blob: 2dc01551e27c7d1e50266e554fe4bb6d378a1482 [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"
Jens Axboe320ae512013-10-24 09:20:05 +010043
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +010044static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
Chengming Zhou660e8022023-07-17 12:00:55 +080045static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd);
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 Hellwig360f2642023-05-19 06:40:46 +020048static void blk_mq_request_bypass_insert(struct request *rq,
49 blk_insert_t flags);
Christoph Hellwig94aa2282023-04-13 08:40:41 +020050static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
51 struct list_head *list);
Keith Buschf6c80cf2023-06-12 12:03:42 -070052static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
53 struct io_comp_batch *iob, unsigned int flags);
Christoph Hellwig3e087732021-10-12 13:12:24 +020054
Jens Axboe320ae512013-10-24 09:20:05 +010055/*
Yufen Yu85fae292019-03-24 17:57:08 +080056 * Check if any of the ctx, dispatch list or elevator
57 * have pending work in this hardware queue.
Jens Axboe320ae512013-10-24 09:20:05 +010058 */
Jens Axboe79f720a2017-11-10 09:13:21 -070059static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
Jens Axboe320ae512013-10-24 09:20:05 +010060{
Jens Axboe79f720a2017-11-10 09:13:21 -070061 return !list_empty_careful(&hctx->dispatch) ||
62 sbitmap_any_bit_set(&hctx->ctx_map) ||
Jens Axboebd166ef2017-01-17 06:03:22 -070063 blk_mq_sched_has_work(hctx);
Jens Axboe320ae512013-10-24 09:20:05 +010064}
65
66/*
67 * Mark this ctx as having pending work in this hardware queue
68 */
69static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
70 struct blk_mq_ctx *ctx)
71{
Jens Axboef31967f2018-10-29 13:13:29 -060072 const int bit = ctx->index_hw[hctx->type];
73
74 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
75 sbitmap_set_bit(&hctx->ctx_map, bit);
Jens Axboe1429d7c2014-05-19 09:23:55 -060076}
77
78static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
79 struct blk_mq_ctx *ctx)
80{
Jens Axboef31967f2018-10-29 13:13:29 -060081 const int bit = ctx->index_hw[hctx->type];
82
83 sbitmap_clear_bit(&hctx->ctx_map, bit);
Jens Axboe320ae512013-10-24 09:20:05 +010084}
85
Jens Axboef299b7c2017-08-08 17:51:45 -060086struct mq_inflight {
Christoph Hellwig8446fe92020-11-24 09:36:54 +010087 struct block_device *part;
Pavel Begunkova2e80f62019-09-30 21:55:34 +030088 unsigned int inflight[2];
Jens Axboef299b7c2017-08-08 17:51:45 -060089};
90
John Garry2dd65322022-07-06 20:03:53 +080091static bool blk_mq_check_inflight(struct request *rq, void *priv)
Jens Axboef299b7c2017-08-08 17:51:45 -060092{
93 struct mq_inflight *mi = priv;
94
Haisu Wangb81c14c2022-05-30 14:40:59 +080095 if (rq->part && blk_do_io_stat(rq) &&
96 (!mi->part->bd_partno || rq->part == mi->part) &&
Jeffle Xub0d97552020-12-02 19:11:45 +080097 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
Pavel Begunkovbb4e6b12019-09-30 21:55:33 +030098 mi->inflight[rq_data_dir(rq)]++;
Jens Axboe7baa8572018-11-08 10:24:07 -070099
100 return true;
Jens Axboef299b7c2017-08-08 17:51:45 -0600101}
102
Christoph Hellwig8446fe92020-11-24 09:36:54 +0100103unsigned int blk_mq_in_flight(struct request_queue *q,
104 struct block_device *part)
Jens Axboef299b7c2017-08-08 17:51:45 -0600105{
Pavel Begunkova2e80f62019-09-30 21:55:34 +0300106 struct mq_inflight mi = { .part = part };
Jens Axboef299b7c2017-08-08 17:51:45 -0600107
Jens Axboef299b7c2017-08-08 17:51:45 -0600108 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
Mikulas Patockae016b782018-12-06 11:41:21 -0500109
Pavel Begunkova2e80f62019-09-30 21:55:34 +0300110 return mi.inflight[0] + mi.inflight[1];
Omar Sandovalbf0ddab2018-04-26 00:21:59 -0700111}
112
Christoph Hellwig8446fe92020-11-24 09:36:54 +0100113void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
114 unsigned int inflight[2])
Omar Sandovalbf0ddab2018-04-26 00:21:59 -0700115{
Pavel Begunkova2e80f62019-09-30 21:55:34 +0300116 struct mq_inflight mi = { .part = part };
Omar Sandovalbf0ddab2018-04-26 00:21:59 -0700117
Pavel Begunkovbb4e6b12019-09-30 21:55:33 +0300118 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
Pavel Begunkova2e80f62019-09-30 21:55:34 +0300119 inflight[0] = mi.inflight[0];
120 inflight[1] = mi.inflight[1];
