blob: f6dad0886a2fa1bacac778c093717ba21e24b33e [file] [log] [blame]
Christoph Hellwig3dcf60bc2019-04-30 14:42:43 -04001// SPDX-License-Identifier: GPL-2.0
Jens Axboe75bb4622014-05-28 10:15:41 -06002/*
3 * Block multiqueue core code
4 *
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
7 */
Jens Axboe320ae512013-10-24 09:20:05 +01008#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>
Christoph Hellwigfe45e632021-09-20 14:33:27 +020013#include <linux/blk-integrity.h>
Catalin Marinasf75782e2015-09-14 18:16:02 +010014#include <linux/kmemleak.h>
Jens Axboe320ae512013-10-24 09:20:05 +010015#include <linux/mm.h>
16#include <linux/init.h>
17#include <linux/slab.h>
18#include <linux/workqueue.h>
19#include <linux/smp.h>
Christoph Hellwige41d12f2021-09-20 14:33:13 +020020#include <linux/interrupt.h>
Jens Axboe320ae512013-10-24 09:20:05 +010021#include <linux/llist.h>
Jens Axboe320ae512013-10-24 09:20:05 +010022#include <linux/cpu.h>
23#include <linux/cache.h>
24#include <linux/sched/sysctl.h>
Ingo Molnar105ab3d82017-02-01 16:36:40 +010025#include <linux/sched/topology.h>
Ingo Molnar174cd4b2017-02-02 19:15:33 +010026#include <linux/sched/signal.h>
Jens Axboe320ae512013-10-24 09:20:05 +010027#include <linux/delay.h>
Jens Axboeaedcd722014-09-17 08:27:03 -060028#include <linux/crash_dump.h>
Jens Axboe88c7b2b2016-08-25 08:07:30 -060029#include <linux/prefetch.h>
Satya Tangiralaa892c8d2020-05-14 00:37:18 +000030#include <linux/blk-crypto.h>
Christoph Hellwig82d981d2021-11-23 19:53:12 +010031#include <linux/part_stat.h>
Jens Axboe320ae512013-10-24 09:20:05 +010032
33#include <trace/events/block.h>
34
Max Gurtovoy54d4e6a2019-09-16 18:44:29 +030035#include <linux/t10-pi.h>
Jens Axboe320ae512013-10-24 09:20:05 +010036#include "blk.h"
37#include "blk-mq.h"
Omar Sandoval9c1051a2017-05-04 08:17:21 -060038#include "blk-mq-debugfs.h"
Bart Van Assche986d4132018-09-26 14:01:10 -070039#include "blk-pm.h"
Jens Axboecf43e6b2016-11-07 21:32:37 -070040#include "blk-stat.h"
Jens Axboebd166ef2017-01-17 06:03:22 -070041#include "blk-mq-sched.h"
Josef Bacikc1c80382018-07-03 11:14:59 -040042#include "blk-rq-qos.h"
Jan Kara82b74ca2022-06-23 09:48:32 +020043#include "blk-ioprio.h"
Jens Axboe320ae512013-10-24 09:20:05 +010044
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +010045static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
Christoph Hellwigc3077b52020-06-11 08:44:41 +020046
Christoph Hellwig710fa372023-04-13 08:40:54 +020047static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
Christoph Hellwig94aa2282023-04-13 08:40:41 +020048static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
49 struct list_head *list);
Christoph Hellwig3e087732021-10-12 13:12:24 +020050
Christoph Hellwigf70299f2021-10-12 13:12:15 +020051static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q,
52 blk_qc_t qc)
53{
Keith Busch54bdd672023-03-20 12:49:26 -070054 return xa_load(&q->hctx_table, qc);
Christoph Hellwigc6699d62021-10-12 13:12:16 +020055}
56
Christoph Hellwig3e087732021-10-12 13:12:24 +020057static inline blk_qc_t blk_rq_to_qc(struct request *rq)
58{
Keith Busch54bdd672023-03-20 12:49:26 -070059 return rq->mq_hctx->queue_num;
Christoph Hellwig3e087732021-10-12 13:12:24 +020060}
61
Jens Axboe320ae512013-10-24 09:20:05 +010062/*
Yufen Yu85fae292019-03-24 17:57:08 +080063 * Check if any of the ctx, dispatch list or elevator
64 * have pending work in this hardware queue.
Jens Axboe320ae512013-10-24 09:20:05 +010065 */
Jens Axboe79f720a2017-11-10 09:13:21 -070066static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
Jens Axboe320ae512013-10-24 09:20:05 +010067{
Jens Axboe79f720a2017-11-10 09:13:21 -070068 return !list_empty_careful(&hctx->dispatch) ||
69 sbitmap_any_bit_set(&hctx->ctx_map) ||
Jens Axboebd166ef2017-01-17 06:03:22 -070070 blk_mq_sched_has_work(hctx);
Jens Axboe320ae512013-10-24 09:20:05 +010071}
72
73/*
74 * Mark this ctx as having pending work in this hardware queue
75 */
76static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
77 struct blk_mq_ctx *ctx)
78{
Jens Axboef31967f2018-10-29 13:13:29 -060079 const int bit = ctx->index_hw[hctx->type];
80
81 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
82 sbitmap_set_bit(&hctx->ctx_map, bit);
Jens Axboe1429d7c2014-05-19 09:23:55 -060083}
84
85static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
86 struct blk_mq_ctx *ctx)
87{
Jens Axboef31967f2018-10-29 13:13:29 -060088 const int bit = ctx->index_hw[hctx->type];
89
90 sbitmap_clear_bit(&hctx->ctx_map, bit);
Jens Axboe320ae512013-10-24 09:20:05 +010091}
92
Jens Axboef299b7c2017-08-08 17:51:45 -060093struct mq_inflight {
Christoph Hellwig8446fe92020-11-24 09:36:54 +010094 struct block_device *part;
Pavel Begunkova2e80f62019-09-30 21:55:34 +030095 unsigned int inflight[2];
Jens Axboef299b7c2017-08-08 17:51:45 -060096};
97
John Garry2dd65322022-07-06 20:03:53 +080098static bool blk_mq_check_inflight(struct request *rq, void *priv)
Jens Axboef299b7c2017-08-08 17:51:45 -060099{
100 struct mq_inflight *mi = priv;
101
Haisu Wangb81c14c2022-05-30 14:40:59 +0800102 if (rq->part && blk_do_io_stat(rq) &&
103 (!mi->part->bd_partno || rq->part == mi->part) &&
Jeffle Xub0d97552020-12-02 19:11:45 +0800104 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
Pavel Begunkovbb4e6b12019-09-30 21:55:33 +0300105 mi->inflight[rq_data_dir(rq)]++;
Jens Axboe7baa8572018-11-08 10:24:07 -0700106
107 return true;
Jens Axboef299b7c2017-08-08 17:51:45 -0600108}
109
Christoph Hellwig8446fe92020-11-24 09:36:54 +0100110unsigned int blk_mq_in_flight(struct request_queue *q,
111 struct block_device *part)
Jens Axboef299b7c2017-08-08 17:51:45 -0600112{
Pavel Begunkova2e80f62019-09-30 21:55:34 +0300113 struct mq_inflight mi = { .part = part };
Jens Axboef299b7c2017-08-08 17:51:45 -0600114
Jens Axboef299b7c2017-08-08 17:51:45 -0600115 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
Mikulas Patockae016b782018-12-06 11:41:21 -0500116
Pavel Begunkova2e80f62019-09-30 21:55:34 +0300117 return mi.inflight[0] + mi.inflight[1];
Omar Sandovalbf0ddab2018-04-26 00:21:59 -0700118}
119
Christoph Hellwig8446fe92020-11-24 09:36:54 +0100120void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
121 unsigned int inflight[2])
Omar Sandovalbf0ddab2018-04-26 00:21:59 -0700122{
Pavel Begunkova2e80f62019-09-30 21:55:34 +0300123 struct mq_inflight mi = { .part = part };
Omar Sandovalbf0ddab2018-04-26 00:21:59 -0700124
Pavel Begunkovbb4e6b12019-09-30 21:55:33 +0300125 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
Pavel Begunkova2e80f62019-09-30 21:55:34 +0300126 inflight[0] = mi.inflight[0];
127 inflight[1] = mi.inflight[1];
Omar Sandovalbf0ddab2018-04-26 00:21:59 -0700128}
129
Ming Lei1671d522017-03-27 20:06:57 +0800130void blk_freeze_queue_start(struct request_queue *q)
Ming Lei43a5e4e2013-12-26 21:31:35 +0800131{
Bob Liu7996a8b2019-05-21 11:25:55 +0800132 mutex_lock(&q->mq_freeze_lock);
133 if (++q->mq_freeze_depth == 1) {
Dan Williams3ef28e82015-10-21 13:20:12 -0400134 percpu_ref_kill(&q->q_usage_counter);
Bob Liu7996a8b2019-05-21 11:25:55 +0800135 mutex_unlock(&q->mq_freeze_lock);
Jens Axboe344e9ff2018-11-15 12:22:51 -0700136 if (queue_is_mq(q))
Ming Lei055f6e12017-11-09 10:49:53 -0800137 blk_mq_run_hw_queues(q, false);
Bob Liu7996a8b2019-05-21 11:25:55 +0800138 } else {
139 mutex_unlock(&q->mq_freeze_lock);
Tejun Heocddd5d12014-08-16 08:02:24 -0400140 }
Tejun Heof3af0202014-11-04 13:52:27 -0500141}
Ming Lei1671d522017-03-27 20:06:57 +0800142EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
Tejun Heof3af0202014-11-04 13:52:27 -0500143
Keith Busch6bae363e2017-03-01 14:22:10 -0500144void blk_mq_freeze_queue_wait(struct request_queue *q)
Tejun Heof3af0202014-11-04 13:52:27 -0500145{
Dan Williams3ef28e82015-10-21 13:20:12 -0400146 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
Ming Lei43a5e4e2013-12-26 21:31:35 +0800147}
Keith Busch6bae363e2017-03-01 14:22:10 -0500148EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
Ming Lei43a5e4e2013-12-26 21:31:35 +0800149
Keith Buschf91328c2017-03-01 14:22:11 -0500150int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
151 unsigned long timeout)
152{
153 return wait_event_timeout(q->mq_freeze_wq,
154 percpu_ref_is_zero(&q->q_usage_counter),
155 timeout);
156}
157EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
Jens Axboe320ae512013-10-24 09:20:05 +0100158
Tejun Heof3af0202014-11-04 13:52:27 -0500159/*
160 * Guarantee no request is in use, so we can change any data structure of
161 * the queue afterward.
162 */
Dan Williams3ef28e82015-10-21 13:20:12 -0400163void blk_freeze_queue(struct request_queue *q)
Tejun Heof3af0202014-11-04 13:52:27 -0500164{
Dan Williams3ef28e82015-10-21 13:20:12 -0400165 /*
166 * In the !blk_mq case we are only calling this to kill the
167 * q_usage_counter, otherwise this increases the freeze depth
168 * and waits for it to return to zero. For this reason there is
169 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
170 * exported to drivers as the only user for unfreeze is blk_mq.
171 */
Ming Lei1671d522017-03-27 20:06:57 +0800172 blk_freeze_queue_start(q);
Tejun Heof3af0202014-11-04 13:52:27 -0500173 blk_mq_freeze_queue_wait(q);
174}
Dan Williams3ef28e82015-10-21 13:20:12 -0400175
176void blk_mq_freeze_queue(struct request_queue *q)
177{
178 /*
179 * ...just an alias to keep freeze and unfreeze actions balanced
180 * in the blk_mq_* namespace
181 */
182 blk_freeze_queue(q);
183}
Jens Axboec761d962015-01-02 15:05:12 -0700184EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
Tejun Heof3af0202014-11-04 13:52:27 -0500185
Christoph Hellwigaec89dc2021-09-29 09:12:41 +0200186void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
Jens Axboe320ae512013-10-24 09:20:05 +0100187{
Bob Liu7996a8b2019-05-21 11:25:55 +0800188 mutex_lock(&q->mq_freeze_lock);
Christoph Hellwigaec89dc2021-09-29 09:12:41 +0200189 if (force_atomic)
190 q->q_usage_counter.data->force_atomic = true;
Bob Liu7996a8b2019-05-21 11:25:55 +0800191 q->mq_freeze_depth--;
192 WARN_ON_ONCE(q->mq_freeze_depth < 0);
193 if (!q->mq_freeze_depth) {
Bart Van Asschebdd63162018-09-26 14:01:08 -0700194 percpu_ref_resurrect(&q->q_usage_counter);
Jens Axboe320ae512013-10-24 09:20:05 +0100195 wake_up_all(&q->mq_freeze_wq);
Tejun Heoadd703f2014-07-01 10:34:38 -0600196 }
Bob Liu7996a8b2019-05-21 11:25:55 +0800197 mutex_unlock(&q->mq_freeze_lock);
Jens Axboe320ae512013-10-24 09:20:05 +0100198}
Christoph Hellwigaec89dc2021-09-29 09:12:41 +0200199
200void blk_mq_unfreeze_queue(struct request_queue *q)
201{
202 __blk_mq_unfreeze_queue(q, false);
203}
Keith Buschb4c6a022014-12-19 17:54:14 -0700204EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
Jens Axboe320ae512013-10-24 09:20:05 +0100205
Bart Van Assche852ec802017-06-21 10:55:47 -0700206/*
207 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
208 * mpt3sas driver such that this function can be removed.
209 */
210void blk_mq_quiesce_queue_nowait(struct request_queue *q)
211{
Ming Leie70feb82021-10-14 16:17:10 +0800212 unsigned long flags;
213
214 spin_lock_irqsave(&q->queue_lock, flags);
215 if (!q->quiesce_depth++)
216 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
217 spin_unlock_irqrestore(&q->queue_lock, flags);
Bart Van Assche852ec802017-06-21 10:55:47 -0700218}
219EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
220
Bart Van Assche6a83e742016-11-02 10:09:51 -0600221/**
Ming Lei9ef4d022021-11-09 15:11:41 +0800222 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
Christoph Hellwig483239c2022-11-01 16:00:48 +0100223 * @set: tag_set to wait on
Ming Lei9ef4d022021-11-09 15:11:41 +0800224 *
225 * Note: it is driver's responsibility for making sure that quiesce has
Christoph Hellwig483239c2022-11-01 16:00:48 +0100226 * been started on or more of the request_queues of the tag_set. This
227 * function only waits for the quiesce on those request_queues that had
228 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
Ming Lei9ef4d022021-11-09 15:11:41 +0800229 */
Christoph Hellwig483239c2022-11-01 16:00:48 +0100230void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
Ming Lei9ef4d022021-11-09 15:11:41 +0800231{
Christoph Hellwig483239c2022-11-01 16:00:48 +0100232 if (set->flags & BLK_MQ_F_BLOCKING)
233 synchronize_srcu(set->srcu);
Ming Lei704b9142021-12-03 21:15:32 +0800234 else
Ming Lei9ef4d022021-11-09 15:11:41 +0800235 synchronize_rcu();
236}
237EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
238
239/**
Ming Lei69e07c42017-06-06 23:22:07 +0800240 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
Bart Van Assche6a83e742016-11-02 10:09:51 -0600241 * @q: request queue.
242 *
243 * Note: this function does not prevent that the struct request end_io()
Ming Lei69e07c42017-06-06 23:22:07 +0800244 * callback function is invoked. Once this function is returned, we make
245 * sure no dispatch can happen until the queue is unquiesced via
246 * blk_mq_unquiesce_queue().
Bart Van Assche6a83e742016-11-02 10:09:51 -0600247 */
248void blk_mq_quiesce_queue(struct request_queue *q)
249{
Ming Lei1d9e9bc2017-06-06 23:22:08 +0800250 blk_mq_quiesce_queue_nowait(q);
Christoph Hellwig85373802022-11-01 16:00:46 +0100251 /* nothing to wait for non-mq queues */
252 if (queue_is_mq(q))
Christoph Hellwig483239c2022-11-01 16:00:48 +0100253 blk_mq_wait_quiesce_done(q->tag_set);
Bart Van Assche6a83e742016-11-02 10:09:51 -0600254}
255EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
256
Ming Leie4e73912017-06-06 23:22:03 +0800257/*
258 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
259 * @q: request queue.
260 *
261 * This function recovers queue into the state before quiescing
262 * which is done by blk_mq_quiesce_queue.
263 */
264void blk_mq_unquiesce_queue(struct request_queue *q)
265{
Ming Leie70feb82021-10-14 16:17:10 +0800266 unsigned long flags;
267 bool run_queue = false;
268
269 spin_lock_irqsave(&q->queue_lock, flags);
270 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
271 ;
272 } else if (!--q->quiesce_depth) {
273 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
274 run_queue = true;
275 }
276 spin_unlock_irqrestore(&q->queue_lock, flags);
Ming Leif4560ff2017-06-18 14:24:27 -0600277
Ming Lei1d9e9bc2017-06-06 23:22:08 +0800278 /* dispatch requests which are inserted during quiescing */
Ming Leie70feb82021-10-14 16:17:10 +0800279 if (run_queue)
280 blk_mq_run_hw_queues(q, true);
Ming Leie4e73912017-06-06 23:22:03 +0800281}
282EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
283
Chao Leng414dd482022-11-01 16:00:49 +0100284void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
285{
286 struct request_queue *q;
287
288 mutex_lock(&set->tag_list_lock);
289 list_for_each_entry(q, &set->tag_list, tag_set_list) {
290 if (!blk_queue_skip_tagset_quiesce(q))
291 blk_mq_quiesce_queue_nowait(q);
292 }
293 blk_mq_wait_quiesce_done(set);
294 mutex_unlock(&set->tag_list_lock);
295}
296EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
297
298void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
299{
300 struct request_queue *q;
301
302 mutex_lock(&set->tag_list_lock);
303 list_for_each_entry(q, &set->tag_list, tag_set_list) {
304 if (!blk_queue_skip_tagset_quiesce(q))
305 blk_mq_unquiesce_queue(q);
306 }
307 mutex_unlock(&set->tag_list_lock);
308}
309EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
310
Jens Axboeaed3ea92014-12-22 14:04:42 -0700311void blk_mq_wake_waiters(struct request_queue *q)
312{
313 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +0800314 unsigned long i;
Jens Axboeaed3ea92014-12-22 14:04:42 -0700315
316 queue_for_each_hw_ctx(q, hctx, i)
317 if (blk_mq_hw_queue_mapped(hctx))
318 blk_mq_tag_wakeup_all(hctx->tags, true);
319}
320
Christoph Hellwig52fdbbc2021-11-17 07:13:59 +0100321void blk_rq_init(struct request_queue *q, struct request *rq)
322{
323 memset(rq, 0, sizeof(*rq));
324
325 INIT_LIST_HEAD(&rq->queuelist);
326 rq->q = q;
327 rq->__sector = (sector_t) -1;
328 INIT_HLIST_NODE(&rq->hash);
329 RB_CLEAR_NODE(&rq->rb_node);
330 rq->tag = BLK_MQ_NO_TAG;
331 rq->internal_tag = BLK_MQ_NO_TAG;
332 rq->start_time_ns = ktime_get_ns();
333 rq->part = NULL;
334 blk_crypto_rq_set_defaults(rq);
335}
336EXPORT_SYMBOL(blk_rq_init);
337
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200338static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
Jens Axboefe6134f2021-10-19 09:32:58 -0600339 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
Jens Axboe320ae512013-10-24 09:20:05 +0100340{
Pavel Begunkov605f7842021-10-18 21:37:28 +0100341 struct blk_mq_ctx *ctx = data->ctx;
342 struct blk_mq_hw_ctx *hctx = data->hctx;
343 struct request_queue *q = data->q;
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200344 struct request *rq = tags->static_rqs[tag];
Bart Van Asschec3a148d2017-06-20 11:15:43 -0700345
Jens Axboec7b84d42021-10-19 09:33:00 -0600346 rq->q = q;
347 rq->mq_ctx = ctx;
348 rq->mq_hctx = hctx;
349 rq->cmd_flags = data->cmd_flags;
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200350
Pavel Begunkov12845902021-10-18 21:37:29 +0100351 if (data->flags & BLK_MQ_REQ_PM)
Jens Axboe56f8da62021-10-19 09:32:57 -0600352 data->rq_flags |= RQF_PM;
Pavel Begunkov12845902021-10-18 21:37:29 +0100353 if (blk_queue_io_stat(q))
Jens Axboe56f8da62021-10-19 09:32:57 -0600354 data->rq_flags |= RQF_IO_STAT;
355 rq->rq_flags = data->rq_flags;
Pavel Begunkov12845902021-10-18 21:37:29 +0100356
Jens Axboec7b84d42021-10-19 09:33:00 -0600357 if (!(data->rq_flags & RQF_ELV)) {
Jens Axboe320ae512013-10-24 09:20:05 +0100358 rq->tag = tag;
359 rq->internal_tag = BLK_MQ_NO_TAG;
Jens Axboec7b84d42021-10-19 09:33:00 -0600360 } else {
361 rq->tag = BLK_MQ_NO_TAG;
362 rq->internal_tag = tag;
Jens Axboe320ae512013-10-24 09:20:05 +0100363 }
Jens Axboec7b84d42021-10-19 09:33:00 -0600364 rq->timeout = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100365
Pavel Begunkov4f266f22021-10-18 21:37:27 +0100366 if (blk_mq_need_time_stamp(rq))
367 rq->start_time_ns = ktime_get_ns();
368 else
369 rq->start_time_ns = 0;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200370 rq->part = NULL;
Tejun Heo6f816b42019-08-28 15:05:57 -0700371#ifdef CONFIG_BLK_RQ_ALLOC_TIME
372 rq->alloc_time_ns = alloc_time_ns;
373#endif
Omar Sandoval544ccc8d2018-05-09 02:08:50 -0700374 rq->io_start_time_ns = 0;
Hou Tao3d244302019-05-21 15:59:03 +0800375 rq->stats_sectors = 0;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200376 rq->nr_phys_segments = 0;
377#if defined(CONFIG_BLK_DEV_INTEGRITY)
378 rq->nr_integrity_segments = 0;
379#endif
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200380 rq->end_io = NULL;
381 rq->end_io_data = NULL;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200382
Pavel Begunkov4f266f22021-10-18 21:37:27 +0100383 blk_crypto_rq_set_defaults(rq);
384 INIT_LIST_HEAD(&rq->queuelist);
385 /* tag was already set */
386 WRITE_ONCE(rq->deadline, 0);
Jens Axboe0a467d02021-10-14 14:39:59 -0600387 req_ref_set(rq, 1);
Christoph Hellwig7ea4d8a2020-05-29 15:53:10 +0200388
Pavel Begunkov4f266f22021-10-18 21:37:27 +0100389 if (rq->rq_flags & RQF_ELV) {
Christoph Hellwig7ea4d8a2020-05-29 15:53:10 +0200390 struct elevator_queue *e = data->q->elevator;
391
Pavel Begunkov4f266f22021-10-18 21:37:27 +0100392 INIT_HLIST_NODE(&rq->hash);
393 RB_CLEAR_NODE(&rq->rb_node);
394
395 if (!op_is_flush(data->cmd_flags) &&
396 e->type->ops.prepare_request) {
Christoph Hellwig7ea4d8a2020-05-29 15:53:10 +0200397 e->type->ops.prepare_request(rq);
398 rq->rq_flags |= RQF_ELVPRIV;
399 }
400 }
401
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200402 return rq;
Jens Axboe320ae512013-10-24 09:20:05 +0100403}
404
Jens Axboe349302d2021-10-09 13:10:39 -0600405static inline struct request *
406__blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
407 u64 alloc_time_ns)
408{
409 unsigned int tag, tag_offset;
Jens Axboefe6134f2021-10-19 09:32:58 -0600410 struct blk_mq_tags *tags;
Jens Axboe349302d2021-10-09 13:10:39 -0600411 struct request *rq;
Jens Axboefe6134f2021-10-19 09:32:58 -0600412 unsigned long tag_mask;
Jens Axboe349302d2021-10-09 13:10:39 -0600413 int i, nr = 0;
414
Jens Axboefe6134f2021-10-19 09:32:58 -0600415 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
416 if (unlikely(!tag_mask))
Jens Axboe349302d2021-10-09 13:10:39 -0600417 return NULL;
418
Jens Axboefe6134f2021-10-19 09:32:58 -0600419 tags = blk_mq_tags_from_data(data);
420 for (i = 0; tag_mask; i++) {
421 if (!(tag_mask & (1UL << i)))
Jens Axboe349302d2021-10-09 13:10:39 -0600422 continue;
423 tag = tag_offset + i;
Jens Axboea22c00b2021-11-01 06:56:09 -0600424 prefetch(tags->static_rqs[tag]);
Jens Axboefe6134f2021-10-19 09:32:58 -0600425 tag_mask &= ~(1UL << i);
426 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
Jens Axboe013a7f92021-10-13 07:58:52 -0600427 rq_list_add(data->cached_rq, rq);
Jens Axboec5fc7b92021-11-03 05:49:07 -0600428 nr++;
Jens Axboe349302d2021-10-09 13:10:39 -0600429 }
Jens Axboec5fc7b92021-11-03 05:49:07 -0600430 /* caller already holds a reference, add for remainder */
431 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
Jens Axboe349302d2021-10-09 13:10:39 -0600432 data->nr_tags -= nr;
433
Jens Axboe013a7f92021-10-13 07:58:52 -0600434 return rq_list_pop(data->cached_rq);
Jens Axboe349302d2021-10-09 13:10:39 -0600435}
436
Christoph Hellwigb90cfae2021-10-12 12:40:44 +0200437static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200438{
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200439 struct request_queue *q = data->q;
Tejun Heo6f816b42019-08-28 15:05:57 -0700440 u64 alloc_time_ns = 0;
Jens Axboe47c122e2021-10-06 06:34:11 -0600441 struct request *rq;
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200442 unsigned int tag;
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200443
Tejun Heo6f816b42019-08-28 15:05:57 -0700444 /* alloc_time includes depth and tag waits */
445 if (blk_queue_rq_alloc_time(q))
446 alloc_time_ns = ktime_get_ns();
447
Jens Axboef9afca42018-10-29 13:11:38 -0600448 if (data->cmd_flags & REQ_NOWAIT)
Goldwyn Rodrigues03a07c92017-06-20 07:05:46 -0500449 data->flags |= BLK_MQ_REQ_NOWAIT;
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200450
Jens Axboe781dd832021-11-02 08:34:09 -0600451 if (q->elevator) {
452 struct elevator_queue *e = q->elevator;
453
454 data->rq_flags |= RQF_ELV;
455
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200456 /*
Lin Feng8d663f32021-04-15 11:39:20 +0800457 * Flush/passthrough requests are special and go directly to the
Jens Axboe17a51192018-05-09 13:28:50 -0600458 * dispatch list. Don't include reserved tags in the
459 * limiting, as it isn't useful.
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200460 */
Jens Axboef9afca42018-10-29 13:11:38 -0600461 if (!op_is_flush(data->cmd_flags) &&
Lin Feng8d663f32021-04-15 11:39:20 +0800462 !blk_op_is_passthrough(data->cmd_flags) &&
Jens Axboef9afca42018-10-29 13:11:38 -0600463 e->type->ops.limit_depth &&
Jens Axboe17a51192018-05-09 13:28:50 -0600464 !(data->flags & BLK_MQ_REQ_RESERVED))
Jens Axboef9afca42018-10-29 13:11:38 -0600465 e->type->ops.limit_depth(data->cmd_flags, data);
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200466 }
467
Ming Leibf0beec2020-05-29 15:53:15 +0200468retry:
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200469 data->ctx = blk_mq_get_ctx(q);
470 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
Jens Axboe781dd832021-11-02 08:34:09 -0600471 if (!(data->rq_flags & RQF_ELV))
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200472 blk_mq_tag_busy(data->hctx);
473
John Garry99e48cd2022-07-06 20:03:50 +0800474 if (data->flags & BLK_MQ_REQ_RESERVED)
475 data->rq_flags |= RQF_RESV;
476
Ming Leibf0beec2020-05-29 15:53:15 +0200477 /*
Jens Axboe349302d2021-10-09 13:10:39 -0600478 * Try batched alloc if we want more than 1 tag.
479 */
480 if (data->nr_tags > 1) {
481 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
482 if (rq)
483 return rq;
484 data->nr_tags = 1;
485 }
486
487 /*
Ming Leibf0beec2020-05-29 15:53:15 +0200488 * Waiting allocations only fail because of an inactive hctx. In that
489 * case just retry the hctx assignment and tag allocation as CPU hotplug
490 * should have migrated us to an online CPU by now.
491 */
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200492 tag = blk_mq_get_tag(data);
Ming Leibf0beec2020-05-29 15:53:15 +0200493 if (tag == BLK_MQ_NO_TAG) {
494 if (data->flags & BLK_MQ_REQ_NOWAIT)
495 return NULL;
Ming Leibf0beec2020-05-29 15:53:15 +0200496 /*
Jens Axboe349302d2021-10-09 13:10:39 -0600497 * Give up the CPU and sleep for a random short time to
498 * ensure that thread using a realtime scheduling class
499 * are migrated off the CPU, and thus off the hctx that
500 * is going away.
