn("updating the poll flag failed\n");4846		blk_mq_map_swqueue(q);4847	}4848 4849reregister:4850	list_for_each_entry(q, &set->tag_list, tag_set_list) {4851		blk_mq_sysfs_register_hctxs(q);4852		blk_mq_debugfs_register_hctxs(q);4853	}4854 4855switch_back:4856	list_for_each_entry(q, &set->tag_list, tag_set_list)4857		blk_mq_elv_switch_back(&head, q);4858 4859	list_for_each_entry(q, &set->tag_list, tag_set_list)4860		blk_mq_unfreeze_queue(q);4861 4862	/* Free the excess tags when nr_hw_queues shrink. */4863	for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)4864		__blk_mq_free_map_and_rqs(set, i);4865}4866 4867void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)4868{4869	mutex_lock(&set->tag_list_lock);4870	__blk_mq_update_nr_hw_queues(set, nr_hw_queues);4871	mutex_unlock(&set->tag_list_lock);4872}4873EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);4874 4875static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,4876			 struct io_comp_batch *iob, unsigned int flags)4877{4878	long state = get_current_state();4879	int ret;4880 4881	do {4882		ret = q->mq_ops->poll(hctx, iob);4883		if (ret > 0) {4884			__set_current_state(TASK_RUNNING);4885			return ret;4886		}4887 4888		if (signal_pending_state(state, current))4889			__set_current_state(TASK_RUNNING);4890		if (task_is_running(current))4891			return 1;4892 4893		if (ret < 0 || (flags & BLK_POLL_ONESHOT))4894			break;4895		cpu_relax();4896	} while (!need_resched());4897 4898	__set_current_state(TASK_RUNNING);4899	return 0;4900}4901 4902int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,4903		struct io_comp_batch *iob, unsigned int flags)4904{4905	struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);4906 4907	return blk_hctx_poll(q, hctx, iob, flags);4908}4909 4910int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,4911		unsigned int poll_flags)4912{4913	struct request_queue *q = rq->q;4914	int ret;4915 4916	if (!blk_rq_is_poll(rq))4917		return 0;4918	if (!percpu_ref_tryget(&q->q_usage_counter))4919		return 0;4920 4921	ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);4922	blk_queue_exit(q);4923 4924	return ret;4925}4926EXPORT_SYMBOL_GPL(blk_rq_poll);4927 4928unsigned int blk_mq_rq_cpu(struct request *rq)4929{4930	return rq->mq_ctx->cpu;4931}4932EXPORT_SYMBOL(blk_mq_rq_cpu);4933 4934void blk_mq_cancel_work_sync(struct request_queue *q)4935{4936	struct blk_mq_hw_ctx *hctx;4937	unsigned long i;4938 4939	cancel_delayed_work_sync(&q->requeue_work);4940 4941	queue_for_each_hw_ctx(q, hctx, i)4942		cancel_delayed_work_sync(&hctx->run_work);4943}4944 4945static int __init blk_mq_init(void)4946{4947	int i;4948 4949	for_each_possible_cpu(i)4950		init_llist_head(&per_cpu(blk_cpu_done, i));4951	for_each_possible_cpu(i)4952		INIT_CSD(&per_cpu(blk_cpu_csd, i),4953			 __blk_mq_complete_request_remote, NULL);4954	open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);4955 4956	cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,4957				  "block/softirq:dead", NULL,4958				  blk_softirq_cpu_dead);4959	cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,4960				blk_mq_hctx_notify_dead);4961	cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",4962				blk_mq_hctx_notify_online,4963				blk_mq_hctx_notify_offline);4964	return 0;4965}4966subsys_initcall(blk_mq_init);4967