Omar Sandovalbf0ddab2018-04-26 00:21:59 -0700121}
122
Ming Lei1671d522017-03-27 20:06:57 +0800123void blk_freeze_queue_start(struct request_queue *q)
Ming Lei43a5e4e2013-12-26 21:31:35 +0800124{
Bob Liu7996a8b2019-05-21 11:25:55 +0800125 mutex_lock(&q->mq_freeze_lock);
126 if (++q->mq_freeze_depth == 1) {
Dan Williams3ef28e82015-10-21 13:20:12 -0400127 percpu_ref_kill(&q->q_usage_counter);
Bob Liu7996a8b2019-05-21 11:25:55 +0800128 mutex_unlock(&q->mq_freeze_lock);
Jens Axboe344e9ff2018-11-15 12:22:51 -0700129 if (queue_is_mq(q))
Ming Lei055f6e12017-11-09 10:49:53 -0800130 blk_mq_run_hw_queues(q, false);
Bob Liu7996a8b2019-05-21 11:25:55 +0800131 } else {
132 mutex_unlock(&q->mq_freeze_lock);
Tejun Heocddd5d12014-08-16 08:02:24 -0400133 }
Tejun Heof3af0202014-11-04 13:52:27 -0500134}
Ming Lei1671d522017-03-27 20:06:57 +0800135EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
Tejun Heof3af0202014-11-04 13:52:27 -0500136
Keith Busch6bae363e2017-03-01 14:22:10 -0500137void blk_mq_freeze_queue_wait(struct request_queue *q)
Tejun Heof3af0202014-11-04 13:52:27 -0500138{
Dan Williams3ef28e82015-10-21 13:20:12 -0400139 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
Ming Lei43a5e4e2013-12-26 21:31:35 +0800140}
Keith Busch6bae363e2017-03-01 14:22:10 -0500141EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
Ming Lei43a5e4e2013-12-26 21:31:35 +0800142
Keith Buschf91328c2017-03-01 14:22:11 -0500143int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
144 unsigned long timeout)
145{
146 return wait_event_timeout(q->mq_freeze_wq,
147 percpu_ref_is_zero(&q->q_usage_counter),
148 timeout);
149}
150EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
Jens Axboe320ae512013-10-24 09:20:05 +0100151
Tejun Heof3af0202014-11-04 13:52:27 -0500152/*
153 * Guarantee no request is in use, so we can change any data structure of
154 * the queue afterward.
155 */
Dan Williams3ef28e82015-10-21 13:20:12 -0400156void blk_freeze_queue(struct request_queue *q)
Tejun Heof3af0202014-11-04 13:52:27 -0500157{
Dan Williams3ef28e82015-10-21 13:20:12 -0400158 /*
159 * In the !blk_mq case we are only calling this to kill the
160 * q_usage_counter, otherwise this increases the freeze depth
161 * and waits for it to return to zero. For this reason there is
162 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
163 * exported to drivers as the only user for unfreeze is blk_mq.
164 */
Ming Lei1671d522017-03-27 20:06:57 +0800165 blk_freeze_queue_start(q);
Tejun Heof3af0202014-11-04 13:52:27 -0500166 blk_mq_freeze_queue_wait(q);
167}
Dan Williams3ef28e82015-10-21 13:20:12 -0400168
169void blk_mq_freeze_queue(struct request_queue *q)
170{
171 /*
172 * ...just an alias to keep freeze and unfreeze actions balanced
173 * in the blk_mq_* namespace
174 */
175 blk_freeze_queue(q);
176}
Jens Axboec761d962015-01-02 15:05:12 -0700177EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
Tejun Heof3af0202014-11-04 13:52:27 -0500178
Christoph Hellwigaec89dc2021-09-29 09:12:41 +0200179void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
Jens Axboe320ae512013-10-24 09:20:05 +0100180{
Bob Liu7996a8b2019-05-21 11:25:55 +0800181 mutex_lock(&q->mq_freeze_lock);
Christoph Hellwigaec89dc2021-09-29 09:12:41 +0200182 if (force_atomic)
183 q->q_usage_counter.data->force_atomic = true;
Bob Liu7996a8b2019-05-21 11:25:55 +0800184 q->mq_freeze_depth--;
185 WARN_ON_ONCE(q->mq_freeze_depth < 0);
186 if (!q->mq_freeze_depth) {
Bart Van Asschebdd63162018-09-26 14:01:08 -0700187 percpu_ref_resurrect(&q->q_usage_counter);
Jens Axboe320ae512013-10-24 09:20:05 +0100188 wake_up_all(&q->mq_freeze_wq);
Tejun Heoadd703f2014-07-01 10:34:38 -0600189 }
Bob Liu7996a8b2019-05-21 11:25:55 +0800190 mutex_unlock(&q->mq_freeze_lock);
Jens Axboe320ae512013-10-24 09:20:05 +0100191}
Christoph Hellwigaec89dc2021-09-29 09:12:41 +0200192
193void blk_mq_unfreeze_queue(struct request_queue *q)
194{
195 __blk_mq_unfreeze_queue(q, false);
196}
Keith Buschb4c6a022014-12-19 17:54:14 -0700197EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
Jens Axboe320ae512013-10-24 09:20:05 +0100198
Bart Van Assche852ec802017-06-21 10:55:47 -0700199/*
200 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
201 * mpt3sas driver such that this function can be removed.