Ming Leibf0beec2020-05-29 15:53:15 +0200501 */
502 msleep(3);
503 goto retry;
504 }
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200505
Jens Axboefe6134f2021-10-19 09:32:58 -0600506 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
507 alloc_time_ns);
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200508}
509
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600510static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
511 struct blk_plug *plug,
512 blk_opf_t opf,
513 blk_mq_req_flags_t flags)
Jens Axboe320ae512013-10-24 09:20:05 +0100514{
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200515 struct blk_mq_alloc_data data = {
516 .q = q,
517 .flags = flags,
Bart Van Assche16458cf2022-07-14 11:06:32 -0700518 .cmd_flags = opf,
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600519 .nr_tags = plug->nr_ios,
520 .cached_rq = &plug->cached_rq,
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200521 };
Jens Axboebd166ef2017-01-17 06:03:22 -0700522 struct request *rq;
Jens Axboe320ae512013-10-24 09:20:05 +0100523
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600524 if (blk_queue_enter(q, flags))
525 return NULL;
526
527 plug->nr_ios = 1;
Jens Axboe320ae512013-10-24 09:20:05 +0100528
Christoph Hellwigb90cfae2021-10-12 12:40:44 +0200529 rq = __blk_mq_alloc_requests(&data);
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600530 if (unlikely(!rq))
531 blk_queue_exit(q);
532 return rq;
533}
534
535static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
536 blk_opf_t opf,
537 blk_mq_req_flags_t flags)
538{
539 struct blk_plug *plug = current->plug;
540 struct request *rq;
541
542 if (!plug)
543 return NULL;
Jinlong Chen40467282022-11-02 10:52:30 +0800544
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600545 if (rq_list_empty(plug->cached_rq)) {
546 if (plug->nr_ios == 1)
547 return NULL;
548 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
Jinlong Chen40467282022-11-02 10:52:30 +0800549 if (!rq)
550 return NULL;
551 } else {
552 rq = rq_list_peek(&plug->cached_rq);
553 if (!rq || rq->q != q)
554 return NULL;
555
556 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
557 return NULL;
558 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
559 return NULL;
560
561 plug->cached_rq = rq_list_next(rq);
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600562 }
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600563
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600564 rq->cmd_flags = opf;
565 INIT_LIST_HEAD(&rq->queuelist);
566 return rq;
567}
568
569struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
570 blk_mq_req_flags_t flags)
571{
572 struct request *rq;
573
574 rq = blk_mq_alloc_cached_request(q, opf, flags);
575 if (!rq) {
576 struct blk_mq_alloc_data data = {
577 .q = q,
578 .flags = flags,
579 .cmd_flags = opf,
580 .nr_tags = 1,
581 };
582 int ret;
583
584 ret = blk_queue_enter(q, flags);
585 if (ret)
586 return ERR_PTR(ret);
587
588 rq = __blk_mq_alloc_requests(&data);
589 if (!rq)
590 goto out_queue_exit;
591 }
Christoph Hellwig0c4de0f2016-07-19 11:31:50 +0200592 rq->__data_len = 0;
593 rq->__sector = (sector_t) -1;
594 rq->bio = rq->biotail = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +0100595 return rq;
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200596out_queue_exit:
597 blk_queue_exit(q);
598 return ERR_PTR(-EWOULDBLOCK);
Jens Axboe320ae512013-10-24 09:20:05 +0100599}
Jens Axboe4bb659b2014-05-09 09:36:49 -0600600EXPORT_SYMBOL(blk_mq_alloc_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100601
Bart Van Asschecd6ce142017-06-20 11:15:39 -0700602struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
Bart Van Assche16458cf2022-07-14 11:06:32 -0700603 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
Ming Lin1f5bd332016-06-13 16:45:21 +0200604{
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200605 struct blk_mq_alloc_data data = {
606 .q = q,
607 .flags = flags,
Bart Van Assche16458cf2022-07-14 11:06:32 -0700608 .cmd_flags = opf,
Jens Axboe47c122e2021-10-06 06:34:11 -0600609 .nr_tags = 1,
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200610 };
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200611 u64 alloc_time_ns = 0;
John Garrye3c5a782022-10-26 18:35:13 +0800612 struct request *rq;
Omar Sandoval6d2809d2017-02-27 10:28:27 -0800613 unsigned int cpu;
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200614 unsigned int tag;
Ming Lin1f5bd332016-06-13 16:45:21 +0200615 int ret;
616
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200617 /* alloc_time includes depth and tag waits */
618 if (blk_queue_rq_alloc_time(q))
619 alloc_time_ns = ktime_get_ns();
620
Ming Lin1f5bd332016-06-13 16:45:21 +0200621 /*
622 * If the tag allocator sleeps we could get an allocation for a
623 * different hardware context. No need to complicate the low level
624 * allocator for this for the rare use case of a command tied to
625 * a specific queue.
626 */
Kemeng Shi6ee858a2023-01-18 17:37:13 +0800627 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
628 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
Ming Lin1f5bd332016-06-13 16:45:21 +0200629 return ERR_PTR(-EINVAL);
630
631 if (hctx_idx >= q->nr_hw_queues)
632 return ERR_PTR(-EIO);
633
Bart Van Assche3a0a5292017-11-09 10:49:58 -0800634 ret = blk_queue_enter(q, flags);
Ming Lin1f5bd332016-06-13 16:45:21 +0200635 if (ret)
636 return ERR_PTR(ret);
637
Christoph Hellwigc8712c62016-09-23 10:25:48 -0600638 /*
639 * Check if the hardware context is actually mapped to anything.
640 * If not tell the caller that it should skip this queue.
641 */
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200642 ret = -EXDEV;
Ming Lei4e5cc992022-03-08 15:32:19 +0800643 data.hctx = xa_load(&q->hctx_table, hctx_idx);
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200644 if (!blk_mq_hw_queue_mapped(data.hctx))
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200645 goto out_queue_exit;
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200646 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
Bart Van Assche14dc7a12022-06-15 14:00:04 -0700647 if (cpu >= nr_cpu_ids)
648 goto out_queue_exit;
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200649 data.ctx = __blk_mq_get_ctx(q, cpu);
Ming Lin1f5bd332016-06-13 16:45:21 +0200650
Christoph Hellwig42fdc5e2020-06-29 17:08:34 +0200651 if (!q->elevator)
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200652 blk_mq_tag_busy(data.hctx);
Jens Axboe781dd832021-11-02 08:34:09 -0600653 else
654 data.rq_flags |= RQF_ELV;
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200655
John Garry99e48cd2022-07-06 20:03:50 +0800656 if (flags & BLK_MQ_REQ_RESERVED)
657 data.rq_flags |= RQF_RESV;
658
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200659 ret = -EWOULDBLOCK;
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200660 tag = blk_mq_get_tag(&data);
661 if (tag == BLK_MQ_NO_TAG)
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200662 goto out_queue_exit;
John Garrye3c5a782022-10-26 18:35:13 +0800663 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
Jens Axboefe6134f2021-10-19 09:32:58 -0600664 alloc_time_ns);
John Garrye3c5a782022-10-26 18:35:13 +0800665 rq->__data_len = 0;
666 rq->__sector = (sector_t) -1;
667 rq->bio = rq->biotail = NULL;
668 return rq;
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200669
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200670out_queue_exit:
671 blk_queue_exit(q);
672 return ERR_PTR(ret);
Ming Lin1f5bd332016-06-13 16:45:21 +0200673}
674EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
675
Keith Busch12f5b932018-05-29 15:52:28 +0200676static void __blk_mq_free_request(struct request *rq)
677{
678 struct request_queue *q = rq->q;
679 struct blk_mq_ctx *ctx = rq->mq_ctx;
Jens Axboeea4f9952018-10-29 15:06:13 -0600680 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Keith Busch12f5b932018-05-29 15:52:28 +0200681 const int sched_tag = rq->internal_tag;
682
Satya Tangiralaa892c8d2020-05-14 00:37:18 +0000683 blk_crypto_free_request(rq);
Bart Van Assche986d4132018-09-26 14:01:10 -0700684 blk_pm_mark_last_busy(rq);
Jens Axboeea4f9952018-10-29 15:06:13 -0600685 rq->mq_hctx = NULL;
Christoph Hellwig766473682020-05-29 15:53:12 +0200686 if (rq->tag != BLK_MQ_NO_TAG)
John Garrycae740a2020-02-26 20:10:15 +0800687 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
Christoph Hellwig766473682020-05-29 15:53:12 +0200688 if (sched_tag != BLK_MQ_NO_TAG)
John Garrycae740a2020-02-26 20:10:15 +0800689 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
Keith Busch12f5b932018-05-29 15:52:28 +0200690 blk_mq_sched_restart(hctx);
691 blk_queue_exit(q);
692}
693
Christoph Hellwig6af54052017-06-16 18:15:22 +0200694void blk_mq_free_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100695{
Jens Axboe320ae512013-10-24 09:20:05 +0100696 struct request_queue *q = rq->q;
Jens Axboeea4f9952018-10-29 15:06:13 -0600697 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Jens Axboe320ae512013-10-24 09:20:05 +0100698
Christoph Hellwig222ee582021-11-26 12:58:11 +0100699 if ((rq->rq_flags & RQF_ELVPRIV) &&
700 q->elevator->type->ops.finish_request)
701 q->elevator->type->ops.finish_request(rq);
Christoph Hellwig6af54052017-06-16 18:15:22 +0200702
Christoph Hellwige8064022016-10-20 15:12:13 +0200703 if (rq->rq_flags & RQF_MQ_INFLIGHT)
John Garrybccf5e22020-08-19 23:20:26 +0800704 __blk_mq_dec_active_requests(hctx);
Jens Axboe87760e52016-11-09 12:38:14 -0700705
Jens Axboe7beb2f82017-09-30 02:08:24 -0600706 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
Christoph Hellwigd152c682021-08-16 15:46:24 +0200707 laptop_io_completion(q->disk->bdi);
Jens Axboe7beb2f82017-09-30 02:08:24 -0600708
Josef Bacika7905042018-07-03 09:32:35 -0600709 rq_qos_done(q, rq);
Jens Axboe0d2602c2014-05-13 15:10:52 -0600710
Keith Busch12f5b932018-05-29 15:52:28 +0200711 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Jens Axboe0a467d02021-10-14 14:39:59 -0600712 if (req_ref_put_and_test(rq))
Keith Busch12f5b932018-05-29 15:52:28 +0200713 __blk_mq_free_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100714}
Jens Axboe1a3b5952014-11-17 10:40:48 -0700715EXPORT_SYMBOL_GPL(blk_mq_free_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100716
Jens Axboe47c122e2021-10-06 06:34:11 -0600717void blk_mq_free_plug_rqs(struct blk_plug *plug)
Jens Axboe320ae512013-10-24 09:20:05 +0100718{
Jens Axboe013a7f92021-10-13 07:58:52 -0600719 struct request *rq;
Jens Axboefe1f4522018-11-28 10:50:07 -0700720
Jens Axboec5fc7b92021-11-03 05:49:07 -0600721 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
Jens Axboe47c122e2021-10-06 06:34:11 -0600722 blk_mq_free_request(rq);
Jens Axboe47c122e2021-10-06 06:34:11 -0600723}
Omar Sandoval522a7772018-05-09 02:08:53 -0700724
Christoph Hellwig22350ad2021-11-17 07:14:02 +0100725void blk_dump_rq_flags(struct request *rq, char *msg)
726{
727 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
Christoph Hellwigf3fa33a2021-11-26 13:18:00 +0100728 rq->q->disk ? rq->q->disk->disk_name : "?",
Bart Van Assche16458cf2022-07-14 11:06:32 -0700729 (__force unsigned long long) rq->cmd_flags);
Christoph Hellwig22350ad2021-11-17 07:14:02 +0100730
731 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
732 (unsigned long long)blk_rq_pos(rq),
733 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
734 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
735 rq->bio, rq->biotail, blk_rq_bytes(rq));
736}
737EXPORT_SYMBOL(blk_dump_rq_flags);
738
Jens Axboe9be3e062021-10-14 09:17:01 -0600739static void req_bio_endio(struct request *rq, struct bio *bio,
740 unsigned int nbytes, blk_status_t error)
741{
Pavel Begunkov478eb722021-10-19 22:24:12 +0100742 if (unlikely(error)) {
Jens Axboe9be3e062021-10-14 09:17:01 -0600743 bio->bi_status = error;
Pavel Begunkov478eb722021-10-19 22:24:12 +0100744 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
Jens Axboe9be3e062021-10-14 09:17:01 -0600745 /*
746 * Partial zone append completions cannot be supported as the
747 * BIO fragments may end up not being written sequentially.
748 */
Pavel Begunkov297db732021-10-22 16:01:44 +0100749 if (bio->bi_iter.bi_size != nbytes)
Jens Axboe9be3e062021-10-14 09:17:01 -0600750 bio->bi_status = BLK_STS_IOERR;
751 else
752 bio->bi_iter.bi_sector = rq->__sector;
753 }
754
Pavel Begunkov478eb722021-10-19 22:24:12 +0100755 bio_advance(bio, nbytes);
756
757 if (unlikely(rq->rq_flags & RQF_QUIET))
758 bio_set_flag(bio, BIO_QUIET);
Jens Axboe9be3e062021-10-14 09:17:01 -0600759 /* don't actually finish bio if it's part of flush sequence */
760 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
761 bio_endio(bio);
762}
763
764static void blk_account_io_completion(struct request *req, unsigned int bytes)
765{
766 if (req->part && blk_do_io_stat(req)) {
767 const int sgrp = op_stat_group(req_op(req));
768
769 part_stat_lock();
770 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
771 part_stat_unlock();
772 }
773}
774
Christoph Hellwig0d7a29a2021-11-17 07:14:03 +0100775static void blk_print_req_error(struct request *req, blk_status_t status)
776{
777 printk_ratelimited(KERN_ERR
778 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
779 "phys_seg %u prio class %u\n",
780 blk_status_to_str(status),
Christoph Hellwigf3fa33a2021-11-26 13:18:00 +0100781 req->q->disk ? req->q->disk->disk_name : "?",
Bart Van Assche16458cf2022-07-14 11:06:32 -0700782 blk_rq_pos(req), (__force u32)req_op(req),
783 blk_op_str(req_op(req)),
784 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
Christoph Hellwig0d7a29a2021-11-17 07:14:03 +0100785 req->nr_phys_segments,
786 IOPRIO_PRIO_CLASS(req->ioprio));
787}
788
Jens Axboe5581a5d2021-12-01 15:01:51 -0700789/*
790 * Fully end IO on a request. Does not support partial completions, or
791 * errors.
792 */
793static void blk_complete_request(struct request *req)
794{
795 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
796 int total_bytes = blk_rq_bytes(req);
797 struct bio *bio = req->bio;
798
799 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
800
801 if (!bio)
802 return;
803
804#ifdef CONFIG_BLK_DEV_INTEGRITY
805 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
806 req->q->integrity.profile->complete_fn(req, total_bytes);
807#endif
808
Eric Biggers9cd1e562023-03-15 11:39:02 -0700809 /*
810 * Upper layers may call blk_crypto_evict_key() anytime after the last
811 * bio_endio(). Therefore, the keyslot must be released before that.
812 */
813 blk_crypto_rq_put_keyslot(req);
814
Jens Axboe5581a5d2021-12-01 15:01:51 -0700815 blk_account_io_completion(req, total_bytes);
816
817 do {
818 struct bio *next = bio->bi_next;
819
820 /* Completion has already been traced */
821 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
Pankaj Raghava12821d2022-02-11 10:34:25 +0100822
823 if (req_op(req) == REQ_OP_ZONE_APPEND)
824 bio->bi_iter.bi_sector = req->__sector;
825
Jens Axboe5581a5d2021-12-01 15:01:51 -0700826 if (!is_flush)
827 bio_endio(bio);
828 bio = next;
829 } while (bio);
830
831 /*
832 * Reset counters so that the request stacking driver
833 * can find how many bytes remain in the request
834 * later.
835 */
Jens Axboeab3e1d32022-09-21 08:24:16 -0600836 if (!req->end_io) {
837 req->bio = NULL;
838 req->__data_len = 0;
839 }
Jens Axboe5581a5d2021-12-01 15:01:51 -0700840}
841
Jens Axboe9be3e062021-10-14 09:17:01 -0600842/**
843 * blk_update_request - Complete multiple bytes without completing the request
844 * @req: the request being processed
845 * @error: block status code
846 * @nr_bytes: number of bytes to complete for @req
847 *
848 * Description:
849 * Ends I/O on a number of bytes attached to @req, but doesn't complete
850 * the request structure even if @req doesn't have leftover.
851 * If @req has leftover, sets it up for the next range of segments.
852 *
853 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
854 * %false return from this function.
855 *
856 * Note:
857 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
858 * except in the consistency check at the end of this function.
859 *
860 * Return:
861 * %false - this request doesn't have any more data
862 * %true - this request has more data
863 **/
864bool blk_update_request(struct request *req, blk_status_t error,
865 unsigned int nr_bytes)
866{
867 int total_bytes;
868
Christoph Hellwig8a7d2672021-10-18 10:45:18 +0200869 trace_block_rq_complete(req, error, nr_bytes);
Jens Axboe9be3e062021-10-14 09:17:01 -0600870
871 if (!req->bio)
872 return false;
873
874#ifdef CONFIG_BLK_DEV_INTEGRITY
875 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
876 error == BLK_STS_OK)
877 req->q->integrity.profile->complete_fn(req, nr_bytes);
878#endif
879
Eric Biggers9cd1e562023-03-15 11:39:02 -0700880 /*
881 * Upper layers may call blk_crypto_evict_key() anytime after the last
882 * bio_endio(). Therefore, the keyslot must be released before that.
883 */
884 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
885 __blk_crypto_rq_put_keyslot(req);
886
Jens Axboe9be3e062021-10-14 09:17:01 -0600887 if (unlikely(error && !blk_rq_is_passthrough(req) &&
Christoph Hellwig3d973a72022-03-23 17:38:15 +0100888 !(req->rq_flags & RQF_QUIET)) &&
889 !test_bit(GD_DEAD, &req->q->disk->state)) {
Jens Axboe9be3e062021-10-14 09:17:01 -0600890 blk_print_req_error(req, error);
Yang Shid5869fd2022-02-10 14:52:22 -0800891 trace_block_rq_error(req, error, nr_bytes);
892 }
Jens Axboe9be3e062021-10-14 09:17:01 -0600893
894 blk_account_io_completion(req, nr_bytes);
895
896 total_bytes = 0;
897 while (req->bio) {
898 struct bio *bio = req->bio;
899 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
900
901 if (bio_bytes == bio->bi_iter.bi_size)
902 req->bio = bio->bi_next;
903
904 /* Completion has already been traced */
905 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
906 req_bio_endio(req, bio, bio_bytes, error);
907
908 total_bytes += bio_bytes;
909 nr_bytes -= bio_bytes;
910
911 if (!nr_bytes)
912 break;
913 }
914
915 /*
916 * completely done
917 */
918 if (!req->bio) {
919 /*
920 * Reset counters so that the request stacking driver
921 * can find how many bytes remain in the request
922 * later.
923 */
924 req->__data_len = 0;
925 return false;
926 }
927
928 req->__data_len -= total_bytes;
929
930 /* update sector only for requests with clear definition of sector */
931 if (!blk_rq_is_passthrough(req))
932 req->__sector += total_bytes >> 9;
933
934 /* mixed attributes always follow the first bio */
935 if (req->rq_flags & RQF_MIXED_MERGE) {
936 req->cmd_flags &= ~REQ_FAILFAST_MASK;
937 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
938 }
939
940 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
941 /*
942 * If total number of sectors is less than the first segment
943 * size, something has gone terribly wrong.
944 */
945 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
946 blk_dump_rq_flags(req, "request botched");
947 req->__data_len = blk_rq_cur_bytes(req);
948 }
949
950 /* recalculate the number of segments */
951 req->nr_phys_segments = blk_recalc_rq_segments(req);
952 }
953
954 return true;
955}
956EXPORT_SYMBOL_GPL(blk_update_request);
957
Christoph Hellwig450b7872021-11-17 07:14:01 +0100958static inline void blk_account_io_done(struct request *req, u64 now)
959{
960 /*
961 * Account IO completion. flush_rq isn't accounted as a
962 * normal IO on queueing nor completion. Accounting the
963 * containing request is enough.
964 */
965 if (blk_do_io_stat(req) && req->part &&
Chaitanya Kulkarni06965032023-03-27 00:34:27 -0700966 !(req->rq_flags & RQF_FLUSH_SEQ)) {
967 const int sgrp = op_stat_group(req_op(req));
Christoph Hellwig450b7872021-11-17 07:14:01 +0100968
Chaitanya Kulkarni06965032023-03-27 00:34:27 -0700969 part_stat_lock();
970 update_io_ticks(req->part, jiffies, true);
971 part_stat_inc(req->part, ios[sgrp]);
972 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
973 part_stat_unlock();
974 }
Christoph Hellwig450b7872021-11-17 07:14:01 +0100975}
976
977static inline void blk_account_io_start(struct request *req)
978{
Chaitanya Kulkarnie165fb42023-03-27 00:34:26 -0700979 if (blk_do_io_stat(req)) {
980 /*
981 * All non-passthrough requests are created from a bio with one
982 * exception: when a flush command that is part of a flush sequence
983 * generated by the state machine in blk-flush.c is cloned onto the
984 * lower device by dm-multipath we can get here without a bio.
985 */
986 if (req->bio)
987 req->part = req->bio->bi_bdev;
988 else
989 req->part = req->q->disk->part0;
990
991 part_stat_lock();
992 update_io_ticks(req->part, jiffies, false);
993 part_stat_unlock();
994 }
Christoph Hellwig450b7872021-11-17 07:14:01 +0100995}
996
Jens Axboef794f332021-10-08 05:50:46 -0600997static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
998{
Keith Busch54bdd672023-03-20 12:49:26 -0700999 if (rq->rq_flags & RQF_STATS)
Omar Sandoval522a7772018-05-09 02:08:53 -07001000 blk_stat_add(rq, now);
Omar Sandoval4bc63392018-05-09 02:08:52 -07001001
Baolin Wang87890092020-07-04 15:28:21 +08001002 blk_mq_sched_completed_request(rq, now);
Omar Sandoval522a7772018-05-09 02:08:53 -07001003 blk_account_io_done(rq, now);
Jens Axboef794f332021-10-08 05:50:46 -06001004}
1005
Ming Lei0d11e6a2013-12-05 10:50:39 -07001006inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
Christoph Hellwig91b63632014-04-16 09:44:53 +02001007{
Jens Axboef794f332021-10-08 05:50:46 -06001008 if (blk_mq_need_time_stamp(rq))
1009 __blk_mq_end_request_acct(rq, ktime_get_ns());
Jens Axboe320ae512013-10-24 09:20:05 +01001010
1011 if (rq->end_io) {
Josef Bacika7905042018-07-03 09:32:35 -06001012 rq_qos_done(rq->q, rq);
Jens Axboede671d62022-09-21 15:19:54 -06001013 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1014 blk_mq_free_request(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +02001015 } else {
Jens Axboe320ae512013-10-24 09:20:05 +01001016 blk_mq_free_request(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +02001017 }
Jens Axboe320ae512013-10-24 09:20:05 +01001018}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -07001019EXPORT_SYMBOL(__blk_mq_end_request);
Christoph Hellwig63151a42014-04-16 09:44:52 +02001020
Christoph Hellwig2a842ac2017-06-03 09:38:04 +02001021void blk_mq_end_request(struct request *rq, blk_status_t error)
Christoph Hellwig63151a42014-04-16 09:44:52 +02001022{
1023 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1024 BUG();
Christoph Hellwigc8a446a2014-09-13 16:40:10 -07001025 __blk_mq_end_request(rq, error);
Christoph Hellwig63151a42014-04-16 09:44:52 +02001026}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -07001027EXPORT_SYMBOL(blk_mq_end_request);
Jens Axboe320ae512013-10-24 09:20:05 +01001028
Jens Axboef794f332021-10-08 05:50:46 -06001029#define TAG_COMP_BATCH 32
1030
1031static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1032 int *tag_array, int nr_tags)
1033{
1034 struct request_queue *q = hctx->queue;
1035
Ming Lei3b87c6e2021-11-02 23:36:19 +08001036 /*
1037 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1038 * update hctx->nr_active in batch
1039 */
1040 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1041 __blk_mq_sub_active_requests(hctx, nr_tags);
1042
Jens Axboef794f332021-10-08 05:50:46 -06001043 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1044 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1045}
1046
1047void blk_mq_end_request_batch(struct io_comp_batch *iob)
1048{
1049 int tags[TAG_COMP_BATCH], nr_tags = 0;
Jens Axboe02f7eab2021-10-28 12:08:34 -06001050 struct blk_mq_hw_ctx *cur_hctx = NULL;
Jens Axboef794f332021-10-08 05:50:46 -06001051 struct request *rq;
1052 u64 now = 0;
1053
1054 if (iob->need_ts)
1055 now = ktime_get_ns();
1056
1057 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1058 prefetch(rq->bio);
1059 prefetch(rq->rq_next);
1060
Jens Axboe5581a5d2021-12-01 15:01:51 -07001061 blk_complete_request(rq);
Jens Axboef794f332021-10-08 05:50:46 -06001062 if (iob->need_ts)
1063 __blk_mq_end_request_acct(rq, now);
1064
Jens Axboe98b26a02021-11-26 09:53:23 -07001065 rq_qos_done(rq->q, rq);
1066
Jens Axboeab3e1d32022-09-21 08:24:16 -06001067 /*
1068 * If end_io handler returns NONE, then it still has
1069 * ownership of the request.
1070 */
1071 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1072 continue;
1073
Jens Axboef794f332021-10-08 05:50:46 -06001074 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Jens Axboe0a467d02021-10-14 14:39:59 -06001075 if (!req_ref_put_and_test(rq))
Jens Axboef794f332021-10-08 05:50:46 -06001076 continue;
1077
1078 blk_crypto_free_request(rq);
1079 blk_pm_mark_last_busy(rq);
Jens Axboef794f332021-10-08 05:50:46 -06001080
Jens Axboe02f7eab2021-10-28 12:08:34 -06001081 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1082 if (cur_hctx)
1083 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
Jens Axboef794f332021-10-08 05:50:46 -06001084 nr_tags = 0;
Jens Axboe02f7eab2021-10-28 12:08:34 -06001085 cur_hctx = rq->mq_hctx;
Jens Axboef794f332021-10-08 05:50:46 -06001086 }
1087 tags[nr_tags++] = rq->tag;
Jens Axboef794f332021-10-08 05:50:46 -06001088 }
1089
1090 if (nr_tags)
Jens Axboe02f7eab2021-10-28 12:08:34 -06001091 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
Jens Axboef794f332021-10-08 05:50:46 -06001092}
1093EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1094
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001095static void blk_complete_reqs(struct llist_head *list)
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001096{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001097 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1098 struct request *rq, *next;
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001099
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001100 llist_for_each_entry_safe(rq, next, entry, ipi_list)
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001101 rq->q->mq_ops->complete(rq);
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001102}
1103
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001104static __latent_entropy void blk_done_softirq(struct softirq_action *h)
Christoph Hellwig115243f2020-06-11 08:44:42 +02001105{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001106 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001107}
1108
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001109static int blk_softirq_cpu_dead(unsigned int cpu)
1110{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001111 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001112 return 0;
1113}
1114
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08001115static void __blk_mq_complete_request_remote(void *data)
Jens Axboe320ae512013-10-24 09:20:05 +01001116{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001117 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
Jens Axboe320ae512013-10-24 09:20:05 +01001118}
1119
Christoph Hellwig963395262020-06-11 08:44:49 +02001120static inline bool blk_mq_complete_need_ipi(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001121{
Christoph Hellwig963395262020-06-11 08:44:49 +02001122 int cpu = raw_smp_processor_id();
Jens Axboe320ae512013-10-24 09:20:05 +01001123
Christoph Hellwig963395262020-06-11 08:44:49 +02001124 if (!IS_ENABLED(CONFIG_SMP) ||
1125 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1126 return false;
Sebastian Andrzej Siewior71425182020-12-04 20:13:54 +01001127 /*
1128 * With force threaded interrupts enabled, raising softirq from an SMP
1129 * function call will always result in waking the ksoftirqd thread.
1130 * This is probably worse than completing the request on a different
1131 * cache domain.
1132 */
Tanner Love91cc4702021-06-02 14:03:38 -04001133 if (force_irqthreads())
Sebastian Andrzej Siewior71425182020-12-04 20:13:54 +01001134 return false;
Christoph Hellwig963395262020-06-11 08:44:49 +02001135
1136 /* same CPU or cache domain? Complete locally */
1137 if (cpu == rq->mq_ctx->cpu ||
1138 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1139 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1140 return false;
1141
1142 /* don't try to IPI to an offline CPU */
1143 return cpu_online(rq->mq_ctx->cpu);
1144}
1145
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001146static void blk_mq_complete_send_ipi(struct request *rq)
1147{
1148 struct llist_head *list;
1149 unsigned int cpu;
1150
1151 cpu = rq->mq_ctx->cpu;
1152 list = &per_cpu(blk_cpu_done, cpu);
1153 if (llist_add(&rq->ipi_list, list)) {
1154 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1155 smp_call_function_single_async(cpu, &rq->csd);
1156 }
1157}
1158
1159static void blk_mq_raise_softirq(struct request *rq)
1160{
1161 struct llist_head *list;
1162
1163 preempt_disable();
1164 list = this_cpu_ptr(&blk_cpu_done);
1165 if (llist_add(&rq->ipi_list, list))
1166 raise_softirq(BLOCK_SOFTIRQ);
1167 preempt_enable();
1168}
1169
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001170bool blk_mq_complete_request_remote(struct request *rq)
1171{
Keith Buschaf78ff72018-11-26 09:54:30 -07001172 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
Ming Lei36e76532018-09-28 16:42:20 +08001173
Jens Axboe4ab32bf2018-11-18 16:15:35 -07001174 /*
Liu Songf1684202022-09-21 11:32:03 +08001175 * For request which hctx has only one ctx mapping,
1176 * or a polled request, always complete locally,
1177 * it's pointless to redirect the completion.
Jens Axboe4ab32bf2018-11-18 16:15:35 -07001178 */
Liu Songf1684202022-09-21 11:32:03 +08001179 if (rq->mq_hctx->nr_ctx == 1 ||
1180 rq->cmd_flags & REQ_POLLED)
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001181 return false;
Jens Axboe320ae512013-10-24 09:20:05 +01001182
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001183 if (blk_mq_complete_need_ipi(rq)) {
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001184 blk_mq_complete_send_ipi(rq);
1185 return true;
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -08001186 }
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001187
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001188 if (rq->q->nr_hw_queues == 1) {
1189 blk_mq_raise_softirq(rq);
1190 return true;
1191 }
1192 return false;
Jens Axboe320ae512013-10-24 09:20:05 +01001193}
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001194EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1195
Jens Axboe320ae512013-10-24 09:20:05 +01001196/**
Christoph Hellwig15f73f52020-06-11 08:44:47 +02001197 * blk_mq_complete_request - end I/O on a request
1198 * @rq: the request being processed
Jens Axboe320ae512013-10-24 09:20:05 +01001199 *
Christoph Hellwig15f73f52020-06-11 08:44:47 +02001200 * Description:
1201 * Complete a request by scheduling the ->complete_rq operation.
1202 **/
1203void blk_mq_complete_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001204{
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001205 if (!blk_mq_complete_request_remote(rq))
Christoph Hellwig963395262020-06-11 08:44:49 +02001206 rq->q->mq_ops->complete(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01001207}
Christoph Hellwig15f73f52020-06-11 08:44:47 +02001208EXPORT_SYMBOL(blk_mq_complete_request);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08001209
1210/**
André Almeida105663f2020-01-06 15:08:18 -03001211 * blk_mq_start_request - Start processing a request
1212 * @rq: Pointer to request to be started
1213 *
1214 * Function used by device drivers to notify the block layer that a request
1215 * is going to be processed now, so blk layer can do proper initializations
1216 * such as starting the timeout timer.
1217 */
Christoph Hellwige2490072014-09-13 16:40:09 -07001218void blk_mq_start_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001219{
1220 struct request_queue *q = rq->q;
1221
Christoph Hellwiga54895f2020-12-03 17:21:39 +01001222 trace_block_rq_issue(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01001223
Jens Axboecf43e6b2016-11-07 21:32:37 -07001224 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
Tejun Heo4cddeac2022-04-27 09:49:12 -10001225 rq->io_start_time_ns = ktime_get_ns();
Hou Tao3d244302019-05-21 15:59:03 +08001226 rq->stats_sectors = blk_rq_sectors(rq);
Jens Axboecf43e6b2016-11-07 21:32:37 -07001227 rq->rq_flags |= RQF_STATS;
Josef Bacika7905042018-07-03 09:32:35 -06001228 rq_qos_issue(q, rq);
Jens Axboecf43e6b2016-11-07 21:32:37 -07001229 }
1230
Tejun Heo1d9bd512018-01-09 08:29:48 -08001231 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
Jens Axboe538b7532014-09-16 10:37:37 -06001232
Tejun Heo1d9bd512018-01-09 08:29:48 -08001233 blk_add_timer(rq);
Keith Busch12f5b932018-05-29 15:52:28 +02001234 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -08001235
Max Gurtovoy54d4e6a2019-09-16 18:44:29 +03001236#ifdef CONFIG_BLK_DEV_INTEGRITY
1237 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1238 q->integrity.profile->prepare_fn(rq);
1239#endif
Christoph Hellwig3e087732021-10-12 13:12:24 +02001240 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1241 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
Jens Axboe320ae512013-10-24 09:20:05 +01001242}
Christoph Hellwige2490072014-09-13 16:40:09 -07001243EXPORT_SYMBOL(blk_mq_start_request);
Jens Axboe320ae512013-10-24 09:20:05 +01001244
Ming Leia327c342022-05-12 22:00:10 +08001245/*
1246 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1247 * queues. This is important for md arrays to benefit from merging
1248 * requests.
1249 */
1250static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1251{
1252 if (plug->multiple_queues)
1253 return BLK_MAX_REQUEST_COUNT * 2;
1254 return BLK_MAX_REQUEST_COUNT;
1255}
1256
1257static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1258{
1259 struct request *last = rq_list_peek(&plug->mq_list);
1260
1261 if (!plug->rq_count) {
1262 trace_block_plug(rq->q);
1263 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1264 (!blk_queue_nomerges(rq->q) &&
1265 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1266 blk_mq_flush_plug_list(plug, false);
Al Viro878eb6e2022-11-01 00:54:13 +00001267 last = NULL;
Ming Leia327c342022-05-12 22:00:10 +08001268 trace_block_plug(rq->q);
1269 }
1270
1271 if (!plug->multiple_queues && last && last->q != rq->q)
1272 plug->multiple_queues = true;
1273 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1274 plug->has_elevator = true;
1275 rq->rq_next = NULL;
1276 rq_list_add(&plug->mq_list, rq);
1277 plug->rq_count++;
1278}
1279
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001280/**
1281 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001282 * @rq: request to insert
1283 * @at_head: insert request at head or tail of queue
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001284 *
1285 * Description:
1286 * Insert a fully prepared request at the back of the I/O scheduler queue
1287 * for execution. Don't wait for completion.
1288 *
1289 * Note:
1290 * This function will invoke @done directly if the queue is dead.
1291 */
Christoph Hellwige2e530862022-05-24 14:15:30 +02001292void blk_execute_rq_nowait(struct request *rq, bool at_head)
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001293{
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001294 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1295
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001296 WARN_ON(irqs_disabled());
1297 WARN_ON(!blk_rq_is_passthrough(rq));
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001298
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001299 blk_account_io_start(rq);
Pankaj Raghav110fdb42022-09-29 16:41:41 +02001300
1301 /*
1302 * As plugging can be enabled for passthrough requests on a zoned
1303 * device, directly accessing the plug instead of using blk_mq_plug()
1304 * should not have any consequences.
1305 */
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001306 if (current->plug && !at_head) {
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001307 blk_add_rq_to_plug(current->plug, rq);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001308 return;
1309 }
1310
Christoph Hellwig710fa372023-04-13 08:40:54 +02001311 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001312 blk_mq_run_hw_queue(hctx, false);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001313}
1314EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1315
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001316struct blk_rq_wait {
1317 struct completion done;
1318 blk_status_t ret;
1319};
1320
Jens Axboede671d62022-09-21 15:19:54 -06001321static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001322{
1323 struct blk_rq_wait *wait = rq->end_io_data;
1324
1325 wait->ret = ret;
1326 complete(&wait->done);
Jens Axboede671d62022-09-21 15:19:54 -06001327 return RQ_END_IO_NONE;
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001328}
1329
Kanchan Joshic6e99ea2022-08-23 21:44:42 +05301330bool blk_rq_is_poll(struct request *rq)
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001331{
1332 if (!rq->mq_hctx)
1333 return false;
1334 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1335 return false;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001336 return true;
1337}
Kanchan Joshic6e99ea2022-08-23 21:44:42 +05301338EXPORT_SYMBOL_GPL(blk_rq_is_poll);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001339
1340static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1341{
1342 do {
Keith Busch38a8c4d12023-03-31 11:00:56 -07001343 blk_mq_poll(rq->q, blk_rq_to_qc(rq), NULL, 0);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001344 cond_resched();
1345 } while (!completion_done(wait));
1346}
1347
1348/**
1349 * blk_execute_rq - insert a request into queue for execution
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001350 * @rq: request to insert
1351 * @at_head: insert request at head or tail of queue
1352 *
1353 * Description:
1354 * Insert a fully prepared request at the back of the I/O scheduler queue
1355 * for execution and wait for completion.
1356 * Return: The blk_status_t result provided to blk_mq_end_request().
1357 */
Christoph Hellwigb84ba302021-11-26 13:18:01 +01001358blk_status_t blk_execute_rq(struct request *rq, bool at_head)
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001359{
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001360 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001361 struct blk_rq_wait wait = {
1362 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1363 };
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001364
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001365 WARN_ON(irqs_disabled());
1366 WARN_ON(!blk_rq_is_passthrough(rq));
1367
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001368 rq->end_io_data = &wait;
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001369 rq->end_io = blk_end_sync_rq;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001370
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001371 blk_account_io_start(rq);
Christoph Hellwig710fa372023-04-13 08:40:54 +02001372 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001373 blk_mq_run_hw_queue(hctx, false);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001374
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001375 if (blk_rq_is_poll(rq)) {
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001376 blk_rq_poll_completion(rq, &wait.done);
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001377 } else {
1378 /*
1379 * Prevent hang_check timer from firing at us during very long
1380 * I/O
1381 */
1382 unsigned long hang_check = sysctl_hung_task_timeout_secs;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001383
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001384 if (hang_check)
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001385 while (!wait_for_completion_io_timeout(&wait.done,
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001386 hang_check * (HZ/2)))
1387 ;
1388 else
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001389 wait_for_completion_io(&wait.done);
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001390 }
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001391
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001392 return wait.ret;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001393}
1394EXPORT_SYMBOL(blk_execute_rq);
1395
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001396static void __blk_mq_requeue_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001397{
1398 struct request_queue *q = rq->q;
1399
Ming Lei923218f2017-11-02 23:24:38 +08001400 blk_mq_put_driver_tag(rq);
1401
Christoph Hellwiga54895f2020-12-03 17:21:39 +01001402 trace_block_rq_requeue(rq);
Josef Bacika7905042018-07-03 09:32:35 -06001403 rq_qos_requeue(q, rq);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -08001404
Keith Busch12f5b932018-05-29 15:52:28 +02001405 if (blk_mq_request_started(rq)) {
1406 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Christoph Hellwigda661262018-06-14 13:58:45 +02001407 rq->rq_flags &= ~RQF_TIMED_OUT;
Christoph Hellwige2490072014-09-13 16:40:09 -07001408 }
Jens Axboe320ae512013-10-24 09:20:05 +01001409}
1410
Bart Van Assche2b053ac2016-10-28 17:21:41 -07001411void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001412{
Christoph Hellwig214a4412023-04-13 08:40:53 +02001413 struct request_queue *q = rq->q;
1414
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001415 __blk_mq_requeue_request(rq);
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001416
Ming Lei105976f2018-02-23 23:36:56 +08001417 /* this request will be re-inserted to io scheduler queue */
1418 blk_mq_sched_requeue_request(rq);
1419
Christoph Hellwigb12e5c62023-04-13 08:40:57 +02001420 blk_mq_add_to_requeue_list(rq, BLK_MQ_INSERT_AT_HEAD);
Christoph Hellwig214a4412023-04-13 08:40:53 +02001421
1422 if (kick_requeue_list)
1423 blk_mq_kick_requeue_list(q);
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001424}
1425EXPORT_SYMBOL(blk_mq_requeue_request);
1426
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001427static void blk_mq_requeue_work(struct work_struct *work)
1428{
1429 struct request_queue *q =
Mike Snitzer28494502016-09-14 13:28:30 -04001430 container_of(work, struct request_queue, requeue_work.work);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001431 LIST_HEAD(rq_list);
1432 struct request *rq, *next;
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001433
Jens Axboe18e97812017-07-27 08:03:57 -06001434 spin_lock_irq(&q->requeue_lock);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001435 list_splice_init(&q->requeue_list, &rq_list);
Jens Axboe18e97812017-07-27 08:03:57 -06001436 spin_unlock_irq(&q->requeue_lock);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001437
1438 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
Jianchao Wangaef18972019-02-12 09:56:25 +08001439 /*
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001440 * If RQF_DONTPREP ist set, the request has been started by the
1441 * driver already and might have driver-specific data allocated
1442 * already. Insert it into the hctx dispatch list to avoid
1443 * block layer merges for the request.
Jianchao Wangaef18972019-02-12 09:56:25 +08001444 */
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001445 if (rq->rq_flags & RQF_DONTPREP) {
1446 rq->rq_flags &= ~RQF_SOFTBARRIER;
1447 list_del_init(&rq->queuelist);
Christoph Hellwig2b597612023-04-13 08:40:55 +02001448 blk_mq_request_bypass_insert(rq, 0);
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001449 } else if (rq->rq_flags & RQF_SOFTBARRIER) {
1450 rq->rq_flags &= ~RQF_SOFTBARRIER;
1451 list_del_init(&rq->queuelist);
Christoph Hellwig710fa372023-04-13 08:40:54 +02001452 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001453 }
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001454 }
1455
1456 while (!list_empty(&rq_list)) {
1457 rq = list_entry(rq_list.next, struct request, queuelist);
1458 list_del_init(&rq->queuelist);
Christoph Hellwig710fa372023-04-13 08:40:54 +02001459 blk_mq_insert_request(rq, 0);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001460 }
1461
Bart Van Assche52d7f1b2016-10-28 17:20:32 -07001462 blk_mq_run_hw_queues(q, false);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001463}
1464
Christoph Hellwigb12e5c62023-04-13 08:40:57 +02001465void blk_mq_add_to_requeue_list(struct request *rq, blk_insert_t insert_flags)
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001466{
1467 struct request_queue *q = rq->q;
1468 unsigned long flags;
1469
1470 /*
1471 * We abuse this flag that is otherwise used by the I/O scheduler to
Jens Axboeff821d22017-11-10 22:05:12 -07001472 * request head insertion from the workqueue.
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001473 */
Christoph Hellwige8064022016-10-20 15:12:13 +02001474 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001475
1476 spin_lock_irqsave(&q->requeue_lock, flags);
Christoph Hellwigb12e5c62023-04-13 08:40:57 +02001477 if (insert_flags & BLK_MQ_INSERT_AT_HEAD) {
Christoph Hellwige8064022016-10-20 15:12:13 +02001478 rq->rq_flags |= RQF_SOFTBARRIER;
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001479 list_add(&rq->queuelist, &q->requeue_list);
1480 } else {
1481 list_add_tail(&rq->queuelist, &q->requeue_list);
1482 }
1483 spin_unlock_irqrestore(&q->requeue_lock, flags);
1484}
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001485
1486void blk_mq_kick_requeue_list(struct request_queue *q)
1487{
Bart Van Asscheae943d22018-01-19 08:58:55 -08001488 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001489}
1490EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1491
Mike Snitzer28494502016-09-14 13:28:30 -04001492void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1493 unsigned long msecs)
1494{
Bart Van Assched4acf362017-08-09 11:28:06 -07001495 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1496 msecs_to_jiffies(msecs));
Mike Snitzer28494502016-09-14 13:28:30 -04001497}
1498EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1499
John Garry2dd65322022-07-06 20:03:53 +08001500static bool blk_mq_rq_inflight(struct request *rq, void *priv)
Jens Axboeae879912018-11-08 09:03:51 -07001501{
1502 /*
John Garry8ab30a32021-12-06 20:49:48 +08001503 * If we find a request that isn't idle we know the queue is busy
1504 * as it's checked in the iter.
1505 * Return false to stop the iteration.
Jens Axboeae879912018-11-08 09:03:51 -07001506 */
John Garry8ab30a32021-12-06 20:49:48 +08001507 if (blk_mq_request_started(rq)) {
Jens Axboeae879912018-11-08 09:03:51 -07001508 bool *busy = priv;
1509
1510 *busy = true;
1511 return false;
1512 }
1513
1514 return true;
1515}
1516
Jens Axboe3c94d832018-12-17 21:11:17 -07001517bool blk_mq_queue_inflight(struct request_queue *q)
Jens Axboeae879912018-11-08 09:03:51 -07001518{
1519 bool busy = false;
1520
Jens Axboe3c94d832018-12-17 21:11:17 -07001521 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
Jens Axboeae879912018-11-08 09:03:51 -07001522 return busy;
1523}
Jens Axboe3c94d832018-12-17 21:11:17 -07001524EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
Jens Axboeae879912018-11-08 09:03:51 -07001525
John Garry9bdb4832022-07-06 20:03:51 +08001526static void blk_mq_rq_timed_out(struct request *req)
Jens Axboe320ae512013-10-24 09:20:05 +01001527{
Christoph Hellwigda661262018-06-14 13:58:45 +02001528 req->rq_flags |= RQF_TIMED_OUT;
Christoph Hellwigd1210d52018-05-29 15:52:39 +02001529 if (req->q->mq_ops->timeout) {
1530 enum blk_eh_timer_return ret;
Jens Axboe87ee7b12014-04-24 08:51:47 -06001531
John Garry9bdb4832022-07-06 20:03:51 +08001532 ret = req->q->mq_ops->timeout(req);
Christoph Hellwigd1210d52018-05-29 15:52:39 +02001533 if (ret == BLK_EH_DONE)
1534 return;
1535 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
Christoph Hellwig46f92d42014-09-13 16:40:12 -07001536 }
Christoph Hellwigd1210d52018-05-29 15:52:39 +02001537
1538 blk_add_timer(req);
Jens Axboe87ee7b12014-04-24 08:51:47 -06001539}
Keith Busch5b3f25f2015-01-07 18:55:46 -07001540
David Jeffery82c22942022-10-26 13:19:57 +08001541struct blk_expired_data {
1542 bool has_timedout_rq;
1543 unsigned long next;
1544 unsigned long timeout_start;
1545};
1546
1547static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
Keith Busch12f5b932018-05-29 15:52:28 +02001548{
1549 unsigned long deadline;
1550
1551 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1552 return false;
Christoph Hellwigda661262018-06-14 13:58:45 +02001553 if (rq->rq_flags & RQF_TIMED_OUT)
1554 return false;
Keith Busch12f5b932018-05-29 15:52:28 +02001555
Christoph Hellwig079076b2018-11-14 17:02:05 +01001556 deadline = READ_ONCE(rq->deadline);
David Jeffery82c22942022-10-26 13:19:57 +08001557 if (time_after_eq(expired->timeout_start, deadline))
Keith Busch12f5b932018-05-29 15:52:28 +02001558 return true;
1559
David Jeffery82c22942022-10-26 13:19:57 +08001560 if (expired->next == 0)
1561 expired->next = deadline;
1562 else if (time_after(expired->next, deadline))
1563 expired->next = deadline;
Keith Busch12f5b932018-05-29 15:52:28 +02001564 return false;
1565}
1566
Ming Lei2e315dc2021-05-11 23:22:34 +08001567void blk_mq_put_rq_ref(struct request *rq)
1568{
Jens Axboede671d62022-09-21 15:19:54 -06001569 if (is_flush_rq(rq)) {
1570 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1571 blk_mq_free_request(rq);
1572 } else if (req_ref_put_and_test(rq)) {
Ming Lei2e315dc2021-05-11 23:22:34 +08001573 __blk_mq_free_request(rq);
Jens Axboede671d62022-09-21 15:19:54 -06001574 }
Ming Lei2e315dc2021-05-11 23:22:34 +08001575}
1576
John Garry2dd65322022-07-06 20:03:53 +08001577static bool blk_mq_check_expired(struct request *rq, void *priv)
Jens Axboe320ae512013-10-24 09:20:05 +01001578{
David Jeffery82c22942022-10-26 13:19:57 +08001579 struct blk_expired_data *expired = priv;
Christoph Hellwig81481eb2014-09-13 16:40:11 -07001580
Keith Busch12f5b932018-05-29 15:52:28 +02001581 /*
Ming Leic797b402021-08-11 23:52:02 +08001582 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1583 * be reallocated underneath the timeout handler's processing, then
1584 * the expire check is reliable. If the request is not expired, then
1585 * it was completed and reallocated as a new request after returning
1586 * from blk_mq_check_expired().
Keith Busch12f5b932018-05-29 15:52:28 +02001587 */
David Jeffery82c22942022-10-26 13:19:57 +08001588 if (blk_mq_req_expired(rq, expired)) {
1589 expired->has_timedout_rq = true;
1590 return false;
1591 }
1592 return true;
1593}
1594
1595static bool blk_mq_handle_expired(struct request *rq, void *priv)
1596{
1597 struct blk_expired_data *expired = priv;
1598
1599 if (blk_mq_req_expired(rq, expired))
John Garry9bdb4832022-07-06 20:03:51 +08001600 blk_mq_rq_timed_out(rq);
Jens Axboe7baa8572018-11-08 10:24:07 -07001601 return true;
Tejun Heo1d9bd512018-01-09 08:29:48 -08001602}
1603
Christoph Hellwig287922e2015-10-30 20:57:30 +08001604static void blk_mq_timeout_work(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +01001605{
Christoph Hellwig287922e2015-10-30 20:57:30 +08001606 struct request_queue *q =
1607 container_of(work, struct request_queue, timeout_work);
David Jeffery82c22942022-10-26 13:19:57 +08001608 struct blk_expired_data expired = {
1609 .timeout_start = jiffies,
1610 };
Tejun Heo1d9bd512018-01-09 08:29:48 -08001611 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08001612 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01001613
Gabriel Krisman Bertazi71f79fb2016-08-01 08:23:39 -06001614 /* A deadlock might occur if a request is stuck requiring a
1615 * timeout at the same time a queue freeze is waiting
1616 * completion, since the timeout code would not be able to
1617 * acquire the queue reference here.
1618 *
1619 * That's why we don't use blk_queue_enter here; instead, we use
1620 * percpu_ref_tryget directly, because we need to be able to
1621 * obtain a reference even in the short window between the queue
1622 * starting to freeze, by dropping the first reference in
Ming Lei1671d522017-03-27 20:06:57 +08001623 * blk_freeze_queue_start, and the moment the last request is
Gabriel Krisman Bertazi71f79fb2016-08-01 08:23:39 -06001624 * consumed, marked by the instant q_usage_counter reaches
1625 * zero.
1626 */
1627 if (!percpu_ref_tryget(&q->q_usage_counter))
Christoph Hellwig287922e2015-10-30 20:57:30 +08001628 return;
1629
David Jeffery82c22942022-10-26 13:19:57 +08001630 /* check if there is any timed-out request */
1631 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1632 if (expired.has_timedout_rq) {
1633 /*
1634 * Before walking tags, we must ensure any submit started
1635 * before the current time has finished. Since the submit
1636 * uses srcu or rcu, wait for a synchronization point to
1637 * ensure all running submits have finished
1638 */
Christoph Hellwig483239c2022-11-01 16:00:48 +01001639 blk_mq_wait_quiesce_done(q->tag_set);
Jens Axboe320ae512013-10-24 09:20:05 +01001640
David Jeffery82c22942022-10-26 13:19:57 +08001641 expired.next = 0;
1642 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1643 }
1644
1645 if (expired.next != 0) {
1646 mod_timer(&q->timeout, expired.next);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001647 } else {
Bart Van Asschefcd36c32018-01-10 08:33:33 -08001648 /*
1649 * Request timeouts are handled as a forward rolling timer. If
1650 * we end up here it means that no requests are pending and
1651 * also that no request has been pending for a while. Mark
1652 * each hctx as idle.
1653 */
Ming Leif054b562015-04-21 10:00:19 +08001654 queue_for_each_hw_ctx(q, hctx, i) {
1655 /* the hctx may be unmapped, so check it here */
1656 if (blk_mq_hw_queue_mapped(hctx))
1657 blk_mq_tag_idle(hctx);
1658 }
Jens Axboe0d2602c2014-05-13 15:10:52 -06001659 }
Christoph Hellwig287922e2015-10-30 20:57:30 +08001660 blk_queue_exit(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001661}
1662
Omar Sandoval88459642016-09-17 08:38:44 -06001663struct flush_busy_ctx_data {
1664 struct blk_mq_hw_ctx *hctx;
1665 struct list_head *list;
1666};
1667
1668static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1669{
1670 struct flush_busy_ctx_data *flush_data = data;
1671 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1672 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
Ming Leic16d6b52018-12-17 08:44:05 -07001673 enum hctx_type type = hctx->type;
Omar Sandoval88459642016-09-17 08:38:44 -06001674
Omar Sandoval88459642016-09-17 08:38:44 -06001675 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07001676 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
Omar Sandovale9a99a62018-02-27 16:56:42 -08001677 sbitmap_clear_bit(sb, bitnr);
Omar Sandoval88459642016-09-17 08:38:44 -06001678 spin_unlock(&ctx->lock);
1679 return true;
1680}
1681
Jens Axboe320ae512013-10-24 09:20:05 +01001682/*
Jens Axboe1429d7c2014-05-19 09:23:55 -06001683 * Process software queues that have been marked busy, splicing them
1684 * to the for-dispatch
1685 */
Jens Axboe2c3ad662016-12-14 14:34:47 -07001686void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
Jens Axboe1429d7c2014-05-19 09:23:55 -06001687{
Omar Sandoval88459642016-09-17 08:38:44 -06001688 struct flush_busy_ctx_data data = {
1689 .hctx = hctx,
1690 .list = list,
1691 };
Jens Axboe1429d7c2014-05-19 09:23:55 -06001692
Omar Sandoval88459642016-09-17 08:38:44 -06001693 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
Jens Axboe1429d7c2014-05-19 09:23:55 -06001694}
Jens Axboe2c3ad662016-12-14 14:34:47 -07001695EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
Jens Axboe1429d7c2014-05-19 09:23:55 -06001696
Ming Leib3476892017-10-14 17:22:30 +08001697struct dispatch_rq_data {
1698 struct blk_mq_hw_ctx *hctx;
1699 struct request *rq;
1700};
1701
1702static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1703 void *data)
1704{
1705 struct dispatch_rq_data *dispatch_data = data;
1706 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1707 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
Ming Leic16d6b52018-12-17 08:44:05 -07001708 enum hctx_type type = hctx->type;
Ming Leib3476892017-10-14 17:22:30 +08001709
1710 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07001711 if (!list_empty(&ctx->rq_lists[type])) {
1712 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
Ming Leib3476892017-10-14 17:22:30 +08001713 list_del_init(&dispatch_data->rq->queuelist);
Ming Leic16d6b52018-12-17 08:44:05 -07001714 if (list_empty(&ctx->rq_lists[type]))
Ming Leib3476892017-10-14 17:22:30 +08001715 sbitmap_clear_bit(sb, bitnr);
1716 }
1717 spin_unlock(&ctx->lock);
1718
1719 return !dispatch_data->rq;
1720}
1721
1722struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1723 struct blk_mq_ctx *start)
1724{
Jens Axboef31967f2018-10-29 13:13:29 -06001725 unsigned off = start ? start->index_hw[hctx->type] : 0;
Ming Leib3476892017-10-14 17:22:30 +08001726 struct dispatch_rq_data data = {
1727 .hctx = hctx,
1728 .rq = NULL,
1729 };
1730
1731 __sbitmap_for_each_set(&hctx->ctx_map, off,
1732 dispatch_rq_from_ctx, &data);
1733
1734 return data.rq;
1735}
1736
Jens Axboea808a9d2021-10-13 08:28:14 -06001737static bool __blk_mq_alloc_driver_tag(struct request *rq)
Jens Axboe703fd1c2016-09-16 13:59:14 -06001738{
John Garryae0f1a72021-10-05 18:23:38 +08001739 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
Ming Lei570e9b72020-06-30 22:03:55 +08001740 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
Ming Lei570e9b72020-06-30 22:03:55 +08001741 int tag;
1742
Ming Lei568f2702020-07-06 22:41:11 +08001743 blk_mq_tag_busy(rq->mq_hctx);
1744
Ming Lei570e9b72020-06-30 22:03:55 +08001745 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
John Garryae0f1a72021-10-05 18:23:38 +08001746 bt = &rq->mq_hctx->tags->breserved_tags;
Ming Lei570e9b72020-06-30 22:03:55 +08001747 tag_offset = 0;
Ming Lei28500852020-09-11 18:41:14 +08001748 } else {
1749 if (!hctx_may_queue(rq->mq_hctx, bt))
1750 return false;
Ming Lei570e9b72020-06-30 22:03:55 +08001751 }
1752
Ming Lei570e9b72020-06-30 22:03:55 +08001753 tag = __sbitmap_queue_get(bt);
1754 if (tag == BLK_MQ_NO_TAG)
1755 return false;
1756
1757 rq->tag = tag + tag_offset;
Ming Lei570e9b72020-06-30 22:03:55 +08001758 return true;
1759}
1760
Jens Axboea808a9d2021-10-13 08:28:14 -06001761bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
Ming Lei570e9b72020-06-30 22:03:55 +08001762{
Jens Axboea808a9d2021-10-13 08:28:14 -06001763 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
Ming Lei568f2702020-07-06 22:41:11 +08001764 return false;
1765
Ming Lei51db1c32020-08-19 23:20:19 +08001766 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
Ming Lei568f2702020-07-06 22:41:11 +08001767 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1768 rq->rq_flags |= RQF_MQ_INFLIGHT;
John Garrybccf5e22020-08-19 23:20:26 +08001769 __blk_mq_inc_active_requests(hctx);
Ming Lei568f2702020-07-06 22:41:11 +08001770 }
1771 hctx->tags->rqs[rq->tag] = rq;
1772 return true;
Ming Lei570e9b72020-06-30 22:03:55 +08001773}
1774
Jens Axboeeb619fd2017-11-09 08:32:43 -07001775static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1776 int flags, void *key)
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001777{
1778 struct blk_mq_hw_ctx *hctx;
1779
1780 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1781
Ming Lei5815839b2018-06-25 19:31:47 +08001782 spin_lock(&hctx->dispatch_wait_lock);
Jens Axboee8618572019-03-25 12:34:10 -06001783 if (!list_empty(&wait->entry)) {
1784 struct sbitmap_queue *sbq;
1785
1786 list_del_init(&wait->entry);
John Garryae0f1a72021-10-05 18:23:38 +08001787 sbq = &hctx->tags->bitmap_tags;
Jens Axboee8618572019-03-25 12:34:10 -06001788 atomic_dec(&sbq->ws_active);
1789 }
Ming Lei5815839b2018-06-25 19:31:47 +08001790 spin_unlock(&hctx->dispatch_wait_lock);
1791
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001792 blk_mq_run_hw_queue(hctx, true);
1793 return 1;
1794}
1795
Jens Axboef906a6a2017-11-09 16:10:13 -07001796/*
1797 * Mark us waiting for a tag. For shared tags, this involves hooking us into
Bart Van Asscheee3e4de2018-01-09 10:09:15 -08001798 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1799 * restart. For both cases, take care to check the condition again after
Jens Axboef906a6a2017-11-09 16:10:13 -07001800 * marking us as waiting.
1801 */
Ming Lei2278d692018-06-25 19:31:46 +08001802static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
Jens Axboef906a6a2017-11-09 16:10:13 -07001803 struct request *rq)
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001804{
Kemeng Shi98b99e92023-01-18 17:37:15 +08001805 struct sbitmap_queue *sbq;
Ming Lei5815839b2018-06-25 19:31:47 +08001806 struct wait_queue_head *wq;
Jens Axboef906a6a2017-11-09 16:10:13 -07001807 wait_queue_entry_t *wait;
1808 bool ret;
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001809
Kemeng Shi47df9ce2023-01-18 17:37:16 +08001810 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1811 !(blk_mq_is_shared_tags(hctx->flags))) {
Yufen Yu684b7322019-03-15 11:05:10 +08001812 blk_mq_sched_mark_restart_hctx(hctx);
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001813
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001814 /*
1815 * It's possible that a tag was freed in the window between the
1816 * allocation failure and adding the hardware queue to the wait
1817 * queue.