202 */
203void blk_mq_quiesce_queue_nowait(struct request_queue *q)
204{
Ming Leie70feb82021-10-14 16:17:10 +0800205 unsigned long flags;
206
207 spin_lock_irqsave(&q->queue_lock, flags);
208 if (!q->quiesce_depth++)
209 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
210 spin_unlock_irqrestore(&q->queue_lock, flags);
Bart Van Assche852ec802017-06-21 10:55:47 -0700211}
212EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
213
Bart Van Assche6a83e742016-11-02 10:09:51 -0600214/**
Ming Lei9ef4d022021-11-09 15:11:41 +0800215 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
Christoph Hellwig483239c2022-11-01 16:00:48 +0100216 * @set: tag_set to wait on
Ming Lei9ef4d022021-11-09 15:11:41 +0800217 *
218 * Note: it is driver's responsibility for making sure that quiesce has
Christoph Hellwig483239c2022-11-01 16:00:48 +0100219 * been started on or more of the request_queues of the tag_set. This
220 * function only waits for the quiesce on those request_queues that had
221 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
Ming Lei9ef4d022021-11-09 15:11:41 +0800222 */
Christoph Hellwig483239c2022-11-01 16:00:48 +0100223void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
Ming Lei9ef4d022021-11-09 15:11:41 +0800224{
Christoph Hellwig483239c2022-11-01 16:00:48 +0100225 if (set->flags & BLK_MQ_F_BLOCKING)
226 synchronize_srcu(set->srcu);
Ming Lei704b9142021-12-03 21:15:32 +0800227 else
Ming Lei9ef4d022021-11-09 15:11:41 +0800228 synchronize_rcu();
229}
230EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
231
232/**
Ming Lei69e07c42017-06-06 23:22:07 +0800233 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
Bart Van Assche6a83e742016-11-02 10:09:51 -0600234 * @q: request queue.
235 *
236 * Note: this function does not prevent that the struct request end_io()
Ming Lei69e07c42017-06-06 23:22:07 +0800237 * callback function is invoked. Once this function is returned, we make
238 * sure no dispatch can happen until the queue is unquiesced via
239 * blk_mq_unquiesce_queue().
Bart Van Assche6a83e742016-11-02 10:09:51 -0600240 */
241void blk_mq_quiesce_queue(struct request_queue *q)
242{
Ming Lei1d9e9bc2017-06-06 23:22:08 +0800243 blk_mq_quiesce_queue_nowait(q);
Christoph Hellwig85373802022-11-01 16:00:46 +0100244 /* nothing to wait for non-mq queues */
245 if (queue_is_mq(q))
Christoph Hellwig483239c2022-11-01 16:00:48 +0100246 blk_mq_wait_quiesce_done(q->tag_set);
Bart Van Assche6a83e742016-11-02 10:09:51 -0600247}
248EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
249
Ming Leie4e73912017-06-06 23:22:03 +0800250/*
251 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
252 * @q: request queue.
253 *
254 * This function recovers queue into the state before quiescing
255 * which is done by blk_mq_quiesce_queue.
256 */
257void blk_mq_unquiesce_queue(struct request_queue *q)
258{
Ming Leie70feb82021-10-14 16:17:10 +0800259 unsigned long flags;
260 bool run_queue = false;
261
262 spin_lock_irqsave(&q->queue_lock, flags);
263 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
264 ;
265 } else if (!--q->quiesce_depth) {
266 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
267 run_queue = true;
268 }
269 spin_unlock_irqrestore(&q->queue_lock, flags);
Ming Leif4560ff2017-06-18 14:24:27 -0600270
Ming Lei1d9e9bc2017-06-06 23:22:08 +0800271 /* dispatch requests which are inserted during quiescing */
Ming Leie70feb82021-10-14 16:17:10 +0800272 if (run_queue)
273 blk_mq_run_hw_queues(q, true);
Ming Leie4e73912017-06-06 23:22:03 +0800274}
275EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
276
Chao Leng414dd482022-11-01 16:00:49 +0100277void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
278{
279 struct request_queue *q;
280
281 mutex_lock(&set->tag_list_lock);
282 list_for_each_entry(q, &set->tag_list, tag_set_list) {
283 if (!blk_queue_skip_tagset_quiesce(q))
284 blk_mq_quiesce_queue_nowait(q);
285 }
286 blk_mq_wait_quiesce_done(set);
287 mutex_unlock(&set->tag_list_lock);
288}
289EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
290
291void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
292{
293 struct request_queue *q;
294
295 mutex_lock(&set->tag_list_lock);
296 list_for_each_entry(q, &set->tag_list, tag_set_list) {
297 if (!blk_queue_skip_tagset_quiesce(q))