1818 *
1819 * Don't clear RESTART here, someone else could have set it.
1820 * At most this will cost an extra queue run.
1821 */
Ming Lei8ab6bb9e2018-06-25 19:31:45 +08001822 return blk_mq_get_driver_tag(rq);
Jens Axboeeb619fd2017-11-09 08:32:43 -07001823 }
1824
Ming Lei2278d692018-06-25 19:31:46 +08001825 wait = &hctx->dispatch_wait;
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001826 if (!list_empty_careful(&wait->entry))
1827 return false;
1828
Kemeng Shi98b99e92023-01-18 17:37:15 +08001829 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1830 sbq = &hctx->tags->breserved_tags;
1831 else
1832 sbq = &hctx->tags->bitmap_tags;
Jens Axboee8618572019-03-25 12:34:10 -06001833 wq = &bt_wait_ptr(sbq, hctx)->wait;
Ming Lei5815839b2018-06-25 19:31:47 +08001834
1835 spin_lock_irq(&wq->lock);
1836 spin_lock(&hctx->dispatch_wait_lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001837 if (!list_empty(&wait->entry)) {
Ming Lei5815839b2018-06-25 19:31:47 +08001838 spin_unlock(&hctx->dispatch_wait_lock);
1839 spin_unlock_irq(&wq->lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001840 return false;
1841 }
1842
Jens Axboee8618572019-03-25 12:34:10 -06001843 atomic_inc(&sbq->ws_active);
Ming Lei5815839b2018-06-25 19:31:47 +08001844 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1845 __add_wait_queue(wq, wait);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001846
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001847 /*
Jens Axboeeb619fd2017-11-09 08:32:43 -07001848 * It's possible that a tag was freed in the window between the
1849 * allocation failure and adding the hardware queue to the wait
1850 * queue.
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001851 */
Ming Lei8ab6bb9e2018-06-25 19:31:45 +08001852 ret = blk_mq_get_driver_tag(rq);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001853 if (!ret) {
Ming Lei5815839b2018-06-25 19:31:47 +08001854 spin_unlock(&hctx->dispatch_wait_lock);
1855 spin_unlock_irq(&wq->lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001856 return false;
Jens Axboef906a6a2017-11-09 16:10:13 -07001857 }
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001858
1859 /*
1860 * We got a tag, remove ourselves from the wait queue to ensure
1861 * someone else gets the wakeup.
1862 */
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001863 list_del_init(&wait->entry);
Jens Axboee8618572019-03-25 12:34:10 -06001864 atomic_dec(&sbq->ws_active);
Ming Lei5815839b2018-06-25 19:31:47 +08001865 spin_unlock(&hctx->dispatch_wait_lock);
1866 spin_unlock_irq(&wq->lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001867
1868 return true;
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001869}
1870
Ming Lei6e7687172018-07-03 09:03:16 -06001871#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1872#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1873/*
1874 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1875 * - EWMA is one simple way to compute running average value
1876 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1877 * - take 4 as factor for avoiding to get too small(0) result, and this
1878 * factor doesn't matter because EWMA decreases exponentially
1879 */
1880static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1881{
1882 unsigned int ewma;
1883
Ming Lei6e7687172018-07-03 09:03:16 -06001884 ewma = hctx->dispatch_busy;
1885
1886 if (!ewma && !busy)
1887 return;
1888
1889 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1890 if (busy)
1891 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1892 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1893
1894 hctx->dispatch_busy = ewma;
1895}
1896
Ming Lei86ff7c22018-01-30 22:04:57 -05001897#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1898
Johannes Thumshirnc92a4102020-03-25 00:24:44 +09001899static void blk_mq_handle_dev_resource(struct request *rq,
1900 struct list_head *list)
1901{
Johannes Thumshirnc92a4102020-03-25 00:24:44 +09001902 list_add(&rq->queuelist, list);
1903 __blk_mq_requeue_request(rq);
1904}
1905
Keith Busch0512a752020-05-12 17:55:47 +09001906static void blk_mq_handle_zone_resource(struct request *rq,
1907 struct list_head *zone_list)
1908{
1909 /*
1910 * If we end up here it is because we cannot dispatch a request to a
1911 * specific zone due to LLD level zone-write locking or other zone
1912 * related resource not being available. In this case, set the request
1913 * aside in zone_list for retrying it later.
1914 */
1915 list_add(&rq->queuelist, zone_list);
1916 __blk_mq_requeue_request(rq);
1917}
1918
Ming Lei75383522020-06-30 18:24:58 +08001919enum prep_dispatch {
1920 PREP_DISPATCH_OK,
1921 PREP_DISPATCH_NO_TAG,
1922 PREP_DISPATCH_NO_BUDGET,
1923};
1924
1925static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1926 bool need_budget)
1927{
1928 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Ming Lei2a5a24a2021-01-22 10:33:12 +08001929 int budget_token = -1;
Ming Lei75383522020-06-30 18:24:58 +08001930
Ming Lei2a5a24a2021-01-22 10:33:12 +08001931 if (need_budget) {
1932 budget_token = blk_mq_get_dispatch_budget(rq->q);
1933 if (budget_token < 0) {
1934 blk_mq_put_driver_tag(rq);
1935 return PREP_DISPATCH_NO_BUDGET;
1936 }
1937 blk_mq_set_rq_budget_token(rq, budget_token);
Ming Lei75383522020-06-30 18:24:58 +08001938 }
1939
1940 if (!blk_mq_get_driver_tag(rq)) {
1941 /*
1942 * The initial allocation attempt failed, so we need to
1943 * rerun the hardware queue when a tag is freed. The
1944 * waitqueue takes care of that. If the queue is run
1945 * before we add this entry back on the dispatch list,
1946 * we'll re-run it below.
1947 */
1948 if (!blk_mq_mark_tag_wait(hctx, rq)) {
Ming Lei1fd40b52020-06-30 18:25:00 +08001949 /*
1950 * All budgets not got from this function will be put
1951 * together during handling partial dispatch
1952 */
1953 if (need_budget)
Ming Lei2a5a24a2021-01-22 10:33:12 +08001954 blk_mq_put_dispatch_budget(rq->q, budget_token);
Ming Lei75383522020-06-30 18:24:58 +08001955 return PREP_DISPATCH_NO_TAG;
1956 }
1957 }
1958
1959 return PREP_DISPATCH_OK;
1960}
1961
Ming Lei1fd40b52020-06-30 18:25:00 +08001962/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1963static void blk_mq_release_budgets(struct request_queue *q,
Ming Lei2a5a24a2021-01-22 10:33:12 +08001964 struct list_head *list)
Ming Lei1fd40b52020-06-30 18:25:00 +08001965{
Ming Lei2a5a24a2021-01-22 10:33:12 +08001966 struct request *rq;
Ming Lei1fd40b52020-06-30 18:25:00 +08001967
Ming Lei2a5a24a2021-01-22 10:33:12 +08001968 list_for_each_entry(rq, list, queuelist) {
1969 int budget_token = blk_mq_get_rq_budget_token(rq);
1970
1971 if (budget_token >= 0)
1972 blk_mq_put_dispatch_budget(q, budget_token);
1973 }
Ming Lei1fd40b52020-06-30 18:25:00 +08001974}
1975
Jens Axboe1f57f8d2018-06-28 11:54:01 -06001976/*
Kemeng Shi34c9f5472023-01-18 17:37:19 +08001977 * blk_mq_commit_rqs will notify driver using bd->last that there is no
1978 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1979 * details)
1980 * Attention, we should explicitly call this in unusual cases:
1981 * 1) did not queue everything initially scheduled to queue
1982 * 2) the last attempt to queue a request failed
1983 */
1984static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
1985 bool from_schedule)
1986{
1987 if (hctx->queue->mq_ops->commit_rqs && queued) {
1988 trace_block_unplug(hctx->queue, queued, !from_schedule);
1989 hctx->queue->mq_ops->commit_rqs(hctx);
1990 }
1991}
1992
1993/*
Jens Axboe1f57f8d2018-06-28 11:54:01 -06001994 * Returns true if we did some work AND can potentially do more.
1995 */
Ming Lei445874e2020-06-30 18:24:57 +08001996bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
Ming Lei1fd40b52020-06-30 18:25:00 +08001997 unsigned int nr_budgets)
Jens Axboef04c3df2016-12-07 08:41:17 -07001998{
Ming Lei75383522020-06-30 18:24:58 +08001999 enum prep_dispatch prep;
Ming Lei445874e2020-06-30 18:24:57 +08002000 struct request_queue *q = hctx->queue;
Kemeng Shif1ce99f2023-01-18 17:37:24 +08002001 struct request *rq;
Kemeng Shi4ea58fe2023-01-18 17:37:23 +08002002 int queued;
Ming Lei86ff7c22018-01-30 22:04:57 -05002003 blk_status_t ret = BLK_STS_OK;
Keith Busch0512a752020-05-12 17:55:47 +09002004 LIST_HEAD(zone_list);
Naohiro Aota9586e672021-10-27 01:51:27 +09002005 bool needs_resource = false;
Jens Axboef04c3df2016-12-07 08:41:17 -07002006
Omar Sandoval81380ca2017-04-07 08:56:26 -06002007 if (list_empty(list))
2008 return false;
2009
Jens Axboef04c3df2016-12-07 08:41:17 -07002010 /*
Jens Axboef04c3df2016-12-07 08:41:17 -07002011 * Now process all the entries, sending them to the driver.
2012 */
Kemeng Shi4ea58fe2023-01-18 17:37:23 +08002013 queued = 0;
Omar Sandoval81380ca2017-04-07 08:56:26 -06002014 do {
Jens Axboef04c3df2016-12-07 08:41:17 -07002015 struct blk_mq_queue_data bd;
2016
2017 rq = list_first_entry(list, struct request, queuelist);
Ming Lei0bca7992018-04-05 00:35:21 +08002018
Ming Lei445874e2020-06-30 18:24:57 +08002019 WARN_ON_ONCE(hctx != rq->mq_hctx);
Ming Lei1fd40b52020-06-30 18:25:00 +08002020 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
Ming Lei75383522020-06-30 18:24:58 +08002021 if (prep != PREP_DISPATCH_OK)
Ming Lei0bca7992018-04-05 00:35:21 +08002022 break;
Ming Leide148292017-10-14 17:22:29 +08002023
Jens Axboef04c3df2016-12-07 08:41:17 -07002024 list_del_init(&rq->queuelist);
2025
2026 bd.rq = rq;
Kemeng Shif1ce99f2023-01-18 17:37:24 +08002027 bd.last = list_empty(list);
Jens Axboef04c3df2016-12-07 08:41:17 -07002028
Ming Lei1fd40b52020-06-30 18:25:00 +08002029 /*
2030 * once the request is queued to lld, no need to cover the
2031 * budget any more
2032 */
2033 if (nr_budgets)
2034 nr_budgets--;
Jens Axboef04c3df2016-12-07 08:41:17 -07002035 ret = q->mq_ops->queue_rq(hctx, &bd);
Ming Lei7bf13722020-07-01 21:58:57 +08002036 switch (ret) {
2037 case BLK_STS_OK:
2038 queued++;
Jens Axboef04c3df2016-12-07 08:41:17 -07002039 break;
Ming Lei7bf13722020-07-01 21:58:57 +08002040 case BLK_STS_RESOURCE:
Naohiro Aota9586e672021-10-27 01:51:27 +09002041 needs_resource = true;
2042 fallthrough;
Ming Lei7bf13722020-07-01 21:58:57 +08002043 case BLK_STS_DEV_RESOURCE:
2044 blk_mq_handle_dev_resource(rq, list);
2045 goto out;
2046 case BLK_STS_ZONE_RESOURCE:
Keith Busch0512a752020-05-12 17:55:47 +09002047 /*
2048 * Move the request to zone_list and keep going through
2049 * the dispatch list to find more requests the drive can
2050 * accept.
2051 */
2052 blk_mq_handle_zone_resource(rq, &zone_list);
Naohiro Aota9586e672021-10-27 01:51:27 +09002053 needs_resource = true;
Ming Lei7bf13722020-07-01 21:58:57 +08002054 break;
2055 default:
Hannes Reineckee21ee5a2020-09-30 10:02:53 +02002056 blk_mq_end_request(rq, ret);
Christoph Hellwigfc17b652017-06-03 09:38:05 +02002057 }
Omar Sandoval81380ca2017-04-07 08:56:26 -06002058 } while (!list_empty(list));
Ming Lei7bf13722020-07-01 21:58:57 +08002059out:
Keith Busch0512a752020-05-12 17:55:47 +09002060 if (!list_empty(&zone_list))
2061 list_splice_tail_init(&zone_list, list);
2062
yangerkun632bfb62020-09-05 19:25:56 +08002063 /* If we didn't flush the entire list, we could have told the driver
2064 * there was more coming, but that turned out to be a lie.
2065 */
Kemeng Shie4ef2e02023-01-18 17:37:22 +08002066 if (!list_empty(list) || ret != BLK_STS_OK)
2067 blk_mq_commit_rqs(hctx, queued, false);
2068
Jens Axboef04c3df2016-12-07 08:41:17 -07002069 /*
2070 * Any items that need requeuing? Stuff them into hctx->dispatch,
2071 * that is where we will continue on next queue run.
2072 */
2073 if (!list_empty(list)) {
Ming Lei86ff7c22018-01-30 22:04:57 -05002074 bool needs_restart;
Ming Lei75383522020-06-30 18:24:58 +08002075 /* For non-shared tags, the RESTART check will suffice */
2076 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
Kemeng Shi47df9ce2023-01-18 17:37:16 +08002077 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2078 blk_mq_is_shared_tags(hctx->flags));
Ming Lei86ff7c22018-01-30 22:04:57 -05002079
Ming Lei2a5a24a2021-01-22 10:33:12 +08002080 if (nr_budgets)
2081 blk_mq_release_budgets(q, list);
Jens Axboef04c3df2016-12-07 08:41:17 -07002082
2083 spin_lock(&hctx->lock);
Ming Lei01e99ae2020-02-25 09:04:32 +08002084 list_splice_tail_init(list, &hctx->dispatch);
Jens Axboef04c3df2016-12-07 08:41:17 -07002085 spin_unlock(&hctx->lock);
2086
2087 /*
Ming Leid7d85352020-08-17 18:01:15 +08002088 * Order adding requests to hctx->dispatch and checking
2089 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2090 * in blk_mq_sched_restart(). Avoid restart code path to
2091 * miss the new added requests to hctx->dispatch, meantime
2092 * SCHED_RESTART is observed here.
2093 */
2094 smp_mb();
2095
2096 /*
Bart Van Assche710c7852017-04-07 11:16:51 -07002097 * If SCHED_RESTART was set by the caller of this function and
2098 * it is no longer set that means that it was cleared by another
2099 * thread and hence that a queue rerun is needed.
Jens Axboef04c3df2016-12-07 08:41:17 -07002100 *
Jens Axboeeb619fd2017-11-09 08:32:43 -07002101 * If 'no_tag' is set, that means that we failed getting
2102 * a driver tag with an I/O scheduler attached. If our dispatch
2103 * waitqueue is no longer active, ensure that we run the queue
2104 * AFTER adding our entries back to the list.
Jens Axboebd166ef2017-01-17 06:03:22 -07002105 *
Bart Van Assche710c7852017-04-07 11:16:51 -07002106 * If no I/O scheduler has been configured it is possible that
2107 * the hardware queue got stopped and restarted before requests
2108 * were pushed back onto the dispatch list. Rerun the queue to
2109 * avoid starvation. Notes:
2110 * - blk_mq_run_hw_queue() checks whether or not a queue has
2111 * been stopped before rerunning a queue.
2112 * - Some but not all block drivers stop a queue before
Christoph Hellwigfc17b652017-06-03 09:38:05 +02002113 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
Bart Van Assche710c7852017-04-07 11:16:51 -07002114 * and dm-rq.
Ming Lei86ff7c22018-01-30 22:04:57 -05002115 *
2116 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2117 * bit is set, run queue after a delay to avoid IO stalls
Douglas Andersonab3cee32020-04-20 09:24:51 -07002118 * that could otherwise occur if the queue is idle. We'll do
Naohiro Aota9586e672021-10-27 01:51:27 +09002119 * similar if we couldn't get budget or couldn't lock a zone
2120 * and SCHED_RESTART is set.
Jens Axboebd166ef2017-01-17 06:03:22 -07002121 */
Ming Lei86ff7c22018-01-30 22:04:57 -05002122 needs_restart = blk_mq_sched_needs_restart(hctx);
Naohiro Aota9586e672021-10-27 01:51:27 +09002123 if (prep == PREP_DISPATCH_NO_BUDGET)
2124 needs_resource = true;
Ming Lei86ff7c22018-01-30 22:04:57 -05002125 if (!needs_restart ||
Jens Axboeeb619fd2017-11-09 08:32:43 -07002126 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
Jens Axboebd166ef2017-01-17 06:03:22 -07002127 blk_mq_run_hw_queue(hctx, true);
Miaohe Lin6d5e8d22022-09-05 18:19:50 +08002128 else if (needs_resource)
Ming Lei86ff7c22018-01-30 22:04:57 -05002129 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
Jens Axboe1f57f8d2018-06-28 11:54:01 -06002130
Ming Lei6e7687172018-07-03 09:03:16 -06002131 blk_mq_update_dispatch_busy(hctx, true);
Jens Axboe1f57f8d2018-06-28 11:54:01 -06002132 return false;
Kemeng Shi4ea58fe2023-01-18 17:37:23 +08002133 }
Jens Axboef04c3df2016-12-07 08:41:17 -07002134
Kemeng Shi4ea58fe2023-01-18 17:37:23 +08002135 blk_mq_update_dispatch_busy(hctx, false);
2136 return true;
Jens Axboef04c3df2016-12-07 08:41:17 -07002137}
2138
Ming Leif82ddf12018-04-08 17:48:10 +08002139static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2140{
2141 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2142
2143 if (cpu >= nr_cpu_ids)
2144 cpu = cpumask_first(hctx->cpumask);
2145 return cpu;
2146}
2147
Jens Axboe506e9312014-05-07 10:26:44 -06002148/*
2149 * It'd be great if the workqueue API had a way to pass
2150 * in a mask and had some smarts for more clever placement.
2151 * For now we just round-robin here, switching for every
2152 * BLK_MQ_CPU_WORK_BATCH queued items.
2153 */
2154static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2155{
Ming Lei7bed4592018-01-18 00:41:51 +08002156 bool tried = false;
Ming Lei476f8c92018-04-08 17:48:09 +08002157 int next_cpu = hctx->next_cpu;
Ming Lei7bed4592018-01-18 00:41:51 +08002158
Christoph Hellwigb657d7e2014-11-24 09:27:23 +01002159 if (hctx->queue->nr_hw_queues == 1)
2160 return WORK_CPU_UNBOUND;
Jens Axboe506e9312014-05-07 10:26:44 -06002161
2162 if (--hctx->next_cpu_batch <= 0) {
Ming Lei7bed4592018-01-18 00:41:51 +08002163select_cpu:
Ming Lei476f8c92018-04-08 17:48:09 +08002164 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
Christoph Hellwig20e4d8132018-01-12 10:53:06 +08002165 cpu_online_mask);
Jens Axboe506e9312014-05-07 10:26:44 -06002166 if (next_cpu >= nr_cpu_ids)
Ming Leif82ddf12018-04-08 17:48:10 +08002167 next_cpu = blk_mq_first_mapped_cpu(hctx);
Jens Axboe506e9312014-05-07 10:26:44 -06002168 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2169 }
2170
Ming Lei7bed4592018-01-18 00:41:51 +08002171 /*
2172 * Do unbound schedule if we can't find a online CPU for this hctx,
2173 * and it should only happen in the path of handling CPU DEAD.
2174 */
Ming Lei476f8c92018-04-08 17:48:09 +08002175 if (!cpu_online(next_cpu)) {
Ming Lei7bed4592018-01-18 00:41:51 +08002176 if (!tried) {
2177 tried = true;
2178 goto select_cpu;
2179 }
2180
2181 /*
2182 * Make sure to re-select CPU next time once after CPUs
2183 * in hctx->cpumask become online again.
2184 */
Ming Lei476f8c92018-04-08 17:48:09 +08002185 hctx->next_cpu = next_cpu;
Ming Lei7bed4592018-01-18 00:41:51 +08002186 hctx->next_cpu_batch = 1;
2187 return WORK_CPU_UNBOUND;
2188 }
Ming Lei476f8c92018-04-08 17:48:09 +08002189
2190 hctx->next_cpu = next_cpu;
2191 return next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -06002192}
2193
André Almeida105663f2020-01-06 15:08:18 -03002194/**
André Almeida105663f2020-01-06 15:08:18 -03002195 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2196 * @hctx: Pointer to the hardware queue to run.
Minwoo Imfa94ba82020-12-05 00:20:55 +09002197 * @msecs: Milliseconds of delay to wait before running the queue.
André Almeida105663f2020-01-06 15:08:18 -03002198 *
2199 * Run a hardware queue asynchronously with a delay of @msecs.
2200 */
Bart Van Assche7587a5a2017-04-07 11:16:52 -07002201void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2202{
Christoph Hellwig1aa8d872023-04-13 08:06:50 +02002203 if (unlikely(blk_mq_hctx_stopped(hctx)))
2204 return;
2205 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2206 msecs_to_jiffies(msecs));
Bart Van Assche7587a5a2017-04-07 11:16:52 -07002207}
2208EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2209
André Almeida105663f2020-01-06 15:08:18 -03002210/**
2211 * blk_mq_run_hw_queue - Start to run a hardware queue.
2212 * @hctx: Pointer to the hardware queue to run.
2213 * @async: If we want to run the queue asynchronously.
2214 *
2215 * Check if the request queue is not in a quiesced state and if there are
2216 * pending requests to be sent. If this is true, run the queue to send requests
2217 * to hardware.
2218 */
John Garry626fb732019-10-30 00:59:30 +08002219void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
Bart Van Assche7587a5a2017-04-07 11:16:52 -07002220{
Ming Lei24f5a902018-01-06 16:27:38 +08002221 bool need_run;
2222
2223 /*
Christoph Hellwig4d5bba52023-04-13 08:06:51 +02002224 * We can't run the queue inline with interrupts disabled.
2225 */
2226 WARN_ON_ONCE(!async && in_interrupt());
2227
2228 /*
Ming Lei24f5a902018-01-06 16:27:38 +08002229 * When queue is quiesced, we may be switching io scheduler, or
2230 * updating nr_hw_queues, or other things, and we can't run queue
2231 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2232 *
2233 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2234 * quiesced.
2235 */
Ming Lei41adf532021-12-06 19:12:13 +08002236 __blk_mq_run_dispatch_ops(hctx->queue, false,
Ming Lei2a904d02021-12-03 21:15:31 +08002237 need_run = !blk_queue_quiesced(hctx->queue) &&
2238 blk_mq_hctx_has_pending(hctx));
Ming Lei24f5a902018-01-06 16:27:38 +08002239
Christoph Hellwig1aa8d872023-04-13 08:06:50 +02002240 if (!need_run)
2241 return;
2242
2243 if (async || (hctx->flags & BLK_MQ_F_BLOCKING) ||
2244 !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2245 blk_mq_delay_run_hw_queue(hctx, 0);
2246 return;
2247 }
2248
Christoph Hellwig4d5bba52023-04-13 08:06:51 +02002249 blk_mq_run_dispatch_ops(hctx->queue,
2250 blk_mq_sched_dispatch_requests(hctx));
Jens Axboe320ae512013-10-24 09:20:05 +01002251}
Omar Sandoval5b727272017-04-14 01:00:00 -07002252EXPORT_SYMBOL(blk_mq_run_hw_queue);
Jens Axboe320ae512013-10-24 09:20:05 +01002253
Jan Karab6e68ee2021-01-11 17:47:17 +01002254/*
Jan Karab6e68ee2021-01-11 17:47:17 +01002255 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2256 * scheduler.
2257 */
2258static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2259{
Ming Lei5d054262022-05-22 20:23:50 +08002260 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
Jan Karab6e68ee2021-01-11 17:47:17 +01002261 /*
2262 * If the IO scheduler does not respect hardware queues when
2263 * dispatching, we just don't bother with multiple HW queues and
2264 * dispatch from hctx for the current CPU since running multiple queues
2265 * just causes lock contention inside the scheduler and pointless cache
2266 * bouncing.
2267 */
Bart Van Assche51ab80f2022-06-15 15:55:49 -07002268 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
Ming Lei5d054262022-05-22 20:23:50 +08002269
Jan Karab6e68ee2021-01-11 17:47:17 +01002270 if (!blk_mq_hctx_stopped(hctx))
2271 return hctx;
2272 return NULL;
2273}
2274
André Almeida105663f2020-01-06 15:08:18 -03002275/**
Mauro Carvalho Chehab24f7bb82020-10-23 18:32:54 +02002276 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
André Almeida105663f2020-01-06 15:08:18 -03002277 * @q: Pointer to the request queue to run.
2278 * @async: If we want to run the queue asynchronously.
2279 */
Mike Snitzerb94ec292015-03-11 23:56:38 -04002280void blk_mq_run_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01002281{
Jan Karab6e68ee2021-01-11 17:47:17 +01002282 struct blk_mq_hw_ctx *hctx, *sq_hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002283 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01002284
Jan Karab6e68ee2021-01-11 17:47:17 +01002285 sq_hctx = NULL;
Ming Lei4d337ceb2022-06-16 09:44:00 +08002286 if (blk_queue_sq_sched(q))
Jan Karab6e68ee2021-01-11 17:47:17 +01002287 sq_hctx = blk_mq_get_sq_hctx(q);
Jens Axboe320ae512013-10-24 09:20:05 +01002288 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboe79f720a2017-11-10 09:13:21 -07002289 if (blk_mq_hctx_stopped(hctx))
Jens Axboe320ae512013-10-24 09:20:05 +01002290 continue;
Jan Karab6e68ee2021-01-11 17:47:17 +01002291 /*
2292 * Dispatch from this hctx either if there's no hctx preferred
2293 * by IO scheduler or if it has requests that bypass the
2294 * scheduler.
2295 */
2296 if (!sq_hctx || sq_hctx == hctx ||
2297 !list_empty_careful(&hctx->dispatch))
2298 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +01002299 }
2300}
Mike Snitzerb94ec292015-03-11 23:56:38 -04002301EXPORT_SYMBOL(blk_mq_run_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +01002302
Bart Van Asschefd001442016-10-28 17:19:37 -07002303/**
Douglas Andersonb9151e72020-04-20 09:24:52 -07002304 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2305 * @q: Pointer to the request queue to run.
Minwoo Imfa94ba82020-12-05 00:20:55 +09002306 * @msecs: Milliseconds of delay to wait before running the queues.
Douglas Andersonb9151e72020-04-20 09:24:52 -07002307 */
2308void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2309{
Jan Karab6e68ee2021-01-11 17:47:17 +01002310 struct blk_mq_hw_ctx *hctx, *sq_hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002311 unsigned long i;
Douglas Andersonb9151e72020-04-20 09:24:52 -07002312
Jan Karab6e68ee2021-01-11 17:47:17 +01002313 sq_hctx = NULL;
Ming Lei4d337ceb2022-06-16 09:44:00 +08002314 if (blk_queue_sq_sched(q))
Jan Karab6e68ee2021-01-11 17:47:17 +01002315 sq_hctx = blk_mq_get_sq_hctx(q);
Douglas Andersonb9151e72020-04-20 09:24:52 -07002316 queue_for_each_hw_ctx(q, hctx, i) {
2317 if (blk_mq_hctx_stopped(hctx))
2318 continue;
Jan Karab6e68ee2021-01-11 17:47:17 +01002319 /*
David Jeffery8f5fea62022-01-31 15:33:37 -05002320 * If there is already a run_work pending, leave the
2321 * pending delay untouched. Otherwise, a hctx can stall
2322 * if another hctx is re-delaying the other's work
2323 * before the work executes.
2324 */
2325 if (delayed_work_pending(&hctx->run_work))
2326 continue;
2327 /*
Jan Karab6e68ee2021-01-11 17:47:17 +01002328 * Dispatch from this hctx either if there's no hctx preferred
2329 * by IO scheduler or if it has requests that bypass the
2330 * scheduler.
2331 */
2332 if (!sq_hctx || sq_hctx == hctx ||
2333 !list_empty_careful(&hctx->dispatch))
2334 blk_mq_delay_run_hw_queue(hctx, msecs);
Douglas Andersonb9151e72020-04-20 09:24:52 -07002335 }
2336}
2337EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2338
Ming Lei39a70c72017-06-06 23:22:09 +08002339/*
2340 * This function is often used for pausing .queue_rq() by driver when
2341 * there isn't enough resource or some conditions aren't satisfied, and
Bart Van Assche4d606212017-08-17 16:23:00 -07002342 * BLK_STS_RESOURCE is usually returned.
Ming Lei39a70c72017-06-06 23:22:09 +08002343 *
2344 * We do not guarantee that dispatch can be drained or blocked
2345 * after blk_mq_stop_hw_queue() returns. Please use
2346 * blk_mq_quiesce_queue() for that requirement.
2347 */
Jens Axboe320ae512013-10-24 09:20:05 +01002348void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2349{
Ming Lei641a9ed2017-06-06 23:22:10 +08002350 cancel_delayed_work(&hctx->run_work);
2351
2352 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboe320ae512013-10-24 09:20:05 +01002353}
2354EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2355
Ming Lei39a70c72017-06-06 23:22:09 +08002356/*
2357 * This function is often used for pausing .queue_rq() by driver when
2358 * there isn't enough resource or some conditions aren't satisfied, and
Bart Van Assche4d606212017-08-17 16:23:00 -07002359 * BLK_STS_RESOURCE is usually returned.
Ming Lei39a70c72017-06-06 23:22:09 +08002360 *
2361 * We do not guarantee that dispatch can be drained or blocked
2362 * after blk_mq_stop_hw_queues() returns. Please use
2363 * blk_mq_quiesce_queue() for that requirement.
2364 */
Jens Axboe2719aa22017-05-03 11:08:14 -06002365void blk_mq_stop_hw_queues(struct request_queue *q)
2366{
Ming Lei641a9ed2017-06-06 23:22:10 +08002367 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002368 unsigned long i;
Ming Lei641a9ed2017-06-06 23:22:10 +08002369
2370 queue_for_each_hw_ctx(q, hctx, i)
2371 blk_mq_stop_hw_queue(hctx);
Christoph Hellwig280d45f2013-10-25 14:45:58 +01002372}
2373EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2374
Jens Axboe320ae512013-10-24 09:20:05 +01002375void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2376{
2377 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboee4043dc2014-04-09 10:18:23 -06002378
Jens Axboe0ffbce82014-06-25 08:22:34 -06002379 blk_mq_run_hw_queue(hctx, false);
Jens Axboe320ae512013-10-24 09:20:05 +01002380}
2381EXPORT_SYMBOL(blk_mq_start_hw_queue);
2382
Christoph Hellwig2f268552014-04-16 09:44:56 +02002383void blk_mq_start_hw_queues(struct request_queue *q)
2384{
2385 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002386 unsigned long i;
Christoph Hellwig2f268552014-04-16 09:44:56 +02002387
2388 queue_for_each_hw_ctx(q, hctx, i)
2389 blk_mq_start_hw_queue(hctx);
2390}
2391EXPORT_SYMBOL(blk_mq_start_hw_queues);
2392
Jens Axboeae911c52016-12-08 13:19:30 -07002393void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2394{
2395 if (!blk_mq_hctx_stopped(hctx))
2396 return;
2397
2398 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2399 blk_mq_run_hw_queue(hctx, async);
2400}
2401EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2402
Christoph Hellwig1b4a3252014-04-16 09:44:54 +02002403void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01002404{
2405 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002406 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01002407
Jens Axboeae911c52016-12-08 13:19:30 -07002408 queue_for_each_hw_ctx(q, hctx, i)
2409 blk_mq_start_stopped_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +01002410}
2411EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2412
Christoph Hellwig70f4db62014-04-16 10:48:08 -06002413static void blk_mq_run_work_fn(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +01002414{
Christoph Hellwigc20a1a22023-04-13 08:06:48 +02002415 struct blk_mq_hw_ctx *hctx =
2416 container_of(work, struct blk_mq_hw_ctx, run_work.work);
Jens Axboe320ae512013-10-24 09:20:05 +01002417
Christoph Hellwig4d5bba52023-04-13 08:06:51 +02002418 blk_mq_run_dispatch_ops(hctx->queue,
2419 blk_mq_sched_dispatch_requests(hctx));
Jens Axboe320ae512013-10-24 09:20:05 +01002420}
2421
André Almeida105663f2020-01-06 15:08:18 -03002422/**
2423 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2424 * @rq: Pointer to request to be inserted.
Christoph Hellwig2b597612023-04-13 08:40:55 +02002425 * @flags: BLK_MQ_INSERT_*
André Almeida105663f2020-01-06 15:08:18 -03002426 *
Jens Axboe157f3772017-09-11 16:43:57 -06002427 * Should only be used carefully, when the caller knows we want to
2428 * bypass a potential IO scheduler on the target device.
2429 */
Christoph Hellwig2b597612023-04-13 08:40:55 +02002430void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
Jens Axboe157f3772017-09-11 16:43:57 -06002431{
Jens Axboeea4f9952018-10-29 15:06:13 -06002432 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Jens Axboe157f3772017-09-11 16:43:57 -06002433
2434 spin_lock(&hctx->lock);
Christoph Hellwig2b597612023-04-13 08:40:55 +02002435 if (flags & BLK_MQ_INSERT_AT_HEAD)
Ming Lei01e99ae2020-02-25 09:04:32 +08002436 list_add(&rq->queuelist, &hctx->dispatch);
2437 else
2438 list_add_tail(&rq->queuelist, &hctx->dispatch);
Jens Axboe157f3772017-09-11 16:43:57 -06002439 spin_unlock(&hctx->lock);
Jens Axboe157f3772017-09-11 16:43:57 -06002440}
2441
Christoph Hellwig05a93112023-04-13 08:40:42 +02002442static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2443 struct blk_mq_ctx *ctx, struct list_head *list,
2444 bool run_queue_async)
Jens Axboe320ae512013-10-24 09:20:05 +01002445{
Ming Lei3f0cedc2018-07-02 17:35:58 +08002446 struct request *rq;
Ming Leic16d6b52018-12-17 08:44:05 -07002447 enum hctx_type type = hctx->type;
Ming Lei3f0cedc2018-07-02 17:35:58 +08002448
Jens Axboe320ae512013-10-24 09:20:05 +01002449 /*
Christoph Hellwig94aa2282023-04-13 08:40:41 +02002450 * Try to issue requests directly if the hw queue isn't busy to save an
2451 * extra enqueue & dequeue to the sw queue.
2452 */
2453 if (!hctx->dispatch_busy && !run_queue_async) {
2454 blk_mq_run_dispatch_ops(hctx->queue,
2455 blk_mq_try_issue_list_directly(hctx, list));
2456 if (list_empty(list))
2457 goto out;
2458 }
2459
2460 /*
Jens Axboe320ae512013-10-24 09:20:05 +01002461 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2462 * offline now
2463 */
Ming Lei3f0cedc2018-07-02 17:35:58 +08002464 list_for_each_entry(rq, list, queuelist) {
Jens Axboee57690f2016-08-24 15:34:35 -06002465 BUG_ON(rq->mq_ctx != ctx);
Christoph Hellwiga54895f2020-12-03 17:21:39 +01002466 trace_block_rq_insert(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01002467 }
Ming Lei3f0cedc2018-07-02 17:35:58 +08002468
2469 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07002470 list_splice_tail_init(list, &ctx->rq_lists[type]);
Ming Leicfd0c552015-10-20 23:13:57 +08002471 blk_mq_hctx_mark_pending(hctx, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01002472 spin_unlock(&ctx->lock);
Christoph Hellwig94aa2282023-04-13 08:40:41 +02002473out:
2474 blk_mq_run_hw_queue(hctx, run_queue_async);
Jens Axboe320ae512013-10-24 09:20:05 +01002475}
2476
Christoph Hellwig710fa372023-04-13 08:40:54 +02002477static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002478{
2479 struct request_queue *q = rq->q;
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002480 struct blk_mq_ctx *ctx = rq->mq_ctx;
2481 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2482
Christoph Hellwig53548d22023-04-13 08:40:47 +02002483 if (blk_rq_is_passthrough(rq)) {
2484 /*
2485 * Passthrough request have to be added to hctx->dispatch
2486 * directly. The device may be in a situation where it can't
2487 * handle FS request, and always returns BLK_STS_RESOURCE for
2488 * them, which gets them added to hctx->dispatch.
2489 *
2490 * If a passthrough request is required to unblock the queues,
2491 * and it is added to the scheduler queue, there is no chance to
2492 * dispatch it given we prioritize requests in hctx->dispatch.
2493 */
Christoph Hellwig2b597612023-04-13 08:40:55 +02002494 blk_mq_request_bypass_insert(rq, flags);
Christoph Hellwig53548d22023-04-13 08:40:47 +02002495 } else if (rq->rq_flags & RQF_FLUSH_SEQ) {
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002496 /*
2497 * Firstly normal IO request is inserted to scheduler queue or
2498 * sw queue, meantime we add flush request to dispatch queue(
2499 * hctx->dispatch) directly and there is at most one in-flight
2500 * flush request for each hw queue, so it doesn't matter to add
2501 * flush request to tail or front of the dispatch queue.
2502 *
2503 * Secondly in case of NCQ, flush request belongs to non-NCQ
2504 * command, and queueing it will fail when there is any
2505 * in-flight normal IO request(NCQ command). When adding flush
2506 * rq to the front of hctx->dispatch, it is easier to introduce
2507 * extra time to flush rq's latency because of S_SCHED_RESTART
2508 * compared with adding to the tail of dispatch queue, then
2509 * chance of flush merge is increased, and less flush requests
2510 * will be issued to controller. It is observed that ~10% time
2511 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2512 * drive when adding flush rq to the front of hctx->dispatch.
2513 *
2514 * Simply queue flush rq to the front of hctx->dispatch so that
2515 * intensive flush workloads can benefit in case of NCQ HW.
2516 */
Christoph Hellwig2b597612023-04-13 08:40:55 +02002517 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
Christoph Hellwig53548d22023-04-13 08:40:47 +02002518 } else if (q->elevator) {
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002519 LIST_HEAD(list);
2520
Christoph Hellwig53548d22023-04-13 08:40:47 +02002521 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2522
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002523 list_add(&rq->queuelist, &list);
Christoph Hellwig93fffe12023-04-13 08:40:56 +02002524 q->elevator->type->ops.insert_requests(hctx, &list, flags);
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002525 } else {
Christoph Hellwig4ec5c052023-04-13 08:40:45 +02002526 trace_block_rq_insert(rq);
2527
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002528 spin_lock(&ctx->lock);
Christoph Hellwig710fa372023-04-13 08:40:54 +02002529 if (flags & BLK_MQ_INSERT_AT_HEAD)
Christoph Hellwig4ec5c052023-04-13 08:40:45 +02002530 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2531 else
2532 list_add_tail(&rq->queuelist,
2533 &ctx->rq_lists[hctx->type]);
Christoph Hellwiga88db1e2023-04-13 08:40:44 +02002534 blk_mq_hctx_mark_pending(hctx, ctx);
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002535 spin_unlock(&ctx->lock);
2536 }
Jens Axboe320ae512013-10-24 09:20:05 +01002537}
2538
Christoph Hellwig14ccb662019-06-06 12:29:01 +02002539static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2540 unsigned int nr_segs)
Jens Axboe320ae512013-10-24 09:20:05 +01002541{
Eric Biggers93f221a2020-09-15 20:53:14 -07002542 int err;
2543
Christoph Hellwigf924cdd2019-06-06 12:29:00 +02002544 if (bio->bi_opf & REQ_RAHEAD)
2545 rq->cmd_flags |= REQ_FAILFAST_MASK;
2546
2547 rq->__sector = bio->bi_iter.bi_sector;
Christoph Hellwig14ccb662019-06-06 12:29:01 +02002548 blk_rq_bio_prep(rq, bio, nr_segs);
Eric Biggers93f221a2020-09-15 20:53:14 -07002549
2550 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2551 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2552 WARN_ON_ONCE(err);
Jens Axboe4b570522014-05-29 11:00:11 -06002553
Konstantin Khlebnikovb5af37a2020-05-27 07:24:16 +02002554 blk_account_io_start(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01002555}
2556
Mike Snitzer0f955492018-01-17 11:25:56 -05002557static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig3e087732021-10-12 13:12:24 +02002558 struct request *rq, bool last)
Shaohua Lif984df12015-05-08 10:51:32 -07002559{
Shaohua Lif984df12015-05-08 10:51:32 -07002560 struct request_queue *q = rq->q;
Shaohua Lif984df12015-05-08 10:51:32 -07002561 struct blk_mq_queue_data bd = {
2562 .rq = rq,
Jens Axboebe94f052018-11-24 10:15:46 -07002563 .last = last,
Shaohua Lif984df12015-05-08 10:51:32 -07002564 };
Jens Axboef06345a2017-06-12 11:22:46 -06002565 blk_status_t ret;
Mike Snitzer0f955492018-01-17 11:25:56 -05002566
Mike Snitzer0f955492018-01-17 11:25:56 -05002567 /*
2568 * For OK queue, we are done. For error, caller may kill it.
2569 * Any other error (busy), just add it to our list as we
2570 * previously would have done.
2571 */
2572 ret = q->mq_ops->queue_rq(hctx, &bd);
2573 switch (ret) {
2574 case BLK_STS_OK:
Ming Lei6ce3dd62018-07-10 09:03:31 +08002575 blk_mq_update_dispatch_busy(hctx, false);
Mike Snitzer0f955492018-01-17 11:25:56 -05002576 break;
2577 case BLK_STS_RESOURCE:
Ming Lei86ff7c22018-01-30 22:04:57 -05002578 case BLK_STS_DEV_RESOURCE:
Ming Lei6ce3dd62018-07-10 09:03:31 +08002579 blk_mq_update_dispatch_busy(hctx, true);
Mike Snitzer0f955492018-01-17 11:25:56 -05002580 __blk_mq_requeue_request(rq);
2581 break;
2582 default:
Ming Lei6ce3dd62018-07-10 09:03:31 +08002583 blk_mq_update_dispatch_busy(hctx, false);
Mike Snitzer0f955492018-01-17 11:25:56 -05002584 break;
2585 }
2586
2587 return ret;
2588}
2589
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002590static bool blk_mq_get_budget_and_tag(struct request *rq)
Mike Snitzer0f955492018-01-17 11:25:56 -05002591{
Ming Lei2a5a24a2021-01-22 10:33:12 +08002592 int budget_token;
Ming Leid964f042017-06-06 23:22:00 +08002593
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002594 budget_token = blk_mq_get_dispatch_budget(rq->q);
Ming Lei2a5a24a2021-01-22 10:33:12 +08002595 if (budget_token < 0)
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002596 return false;
Ming Lei2a5a24a2021-01-22 10:33:12 +08002597 blk_mq_set_rq_budget_token(rq, budget_token);
Ming Lei8ab6bb9e2018-06-25 19:31:45 +08002598 if (!blk_mq_get_driver_tag(rq)) {
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002599 blk_mq_put_dispatch_budget(rq->q, budget_token);
2600 return false;
Ming Lei88022d72017-11-05 02:21:12 +08002601 }
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002602 return true;
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002603}
2604
André Almeida105663f2020-01-06 15:08:18 -03002605/**
2606 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2607 * @hctx: Pointer of the associated hardware queue.
2608 * @rq: Pointer to request to be sent.
André Almeida105663f2020-01-06 15:08:18 -03002609 *
2610 * If the device has enough resources to accept a new request now, send the
2611 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2612 * we can try send it another time in the future. Requests inserted at this
2613 * queue have higher priority.
2614 */
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002615static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig3e087732021-10-12 13:12:24 +02002616 struct request *rq)
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002617{
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002618 blk_status_t ret;
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002619
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002620 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
Christoph Hellwig710fa372023-04-13 08:40:54 +02002621 blk_mq_insert_request(rq, 0);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002622 return;
2623 }
2624
2625 if ((rq->rq_flags & RQF_ELV) || !blk_mq_get_budget_and_tag(rq)) {
Christoph Hellwig710fa372023-04-13 08:40:54 +02002626 blk_mq_insert_request(rq, 0);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02002627 blk_mq_run_hw_queue(hctx, false);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002628 return;
2629 }
2630
2631 ret = __blk_mq_issue_directly(hctx, rq, true);
2632 switch (ret) {
2633 case BLK_STS_OK:
2634 break;
2635 case BLK_STS_RESOURCE:
2636 case BLK_STS_DEV_RESOURCE:
Christoph Hellwig2b597612023-04-13 08:40:55 +02002637 blk_mq_request_bypass_insert(rq, 0);
Christoph Hellwig23943952023-04-13 08:40:52 +02002638 blk_mq_run_hw_queue(hctx, false);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002639 break;
2640 default:
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002641 blk_mq_end_request(rq, ret);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002642 break;
2643 }
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002644}
2645
Christoph Hellwig06c8c692021-11-17 07:13:58 +01002646static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002647{
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002648 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2649
2650 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
Christoph Hellwig710fa372023-04-13 08:40:54 +02002651 blk_mq_insert_request(rq, 0);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002652 return BLK_STS_OK;
2653 }
2654
2655 if (!blk_mq_get_budget_and_tag(rq))
2656 return BLK_STS_RESOURCE;
2657 return __blk_mq_issue_directly(hctx, rq, last);
Christoph Hellwig5eb61262017-03-22 15:01:51 -04002658}
2659
Kemeng Shi3e368fb2023-01-18 17:37:18 +08002660static void blk_mq_plug_issue_direct(struct blk_plug *plug)
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002661{
2662 struct blk_mq_hw_ctx *hctx = NULL;
2663 struct request *rq;
2664 int queued = 0;
Kemeng Shi0d617a82023-01-18 17:37:20 +08002665 blk_status_t ret = BLK_STS_OK;
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002666
2667 while ((rq = rq_list_pop(&plug->mq_list))) {
2668 bool last = rq_list_empty(plug->mq_list);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002669
2670 if (hctx != rq->mq_hctx) {
Kemeng Shi34c9f5472023-01-18 17:37:19 +08002671 if (hctx) {
2672 blk_mq_commit_rqs(hctx, queued, false);
2673 queued = 0;
2674 }
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002675 hctx = rq->mq_hctx;
2676 }
2677
2678 ret = blk_mq_request_issue_directly(rq, last);
2679 switch (ret) {
2680 case BLK_STS_OK:
2681 queued++;
2682 break;
2683 case BLK_STS_RESOURCE:
2684 case BLK_STS_DEV_RESOURCE:
Christoph Hellwig2b597612023-04-13 08:40:55 +02002685 blk_mq_request_bypass_insert(rq, 0);
Christoph Hellwig23943952023-04-13 08:40:52 +02002686 blk_mq_run_hw_queue(hctx, false);
Kemeng Shi0d617a82023-01-18 17:37:20 +08002687 goto out;
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002688 default:
2689 blk_mq_end_request(rq, ret);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002690 break;
2691 }
2692 }
2693
Kemeng Shi0d617a82023-01-18 17:37:20 +08002694out:
2695 if (ret != BLK_STS_OK)
Kemeng Shi34c9f5472023-01-18 17:37:19 +08002696 blk_mq_commit_rqs(hctx, queued, false);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002697}
2698
Keith Busch518579a2021-12-20 12:59:19 -08002699static void __blk_mq_flush_plug_list(struct request_queue *q,
2700 struct blk_plug *plug)
2701{
2702 if (blk_queue_quiesced(q))
2703 return;
2704 q->mq_ops->queue_rqs(&plug->mq_list);
2705}
2706
Jens Axboe26fed4a2022-03-11 10:24:17 -07002707static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2708{
2709 struct blk_mq_hw_ctx *this_hctx = NULL;
2710 struct blk_mq_ctx *this_ctx = NULL;
2711 struct request *requeue_list = NULL;
Jan Kara34e0a272023-03-13 10:30:02 +01002712 struct request **requeue_lastp = &requeue_list;
Jens Axboe26fed4a2022-03-11 10:24:17 -07002713 unsigned int depth = 0;
2714 LIST_HEAD(list);
2715
2716 do {
2717 struct request *rq = rq_list_pop(&plug->mq_list);
2718
2719 if (!this_hctx) {
2720 this_hctx = rq->mq_hctx;
2721 this_ctx = rq->mq_ctx;
2722 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
Jan Kara34e0a272023-03-13 10:30:02 +01002723 rq_list_add_tail(&requeue_lastp, rq);
Jens Axboe26fed4a2022-03-11 10:24:17 -07002724 continue;
2725 }
Jan Kara34e0a272023-03-13 10:30:02 +01002726 list_add(&rq->queuelist, &list);
Jens Axboe26fed4a2022-03-11 10:24:17 -07002727 depth++;
2728 } while (!rq_list_empty(plug->mq_list));
2729
2730 plug->mq_list = requeue_list;
2731 trace_block_unplug(this_hctx->queue, depth, !from_sched);
Christoph Hellwig05a93112023-04-13 08:40:42 +02002732
2733 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2734 if (this_hctx->queue->elevator) {
2735 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
Christoph Hellwig93fffe12023-04-13 08:40:56 +02002736 &list, 0);
Christoph Hellwig05a93112023-04-13 08:40:42 +02002737 blk_mq_run_hw_queue(this_hctx, from_sched);
2738 } else {
2739 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2740 }
2741 percpu_ref_put(&this_hctx->queue->q_usage_counter);
Jens Axboe26fed4a2022-03-11 10:24:17 -07002742}
2743
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002744void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2745{
Jens Axboe3c67d442021-12-03 06:48:53 -07002746 struct request *rq;
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002747
2748 if (rq_list_empty(plug->mq_list))
2749 return;
2750 plug->rq_count = 0;
2751
2752 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
Jens Axboe3c67d442021-12-03 06:48:53 -07002753 struct request_queue *q;
2754
2755 rq = rq_list_peek(&plug->mq_list);
2756 q = rq->q;
2757
2758 /*
2759 * Peek first request and see if we have a ->queue_rqs() hook.
2760 * If we do, we can dispatch the whole plug list in one go. We
2761 * already know at this point that all requests belong to the
2762 * same queue, caller must ensure that's the case.
2763 *
2764 * Since we pass off the full list to the driver at this point,
2765 * we do not increment the active request count for the queue.
2766 * Bypass shared tags for now because of that.
2767 */
2768 if (q->mq_ops->queue_rqs &&
2769 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2770 blk_mq_run_dispatch_ops(q,
Keith Busch518579a2021-12-20 12:59:19 -08002771 __blk_mq_flush_plug_list(q, plug));
Jens Axboe3c67d442021-12-03 06:48:53 -07002772 if (rq_list_empty(plug->mq_list))
2773 return;
2774 }
Ming Lei73f37602021-12-06 11:33:50 +08002775
2776 blk_mq_run_dispatch_ops(q,
Kemeng Shi3e368fb2023-01-18 17:37:18 +08002777 blk_mq_plug_issue_direct(plug));
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002778 if (rq_list_empty(plug->mq_list))
2779 return;
2780 }
2781
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002782 do {
Jens Axboe26fed4a2022-03-11 10:24:17 -07002783 blk_mq_dispatch_plug_list(plug, from_schedule);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002784 } while (!rq_list_empty(plug->mq_list));
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002785}
2786
Christoph Hellwig94aa2282023-04-13 08:40:41 +02002787static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
Ming Lei6ce3dd62018-07-10 09:03:31 +08002788 struct list_head *list)
2789{
Keith Busch536167d42020-04-07 03:13:48 +09002790 int queued = 0;
Kemeng Shi984ce0a2023-01-18 17:37:21 +08002791 blk_status_t ret = BLK_STS_OK;
Keith Busch536167d42020-04-07 03:13:48 +09002792
Ming Lei6ce3dd62018-07-10 09:03:31 +08002793 while (!list_empty(list)) {
Ming Lei6ce3dd62018-07-10 09:03:31 +08002794 struct request *rq = list_first_entry(list, struct request,
2795 queuelist);
2796
2797 list_del_init(&rq->queuelist);
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002798 ret = blk_mq_request_issue_directly(rq, list_empty(list));
Kemeng Shi27e8b2b2023-01-18 17:37:25 +08002799 switch (ret) {
2800 case BLK_STS_OK:
Keith Busch536167d42020-04-07 03:13:48 +09002801 queued++;
Kemeng Shi27e8b2b2023-01-18 17:37:25 +08002802 break;
2803 case BLK_STS_RESOURCE:
2804 case BLK_STS_DEV_RESOURCE:
Christoph Hellwig2b597612023-04-13 08:40:55 +02002805 blk_mq_request_bypass_insert(rq, 0);
Christoph Hellwig23943952023-04-13 08:40:52 +02002806 if (list_empty(list))
2807 blk_mq_run_hw_queue(hctx, false);
Kemeng Shi27e8b2b2023-01-18 17:37:25 +08002808 goto out;
2809 default:
2810 blk_mq_end_request(rq, ret);
2811 break;
2812 }
Ming Lei6ce3dd62018-07-10 09:03:31 +08002813 }
Jens Axboed666ba92018-11-27 17:02:25 -07002814
Kemeng Shi27e8b2b2023-01-18 17:37:25 +08002815out:
Kemeng Shi984ce0a2023-01-18 17:37:21 +08002816 if (ret != BLK_STS_OK)
2817 blk_mq_commit_rqs(hctx, queued, false);
Ming Lei6ce3dd62018-07-10 09:03:31 +08002818}
2819
Ming Leib131f202021-11-11 16:51:34 +08002820static bool blk_mq_attempt_bio_merge(struct request_queue *q,
Christoph Hellwig0c5bcc92021-11-23 17:04:41 +01002821 struct bio *bio, unsigned int nr_segs)
Jens Axboe900e0802021-11-03 05:47:09 -06002822{
2823 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
Christoph Hellwig0c5bcc92021-11-23 17:04:41 +01002824 if (blk_attempt_plug_merge(q, bio, nr_segs))
Jens Axboe900e0802021-11-03 05:47:09 -06002825 return true;
2826 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2827 return true;
2828 }
2829 return false;
2830}
2831
Jens Axboe71539712021-11-03 05:52:45 -06002832static struct request *blk_mq_get_new_requests(struct request_queue *q,
2833 struct blk_plug *plug,
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002834 struct bio *bio,
2835 unsigned int nsegs)
Jens Axboe71539712021-11-03 05:52:45 -06002836{
2837 struct blk_mq_alloc_data data = {
2838 .q = q,
2839 .nr_tags = 1,
Ming Lei9d497e22022-01-04 21:42:23 +08002840 .cmd_flags = bio->bi_opf,
Jens Axboe71539712021-11-03 05:52:45 -06002841 };
2842 struct request *rq;
2843
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002844 if (unlikely(bio_queue_enter(bio)))
Ming Leib6371082021-11-12 20:47:15 +08002845 return NULL;
Jens Axboe900e0802021-11-03 05:47:09 -06002846
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002847 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2848 goto queue_exit;
2849
2850 rq_qos_throttle(q, bio);
2851
Jens Axboe71539712021-11-03 05:52:45 -06002852 if (plug) {
2853 data.nr_tags = plug->nr_ios;
2854 plug->nr_ios = 1;
2855 data.cached_rq = &plug->cached_rq;
2856 }
2857
2858 rq = __blk_mq_alloc_requests(&data);
Jens Axboe373b5412021-12-02 12:42:58 -07002859 if (rq)
2860 return rq;
Jens Axboe71539712021-11-03 05:52:45 -06002861 rq_qos_cleanup(q, bio);
2862 if (bio->bi_opf & REQ_NOWAIT)
2863 bio_wouldblock_error(bio);
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002864queue_exit:
Jens Axboe95febeb2021-11-15 14:23:08 -07002865 blk_queue_exit(q);
2866 return NULL;
Jens Axboe71539712021-11-03 05:52:45 -06002867}
2868
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002869static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002870 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002871{
2872 struct request *rq;
Pavel Begunkov77465642023-01-17 11:42:15 +00002873 enum hctx_type type, hctx_type;
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002874
2875 if (!plug)
2876 return NULL;
Ming Lei81ea1222023-04-20 19:20:18 +08002877 rq = rq_list_peek(&plug->cached_rq);
2878 if (!rq || rq->q != q)
2879 return NULL;
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002880
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002881 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2882 *bio = NULL;
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002883 return NULL;
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002884 }
2885
Pavel Begunkov77465642023-01-17 11:42:15 +00002886 type = blk_mq_get_hctx_type((*bio)->bi_opf);
2887 hctx_type = rq->mq_hctx->type;
2888 if (type != hctx_type &&
2889 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002890 return NULL;
2891 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002892 return NULL;
2893
Jens Axboe26456722022-06-21 10:03:57 -06002894 /*
2895 * If any qos ->throttle() end up blocking, we will have flushed the
2896 * plug and hence killed the cached_rq list as well. Pop this entry
2897 * before we throttle.
2898 */
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002899 plug->cached_rq = rq_list_next(rq);
Jens Axboe26456722022-06-21 10:03:57 -06002900 rq_qos_throttle(q, *bio);
2901
2902 rq->cmd_flags = (*bio)->bi_opf;
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002903 INIT_LIST_HEAD(&rq->queuelist);
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002904 return rq;
2905}
2906
Jan Kara82b74ca2022-06-23 09:48:32 +02002907static void bio_set_ioprio(struct bio *bio)
2908{
Jan Karaa78418e2022-06-23 09:48:34 +02002909 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2910 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2911 bio->bi_ioprio = get_current_ioprio();
Jan Kara82b74ca2022-06-23 09:48:32 +02002912 blkcg_set_ioprio(bio);
2913}
2914
André Almeida105663f2020-01-06 15:08:18 -03002915/**
Christoph Hellwigc62b37d2020-07-01 10:59:43 +02002916 * blk_mq_submit_bio - Create and send a request to block device.
André Almeida105663f2020-01-06 15:08:18 -03002917 * @bio: Bio pointer.
2918 *
2919 * Builds up a request structure from @q and @bio and send to the device. The
2920 * request may not be queued directly to hardware if:
2921 * * This request can be merged with another one
2922 * * We want to place request at plug queue for possible future merging
2923 * * There is an IO scheduler active at this queue
2924 *
2925 * It will not queue the request if there is an error with the bio, or at the
2926 * request creation.