298 blk_mq_unquiesce_queue(q);
299 }
300 mutex_unlock(&set->tag_list_lock);
301}
302EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
303
Jens Axboeaed3ea92014-12-22 14:04:42 -0700304void blk_mq_wake_waiters(struct request_queue *q)
305{
306 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +0800307 unsigned long i;
Jens Axboeaed3ea92014-12-22 14:04:42 -0700308
309 queue_for_each_hw_ctx(q, hctx, i)
310 if (blk_mq_hw_queue_mapped(hctx))
311 blk_mq_tag_wakeup_all(hctx->tags, true);
312}
313
Christoph Hellwig52fdbbc2021-11-17 07:13:59 +0100314void blk_rq_init(struct request_queue *q, struct request *rq)
315{
316 memset(rq, 0, sizeof(*rq));
317
318 INIT_LIST_HEAD(&rq->queuelist);
319 rq->q = q;
320 rq->__sector = (sector_t) -1;
321 INIT_HLIST_NODE(&rq->hash);
322 RB_CLEAR_NODE(&rq->rb_node);
323 rq->tag = BLK_MQ_NO_TAG;
324 rq->internal_tag = BLK_MQ_NO_TAG;
325 rq->start_time_ns = ktime_get_ns();
326 rq->part = NULL;
327 blk_crypto_rq_set_defaults(rq);
328}
329EXPORT_SYMBOL(blk_rq_init);
330
Chengming Zhou5c17f452023-07-10 18:55:16 +0800331/* Set start and alloc time when the allocated request is actually used */
332static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
333{
334 if (blk_mq_need_time_stamp(rq))
335 rq->start_time_ns = ktime_get_ns();
336 else
337 rq->start_time_ns = 0;
338
339#ifdef CONFIG_BLK_RQ_ALLOC_TIME
340 if (blk_queue_rq_alloc_time(rq->q))
341 rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;
342 else
343 rq->alloc_time_ns = 0;
344#endif
345}
346
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200347static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
Chengming Zhou5c17f452023-07-10 18:55:16 +0800348 struct blk_mq_tags *tags, unsigned int tag)
Jens Axboe320ae512013-10-24 09:20:05 +0100349{
Pavel Begunkov605f7842021-10-18 21:37:28 +0100350 struct blk_mq_ctx *ctx = data->ctx;
351 struct blk_mq_hw_ctx *hctx = data->hctx;
352 struct request_queue *q = data->q;
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200353 struct request *rq = tags->static_rqs[tag];
Bart Van Asschec3a148d2017-06-20 11:15:43 -0700354
Jens Axboec7b84d42021-10-19 09:33:00 -0600355 rq->q = q;
356 rq->mq_ctx = ctx;
357 rq->mq_hctx = hctx;
358 rq->cmd_flags = data->cmd_flags;
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200359
Pavel Begunkov12845902021-10-18 21:37:29 +0100360 if (data->flags & BLK_MQ_REQ_PM)
Jens Axboe56f8da62021-10-19 09:32:57 -0600361 data->rq_flags |= RQF_PM;
Pavel Begunkov12845902021-10-18 21:37:29 +0100362 if (blk_queue_io_stat(q))
Jens Axboe56f8da62021-10-19 09:32:57 -0600363 data->rq_flags |= RQF_IO_STAT;
364 rq->rq_flags = data->rq_flags;
Pavel Begunkov12845902021-10-18 21:37:29 +0100365
Christoph Hellwigdd6216b2023-05-18 07:31:01 +0200366 if (data->rq_flags & RQF_SCHED_TAGS) {
Jens Axboec7b84d42021-10-19 09:33:00 -0600367 rq->tag = BLK_MQ_NO_TAG;
368 rq->internal_tag = tag;
Christoph Hellwigdd6216b2023-05-18 07:31:01 +0200369 } else {
370 rq->tag = tag;
371 rq->internal_tag = BLK_MQ_NO_TAG;
Jens Axboe320ae512013-10-24 09:20:05 +0100372 }
Jens Axboec7b84d42021-10-19 09:33:00 -0600373 rq->timeout = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100374
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200375 rq->part = NULL;
Omar Sandoval544ccc8d2018-05-09 02:08:50 -0700376 rq->io_start_time_ns = 0;
Hou Tao3d244302019-05-21 15:59:03 +0800377 rq->stats_sectors = 0;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200378 rq->nr_phys_segments = 0;
379#if defined(CONFIG_BLK_DEV_INTEGRITY)
380 rq->nr_integrity_segments = 0;
381#endif
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200382 rq->end_io = NULL;
383 rq->end_io_data = NULL;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200384
Pavel Begunkov4f266f22021-10-18 21:37:27 +0100385 blk_crypto_rq_set_defaults(rq);
386 INIT_LIST_HEAD(&rq->queuelist);
387 /* tag was already set */
388 WRITE_ONCE(rq->deadline, 0);
Jens Axboe0a467d02021-10-14 14:39:59 -0600389 req_ref_set(rq, 1);
Christoph Hellwig7ea4d8a2020-05-29 15:53:10 +0200390
Christoph Hellwigdd6216b2023-05-18 07:31:01 +0200391 if (rq->rq_flags & RQF_USE_SCHED) {
Christoph Hellwig7ea4d8a2020-05-29 15:53:10 +0200392 struct elevator_queue *e = data->q->elevator;
393
Pavel Begunkov4f266f22021-10-18 21:37:27 +0100394 INIT_HLIST_NODE(&rq->hash);