André Almeida105663f2020-01-06 15:08:18 -03002927 */
Christoph Hellwig3e087732021-10-12 13:12:24 +02002928void blk_mq_submit_bio(struct bio *bio)
Jens Axboe07068d52014-05-22 10:40:51 -06002929{
Pavel Begunkoved6cdde2021-10-14 15:03:30 +01002930 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
Christoph Hellwig6deacb32022-07-06 09:03:38 +02002931 struct blk_plug *plug = blk_mq_plug(bio);
Christoph Hellwigef295ec2016-10-28 08:48:16 -06002932 const int is_sync = op_is_sync(bio->bi_opf);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02002933 struct blk_mq_hw_ctx *hctx;
Jens Axboe07068d52014-05-22 10:40:51 -06002934 struct request *rq;
Jens Axboeabd45c12021-10-13 12:43:41 -06002935 unsigned int nr_segs = 1;
Satya Tangiralaa892c8d2020-05-14 00:37:18 +00002936 blk_status_t ret;
Jens Axboe07068d52014-05-22 10:40:51 -06002937
Christoph Hellwig51d798c2022-07-27 12:22:56 -04002938 bio = blk_queue_bounce(bio, q);
Jens Axboe613b1482023-01-04 08:51:19 -07002939 if (bio_may_exceed_limits(bio, &q->limits)) {
Christoph Hellwigc55ddd92022-07-27 12:23:00 -04002940 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
Jens Axboe613b1482023-01-04 08:51:19 -07002941 if (!bio)
2942 return;
2943 }
Wen Xiongf36ea502017-05-10 08:54:11 -05002944
Dmitry Monakhove23947b2017-06-29 11:31:11 -07002945 if (!bio_integrity_prep(bio))
Jens Axboe900e0802021-11-03 05:47:09 -06002946 return;
Jens Axboe87760e52016-11-09 12:38:14 -07002947
Jan Kara9c6227e2022-06-23 09:48:33 +02002948 bio_set_ioprio(bio);
2949
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002950 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002951 if (!rq) {
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002952 if (!bio)
2953 return;
2954 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002955 if (unlikely(!rq))
2956 return;
2957 }
Jens Axboe87760e52016-11-09 12:38:14 -07002958
Christoph Hellwige8a676d2020-12-03 17:21:36 +01002959 trace_block_getrq(bio);
Xiaoguang Wangd6f1dda2018-10-23 22:30:50 +08002960
Josef Bacikc1c80382018-07-03 11:14:59 -04002961 rq_qos_track(q, rq, bio);
Jens Axboe07068d52014-05-22 10:40:51 -06002962
Bart Van Assche970d1682019-07-01 08:47:30 -07002963 blk_mq_bio_to_request(rq, bio, nr_segs);
2964
Eric Biggers9cd1e562023-03-15 11:39:02 -07002965 ret = blk_crypto_rq_get_keyslot(rq);
Satya Tangiralaa892c8d2020-05-14 00:37:18 +00002966 if (ret != BLK_STS_OK) {
2967 bio->bi_status = ret;
2968 bio_endio(bio);
2969 blk_mq_free_request(rq);
Christoph Hellwig3e087732021-10-12 13:12:24 +02002970 return;
Satya Tangiralaa892c8d2020-05-14 00:37:18 +00002971 }
2972
Ming Lei2b504bd2021-11-18 23:30:41 +08002973 if (op_is_flush(bio->bi_opf)) {
2974 blk_insert_flush(rq);
Christoph Hellwigd92ca9d82021-10-19 14:25:53 +02002975 return;
Ming Lei2b504bd2021-11-18 23:30:41 +08002976 }
Christoph Hellwigd92ca9d82021-10-19 14:25:53 +02002977
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02002978 if (plug) {
Jens Axboece5b0092018-11-27 17:13:56 -07002979 blk_add_rq_to_plug(plug, rq);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02002980 return;
2981 }
2982
2983 hctx = rq->mq_hctx;
2984 if ((rq->rq_flags & RQF_ELV) ||
2985 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
Christoph Hellwig710fa372023-04-13 08:40:54 +02002986 blk_mq_insert_request(rq, 0);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02002987 blk_mq_run_hw_queue(hctx, true);
2988 } else {
2989 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
2990 }
Jens Axboe320ae512013-10-24 09:20:05 +01002991}
2992
Christoph Hellwig248c7932022-02-15 11:05:36 +01002993#ifdef CONFIG_BLK_MQ_STACKING
Christoph Hellwig06c8c692021-11-17 07:13:58 +01002994/**
Christoph Hellwiga5efda32022-02-15 11:05:37 +01002995 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
Christoph Hellwiga5efda32022-02-15 11:05:37 +01002996 * @rq: the request being queued
Christoph Hellwig06c8c692021-11-17 07:13:58 +01002997 */
Christoph Hellwig28db4712022-02-15 11:05:38 +01002998blk_status_t blk_insert_cloned_request(struct request *rq)
Christoph Hellwig06c8c692021-11-17 07:13:58 +01002999{
Christoph Hellwig28db4712022-02-15 11:05:38 +01003000 struct request_queue *q = rq->q;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003001 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
Uday Shankar49d24392023-02-28 17:06:55 -07003002 unsigned int max_segments = blk_rq_get_max_segments(rq);
Christoph Hellwiga5efda32022-02-15 11:05:37 +01003003 blk_status_t ret;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003004
3005 if (blk_rq_sectors(rq) > max_sectors) {
3006 /*
3007 * SCSI device does not have a good way to return if
3008 * Write Same/Zero is actually supported. If a device rejects
3009 * a non-read/write command (discard, write same,etc.) the
3010 * low-level device driver will set the relevant queue limit to
3011 * 0 to prevent blk-lib from issuing more of the offending
3012 * operations. Commands queued prior to the queue limit being
3013 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3014 * errors being propagated to upper layers.
3015 */
3016 if (max_sectors == 0)
3017 return BLK_STS_NOTSUPP;
3018
3019 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3020 __func__, blk_rq_sectors(rq), max_sectors);
3021 return BLK_STS_IOERR;
3022 }
3023
3024 /*
3025 * The queue settings related to segment counting may differ from the
3026 * original queue.
3027 */
3028 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
Uday Shankar49d24392023-02-28 17:06:55 -07003029 if (rq->nr_phys_segments > max_segments) {
3030 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3031 __func__, rq->nr_phys_segments, max_segments);
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003032 return BLK_STS_IOERR;
3033 }
3034
Christoph Hellwig28db4712022-02-15 11:05:38 +01003035 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003036 return BLK_STS_IOERR;
3037
Eric Biggers5b8562f2023-03-15 11:39:06 -07003038 ret = blk_crypto_rq_get_keyslot(rq);
3039 if (ret != BLK_STS_OK)
3040 return ret;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003041
3042 blk_account_io_start(rq);
3043
3044 /*
3045 * Since we have a scheduler attached on the top device,
3046 * bypass a potential scheduler on the bottom device for
3047 * insert.
3048 */
Christoph Hellwig28db4712022-02-15 11:05:38 +01003049 blk_mq_run_dispatch_ops(q,
Ming Lei4cafe862021-12-03 21:15:34 +08003050 ret = blk_mq_request_issue_directly(rq, true));
Yu Kuai592ee112022-01-26 09:21:32 +08003051 if (ret)
3052 blk_account_io_done(rq, ktime_get_ns());
Ming Lei4cafe862021-12-03 21:15:34 +08003053 return ret;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003054}
3055EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3056
3057/**
3058 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3059 * @rq: the clone request to be cleaned up
3060 *
3061 * Description:
3062 * Free all bios in @rq for a cloned request.
3063 */
3064void blk_rq_unprep_clone(struct request *rq)
3065{
3066 struct bio *bio;
3067
3068 while ((bio = rq->bio) != NULL) {
3069 rq->bio = bio->bi_next;
3070
3071 bio_put(bio);
3072 }
3073}
3074EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3075
3076/**
3077 * blk_rq_prep_clone - Helper function to setup clone request
3078 * @rq: the request to be setup
3079 * @rq_src: original request to be cloned
3080 * @bs: bio_set that bios for clone are allocated from
3081 * @gfp_mask: memory allocation mask for bio
3082 * @bio_ctr: setup function to be called for each clone bio.
3083 * Returns %0 for success, non %0 for failure.
3084 * @data: private data to be passed to @bio_ctr
3085 *
3086 * Description:
3087 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3088 * Also, pages which the original bios are pointing to are not copied
3089 * and the cloned bios just point same pages.
3090 * So cloned bios must be completed before original bios, which means
3091 * the caller must complete @rq before @rq_src.
3092 */
3093int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3094 struct bio_set *bs, gfp_t gfp_mask,
3095 int (*bio_ctr)(struct bio *, struct bio *, void *),
3096 void *data)
3097{
3098 struct bio *bio, *bio_src;
3099
3100 if (!bs)
3101 bs = &fs_bio_set;
3102
3103 __rq_for_each_bio(bio_src, rq_src) {
Christoph Hellwigabfc4262022-02-02 17:01:09 +01003104 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3105 bs);
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003106 if (!bio)
3107 goto free_and_out;
3108
3109 if (bio_ctr && bio_ctr(bio, bio_src, data))
3110 goto free_and_out;
3111
3112 if (rq->bio) {
3113 rq->biotail->bi_next = bio;
3114 rq->biotail = bio;
3115 } else {
3116 rq->bio = rq->biotail = bio;
3117 }
3118 bio = NULL;
3119 }
3120
3121 /* Copy attributes of the original request to the clone request. */
3122 rq->__sector = blk_rq_pos(rq_src);
3123 rq->__data_len = blk_rq_bytes(rq_src);
3124 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3125 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3126 rq->special_vec = rq_src->special_vec;
3127 }
3128 rq->nr_phys_segments = rq_src->nr_phys_segments;
3129 rq->ioprio = rq_src->ioprio;
3130
3131 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3132 goto free_and_out;
3133
3134 return 0;
3135
3136free_and_out:
3137 if (bio)
3138 bio_put(bio);
3139 blk_rq_unprep_clone(rq);
3140
3141 return -ENOMEM;
3142}
3143EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
Christoph Hellwig248c7932022-02-15 11:05:36 +01003144#endif /* CONFIG_BLK_MQ_STACKING */
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003145
Christoph Hellwigf2b8f3c2021-11-17 07:14:00 +01003146/*
3147 * Steal bios from a request and add them to a bio list.
3148 * The request must not have been partially completed before.
3149 */
3150void blk_steal_bios(struct bio_list *list, struct request *rq)
3151{
3152 if (rq->bio) {
3153 if (list->tail)
3154 list->tail->bi_next = rq->bio;
3155 else
3156 list->head = rq->bio;
3157 list->tail = rq->biotail;
3158
3159 rq->bio = NULL;
3160 rq->biotail = NULL;
3161 }
3162
3163 rq->__data_len = 0;
3164}
3165EXPORT_SYMBOL_GPL(blk_steal_bios);
3166
Ming Leibd631412021-05-11 23:22:35 +08003167static size_t order_to_size(unsigned int order)
3168{
3169 return (size_t)PAGE_SIZE << order;
3170}
3171
3172/* called before freeing request pool in @tags */
John Garryf32e4ea2021-10-05 18:23:32 +08003173static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3174 struct blk_mq_tags *tags)
Ming Leibd631412021-05-11 23:22:35 +08003175{
Ming Leibd631412021-05-11 23:22:35 +08003176 struct page *page;
3177 unsigned long flags;
3178
Yu Kuai76dd2982022-10-11 22:22:53 +08003179 /*
3180 * There is no need to clear mapping if driver tags is not initialized
3181 * or the mapping belongs to the driver tags.
3182 */
3183 if (!drv_tags || drv_tags == tags)
John Garry4f245d52021-10-05 18:23:33 +08003184 return;
3185
Ming Leibd631412021-05-11 23:22:35 +08003186 list_for_each_entry(page, &tags->page_list, lru) {
3187 unsigned long start = (unsigned long)page_address(page);
3188 unsigned long end = start + order_to_size(page->private);
3189 int i;
3190
John Garryf32e4ea2021-10-05 18:23:32 +08003191 for (i = 0; i < drv_tags->nr_tags; i++) {
Ming Leibd631412021-05-11 23:22:35 +08003192 struct request *rq = drv_tags->rqs[i];
3193 unsigned long rq_addr = (unsigned long)rq;
3194
3195 if (rq_addr >= start && rq_addr < end) {
Jens Axboe0a467d02021-10-14 14:39:59 -06003196 WARN_ON_ONCE(req_ref_read(rq) != 0);
Ming Leibd631412021-05-11 23:22:35 +08003197 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3198 }
3199 }
3200 }
3201
3202 /*
3203 * Wait until all pending iteration is done.
3204 *
3205 * Request reference is cleared and it is guaranteed to be observed
3206 * after the ->lock is released.
3207 */
3208 spin_lock_irqsave(&drv_tags->lock, flags);
3209 spin_unlock_irqrestore(&drv_tags->lock, flags);
3210}
3211
Jens Axboecc71a6f2017-01-11 14:29:56 -07003212void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3213 unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01003214{
John Garryf32e4ea2021-10-05 18:23:32 +08003215 struct blk_mq_tags *drv_tags;
Jens Axboe320ae512013-10-24 09:20:05 +01003216 struct page *page;
3217
Ming Leie02657e2022-03-08 06:51:48 +01003218 if (list_empty(&tags->page_list))
3219 return;
3220
John Garry079a2e32021-10-05 18:23:39 +08003221 if (blk_mq_is_shared_tags(set->flags))
3222 drv_tags = set->shared_tags;
John Garrye155b0c2021-10-05 18:23:37 +08003223 else
3224 drv_tags = set->tags[hctx_idx];
John Garryf32e4ea2021-10-05 18:23:32 +08003225
John Garry65de57b2021-10-05 18:23:26 +08003226 if (tags->static_rqs && set->ops->exit_request) {
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003227 int i;
3228
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003229 for (i = 0; i < tags->nr_tags; i++) {
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003230 struct request *rq = tags->static_rqs[i];
3231
3232 if (!rq)
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003233 continue;
Christoph Hellwigd6296d392017-05-01 10:19:08 -06003234 set->ops->exit_request(set, rq, hctx_idx);
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003235 tags->static_rqs[i] = NULL;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003236 }
3237 }
3238
John Garryf32e4ea2021-10-05 18:23:32 +08003239 blk_mq_clear_rq_mapping(drv_tags, tags);
Ming Leibd631412021-05-11 23:22:35 +08003240
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003241 while (!list_empty(&tags->page_list)) {
3242 page = list_first_entry(&tags->page_list, struct page, lru);
Dave Hansen67534712014-01-08 20:17:46 -07003243 list_del_init(&page->lru);
Catalin Marinasf75782e2015-09-14 18:16:02 +01003244 /*
3245 * Remove kmemleak object previously allocated in
Raul E Rangel273938b2019-05-02 13:48:11 -06003246 * blk_mq_alloc_rqs().
Catalin Marinasf75782e2015-09-14 18:16:02 +01003247 */
3248 kmemleak_free(page_address(page));
Jens Axboe320ae512013-10-24 09:20:05 +01003249 __free_pages(page, page->private);
3250 }
Jens Axboecc71a6f2017-01-11 14:29:56 -07003251}
Jens Axboe320ae512013-10-24 09:20:05 +01003252
John Garrye155b0c2021-10-05 18:23:37 +08003253void blk_mq_free_rq_map(struct blk_mq_tags *tags)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003254{
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003255 kfree(tags->rqs);
Jens Axboecc71a6f2017-01-11 14:29:56 -07003256 tags->rqs = NULL;
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003257 kfree(tags->static_rqs);
3258 tags->static_rqs = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003259
John Garrye155b0c2021-10-05 18:23:37 +08003260 blk_mq_free_tags(tags);
Jens Axboe320ae512013-10-24 09:20:05 +01003261}
3262
Ming Lei4d805132022-03-08 15:32:14 +08003263static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3264 unsigned int hctx_idx)
3265{
3266 int i;
3267
3268 for (i = 0; i < set->nr_maps; i++) {
3269 unsigned int start = set->map[i].queue_offset;
3270 unsigned int end = start + set->map[i].nr_queues;
3271
3272 if (hctx_idx >= start && hctx_idx < end)
3273 break;
3274 }
3275
3276 if (i >= set->nr_maps)
3277 i = HCTX_TYPE_DEFAULT;
3278
3279 return i;
3280}
3281
3282static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3283 unsigned int hctx_idx)
3284{
3285 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3286
3287 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3288}
3289
John Garry63064be2021-10-05 18:23:35 +08003290static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3291 unsigned int hctx_idx,
3292 unsigned int nr_tags,
John Garrye155b0c2021-10-05 18:23:37 +08003293 unsigned int reserved_tags)
Jens Axboe320ae512013-10-24 09:20:05 +01003294{
Ming Lei4d805132022-03-08 15:32:14 +08003295 int node = blk_mq_get_hctx_node(set, hctx_idx);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003296 struct blk_mq_tags *tags;
Jens Axboe320ae512013-10-24 09:20:05 +01003297
Shaohua Li59f082e2017-02-01 09:53:14 -08003298 if (node == NUMA_NO_NODE)
3299 node = set->numa_node;
3300
John Garrye155b0c2021-10-05 18:23:37 +08003301 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3302 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003303 if (!tags)
3304 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003305
Kees Cook590b5b72018-06-12 14:04:20 -07003306 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
Gabriel Krisman Bertazi36e1f3d12016-12-06 13:31:44 -02003307 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
Shaohua Li59f082e2017-02-01 09:53:14 -08003308 node);
Jinlong Chen7edfd682022-11-02 10:52:29 +08003309 if (!tags->rqs)
3310 goto err_free_tags;
Jens Axboe320ae512013-10-24 09:20:05 +01003311
Kees Cook590b5b72018-06-12 14:04:20 -07003312 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3313 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3314 node);
Jinlong Chen7edfd682022-11-02 10:52:29 +08003315 if (!tags->static_rqs)
3316 goto err_free_rqs;
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003317
Jens Axboecc71a6f2017-01-11 14:29:56 -07003318 return tags;
Jinlong Chen7edfd682022-11-02 10:52:29 +08003319
3320err_free_rqs:
3321 kfree(tags->rqs);
3322err_free_tags:
3323 blk_mq_free_tags(tags);
3324 return NULL;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003325}
3326
Tejun Heo1d9bd512018-01-09 08:29:48 -08003327static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3328 unsigned int hctx_idx, int node)
3329{
3330 int ret;
3331
3332 if (set->ops->init_request) {
3333 ret = set->ops->init_request(set, rq, hctx_idx, node);
3334 if (ret)
3335 return ret;
3336 }
3337
Keith Busch12f5b932018-05-29 15:52:28 +02003338 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Tejun Heo1d9bd512018-01-09 08:29:48 -08003339 return 0;
3340}
3341
John Garry63064be2021-10-05 18:23:35 +08003342static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3343 struct blk_mq_tags *tags,
3344 unsigned int hctx_idx, unsigned int depth)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003345{
3346 unsigned int i, j, entries_per_page, max_order = 4;
Ming Lei4d805132022-03-08 15:32:14 +08003347 int node = blk_mq_get_hctx_node(set, hctx_idx);
Jens Axboecc71a6f2017-01-11 14:29:56 -07003348 size_t rq_size, left;
Shaohua Li59f082e2017-02-01 09:53:14 -08003349
Shaohua Li59f082e2017-02-01 09:53:14 -08003350 if (node == NUMA_NO_NODE)
3351 node = set->numa_node;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003352
3353 INIT_LIST_HEAD(&tags->page_list);
3354
Jens Axboe320ae512013-10-24 09:20:05 +01003355 /*
3356 * rq_size is the size of the request plus driver payload, rounded
3357 * to the cacheline size
3358 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003359 rq_size = round_up(sizeof(struct request) + set->cmd_size,
Jens Axboe320ae512013-10-24 09:20:05 +01003360 cache_line_size());
Jens Axboecc71a6f2017-01-11 14:29:56 -07003361 left = rq_size * depth;
Jens Axboe320ae512013-10-24 09:20:05 +01003362
Jens Axboecc71a6f2017-01-11 14:29:56 -07003363 for (i = 0; i < depth; ) {
Jens Axboe320ae512013-10-24 09:20:05 +01003364 int this_order = max_order;
3365 struct page *page;
3366 int to_do;
3367 void *p;
3368
Bartlomiej Zolnierkiewiczb3a834b2016-05-16 09:54:47 -06003369 while (this_order && left < order_to_size(this_order - 1))
Jens Axboe320ae512013-10-24 09:20:05 +01003370 this_order--;
3371
3372 do {
Shaohua Li59f082e2017-02-01 09:53:14 -08003373 page = alloc_pages_node(node,
Gabriel Krisman Bertazi36e1f3d12016-12-06 13:31:44 -02003374 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
Jens Axboea5164402014-09-10 09:02:03 -06003375 this_order);
Jens Axboe320ae512013-10-24 09:20:05 +01003376 if (page)
3377 break;
3378 if (!this_order--)
3379 break;
3380 if (order_to_size(this_order) < rq_size)
3381 break;
3382 } while (1);
3383
3384 if (!page)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003385 goto fail;
Jens Axboe320ae512013-10-24 09:20:05 +01003386
3387 page->private = this_order;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003388 list_add_tail(&page->lru, &tags->page_list);
Jens Axboe320ae512013-10-24 09:20:05 +01003389
3390 p = page_address(page);
Catalin Marinasf75782e2015-09-14 18:16:02 +01003391 /*
3392 * Allow kmemleak to scan these pages as they contain pointers
3393 * to additional allocations like via ops->init_request().
3394 */
Gabriel Krisman Bertazi36e1f3d12016-12-06 13:31:44 -02003395 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
Jens Axboe320ae512013-10-24 09:20:05 +01003396 entries_per_page = order_to_size(this_order) / rq_size;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003397 to_do = min(entries_per_page, depth - i);
Jens Axboe320ae512013-10-24 09:20:05 +01003398 left -= to_do * rq_size;
3399 for (j = 0; j < to_do; j++) {
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003400 struct request *rq = p;
3401
3402 tags->static_rqs[i] = rq;
Tejun Heo1d9bd512018-01-09 08:29:48 -08003403 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3404 tags->static_rqs[i] = NULL;
3405 goto fail;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003406 }
3407
Jens Axboe320ae512013-10-24 09:20:05 +01003408 p += rq_size;
3409 i++;
3410 }
3411 }
Jens Axboecc71a6f2017-01-11 14:29:56 -07003412 return 0;
Jens Axboe320ae512013-10-24 09:20:05 +01003413
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003414fail:
Jens Axboecc71a6f2017-01-11 14:29:56 -07003415 blk_mq_free_rqs(set, tags, hctx_idx);
3416 return -ENOMEM;
Jens Axboe320ae512013-10-24 09:20:05 +01003417}
3418
Ming Leibf0beec2020-05-29 15:53:15 +02003419struct rq_iter_data {
3420 struct blk_mq_hw_ctx *hctx;
3421 bool has_rq;
3422};
3423
John Garry2dd65322022-07-06 20:03:53 +08003424static bool blk_mq_has_request(struct request *rq, void *data)
Ming Leibf0beec2020-05-29 15:53:15 +02003425{
3426 struct rq_iter_data *iter_data = data;
3427
3428 if (rq->mq_hctx != iter_data->hctx)
3429 return true;
3430 iter_data->has_rq = true;
3431 return false;
3432}
3433
3434static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3435{
3436 struct blk_mq_tags *tags = hctx->sched_tags ?
3437 hctx->sched_tags : hctx->tags;
3438 struct rq_iter_data data = {
3439 .hctx = hctx,
3440 };
3441
3442 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3443 return data.has_rq;
3444}
3445
3446static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3447 struct blk_mq_hw_ctx *hctx)
3448{
Yury Norov9b51d9d2021-08-14 14:17:05 -07003449 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
Ming Leibf0beec2020-05-29 15:53:15 +02003450 return false;
3451 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3452 return false;
3453 return true;
3454}
3455
3456static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3457{
3458 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3459 struct blk_mq_hw_ctx, cpuhp_online);
3460
3461 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3462 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3463 return 0;
3464
3465 /*
3466 * Prevent new request from being allocated on the current hctx.
3467 *
3468 * The smp_mb__after_atomic() Pairs with the implied barrier in
3469 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3470 * seen once we return from the tag allocator.
3471 */
3472 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3473 smp_mb__after_atomic();
3474
3475 /*
3476 * Try to grab a reference to the queue and wait for any outstanding
3477 * requests. If we could not grab a reference the queue has been
3478 * frozen and there are no requests.
3479 */
3480 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3481 while (blk_mq_hctx_has_requests(hctx))
3482 msleep(5);
3483 percpu_ref_put(&hctx->queue->q_usage_counter);
3484 }
3485
3486 return 0;
3487}
3488
3489static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3490{
3491 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3492 struct blk_mq_hw_ctx, cpuhp_online);
3493
3494 if (cpumask_test_cpu(cpu, hctx->cpumask))
3495 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3496 return 0;
3497}
3498
Jens Axboee57690f2016-08-24 15:34:35 -06003499/*
3500 * 'cpu' is going away. splice any existing rq_list entries from this
3501 * software queue to the hw queue dispatch list, and ensure that it
3502 * gets run.
3503 */
Thomas Gleixner9467f852016-09-22 08:05:17 -06003504static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
Jens Axboe484b4062014-05-21 14:01:15 -06003505{
Thomas Gleixner9467f852016-09-22 08:05:17 -06003506 struct blk_mq_hw_ctx *hctx;
Jens Axboe484b4062014-05-21 14:01:15 -06003507 struct blk_mq_ctx *ctx;
3508 LIST_HEAD(tmp);
Ming Leic16d6b52018-12-17 08:44:05 -07003509 enum hctx_type type;
Jens Axboe484b4062014-05-21 14:01:15 -06003510
Thomas Gleixner9467f852016-09-22 08:05:17 -06003511 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
Ming Leibf0beec2020-05-29 15:53:15 +02003512 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3513 return 0;
3514
Jens Axboee57690f2016-08-24 15:34:35 -06003515 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
Ming Leic16d6b52018-12-17 08:44:05 -07003516 type = hctx->type;
Jens Axboe484b4062014-05-21 14:01:15 -06003517
3518 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07003519 if (!list_empty(&ctx->rq_lists[type])) {
3520 list_splice_init(&ctx->rq_lists[type], &tmp);
Jens Axboe484b4062014-05-21 14:01:15 -06003521 blk_mq_hctx_clear_pending(hctx, ctx);
3522 }
3523 spin_unlock(&ctx->lock);
3524
3525 if (list_empty(&tmp))
Thomas Gleixner9467f852016-09-22 08:05:17 -06003526 return 0;
Jens Axboe484b4062014-05-21 14:01:15 -06003527
Jens Axboee57690f2016-08-24 15:34:35 -06003528 spin_lock(&hctx->lock);
3529 list_splice_tail_init(&tmp, &hctx->dispatch);
3530 spin_unlock(&hctx->lock);
Jens Axboe484b4062014-05-21 14:01:15 -06003531
3532 blk_mq_run_hw_queue(hctx, true);
Thomas Gleixner9467f852016-09-22 08:05:17 -06003533 return 0;
Jens Axboe484b4062014-05-21 14:01:15 -06003534}
3535
Thomas Gleixner9467f852016-09-22 08:05:17 -06003536static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
Jens Axboe484b4062014-05-21 14:01:15 -06003537{
Ming Leibf0beec2020-05-29 15:53:15 +02003538 if (!(hctx->flags & BLK_MQ_F_STACKING))
3539 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3540 &hctx->cpuhp_online);
Thomas Gleixner9467f852016-09-22 08:05:17 -06003541 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3542 &hctx->cpuhp_dead);
Jens Axboe484b4062014-05-21 14:01:15 -06003543}
3544
Ming Lei364b6182021-05-11 23:22:36 +08003545/*
3546 * Before freeing hw queue, clearing the flush request reference in
3547 * tags->rqs[] for avoiding potential UAF.
3548 */
3549static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3550 unsigned int queue_depth, struct request *flush_rq)
3551{
3552 int i;
3553 unsigned long flags;
3554
3555 /* The hw queue may not be mapped yet */
3556 if (!tags)
3557 return;
3558
Jens Axboe0a467d02021-10-14 14:39:59 -06003559 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
Ming Lei364b6182021-05-11 23:22:36 +08003560
3561 for (i = 0; i < queue_depth; i++)
3562 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3563
3564 /*
3565 * Wait until all pending iteration is done.
3566 *
3567 * Request reference is cleared and it is guaranteed to be observed
3568 * after the ->lock is released.
3569 */
3570 spin_lock_irqsave(&tags->lock, flags);
3571 spin_unlock_irqrestore(&tags->lock, flags);
3572}
3573
Ming Leic3b4afc2015-06-04 22:25:04 +08003574/* hctx->ctxs will be freed in queue's release handler */
Ming Lei08e98fc2014-09-25 23:23:38 +08003575static void blk_mq_exit_hctx(struct request_queue *q,
3576 struct blk_mq_tag_set *set,
3577 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3578{
Ming Lei364b6182021-05-11 23:22:36 +08003579 struct request *flush_rq = hctx->fq->flush_rq;
3580
Ming Lei8ab0b7d2018-01-09 21:28:29 +08003581 if (blk_mq_hw_queue_mapped(hctx))
3582 blk_mq_tag_idle(hctx);
Ming Lei08e98fc2014-09-25 23:23:38 +08003583
Ming Lei6cfeadb2022-06-16 09:44:01 +08003584 if (blk_queue_init_done(q))
3585 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3586 set->queue_depth, flush_rq);
Ming Leif70ced02014-09-25 23:23:47 +08003587 if (set->ops->exit_request)
Ming Lei364b6182021-05-11 23:22:36 +08003588 set->ops->exit_request(set, flush_rq, hctx_idx);
Ming Leif70ced02014-09-25 23:23:47 +08003589
Ming Lei08e98fc2014-09-25 23:23:38 +08003590 if (set->ops->exit_hctx)
3591 set->ops->exit_hctx(hctx, hctx_idx);
3592
Thomas Gleixner9467f852016-09-22 08:05:17 -06003593 blk_mq_remove_cpuhp(hctx);
Ming Lei2f8f1332019-04-30 09:52:27 +08003594
Ming Lei4e5cc992022-03-08 15:32:19 +08003595 xa_erase(&q->hctx_table, hctx_idx);
3596
Ming Lei2f8f1332019-04-30 09:52:27 +08003597 spin_lock(&q->unused_hctx_lock);
3598 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3599 spin_unlock(&q->unused_hctx_lock);
Ming Lei08e98fc2014-09-25 23:23:38 +08003600}
3601
Ming Lei624dbe42014-05-27 23:35:13 +08003602static void blk_mq_exit_hw_queues(struct request_queue *q,
3603 struct blk_mq_tag_set *set, int nr_queue)
3604{
3605 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08003606 unsigned long i;
Ming Lei624dbe42014-05-27 23:35:13 +08003607
3608 queue_for_each_hw_ctx(q, hctx, i) {
3609 if (i == nr_queue)
3610 break;
Ming Lei08e98fc2014-09-25 23:23:38 +08003611 blk_mq_exit_hctx(q, set, hctx, i);
Ming Lei624dbe42014-05-27 23:35:13 +08003612 }
Ming Lei624dbe42014-05-27 23:35:13 +08003613}
3614
Ming Lei08e98fc2014-09-25 23:23:38 +08003615static int blk_mq_init_hctx(struct request_queue *q,
3616 struct blk_mq_tag_set *set,
3617 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3618{
Ming Lei7c6c5b72019-04-30 09:52:26 +08003619 hctx->queue_num = hctx_idx;
Ming Lei08e98fc2014-09-25 23:23:38 +08003620
Ming Leibf0beec2020-05-29 15:53:15 +02003621 if (!(hctx->flags & BLK_MQ_F_STACKING))
3622 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3623 &hctx->cpuhp_online);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003624 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3625
3626 hctx->tags = set->tags[hctx_idx];
3627
3628 if (set->ops->init_hctx &&
3629 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3630 goto unregister_cpu_notifier;
3631
3632 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3633 hctx->numa_node))
3634 goto exit_hctx;
Ming Lei4e5cc992022-03-08 15:32:19 +08003635
3636 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3637 goto exit_flush_rq;
3638
Ming Lei7c6c5b72019-04-30 09:52:26 +08003639 return 0;
3640
Ming Lei4e5cc992022-03-08 15:32:19 +08003641 exit_flush_rq:
3642 if (set->ops->exit_request)
3643 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003644 exit_hctx:
3645 if (set->ops->exit_hctx)
3646 set->ops->exit_hctx(hctx, hctx_idx);
3647 unregister_cpu_notifier:
3648 blk_mq_remove_cpuhp(hctx);
3649 return -1;
3650}
3651
3652static struct blk_mq_hw_ctx *
3653blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3654 int node)
3655{
3656 struct blk_mq_hw_ctx *hctx;
3657 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3658
Ming Lei704b9142021-12-03 21:15:32 +08003659 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003660 if (!hctx)
3661 goto fail_alloc_hctx;
3662
3663 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3664 goto free_hctx;
3665
3666 atomic_set(&hctx->nr_active, 0);
Ming Lei08e98fc2014-09-25 23:23:38 +08003667 if (node == NUMA_NO_NODE)
Ming Lei7c6c5b72019-04-30 09:52:26 +08003668 node = set->numa_node;
3669 hctx->numa_node = node;
Ming Lei08e98fc2014-09-25 23:23:38 +08003670
Jens Axboe9f993732017-04-10 09:54:54 -06003671 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
Ming Lei08e98fc2014-09-25 23:23:38 +08003672 spin_lock_init(&hctx->lock);
3673 INIT_LIST_HEAD(&hctx->dispatch);
3674 hctx->queue = q;
Ming Lei51db1c32020-08-19 23:20:19 +08003675 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
Ming Lei08e98fc2014-09-25 23:23:38 +08003676
Ming Lei2f8f1332019-04-30 09:52:27 +08003677 INIT_LIST_HEAD(&hctx->hctx_list);
3678
Ming Lei08e98fc2014-09-25 23:23:38 +08003679 /*
3680 * Allocate space for all possible cpus to avoid allocation at
3681 * runtime
3682 */
Johannes Thumshirnd904bfa2017-11-15 17:32:33 -08003683 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
Ming Lei7c6c5b72019-04-30 09:52:26 +08003684 gfp, node);
Ming Lei08e98fc2014-09-25 23:23:38 +08003685 if (!hctx->ctxs)
Ming Lei7c6c5b72019-04-30 09:52:26 +08003686 goto free_cpumask;
Ming Lei08e98fc2014-09-25 23:23:38 +08003687
Jianchao Wang5b202852018-10-12 18:07:26 +08003688 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
Ming Leic548e622021-01-22 10:33:08 +08003689 gfp, node, false, false))
Ming Lei08e98fc2014-09-25 23:23:38 +08003690 goto free_ctxs;
Ming Lei08e98fc2014-09-25 23:23:38 +08003691 hctx->nr_ctx = 0;
3692
Ming Lei5815839b2018-06-25 19:31:47 +08003693 spin_lock_init(&hctx->dispatch_wait_lock);
Jens Axboeeb619fd2017-11-09 08:32:43 -07003694 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3695 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3696
Guoqing Jiang754a1572020-03-09 22:41:37 +01003697 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
Ming Leif70ced02014-09-25 23:23:47 +08003698 if (!hctx->fq)
Ming Lei7c6c5b72019-04-30 09:52:26 +08003699 goto free_bitmap;
Ming Leif70ced02014-09-25 23:23:47 +08003700
Ming Lei7c6c5b72019-04-30 09:52:26 +08003701 blk_mq_hctx_kobj_init(hctx);
Bart Van Assche6a83e742016-11-02 10:09:51 -06003702
Ming Lei7c6c5b72019-04-30 09:52:26 +08003703 return hctx;
Ming Lei08e98fc2014-09-25 23:23:38 +08003704
3705 free_bitmap:
Omar Sandoval88459642016-09-17 08:38:44 -06003706 sbitmap_free(&hctx->ctx_map);
Ming Lei08e98fc2014-09-25 23:23:38 +08003707 free_ctxs:
3708 kfree(hctx->ctxs);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003709 free_cpumask:
3710 free_cpumask_var(hctx->cpumask);
3711 free_hctx:
3712 kfree(hctx);
3713 fail_alloc_hctx:
3714 return NULL;
Ming Lei08e98fc2014-09-25 23:23:38 +08003715}
3716
Jens Axboe320ae512013-10-24 09:20:05 +01003717static void blk_mq_init_cpu_queues(struct request_queue *q,
3718 unsigned int nr_hw_queues)
3719{
Jens Axboeb3c661b2018-10-30 10:36:06 -06003720 struct blk_mq_tag_set *set = q->tag_set;
3721 unsigned int i, j;
Jens Axboe320ae512013-10-24 09:20:05 +01003722
3723 for_each_possible_cpu(i) {
3724 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3725 struct blk_mq_hw_ctx *hctx;
Ming Leic16d6b52018-12-17 08:44:05 -07003726 int k;
Jens Axboe320ae512013-10-24 09:20:05 +01003727
Jens Axboe320ae512013-10-24 09:20:05 +01003728 __ctx->cpu = i;
3729 spin_lock_init(&__ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07003730 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3731 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3732
Jens Axboe320ae512013-10-24 09:20:05 +01003733 __ctx->queue = q;
3734
Jens Axboe320ae512013-10-24 09:20:05 +01003735 /*
3736 * Set local node, IFF we have more than one hw queue. If
3737 * not, we remain on the home node of the device
3738 */
Jens Axboeb3c661b2018-10-30 10:36:06 -06003739 for (j = 0; j < set->nr_maps; j++) {
3740 hctx = blk_mq_map_queue_type(q, j, i);
3741 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
Xianting Tian576e85c2020-10-19 16:20:47 +08003742 hctx->numa_node = cpu_to_node(i);
Jens Axboeb3c661b2018-10-30 10:36:06 -06003743 }
Jens Axboe320ae512013-10-24 09:20:05 +01003744 }
3745}
3746
John Garry63064be2021-10-05 18:23:35 +08003747struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3748 unsigned int hctx_idx,
3749 unsigned int depth)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003750{
John Garry63064be2021-10-05 18:23:35 +08003751 struct blk_mq_tags *tags;
3752 int ret;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003753
John Garrye155b0c2021-10-05 18:23:37 +08003754 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
John Garry63064be2021-10-05 18:23:35 +08003755 if (!tags)
3756 return NULL;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003757
John Garry63064be2021-10-05 18:23:35 +08003758 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3759 if (ret) {
John Garrye155b0c2021-10-05 18:23:37 +08003760 blk_mq_free_rq_map(tags);
John Garry63064be2021-10-05 18:23:35 +08003761 return NULL;
3762 }
Jens Axboecc71a6f2017-01-11 14:29:56 -07003763
John Garry63064be2021-10-05 18:23:35 +08003764 return tags;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003765}
3766
John Garry63064be2021-10-05 18:23:35 +08003767static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3768 int hctx_idx)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003769{
John Garry079a2e32021-10-05 18:23:39 +08003770 if (blk_mq_is_shared_tags(set->flags)) {
3771 set->tags[hctx_idx] = set->shared_tags;
John Garry1c0706a2020-08-19 23:20:22 +08003772
John Garrye155b0c2021-10-05 18:23:37 +08003773 return true;
Jens Axboebd166ef2017-01-17 06:03:22 -07003774 }
John Garrye155b0c2021-10-05 18:23:37 +08003775
John Garry63064be2021-10-05 18:23:35 +08003776 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3777 set->queue_depth);
3778
3779 return set->tags[hctx_idx];
Jens Axboecc71a6f2017-01-11 14:29:56 -07003780}
3781
John Garry645db342021-10-05 18:23:36 +08003782void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3783 struct blk_mq_tags *tags,
3784 unsigned int hctx_idx)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003785{
John Garry645db342021-10-05 18:23:36 +08003786 if (tags) {
3787 blk_mq_free_rqs(set, tags, hctx_idx);
John Garrye155b0c2021-10-05 18:23:37 +08003788 blk_mq_free_rq_map(tags);
Jens Axboecc71a6f2017-01-11 14:29:56 -07003789 }
3790}
3791
John Garrye155b0c2021-10-05 18:23:37 +08003792static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3793 unsigned int hctx_idx)
3794{
John Garry079a2e32021-10-05 18:23:39 +08003795 if (!blk_mq_is_shared_tags(set->flags))
John Garrye155b0c2021-10-05 18:23:37 +08003796 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3797
3798 set->tags[hctx_idx] = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003799}
3800
Christoph Hellwig4b855ad2017-06-26 12:20:57 +02003801static void blk_mq_map_swqueue(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01003802{
Ming Lei4f481202022-03-08 15:32:18 +08003803 unsigned int j, hctx_idx;
3804 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01003805 struct blk_mq_hw_ctx *hctx;
3806 struct blk_mq_ctx *ctx;
Ming Lei2a34c082015-04-21 10:00:20 +08003807 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01003808
3809 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboee4043dc2014-04-09 10:18:23 -06003810 cpumask_clear(hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01003811 hctx->nr_ctx = 0;
huhaid416c922018-05-18 08:32:30 -06003812 hctx->dispatch_from = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003813 }
3814
3815 /*
Christoph Hellwig4b855ad2017-06-26 12:20:57 +02003816 * Map software to hardware queues.
Ming Lei4412efe2018-04-25 04:01:44 +08003817 *
3818 * If the cpu isn't present, the cpu is mapped to first hctx.
Jens Axboe320ae512013-10-24 09:20:05 +01003819 */
Christoph Hellwig20e4d8132018-01-12 10:53:06 +08003820 for_each_possible_cpu(i) {
Ming Lei4412efe2018-04-25 04:01:44 +08003821
Thomas Gleixner897bb0c2016-03-19 11:30:33 +01003822 ctx = per_cpu_ptr(q->queue_ctx, i);
Jens Axboeb3c661b2018-10-30 10:36:06 -06003823 for (j = 0; j < set->nr_maps; j++) {
Jianchao Wangbb94aea2019-01-24 18:25:33 +08003824 if (!set->map[j].nr_queues) {
3825 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3826 HCTX_TYPE_DEFAULT, i);
Ming Leie5edd5f2018-12-18 01:28:56 +08003827 continue;
Jianchao Wangbb94aea2019-01-24 18:25:33 +08003828 }
Ming Leifd689872020-05-07 21:04:08 +08003829 hctx_idx = set->map[j].mq_map[i];
3830 /* unmapped hw queue can be remapped after CPU topo changed */
3831 if (!set->tags[hctx_idx] &&
John Garry63064be2021-10-05 18:23:35 +08003832 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
Ming Leifd689872020-05-07 21:04:08 +08003833 /*
3834 * If tags initialization fail for some hctx,
3835 * that hctx won't be brought online. In this
3836 * case, remap the current ctx to hctx[0] which
3837 * is guaranteed to always have tags allocated
3838 */
3839 set->map[j].mq_map[i] = 0;
3840 }
Ming Leie5edd5f2018-12-18 01:28:56 +08003841
Jens Axboeb3c661b2018-10-30 10:36:06 -06003842 hctx = blk_mq_map_queue_type(q, j, i);
Jianchao Wang8ccdf4a2019-01-24 18:25:32 +08003843 ctx->hctxs[j] = hctx;
Jens Axboeb3c661b2018-10-30 10:36:06 -06003844 /*
3845 * If the CPU is already set in the mask, then we've
3846 * mapped this one already. This can happen if
3847 * devices share queues across queue maps.
3848 */
3849 if (cpumask_test_cpu(i, hctx->cpumask))
3850 continue;
3851
3852 cpumask_set_cpu(i, hctx->cpumask);
3853 hctx->type = j;
3854 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3855 hctx->ctxs[hctx->nr_ctx++] = ctx;
3856
3857 /*
3858 * If the nr_ctx type overflows, we have exceeded the
3859 * amount of sw queues we can support.
3860 */
3861 BUG_ON(!hctx->nr_ctx);
3862 }
Jianchao Wangbb94aea2019-01-24 18:25:33 +08003863
3864 for (; j < HCTX_MAX_TYPES; j++)
3865 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3866 HCTX_TYPE_DEFAULT, i);
Jens Axboe320ae512013-10-24 09:20:05 +01003867 }
Jens Axboe506e9312014-05-07 10:26:44 -06003868
3869 queue_for_each_hw_ctx(q, hctx, i) {
Ming Lei4412efe2018-04-25 04:01:44 +08003870 /*
3871 * If no software queues are mapped to this hardware queue,
3872 * disable it and free the request entries.
3873 */
3874 if (!hctx->nr_ctx) {
3875 /* Never unmap queue 0. We need it as a
3876 * fallback in case of a new remap fails
3877 * allocation
3878 */
John Garrye155b0c2021-10-05 18:23:37 +08003879 if (i)
3880 __blk_mq_free_map_and_rqs(set, i);
Ming Lei4412efe2018-04-25 04:01:44 +08003881
3882 hctx->tags = NULL;
3883 continue;
3884 }
Jens Axboe484b4062014-05-21 14:01:15 -06003885
Ming Lei2a34c082015-04-21 10:00:20 +08003886 hctx->tags = set->tags[i];
3887 WARN_ON(!hctx->tags);
3888
Jens Axboe484b4062014-05-21 14:01:15 -06003889 /*
Chong Yuan889fa312015-04-15 11:39:29 -06003890 * Set the map size to the number of mapped software queues.
3891 * This is more accurate and more efficient than looping
3892 * over all possibly mapped software queues.
3893 */
Omar Sandoval88459642016-09-17 08:38:44 -06003894 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
Chong Yuan889fa312015-04-15 11:39:29 -06003895
3896 /*
Jens Axboe484b4062014-05-21 14:01:15 -06003897 * Initialize batch roundrobin counts
3898 */
Ming Leif82ddf12018-04-08 17:48:10 +08003899 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
Jens Axboe506e9312014-05-07 10:26:44 -06003900 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3901 }
Jens Axboe320ae512013-10-24 09:20:05 +01003902}
3903
Jens Axboe8e8320c2017-06-20 17:56:13 -06003904/*
3905 * Caller needs to ensure that we're either frozen/quiesced, or that
3906 * the queue isn't live yet.
3907 */
Jeff Moyer2404e602015-11-03 10:40:06 -05003908static void queue_set_hctx_shared(struct request_queue *q, bool shared)
Jens Axboe0d2602c2014-05-13 15:10:52 -06003909{
3910 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08003911 unsigned long i;
Jens Axboe0d2602c2014-05-13 15:10:52 -06003912
Jeff Moyer2404e602015-11-03 10:40:06 -05003913 queue_for_each_hw_ctx(q, hctx, i) {
Yu Kuai454bb672021-07-31 14:21:30 +08003914 if (shared) {
Ming Lei51db1c32020-08-19 23:20:19 +08003915 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
Yu Kuai454bb672021-07-31 14:21:30 +08003916 } else {
3917 blk_mq_tag_idle(hctx);
Ming Lei51db1c32020-08-19 23:20:19 +08003918 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
Yu Kuai454bb672021-07-31 14:21:30 +08003919 }
Jeff Moyer2404e602015-11-03 10:40:06 -05003920 }
3921}
3922
Hannes Reinecke655ac302020-08-19 23:20:20 +08003923static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3924 bool shared)
Jeff Moyer2404e602015-11-03 10:40:06 -05003925{
3926 struct request_queue *q;
Jens Axboe0d2602c2014-05-13 15:10:52 -06003927
Bart Van Assche705cda92017-04-07 11:16:49 -07003928 lockdep_assert_held(&set->tag_list_lock);
3929
Jens Axboe0d2602c2014-05-13 15:10:52 -06003930 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3931 blk_mq_freeze_queue(q);
Jeff Moyer2404e602015-11-03 10:40:06 -05003932 queue_set_hctx_shared(q, shared);
Jens Axboe0d2602c2014-05-13 15:10:52 -06003933 blk_mq_unfreeze_queue(q);
3934 }
3935}
3936
3937static void blk_mq_del_queue_tag_set(struct request_queue *q)
3938{
3939 struct blk_mq_tag_set *set = q->tag_set;
3940
Jens Axboe0d2602c2014-05-13 15:10:52 -06003941 mutex_lock(&set->tag_list_lock);
Daniel Wagner08c875c2020-07-28 15:29:51 +02003942 list_del(&q->tag_set_list);
Jeff Moyer2404e602015-11-03 10:40:06 -05003943 if (list_is_singular(&set->tag_list)) {
3944 /* just transitioned to unshared */
Ming Lei51db1c32020-08-19 23:20:19 +08003945 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
Jeff Moyer2404e602015-11-03 10:40:06 -05003946 /* update existing queue */
Hannes Reinecke655ac302020-08-19 23:20:20 +08003947 blk_mq_update_tag_set_shared(set, false);
Jeff Moyer2404e602015-11-03 10:40:06 -05003948 }
Jens Axboe0d2602c2014-05-13 15:10:52 -06003949 mutex_unlock(&set->tag_list_lock);
Roman Pena347c7a2018-06-10 22:38:24 +02003950 INIT_LIST_HEAD(&q->tag_set_list);
Jens Axboe0d2602c2014-05-13 15:10:52 -06003951}
3952
3953static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3954 struct request_queue *q)
3955{
Jens Axboe0d2602c2014-05-13 15:10:52 -06003956 mutex_lock(&set->tag_list_lock);
Jeff Moyer2404e602015-11-03 10:40:06 -05003957
Jens Axboeff821d22017-11-10 22:05:12 -07003958 /*
3959 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3960 */
3961 if (!list_empty(&set->tag_list) &&
Ming Lei51db1c32020-08-19 23:20:19 +08003962 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3963 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
Jeff Moyer2404e602015-11-03 10:40:06 -05003964 /* update existing queue */
Hannes Reinecke655ac302020-08-19 23:20:20 +08003965 blk_mq_update_tag_set_shared(set, true);
Jeff Moyer2404e602015-11-03 10:40:06 -05003966 }
Ming Lei51db1c32020-08-19 23:20:19 +08003967 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
Jeff Moyer2404e602015-11-03 10:40:06 -05003968 queue_set_hctx_shared(q, true);
Daniel Wagner08c875c2020-07-28 15:29:51 +02003969 list_add_tail(&q->tag_set_list, &set->tag_list);
Jeff Moyer2404e602015-11-03 10:40:06 -05003970
Jens Axboe0d2602c2014-05-13 15:10:52 -06003971 mutex_unlock(&set->tag_list_lock);
3972}
3973
Ming Lei1db49092018-11-20 09:44:35 +08003974/* All allocations will be freed in release handler of q->mq_kobj */
3975static int blk_mq_alloc_ctxs(struct request_queue *q)
3976{
3977 struct blk_mq_ctxs *ctxs;
3978 int cpu;
3979
3980 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3981 if (!ctxs)
3982 return -ENOMEM;
3983
3984 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3985 if (!ctxs->queue_ctx)
3986 goto fail;
3987
3988 for_each_possible_cpu(cpu) {
3989 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3990 ctx->ctxs = ctxs;
3991 }
3992
3993 q->mq_kobj = &ctxs->kobj;
3994 q->queue_ctx = ctxs->queue_ctx;
3995
3996 return 0;
3997 fail:
3998 kfree(ctxs);
3999 return -ENOMEM;
4000}
4001
Ming Leie09aae72015-01-29 20:17:27 +08004002/*
4003 * It is the actual release handler for mq, but we do it from
4004 * request queue's release handler for avoiding use-after-free
4005 * and headache because q->mq_kobj shouldn't have been introduced,
4006 * but we can't group ctx/kctx kobj without it.
4007 */
4008void blk_mq_release(struct request_queue *q)
4009{
Ming Lei2f8f1332019-04-30 09:52:27 +08004010 struct blk_mq_hw_ctx *hctx, *next;
Ming Lei4f481202022-03-08 15:32:18 +08004011 unsigned long i;
Ming Leie09aae72015-01-29 20:17:27 +08004012
Ming Lei2f8f1332019-04-30 09:52:27 +08004013 queue_for_each_hw_ctx(q, hctx, i)
4014 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4015
4016 /* all hctx are in .unused_hctx_list now */
4017 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4018 list_del_init(&hctx->hctx_list);
Ming Lei6c8b2322017-02-22 18:14:01 +08004019 kobject_put(&hctx->kobj);
Ming Leic3b4afc2015-06-04 22:25:04 +08004020 }
Ming Leie09aae72015-01-29 20:17:27 +08004021
Ming Lei4e5cc992022-03-08 15:32:19 +08004022 xa_destroy(&q->hctx_table);
Ming Leie09aae72015-01-29 20:17:27 +08004023
Ming Lei7ea5fe32017-02-22 18:14:00 +08004024 /*
4025 * release .mq_kobj and sw queue's kobject now because
4026 * both share lifetime with request queue.
4027 */
4028 blk_mq_sysfs_deinit(q);
Ming Leie09aae72015-01-29 20:17:27 +08004029}
4030
Christoph Hellwig5ec780a2021-06-24 10:10:12 +02004031static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
Christoph Hellwig2f227bb2020-03-27 09:30:08 +01004032 void *queuedata)
Jens Axboe320ae512013-10-24 09:20:05 +01004033{
Christoph Hellwig26a97502021-06-02 09:53:17 +03004034 struct request_queue *q;
4035 int ret;
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004036
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004037 q = blk_alloc_queue(set->numa_node);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004038 if (!q)
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004039 return ERR_PTR(-ENOMEM);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004040 q->queuedata = queuedata;
4041 ret = blk_mq_init_allocated_queue(set, q);
4042 if (ret) {
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004043 blk_put_queue(q);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004044 return ERR_PTR(ret);
4045 }
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004046 return q;
4047}
Christoph Hellwig2f227bb2020-03-27 09:30:08 +01004048
4049struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4050{
4051 return blk_mq_init_queue_data(set, NULL);
4052}
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004053EXPORT_SYMBOL(blk_mq_init_queue);
4054
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004055/**
4056 * blk_mq_destroy_queue - shutdown a request queue
4057 * @q: request queue to shutdown
4058 *
Bart Van Assche81ea42b2023-01-30 13:12:33 -08004059 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4060 * requests will be failed with -ENODEV. The caller is responsible for dropping
4061 * the reference from blk_mq_init_queue() by calling blk_put_queue().
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004062 *
4063 * Context: can sleep
4064 */
4065void blk_mq_destroy_queue(struct request_queue *q)
4066{
4067 WARN_ON_ONCE(!queue_is_mq(q));
4068 WARN_ON_ONCE(blk_queue_registered(q));
4069
4070 might_sleep();
4071
4072 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4073 blk_queue_start_drain(q);
Jinlong Chen56c1ee92022-10-30 16:32:12 +08004074 blk_mq_freeze_queue_wait(q);
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004075
4076 blk_sync_queue(q);
4077 blk_mq_cancel_work_sync(q);
4078 blk_mq_exit_queue(q);
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004079}
4080EXPORT_SYMBOL(blk_mq_destroy_queue);
4081
Christoph Hellwig4dcc48742021-08-16 15:19:05 +02004082struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4083 struct lock_class_key *lkclass)
Jens Axboe9316a9e2018-10-15 08:40:37 -06004084{
4085 struct request_queue *q;
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004086 struct gendisk *disk;
Jens Axboe9316a9e2018-10-15 08:40:37 -06004087
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004088 q = blk_mq_init_queue_data(set, queuedata);
4089 if (IS_ERR(q))
4090 return ERR_CAST(q);
Jens Axboe9316a9e2018-10-15 08:40:37 -06004091
Christoph Hellwig4a1fa412021-08-16 15:19:08 +02004092 disk = __alloc_disk_node(q, set->numa_node, lkclass);
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004093 if (!disk) {
Christoph Hellwig0a3e5cc2022-07-20 15:05:40 +02004094 blk_mq_destroy_queue(q);
Christoph Hellwig2b3f0562022-10-18 15:57:17 +02004095 blk_put_queue(q);
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004096 return ERR_PTR(-ENOMEM);
Jens Axboe9316a9e2018-10-15 08:40:37 -06004097 }
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004098 set_bit(GD_OWNS_QUEUE, &disk->state);
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004099 return disk;
Jens Axboe9316a9e2018-10-15 08:40:37 -06004100}
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004101EXPORT_SYMBOL(__blk_mq_alloc_disk);
Jens Axboe9316a9e2018-10-15 08:40:37 -06004102
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004103struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4104 struct lock_class_key *lkclass)
4105{
Christoph Hellwig22c17e22022-11-22 08:27:53 +01004106 struct gendisk *disk;
4107
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004108 if (!blk_get_queue(q))
4109 return NULL;
Christoph Hellwig22c17e22022-11-22 08:27:53 +01004110 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4111 if (!disk)
4112 blk_put_queue(q);
4113 return disk;
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004114}
4115EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4116
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004117static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4118 struct blk_mq_tag_set *set, struct request_queue *q,
4119 int hctx_idx, int node)
4120{
Ming Lei2f8f1332019-04-30 09:52:27 +08004121 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004122
Ming Lei2f8f1332019-04-30 09:52:27 +08004123 /* reuse dead hctx first */
4124 spin_lock(&q->unused_hctx_lock);
4125 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4126 if (tmp->numa_node == node) {
4127 hctx = tmp;
4128 break;
4129 }
4130 }
4131 if (hctx)
4132 list_del_init(&hctx->hctx_list);
4133 spin_unlock(&q->unused_hctx_lock);
4134
4135 if (!hctx)
4136 hctx = blk_mq_alloc_hctx(q, set, node);
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004137 if (!hctx)
Ming Lei7c6c5b72019-04-30 09:52:26 +08004138 goto fail;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004139
Ming Lei7c6c5b72019-04-30 09:52:26 +08004140 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4141 goto free_hctx;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004142
4143 return hctx;
Ming Lei7c6c5b72019-04-30 09:52:26 +08004144
4145 free_hctx:
4146 kobject_put(&hctx->kobj);
4147 fail:
4148 return NULL;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004149}
4150
Keith Busch868f2f02015-12-17 17:08:14 -07004151static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4152 struct request_queue *q)
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004153{
Ming Lei4e5cc992022-03-08 15:32:19 +08004154 struct blk_mq_hw_ctx *hctx;
4155 unsigned long i, j;
Bart Van Asscheac0d6b92019-10-25 09:50:09 -07004156
Ming Leifb350e02018-01-06 16:27:40 +08004157 /* protect against switching io scheduler */
4158 mutex_lock(&q->sysfs_lock);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004159 for (i = 0; i < set->nr_hw_queues; i++) {
Ming Lei306f13ee2022-03-08 15:32:15 +08004160 int old_node;
Ming Lei4d805132022-03-08 15:32:14 +08004161 int node = blk_mq_get_hctx_node(set, i);
Ming Lei4e5cc992022-03-08 15:32:19 +08004162 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
Keith Busch868f2f02015-12-17 17:08:14 -07004163
Ming Lei306f13ee2022-03-08 15:32:15 +08004164 if (old_hctx) {
4165 old_node = old_hctx->numa_node;
4166 blk_mq_exit_hctx(q, set, old_hctx, i);
4167 }
Jens Axboe320ae512013-10-24 09:20:05 +01004168
Ming Lei4e5cc992022-03-08 15:32:19 +08004169 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
Ming Lei306f13ee2022-03-08 15:32:15 +08004170 if (!old_hctx)
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004171 break;
Ming Lei306f13ee2022-03-08 15:32:15 +08004172 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4173 node, old_node);
Ming Lei4e5cc992022-03-08 15:32:19 +08004174 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4175 WARN_ON_ONCE(!hctx);
Keith Busch868f2f02015-12-17 17:08:14 -07004176 }
Jens Axboe320ae512013-10-24 09:20:05 +01004177 }
Jianchao Wange01ad462018-10-12 18:07:28 +08004178 /*
4179 * Increasing nr_hw_queues fails. Free the newly allocated
4180 * hctxs and keep the previous q->nr_hw_queues.