395 RB_CLEAR_NODE(&rq->rb_node);
396
Christoph Hellwigdd6216b2023-05-18 07:31:01 +0200397 if (e->type->ops.prepare_request)
Christoph Hellwig7ea4d8a2020-05-29 15:53:10 +0200398 e->type->ops.prepare_request(rq);
Christoph Hellwig7ea4d8a2020-05-29 15:53:10 +0200399 }
400
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200401 return rq;
Jens Axboe320ae512013-10-24 09:20:05 +0100402}
403
Jens Axboe349302d2021-10-09 13:10:39 -0600404static inline struct request *
Chengming Zhou5c17f452023-07-10 18:55:16 +0800405__blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
Jens Axboe349302d2021-10-09 13:10:39 -0600406{
407 unsigned int tag, tag_offset;
Jens Axboefe6134f2021-10-19 09:32:58 -0600408 struct blk_mq_tags *tags;
Jens Axboe349302d2021-10-09 13:10:39 -0600409 struct request *rq;
Jens Axboefe6134f2021-10-19 09:32:58 -0600410 unsigned long tag_mask;
Jens Axboe349302d2021-10-09 13:10:39 -0600411 int i, nr = 0;
412
Jens Axboefe6134f2021-10-19 09:32:58 -0600413 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
414 if (unlikely(!tag_mask))
Jens Axboe349302d2021-10-09 13:10:39 -0600415 return NULL;
416
Jens Axboefe6134f2021-10-19 09:32:58 -0600417 tags = blk_mq_tags_from_data(data);
418 for (i = 0; tag_mask; i++) {
419 if (!(tag_mask & (1UL << i)))
Jens Axboe349302d2021-10-09 13:10:39 -0600420 continue;
421 tag = tag_offset + i;
Jens Axboea22c00b2021-11-01 06:56:09 -0600422 prefetch(tags->static_rqs[tag]);
Jens Axboefe6134f2021-10-19 09:32:58 -0600423 tag_mask &= ~(1UL << i);
Chengming Zhou5c17f452023-07-10 18:55:16 +0800424 rq = blk_mq_rq_ctx_init(data, tags, tag);
Jens Axboe013a7f92021-10-13 07:58:52 -0600425 rq_list_add(data->cached_rq, rq);
Jens Axboec5fc7b92021-11-03 05:49:07 -0600426 nr++;
Jens Axboe349302d2021-10-09 13:10:39 -0600427 }
Chengming Zhoub8643d62023-09-13 15:16:12 +0000428 if (!(data->rq_flags & RQF_SCHED_TAGS))
429 blk_mq_add_active_requests(data->hctx, nr);
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) {
Christoph Hellwigdd6216b2023-05-18 07:31:01 +0200452 /*
453 * All requests use scheduler tags when an I/O scheduler is
454 * enabled for the queue.
455 */
456 data->rq_flags |= RQF_SCHED_TAGS;
Jens Axboe781dd832021-11-02 08:34:09 -0600457
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200458 /*
Lin Feng8d663f32021-04-15 11:39:20 +0800459 * Flush/passthrough requests are special and go directly to the
Christoph Hellwigdd6216b2023-05-18 07:31:01 +0200460 * dispatch list.
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200461 */
Bart Van Asschebe4c4272023-05-19 06:40:47 +0200462 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
Christoph Hellwigdd6216b2023-05-18 07:31:01 +0200463 !blk_op_is_passthrough(data->cmd_flags)) {
464 struct elevator_mq_ops *ops = &q->elevator->type->ops;
465
466 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
467
468 data->rq_flags |= RQF_USE_SCHED;
469 if (ops->limit_depth)
470 ops->limit_depth(data->cmd_flags, data);
471 }
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200472 }
473
Ming Leibf0beec2020-05-29 15:53:15 +0200474retry:
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200475 data->ctx = blk_mq_get_ctx(q);
476 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
Christoph Hellwigdd6216b2023-05-18 07:31:01 +0200477 if (!(data->rq_flags & RQF_SCHED_TAGS))
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200478 blk_mq_tag_busy(data->hctx);
479
John Garry99e48cd2022-07-06 20:03:50 +0800480 if (data->flags & BLK_MQ_REQ_RESERVED)
481 data->rq_flags |= RQF_RESV;
482
Ming Leibf0beec2020-05-29 15:53:15 +0200483 /*
Jens Axboe349302d2021-10-09 13:10:39 -0600484 * Try batched alloc if we want more than 1 tag.
485 */
486 if (data->nr_tags > 1) {
Chengming Zhou5c17f452023-07-10 18:55:16 +0800487 rq = __blk_mq_alloc_requests_batch(data);
488 if (rq) {
489 blk_mq_rq_time_init(rq, alloc_time_ns);
Jens Axboe349302d2021-10-09 13:10:39 -0600490 return rq;
Chengming Zhou5c17f452023-07-10 18:55:16 +0800491 }
Jens Axboe349302d2021-10-09 13:10:39 -0600492 data->nr_tags = 1;
493 }
494
495 /*
Ming Leibf0beec2020-05-29 15:53:15 +0200496 * Waiting allocations only fail because of an inactive hctx. In that
497 * case just retry the hctx assignment and tag allocation as CPU hotplug
498 * should have migrated us to an online CPU by now.