4181 */
4182 if (i != set->nr_hw_queues) {
4183 j = q->nr_hw_queues;
Jianchao Wange01ad462018-10-12 18:07:28 +08004184 } else {
4185 j = i;
Jianchao Wange01ad462018-10-12 18:07:28 +08004186 q->nr_hw_queues = set->nr_hw_queues;
4187 }
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004188
Ming Lei4e5cc992022-03-08 15:32:19 +08004189 xa_for_each_start(&q->hctx_table, j, hctx, j)
4190 blk_mq_exit_hctx(q, set, hctx, j);
Ming Leifb350e02018-01-06 16:27:40 +08004191 mutex_unlock(&q->sysfs_lock);
Keith Busch868f2f02015-12-17 17:08:14 -07004192}
4193
Ming Lei42ee3062022-03-08 15:32:16 +08004194static void blk_mq_update_poll_flag(struct request_queue *q)
4195{
4196 struct blk_mq_tag_set *set = q->tag_set;
4197
4198 if (set->nr_maps > HCTX_TYPE_POLL &&
4199 set->map[HCTX_TYPE_POLL].nr_queues)
4200 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4201 else
4202 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4203}
4204
Christoph Hellwig26a97502021-06-02 09:53:17 +03004205int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4206 struct request_queue *q)
Keith Busch868f2f02015-12-17 17:08:14 -07004207{
Ming Lei66841672016-02-12 15:27:00 +08004208 /* mark the queue as mq asap */
4209 q->mq_ops = set->ops;
4210
Ming Lei1db49092018-11-20 09:44:35 +08004211 if (blk_mq_alloc_ctxs(q))
Keith Busch54bdd672023-03-20 12:49:26 -07004212 goto err_exit;
Keith Busch868f2f02015-12-17 17:08:14 -07004213
Ming Lei737f98c2017-02-22 18:13:59 +08004214 /* init q->mq_kobj and sw queues' kobjects */
4215 blk_mq_sysfs_init(q);
4216
Ming Lei2f8f1332019-04-30 09:52:27 +08004217 INIT_LIST_HEAD(&q->unused_hctx_list);
4218 spin_lock_init(&q->unused_hctx_lock);
4219
Ming Lei4e5cc992022-03-08 15:32:19 +08004220 xa_init(&q->hctx_table);
4221
Keith Busch868f2f02015-12-17 17:08:14 -07004222 blk_mq_realloc_hw_ctxs(set, q);
4223 if (!q->nr_hw_queues)
4224 goto err_hctxs;
Jens Axboe320ae512013-10-24 09:20:05 +01004225
Christoph Hellwig287922e2015-10-30 20:57:30 +08004226 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
Ming Leie56f6982015-07-16 19:53:22 +08004227 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
Jens Axboe320ae512013-10-24 09:20:05 +01004228
Jens Axboea8908932018-10-16 14:23:06 -06004229 q->tag_set = set;
Jens Axboe320ae512013-10-24 09:20:05 +01004230
Jens Axboe94eddfb2013-11-19 09:25:07 -07004231 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
Ming Lei42ee3062022-03-08 15:32:16 +08004232 blk_mq_update_poll_flag(q);
Jens Axboe320ae512013-10-24 09:20:05 +01004233
Mike Snitzer28494502016-09-14 13:28:30 -04004234 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06004235 INIT_LIST_HEAD(&q->requeue_list);
4236 spin_lock_init(&q->requeue_lock);
4237
Jens Axboeeba71762014-05-20 15:17:27 -06004238 q->nr_requests = set->queue_depth;
4239
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004240 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
Jens Axboe0d2602c2014-05-13 15:10:52 -06004241 blk_mq_add_queue_tag_set(set, q);
Christoph Hellwig4b855ad2017-06-26 12:20:57 +02004242 blk_mq_map_swqueue(q);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004243 return 0;
Christoph Hellwig18741982014-02-10 09:29:00 -07004244
Jens Axboe320ae512013-10-24 09:20:05 +01004245err_hctxs:
Chen Jun943f45b2022-10-31 03:12:42 +00004246 blk_mq_release(q);
Ming Linc7de5722016-05-25 23:23:27 -07004247err_exit:
4248 q->mq_ops = NULL;
Christoph Hellwig26a97502021-06-02 09:53:17 +03004249 return -ENOMEM;
Jens Axboe320ae512013-10-24 09:20:05 +01004250}
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004251EXPORT_SYMBOL(blk_mq_init_allocated_queue);
Jens Axboe320ae512013-10-24 09:20:05 +01004252
Ming Leic7e2d942019-04-30 09:52:25 +08004253/* tags can _not_ be used after returning from blk_mq_exit_queue */
4254void blk_mq_exit_queue(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01004255{
Bart Van Assche630ef622021-05-13 10:15:29 -07004256 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01004257
Bart Van Assche630ef622021-05-13 10:15:29 -07004258 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
Ming Lei624dbe42014-05-27 23:35:13 +08004259 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
Bart Van Assche630ef622021-05-13 10:15:29 -07004260 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4261 blk_mq_del_queue_tag_set(q);
Jens Axboe320ae512013-10-24 09:20:05 +01004262}
Jens Axboe320ae512013-10-24 09:20:05 +01004263
Jens Axboea5164402014-09-10 09:02:03 -06004264static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4265{
4266 int i;
4267
John Garry079a2e32021-10-05 18:23:39 +08004268 if (blk_mq_is_shared_tags(set->flags)) {
4269 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
John Garrye155b0c2021-10-05 18:23:37 +08004270 BLK_MQ_NO_HCTX_IDX,
4271 set->queue_depth);
John Garry079a2e32021-10-05 18:23:39 +08004272 if (!set->shared_tags)
John Garrye155b0c2021-10-05 18:23:37 +08004273 return -ENOMEM;
4274 }
4275
Xianting Tian8229cca2020-09-26 10:39:47 +08004276 for (i = 0; i < set->nr_hw_queues; i++) {
John Garry63064be2021-10-05 18:23:35 +08004277 if (!__blk_mq_alloc_map_and_rqs(set, i))
Jens Axboea5164402014-09-10 09:02:03 -06004278 goto out_unwind;
Xianting Tian8229cca2020-09-26 10:39:47 +08004279 cond_resched();
4280 }
Jens Axboea5164402014-09-10 09:02:03 -06004281
4282 return 0;
4283
4284out_unwind:
4285 while (--i >= 0)
John Garrye155b0c2021-10-05 18:23:37 +08004286 __blk_mq_free_map_and_rqs(set, i);
4287
John Garry079a2e32021-10-05 18:23:39 +08004288 if (blk_mq_is_shared_tags(set->flags)) {
4289 blk_mq_free_map_and_rqs(set, set->shared_tags,
John Garrye155b0c2021-10-05 18:23:37 +08004290 BLK_MQ_NO_HCTX_IDX);
John Garry645db342021-10-05 18:23:36 +08004291 }
Jens Axboea5164402014-09-10 09:02:03 -06004292
Jens Axboea5164402014-09-10 09:02:03 -06004293 return -ENOMEM;
4294}
4295
4296/*
4297 * Allocate the request maps associated with this tag_set. Note that this
4298 * may reduce the depth asked for, if memory is tight. set->queue_depth
4299 * will be updated to reflect the allocated depth.
4300 */
John Garry63064be2021-10-05 18:23:35 +08004301static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
Jens Axboea5164402014-09-10 09:02:03 -06004302{
4303 unsigned int depth;
4304 int err;
4305
4306 depth = set->queue_depth;
4307 do {
4308 err = __blk_mq_alloc_rq_maps(set);
4309 if (!err)
4310 break;
4311
4312 set->queue_depth >>= 1;
4313 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4314 err = -ENOMEM;
4315 break;
4316 }
4317 } while (set->queue_depth);
4318
4319 if (!set->queue_depth || err) {
4320 pr_err("blk-mq: failed to allocate request map\n");
4321 return -ENOMEM;
4322 }
4323
4324 if (depth != set->queue_depth)
4325 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4326 depth, set->queue_depth);
4327
4328 return 0;
4329}
4330
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004331static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
Omar Sandovalebe8bdd2017-04-07 08:53:11 -06004332{
Bart Van Assche6e66b492020-03-09 21:26:17 -07004333 /*
4334 * blk_mq_map_queues() and multiple .map_queues() implementations
4335 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4336 * number of hardware queues.
4337 */
4338 if (set->nr_maps == 1)
4339 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4340
Ming Lei59388702018-12-07 11:03:53 +08004341 if (set->ops->map_queues && !is_kdump_kernel()) {
Jens Axboeb3c661b2018-10-30 10:36:06 -06004342 int i;
4343
Ming Lei7d4901a2018-01-06 16:27:39 +08004344 /*
4345 * transport .map_queues is usually done in the following
4346 * way:
4347 *
4348 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4349 * mask = get_cpu_mask(queue)
4350 * for_each_cpu(cpu, mask)
Jens Axboeb3c661b2018-10-30 10:36:06 -06004351 * set->map[x].mq_map[cpu] = queue;
Ming Lei7d4901a2018-01-06 16:27:39 +08004352 * }
4353 *
4354 * When we need to remap, the table has to be cleared for
4355 * killing stale mapping since one CPU may not be mapped
4356 * to any hw queue.
4357 */
Jens Axboeb3c661b2018-10-30 10:36:06 -06004358 for (i = 0; i < set->nr_maps; i++)
4359 blk_mq_clear_mq_map(&set->map[i]);
Ming Lei7d4901a2018-01-06 16:27:39 +08004360
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004361 set->ops->map_queues(set);
Jens Axboeb3c661b2018-10-30 10:36:06 -06004362 } else {
4363 BUG_ON(set->nr_maps > 1);
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004364 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
Jens Axboeb3c661b2018-10-30 10:36:06 -06004365 }
Omar Sandovalebe8bdd2017-04-07 08:53:11 -06004366}
4367
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004368static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004369 int new_nr_hw_queues)
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004370{
4371 struct blk_mq_tags **new_tags;
4372
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004373 if (set->nr_hw_queues >= new_nr_hw_queues)
Shin'ichiro Kawasakid4b2e0d2022-11-22 17:49:17 +09004374 goto done;
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004375
4376 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4377 GFP_KERNEL, set->numa_node);
4378 if (!new_tags)
4379 return -ENOMEM;
4380
4381 if (set->tags)
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004382 memcpy(new_tags, set->tags, set->nr_hw_queues *
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004383 sizeof(*set->tags));
4384 kfree(set->tags);
4385 set->tags = new_tags;
Shin'ichiro Kawasakid4b2e0d2022-11-22 17:49:17 +09004386done:
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004387 set->nr_hw_queues = new_nr_hw_queues;
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004388 return 0;
4389}
4390
Jens Axboea4391c62014-06-05 15:21:56 -06004391/*
4392 * Alloc a tag set to be associated with one or more request queues.
4393 * May fail with EINVAL for various error conditions. May adjust the
Minwoo Imc018c842018-06-30 22:12:41 +09004394 * requested depth down, if it's too large. In that case, the set
Jens Axboea4391c62014-06-05 15:21:56 -06004395 * value will be stored in set->queue_depth.
4396 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004397int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4398{
Jens Axboeb3c661b2018-10-30 10:36:06 -06004399 int i, ret;
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004400
Bart Van Assche205fb5f52014-10-30 14:45:11 +01004401 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4402
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004403 if (!set->nr_hw_queues)
4404 return -EINVAL;
Jens Axboea4391c62014-06-05 15:21:56 -06004405 if (!set->queue_depth)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004406 return -EINVAL;
4407 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4408 return -EINVAL;
4409
Christoph Hellwig7d7e0f92016-09-14 16:18:54 +02004410 if (!set->ops->queue_rq)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004411 return -EINVAL;
4412
Ming Leide148292017-10-14 17:22:29 +08004413 if (!set->ops->get_budget ^ !set->ops->put_budget)
4414 return -EINVAL;
4415
Jens Axboea4391c62014-06-05 15:21:56 -06004416 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4417 pr_info("blk-mq: reduced tag depth to %u\n",
4418 BLK_MQ_MAX_DEPTH);
4419 set->queue_depth = BLK_MQ_MAX_DEPTH;
4420 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004421
Jens Axboeb3c661b2018-10-30 10:36:06 -06004422 if (!set->nr_maps)
4423 set->nr_maps = 1;
4424 else if (set->nr_maps > HCTX_MAX_TYPES)
4425 return -EINVAL;
4426
Shaohua Li6637fad2014-11-30 16:00:58 -08004427 /*
4428 * If a crashdump is active, then we are potentially in a very
4429 * memory constrained environment. Limit us to 1 queue and
4430 * 64 tags to prevent using too much memory.
4431 */
4432 if (is_kdump_kernel()) {
4433 set->nr_hw_queues = 1;
Ming Lei59388702018-12-07 11:03:53 +08004434 set->nr_maps = 1;
Shaohua Li6637fad2014-11-30 16:00:58 -08004435 set->queue_depth = min(64U, set->queue_depth);
4436 }
Keith Busch868f2f02015-12-17 17:08:14 -07004437 /*
Jens Axboe392546a2018-10-29 13:25:27 -06004438 * There is no use for more h/w queues than cpus if we just have
4439 * a single map
Keith Busch868f2f02015-12-17 17:08:14 -07004440 */
Jens Axboe392546a2018-10-29 13:25:27 -06004441 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
Keith Busch868f2f02015-12-17 17:08:14 -07004442 set->nr_hw_queues = nr_cpu_ids;
Shaohua Li6637fad2014-11-30 16:00:58 -08004443
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004444 if (set->flags & BLK_MQ_F_BLOCKING) {
4445 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4446 if (!set->srcu)
4447 return -ENOMEM;
4448 ret = init_srcu_struct(set->srcu);
4449 if (ret)
4450 goto out_free_srcu;
4451 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004452
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004453 ret = -ENOMEM;
Christoph Hellwig5ee20292022-11-09 11:08:10 +01004454 set->tags = kcalloc_node(set->nr_hw_queues,
4455 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4456 set->numa_node);
4457 if (!set->tags)
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004458 goto out_cleanup_srcu;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004459
Jens Axboeb3c661b2018-10-30 10:36:06 -06004460 for (i = 0; i < set->nr_maps; i++) {
4461 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
Ming Lei07b35eb2018-12-17 18:42:45 +08004462 sizeof(set->map[i].mq_map[0]),
Jens Axboeb3c661b2018-10-30 10:36:06 -06004463 GFP_KERNEL, set->numa_node);
4464 if (!set->map[i].mq_map)
4465 goto out_free_mq_map;
Ming Lei59388702018-12-07 11:03:53 +08004466 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
Jens Axboeb3c661b2018-10-30 10:36:06 -06004467 }
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004468
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004469 blk_mq_update_queue_map(set);
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004470
John Garry63064be2021-10-05 18:23:35 +08004471 ret = blk_mq_alloc_set_map_and_rqs(set);
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004472 if (ret)
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004473 goto out_free_mq_map;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004474
Jens Axboe0d2602c2014-05-13 15:10:52 -06004475 mutex_init(&set->tag_list_lock);
4476 INIT_LIST_HEAD(&set->tag_list);
4477
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004478 return 0;
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004479
4480out_free_mq_map:
Jens Axboeb3c661b2018-10-30 10:36:06 -06004481 for (i = 0; i < set->nr_maps; i++) {
4482 kfree(set->map[i].mq_map);
4483 set->map[i].mq_map = NULL;
4484 }
Robert Elliott5676e7b2014-09-02 11:38:44 -05004485 kfree(set->tags);
4486 set->tags = NULL;
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004487out_cleanup_srcu:
4488 if (set->flags & BLK_MQ_F_BLOCKING)
4489 cleanup_srcu_struct(set->srcu);
4490out_free_srcu:
4491 if (set->flags & BLK_MQ_F_BLOCKING)
4492 kfree(set->srcu);
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004493 return ret;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004494}
4495EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4496
Christoph Hellwigcdb14e02021-06-02 09:53:16 +03004497/* allocate and initialize a tagset for a simple single-queue device */
4498int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4499 const struct blk_mq_ops *ops, unsigned int queue_depth,
4500 unsigned int set_flags)
4501{
4502 memset(set, 0, sizeof(*set));
4503 set->ops = ops;
4504 set->nr_hw_queues = 1;
4505 set->nr_maps = 1;
4506 set->queue_depth = queue_depth;
4507 set->numa_node = NUMA_NO_NODE;
4508 set->flags = set_flags;
4509 return blk_mq_alloc_tag_set(set);
4510}
4511EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4512
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004513void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4514{
Jens Axboeb3c661b2018-10-30 10:36:06 -06004515 int i, j;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004516
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004517 for (i = 0; i < set->nr_hw_queues; i++)
John Garrye155b0c2021-10-05 18:23:37 +08004518 __blk_mq_free_map_and_rqs(set, i);
Jens Axboe484b4062014-05-21 14:01:15 -06004519
John Garry079a2e32021-10-05 18:23:39 +08004520 if (blk_mq_is_shared_tags(set->flags)) {
4521 blk_mq_free_map_and_rqs(set, set->shared_tags,
John Garrye155b0c2021-10-05 18:23:37 +08004522 BLK_MQ_NO_HCTX_IDX);
4523 }
John Garry32bc15a2020-08-19 23:20:24 +08004524
Jens Axboeb3c661b2018-10-30 10:36:06 -06004525 for (j = 0; j < set->nr_maps; j++) {
4526 kfree(set->map[j].mq_map);
4527 set->map[j].mq_map = NULL;
4528 }
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004529
Ming Lei981bd182014-04-24 00:07:34 +08004530 kfree(set->tags);
Robert Elliott5676e7b2014-09-02 11:38:44 -05004531 set->tags = NULL;
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004532 if (set->flags & BLK_MQ_F_BLOCKING) {
4533 cleanup_srcu_struct(set->srcu);
4534 kfree(set->srcu);
4535 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004536}
4537EXPORT_SYMBOL(blk_mq_free_tag_set);
4538
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004539int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4540{
4541 struct blk_mq_tag_set *set = q->tag_set;
4542 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08004543 int ret;
4544 unsigned long i;
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004545
Jens Axboebd166ef2017-01-17 06:03:22 -07004546 if (!set)
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004547 return -EINVAL;
4548
Aleksei Zakharove5fa8142019-02-08 19:14:05 +03004549 if (q->nr_requests == nr)
4550 return 0;
4551
Jens Axboe70f36b62017-01-19 10:59:07 -07004552 blk_mq_freeze_queue(q);
Ming Lei24f5a902018-01-06 16:27:38 +08004553 blk_mq_quiesce_queue(q);
Jens Axboe70f36b62017-01-19 10:59:07 -07004554
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004555 ret = 0;
4556 queue_for_each_hw_ctx(q, hctx, i) {
Keith Busche9137d42016-02-18 14:56:35 -07004557 if (!hctx->tags)
4558 continue;
Jens Axboebd166ef2017-01-17 06:03:22 -07004559 /*
4560 * If we're using an MQ scheduler, just update the scheduler
4561 * queue depth. This is similar to what the old code would do.
4562 */
John Garryf6adcef2021-10-05 18:23:29 +08004563 if (hctx->sched_tags) {
Jens Axboe70f36b62017-01-19 10:59:07 -07004564 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
John Garryf6adcef2021-10-05 18:23:29 +08004565 nr, true);
John Garryf6adcef2021-10-05 18:23:29 +08004566 } else {
4567 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4568 false);
Jens Axboe70f36b62017-01-19 10:59:07 -07004569 }
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004570 if (ret)
4571 break;
Jens Axboe77f1e0a2019-01-18 10:34:16 -07004572 if (q->elevator && q->elevator->type->ops.depth_updated)
4573 q->elevator->type->ops.depth_updated(hctx);
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004574 }
John Garryd97e5942021-05-13 20:00:58 +08004575 if (!ret) {
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004576 q->nr_requests = nr;
John Garry079a2e32021-10-05 18:23:39 +08004577 if (blk_mq_is_shared_tags(set->flags)) {
John Garry8fa04462021-10-05 18:23:28 +08004578 if (q->elevator)
John Garry079a2e32021-10-05 18:23:39 +08004579 blk_mq_tag_update_sched_shared_tags(q);
John Garry8fa04462021-10-05 18:23:28 +08004580 else
John Garry079a2e32021-10-05 18:23:39 +08004581 blk_mq_tag_resize_shared_tags(set, nr);
John Garry8fa04462021-10-05 18:23:28 +08004582 }
John Garryd97e5942021-05-13 20:00:58 +08004583 }
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004584
Ming Lei24f5a902018-01-06 16:27:38 +08004585 blk_mq_unquiesce_queue(q);
Jens Axboe70f36b62017-01-19 10:59:07 -07004586 blk_mq_unfreeze_queue(q);
Jens Axboe70f36b62017-01-19 10:59:07 -07004587
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004588 return ret;
4589}
4590
Jianchao Wangd48ece22018-08-21 15:15:03 +08004591/*
4592 * request_queue and elevator_type pair.
4593 * It is just used by __blk_mq_update_nr_hw_queues to cache
4594 * the elevator_type associated with a request_queue.
4595 */
4596struct blk_mq_qe_pair {
4597 struct list_head node;
4598 struct request_queue *q;
4599 struct elevator_type *type;
4600};
4601
4602/*
4603 * Cache the elevator_type in qe pair list and switch the
4604 * io scheduler to 'none'
4605 */
4606static bool blk_mq_elv_switch_none(struct list_head *head,
4607 struct request_queue *q)
4608{
4609 struct blk_mq_qe_pair *qe;
4610
4611 if (!q->elevator)
4612 return true;
4613
4614 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4615 if (!qe)
4616 return false;
4617
Ming Lei5fd7a842022-06-16 09:43:59 +08004618 /* q->elevator needs protection from ->sysfs_lock */
4619 mutex_lock(&q->sysfs_lock);
4620
Jianchao Wangd48ece22018-08-21 15:15:03 +08004621 INIT_LIST_HEAD(&qe->node);
4622 qe->q = q;
4623 qe->type = q->elevator->type;
Christoph Hellwigdd6f7f12022-10-20 08:48:16 +02004624 /* keep a reference to the elevator module as we'll switch back */
4625 __elevator_get(qe->type);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004626 list_add(&qe->node, head);
Christoph Hellwig64b36072022-10-30 11:07:14 +01004627 elevator_disable(q);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004628 mutex_unlock(&q->sysfs_lock);
4629
4630 return true;
4631}
4632
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004633static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4634 struct request_queue *q)
Jianchao Wangd48ece22018-08-21 15:15:03 +08004635{
4636 struct blk_mq_qe_pair *qe;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004637
4638 list_for_each_entry(qe, head, node)
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004639 if (qe->q == q)
4640 return qe;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004641
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004642 return NULL;
4643}
4644
4645static void blk_mq_elv_switch_back(struct list_head *head,
4646 struct request_queue *q)
4647{
4648 struct blk_mq_qe_pair *qe;
4649 struct elevator_type *t;
4650
4651 qe = blk_lookup_qe_pair(head, q);
4652 if (!qe)
Jianchao Wangd48ece22018-08-21 15:15:03 +08004653 return;
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004654 t = qe->type;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004655 list_del(&qe->node);
4656 kfree(qe);
4657
4658 mutex_lock(&q->sysfs_lock);
Keith Busch8237c012022-09-27 08:56:52 -07004659 elevator_switch(q, t);
Jinlong Chen8ed40ee2022-10-20 08:48:19 +02004660 /* drop the reference acquired in blk_mq_elv_switch_none */
4661 elevator_put(t);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004662 mutex_unlock(&q->sysfs_lock);
4663}
4664
Keith Busche4dc2b32017-05-30 14:39:11 -04004665static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4666 int nr_hw_queues)
Keith Busch868f2f02015-12-17 17:08:14 -07004667{
4668 struct request_queue *q;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004669 LIST_HEAD(head);
Jianchao Wange01ad462018-10-12 18:07:28 +08004670 int prev_nr_hw_queues;
Keith Busch868f2f02015-12-17 17:08:14 -07004671
Bart Van Assche705cda92017-04-07 11:16:49 -07004672 lockdep_assert_held(&set->tag_list_lock);
4673
Jens Axboe392546a2018-10-29 13:25:27 -06004674 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
Keith Busch868f2f02015-12-17 17:08:14 -07004675 nr_hw_queues = nr_cpu_ids;
Weiping Zhangfe35ec52020-06-17 14:18:37 +08004676 if (nr_hw_queues < 1)
4677 return;
4678 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
Keith Busch868f2f02015-12-17 17:08:14 -07004679 return;
4680
4681 list_for_each_entry(q, &set->tag_list, tag_set_list)
4682 blk_mq_freeze_queue(q);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004683 /*
4684 * Switch IO scheduler to 'none', cleaning up the data associated
4685 * with the previous scheduler. We will switch back once we are done
4686 * updating the new sw to hw queue mappings.
4687 */
4688 list_for_each_entry(q, &set->tag_list, tag_set_list)
4689 if (!blk_mq_elv_switch_none(&head, q))
4690 goto switch_back;
Keith Busch868f2f02015-12-17 17:08:14 -07004691
Jianchao Wang477e19d2018-10-12 18:07:25 +08004692 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4693 blk_mq_debugfs_unregister_hctxs(q);
Christoph Hellwigeaa870f2022-06-28 19:18:49 +02004694 blk_mq_sysfs_unregister_hctxs(q);
Jianchao Wang477e19d2018-10-12 18:07:25 +08004695 }
4696
Weiping Zhanga2584e42020-05-07 21:03:56 +08004697 prev_nr_hw_queues = set->nr_hw_queues;
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004698 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004699 goto reregister;
4700
Jianchao Wange01ad462018-10-12 18:07:28 +08004701fallback:
Weiping Zhangaa880ad2020-05-13 08:44:05 +08004702 blk_mq_update_queue_map(set);
Keith Busch868f2f02015-12-17 17:08:14 -07004703 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4704 blk_mq_realloc_hw_ctxs(set, q);
Ming Lei42ee3062022-03-08 15:32:16 +08004705 blk_mq_update_poll_flag(q);
Jianchao Wange01ad462018-10-12 18:07:28 +08004706 if (q->nr_hw_queues != set->nr_hw_queues) {
Ye Bina846a8e2021-11-08 15:40:19 +08004707 int i = prev_nr_hw_queues;
4708
Jianchao Wange01ad462018-10-12 18:07:28 +08004709 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4710 nr_hw_queues, prev_nr_hw_queues);
Ye Bina846a8e2021-11-08 15:40:19 +08004711 for (; i < set->nr_hw_queues; i++)
4712 __blk_mq_free_map_and_rqs(set, i);
4713
Jianchao Wange01ad462018-10-12 18:07:28 +08004714 set->nr_hw_queues = prev_nr_hw_queues;
Dongli Zhang7d76f852019-02-27 21:35:01 +08004715 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
Jianchao Wange01ad462018-10-12 18:07:28 +08004716 goto fallback;
4717 }
Jianchao Wang477e19d2018-10-12 18:07:25 +08004718 blk_mq_map_swqueue(q);
4719 }
4720
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004721reregister:
Jianchao Wang477e19d2018-10-12 18:07:25 +08004722 list_for_each_entry(q, &set->tag_list, tag_set_list) {
Christoph Hellwigeaa870f2022-06-28 19:18:49 +02004723 blk_mq_sysfs_register_hctxs(q);
Jianchao Wang477e19d2018-10-12 18:07:25 +08004724 blk_mq_debugfs_register_hctxs(q);
Keith Busch868f2f02015-12-17 17:08:14 -07004725 }
4726
Jianchao Wangd48ece22018-08-21 15:15:03 +08004727switch_back:
4728 list_for_each_entry(q, &set->tag_list, tag_set_list)
4729 blk_mq_elv_switch_back(&head, q);
4730
Keith Busch868f2f02015-12-17 17:08:14 -07004731 list_for_each_entry(q, &set->tag_list, tag_set_list)
4732 blk_mq_unfreeze_queue(q);
4733}
Keith Busche4dc2b32017-05-30 14:39:11 -04004734
4735void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4736{
4737 mutex_lock(&set->tag_list_lock);
4738 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4739 mutex_unlock(&set->tag_list_lock);
4740}
Keith Busch868f2f02015-12-17 17:08:14 -07004741EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4742
Keith Busch54bdd672023-03-20 12:49:26 -07004743int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4744 unsigned int flags)
Jens Axboebbd7bb72016-11-04 09:34:34 -06004745{
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004746 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4747 long state = get_current_state();
4748 int ret;
Jens Axboe1052b8a2018-11-26 08:21:49 -07004749
Jens Axboeaa61bec2018-11-13 21:32:10 -07004750 do {
Jens Axboe5a72e892021-10-12 09:24:29 -06004751 ret = q->mq_ops->poll(hctx, iob);
Jens Axboebbd7bb72016-11-04 09:34:34 -06004752 if (ret > 0) {
Jens Axboe849a3702018-11-16 08:37:34 -07004753 __set_current_state(TASK_RUNNING);
Jens Axboe85f4d4b2018-11-06 13:30:55 -07004754 return ret;
Jens Axboebbd7bb72016-11-04 09:34:34 -06004755 }
4756
4757 if (signal_pending_state(state, current))
Jens Axboe849a3702018-11-16 08:37:34 -07004758 __set_current_state(TASK_RUNNING);
Peter Zijlstrab03fbd42021-06-11 10:28:12 +02004759 if (task_is_running(current))
Jens Axboe85f4d4b2018-11-06 13:30:55 -07004760 return 1;
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004761
Christoph Hellwigef99b2d2021-10-12 13:12:19 +02004762 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
Jens Axboebbd7bb72016-11-04 09:34:34 -06004763 break;
4764 cpu_relax();
Jens Axboeaa61bec2018-11-13 21:32:10 -07004765 } while (!need_resched());
Jens Axboebbd7bb72016-11-04 09:34:34 -06004766
Nitesh Shetty67b41102018-02-13 21:18:12 +05304767 __set_current_state(TASK_RUNNING);
Jens Axboe85f4d4b2018-11-06 13:30:55 -07004768 return 0;
Jens Axboebbd7bb72016-11-04 09:34:34 -06004769}
Jens Axboebbd7bb72016-11-04 09:34:34 -06004770
Jens Axboe9cf2bab2018-10-31 17:01:22 -06004771unsigned int blk_mq_rq_cpu(struct request *rq)
4772{
4773 return rq->mq_ctx->cpu;
4774}
4775EXPORT_SYMBOL(blk_mq_rq_cpu);
4776
Ming Lei2a19b282021-11-16 09:43:43 +08004777void blk_mq_cancel_work_sync(struct request_queue *q)
4778{
Jinlong Chen219cf432022-10-30 17:47:30 +08004779 struct blk_mq_hw_ctx *hctx;
4780 unsigned long i;
Ming Lei2a19b282021-11-16 09:43:43 +08004781
Jinlong Chen219cf432022-10-30 17:47:30 +08004782 cancel_delayed_work_sync(&q->requeue_work);
Ming Lei2a19b282021-11-16 09:43:43 +08004783
Jinlong Chen219cf432022-10-30 17:47:30 +08004784 queue_for_each_hw_ctx(q, hctx, i)
4785 cancel_delayed_work_sync(&hctx->run_work);
Ming Lei2a19b282021-11-16 09:43:43 +08004786}
4787
Jens Axboe320ae512013-10-24 09:20:05 +01004788static int __init blk_mq_init(void)
4789{
Christoph Hellwigc3077b52020-06-11 08:44:41 +02004790 int i;
4791
4792 for_each_possible_cpu(i)
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01004793 init_llist_head(&per_cpu(blk_cpu_done, i));
Christoph Hellwigc3077b52020-06-11 08:44:41 +02004794 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4795
4796 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4797 "block/softirq:dead", NULL,
4798 blk_softirq_cpu_dead);
Thomas Gleixner9467f852016-09-22 08:05:17 -06004799 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4800 blk_mq_hctx_notify_dead);
Ming Leibf0beec2020-05-29 15:53:15 +02004801 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4802 blk_mq_hctx_notify_online,
4803 blk_mq_hctx_notify_offline);
Jens Axboe320ae512013-10-24 09:20:05 +01004804 return 0;
4805}
4806subsys_initcall(blk_mq_init);