499 */
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200500 tag = blk_mq_get_tag(data);
Ming Leibf0beec2020-05-29 15:53:15 +0200501 if (tag == BLK_MQ_NO_TAG) {
502 if (data->flags & BLK_MQ_REQ_NOWAIT)
503 return NULL;
Ming Leibf0beec2020-05-29 15:53:15 +0200504 /*
Jens Axboe349302d2021-10-09 13:10:39 -0600505 * Give up the CPU and sleep for a random short time to
506 * ensure that thread using a realtime scheduling class
507 * are migrated off the CPU, and thus off the hctx that
508 * is going away.
Ming Leibf0beec2020-05-29 15:53:15 +0200509 */
510 msleep(3);
511 goto retry;
512 }
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200513
Chengming Zhoub8643d62023-09-13 15:16:12 +0000514 if (!(data->rq_flags & RQF_SCHED_TAGS))
515 blk_mq_inc_active_requests(data->hctx);
Chengming Zhou5c17f452023-07-10 18:55:16 +0800516 rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
517 blk_mq_rq_time_init(rq, alloc_time_ns);
518 return rq;
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200519}
520
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600521static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
522 struct blk_plug *plug,
523 blk_opf_t opf,
524 blk_mq_req_flags_t flags)
Jens Axboe320ae512013-10-24 09:20:05 +0100525{
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200526 struct blk_mq_alloc_data data = {
527 .q = q,
528 .flags = flags,
Bart Van Assche16458cf2022-07-14 11:06:32 -0700529 .cmd_flags = opf,
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600530 .nr_tags = plug->nr_ios,
531 .cached_rq = &plug->cached_rq,
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200532 };
Jens Axboebd166ef2017-01-17 06:03:22 -0700533 struct request *rq;
Jens Axboe320ae512013-10-24 09:20:05 +0100534
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600535 if (blk_queue_enter(q, flags))
536 return NULL;
537
538 plug->nr_ios = 1;
Jens Axboe320ae512013-10-24 09:20:05 +0100539
Christoph Hellwigb90cfae2021-10-12 12:40:44 +0200540 rq = __blk_mq_alloc_requests(&data);
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600541 if (unlikely(!rq))
542 blk_queue_exit(q);
543 return rq;
544}
545
546static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
547 blk_opf_t opf,
548 blk_mq_req_flags_t flags)
549{
550 struct blk_plug *plug = current->plug;
551 struct request *rq;
552
553 if (!plug)
554 return NULL;
Jinlong Chen40467282022-11-02 10:52:30 +0800555
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600556 if (rq_list_empty(plug->cached_rq)) {
557 if (plug->nr_ios == 1)
558 return NULL;
559 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
Jinlong Chen40467282022-11-02 10:52:30 +0800560 if (!rq)
561 return NULL;
562 } else {
563 rq = rq_list_peek(&plug->cached_rq);
564 if (!rq || rq->q != q)
565 return NULL;
566
567 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
568 return NULL;
569 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
570 return NULL;
571
572 plug->cached_rq = rq_list_next(rq);
Chengming Zhou5c17f452023-07-10 18:55:16 +0800573 blk_mq_rq_time_init(rq, 0);
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600574 }
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600575
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600576 rq->cmd_flags = opf;
577 INIT_LIST_HEAD(&rq->queuelist);
578 return rq;
579}
580
581struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
582 blk_mq_req_flags_t flags)
583{
584 struct request *rq;
585
586 rq = blk_mq_alloc_cached_request(q, opf, flags);
587 if (!rq) {
588 struct blk_mq_alloc_data data = {
589 .q = q,
590 .flags = flags,
591 .cmd_flags = opf,
592 .nr_tags = 1,
593 };
594 int ret;
595
596 ret = blk_queue_enter(q, flags);
597 if (ret)
598 return ERR_PTR(ret);
599
600 rq = __blk_mq_alloc_requests(&data);
601 if (!rq)
602 goto out_queue_exit;
603 }
Christoph Hellwig0c4de0f2016-07-19 11:31:50 +0200604 rq->__data_len = 0;
605 rq->__sector = (sector_t) -1;
606 rq->bio = rq->biotail = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +0100607 return rq;
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200608out_queue_exit:
609 blk_queue_exit(q);
610 return ERR_PTR(-EWOULDBLOCK);
Jens Axboe320ae512013-10-24 09:20:05 +0100611}
Jens Axboe4bb659b2014-05-09 09:36:49 -0600612EXPORT_SYMBOL(blk_mq_alloc_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100613
Bart Van Asschecd6ce142017-06-20 11:15:39 -0700614struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
Bart Van Assche16458cf2022-07-14 11:06:32 -0700615 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
Ming Lin1f5bd332016-06-13 16:45:21 +0200616{
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200617 struct blk_mq_alloc_data data = {
618 .q = q,
619 .flags = flags,
Bart Van Assche16458cf2022-07-14 11:06:32 -0700620 .cmd_flags = opf,
Jens Axboe47c122e2021-10-06 06:34:11 -0600621 .nr_tags = 1,
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200622 };
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200623 u64 alloc_time_ns = 0;
John Garrye3c5a782022-10-26 18:35:13 +0800624 struct request *rq;
Omar Sandoval6d2809d2017-02-27 10:28:27 -0800625 unsigned int cpu;
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200626 unsigned int tag;
Ming Lin1f5bd332016-06-13 16:45:21 +0200627 int ret;
628
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200629 /* alloc_time includes depth and tag waits */
630 if (blk_queue_rq_alloc_time(q))
631 alloc_time_ns = ktime_get_ns();
632
Ming Lin1f5bd332016-06-13 16:45:21 +0200633 /*
634 * If the tag allocator sleeps we could get an allocation for a
635 * different hardware context. No need to complicate the low level
636 * allocator for this for the rare use case of a command tied to
637 * a specific queue.
638 */
Kemeng Shi6ee858a2023-01-18 17:37:13 +0800639 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
640 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
Ming Lin1f5bd332016-06-13 16:45:21 +0200641 return ERR_PTR(-EINVAL);
642
643 if (hctx_idx >= q->nr_hw_queues)
644 return ERR_PTR(-EIO);
645
Bart Van Assche3a0a5292017-11-09 10:49:58 -0800646 ret = blk_queue_enter(q, flags);
Ming Lin1f5bd332016-06-13 16:45:21 +0200647 if (ret)
648 return ERR_PTR(ret);
649
Christoph Hellwigc8712c62016-09-23 10:25:48 -0600650 /*
651 * Check if the hardware context is actually mapped to anything.
652 * If not tell the caller that it should skip this queue.
653 */
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200654 ret = -EXDEV;
Ming Lei4e5cc992022-03-08 15:32:19 +0800655 data.hctx = xa_load(&q->hctx_table, hctx_idx);
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200656 if (!blk_mq_hw_queue_mapped(data.hctx))
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200657 goto out_queue_exit;
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200658 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
Bart Van Assche14dc7a12022-06-15 14:00:04 -0700659 if (cpu >= nr_cpu_ids)
660 goto out_queue_exit;
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200661 data.ctx = __blk_mq_get_ctx(q, cpu);
Ming Lin1f5bd332016-06-13 16:45:21 +0200662
Christoph Hellwigdd6216b2023-05-18 07:31:01 +0200663 if (q->elevator)
664 data.rq_flags |= RQF_SCHED_TAGS;
Jens Axboe781dd832021-11-02 08:34:09 -0600665 else
Christoph Hellwigdd6216b2023-05-18 07:31:01 +0200666 blk_mq_tag_busy(data.hctx);
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200667
John Garry99e48cd2022-07-06 20:03:50 +0800668 if (flags & BLK_MQ_REQ_RESERVED)
669 data.rq_flags |= RQF_RESV;
670
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200671 ret = -EWOULDBLOCK;
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200672 tag = blk_mq_get_tag(&data);
673 if (tag == BLK_MQ_NO_TAG)
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200674 goto out_queue_exit;
Chengming Zhoub8643d62023-09-13 15:16:12 +0000675 if (!(data.rq_flags & RQF_SCHED_TAGS))
676 blk_mq_inc_active_requests(data.hctx);
Chengming Zhou5c17f452023-07-10 18:55:16 +0800677 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
678 blk_mq_rq_time_init(rq, alloc_time_ns);
John Garrye3c5a782022-10-26 18:35:13 +0800679 rq->__data_len = 0;
680 rq->__sector = (sector_t) -1;
681 rq->bio = rq->biotail = NULL;
682 return rq;
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200683
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200684out_queue_exit:
685 blk_queue_exit(q);
686 return ERR_PTR(ret);
Ming Lin1f5bd332016-06-13 16:45:21 +0200687}
688EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
689
Chengming Zhoue5c0ca12023-08-13 23:23:25 +0800690static void blk_mq_finish_request(struct request *rq)
691{
692 struct request_queue *q = rq->q;
693
694 if (rq->rq_flags & RQF_USE_SCHED) {
695 q->elevator->type->ops.finish_request(rq);
696 /*
697 * For postflush request that may need to be
698 * completed twice, we should clear this flag
699 * to avoid double finish_request() on the rq.
700 */
701 rq->rq_flags &= ~RQF_USE_SCHED;
702 }
703}
704
Keith Busch12f5b932018-05-29 15:52:28 +0200705static void __blk_mq_free_request(struct request *rq)
706{
707 struct request_queue *q = rq->q;
708 struct blk_mq_ctx *ctx = rq->mq_ctx;
Jens Axboeea4f9952018-10-29 15:06:13 -0600709 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Keith Busch12f5b932018-05-29 15:52:28 +0200710 const int sched_tag = rq->internal_tag;
711
Satya Tangiralaa892c8d2020-05-14 00:37:18 +0000712 blk_crypto_free_request(rq);
Bart Van Assche986d4132018-09-26 14:01:10 -0700713 blk_pm_mark_last_busy(rq);
Jens Axboeea4f9952018-10-29 15:06:13 -0600714 rq->mq_hctx = NULL;
Tian Landdad59332023-05-13 18:12:27 -0400715
Chengming Zhoub8643d62023-09-13 15:16:12 +0000716 if (rq->tag != BLK_MQ_NO_TAG) {
717 blk_mq_dec_active_requests(hctx);
John Garrycae740a2020-02-26 20:10:15 +0800718 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
Chengming Zhoub8643d62023-09-13 15:16:12 +0000719 }
Christoph Hellwig766473682020-05-29 15:53:12 +0200720 if (sched_tag != BLK_MQ_NO_TAG)
John Garrycae740a2020-02-26 20:10:15 +0800721 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
Keith Busch12f5b932018-05-29 15:52:28 +0200722 blk_mq_sched_restart(hctx);
723 blk_queue_exit(q);
724}
725
Christoph Hellwig6af54052017-06-16 18:15:22 +0200726void blk_mq_free_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100727{
Jens Axboe320ae512013-10-24 09:20:05 +0100728 struct request_queue *q = rq->q;
729
Chengming Zhoue5c0ca12023-08-13 23:23:25 +0800730 blk_mq_finish_request(rq);
Christoph Hellwig6af54052017-06-16 18:15:22 +0200731
Jens Axboe7beb2f82017-09-30 02:08:24 -0600732 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
Christoph Hellwigd152c682021-08-16 15:46:24 +0200733 laptop_io_completion(q->disk->bdi);
Jens Axboe7beb2f82017-09-30 02:08:24 -0600734
Josef Bacika7905042018-07-03 09:32:35 -0600735 rq_qos_done(q, rq);
Jens Axboe0d2602c2014-05-13 15:10:52 -0600736
Keith Busch12f5b932018-05-29 15:52:28 +0200737 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Jens Axboe0a467d02021-10-14 14:39:59 -0600738 if (req_ref_put_and_test(rq))
Keith Busch12f5b932018-05-29 15:52:28 +0200739 __blk_mq_free_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100740}
Jens Axboe1a3b5952014-11-17 10:40:48 -0700741EXPORT_SYMBOL_GPL(blk_mq_free_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100742
Jens Axboe47c122e2021-10-06 06:34:11 -0600743void blk_mq_free_plug_rqs(struct blk_plug *plug)
Jens Axboe320ae512013-10-24 09:20:05 +0100744{
Jens Axboe013a7f92021-10-13 07:58:52 -0600745 struct request *rq;
Jens Axboefe1f4522018-11-28 10:50:07 -0700746
Jens Axboec5fc7b92021-11-03 05:49:07 -0600747 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
Jens Axboe47c122e2021-10-06 06:34:11 -0600748 blk_mq_free_request(rq);
Jens Axboe47c122e2021-10-06 06:34:11 -0600749}
Omar Sandoval522a7772018-05-09 02:08:53 -0700750
Christoph Hellwig22350ad2021-11-17 07:14:02 +0100751void blk_dump_rq_flags(struct request *rq, char *msg)
752{
753 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
Christoph Hellwigf3fa33a2021-11-26 13:18:00 +0100754 rq->q->disk ? rq->q->disk->disk_name : "?",
Bart Van Assche16458cf2022-07-14 11:06:32 -0700755 (__force unsigned long long) rq->cmd_flags);
Christoph Hellwig22350ad2021-11-17 07:14:02 +0100756
757 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
758 (unsigned long long)blk_rq_pos(rq),
759 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
760 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
761 rq->bio, rq->biotail, blk_rq_bytes(rq));
762}
763EXPORT_SYMBOL(blk_dump_rq_flags);
764
Jens Axboe9be3e062021-10-14 09:17:01 -0600765static void req_bio_endio(struct request *rq, struct bio *bio,
766 unsigned int nbytes, blk_status_t error)
767{
Pavel Begunkov478eb722021-10-19 22:24:12 +0100768 if (unlikely(error)) {
Jens Axboe9be3e062021-10-14 09:17:01 -0600769 bio->bi_status = error;
Pavel Begunkov478eb722021-10-19 22:24:12 +0100770 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
Jens Axboe9be3e062021-10-14 09:17:01 -0600771 /*
772 * Partial zone append completions cannot be supported as the
773 * BIO fragments may end up not being written sequentially.
Damien Le Moal748dc0b2024-01-10 18:29:42 +0900774 * For such case, force the completed nbytes to be equal to
775 * the BIO size so that bio_advance() sets the BIO remaining
776 * size to 0 and we end up calling bio_endio() before returning.
Jens Axboe9be3e062021-10-14 09:17:01 -0600777 */
Damien Le Moal748dc0b2024-01-10 18:29:42 +0900778 if (bio->bi_iter.bi_size != nbytes) {
Jens Axboe9be3e062021-10-14 09:17:01 -0600779 bio->bi_status = BLK_STS_IOERR;
Damien Le Moal748dc0b2024-01-10 18:29:42 +0900780 nbytes = bio->bi_iter.bi_size;
781 } else {
Jens Axboe9be3e062021-10-14 09:17:01 -0600782 bio->bi_iter.bi_sector = rq->__sector;
Damien Le Moal748dc0b2024-01-10 18:29:42 +0900783 }
Jens Axboe9be3e062021-10-14 09:17:01 -0600784 }
785
Pavel Begunkov478eb722021-10-19 22:24:12 +0100786 bio_advance(bio, nbytes);
787
788 if (unlikely(rq->rq_flags & RQF_QUIET))
789 bio_set_flag(bio, BIO_QUIET);
Jens Axboe9be3e062021-10-14 09:17:01 -0600790 /* don't actually finish bio if it's part of flush sequence */
791 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
792 bio_endio(bio);
793}
794
795static void blk_account_io_completion(struct request *req, unsigned int bytes)
796{
797 if (req->part && blk_do_io_stat(req)) {
798 const int sgrp = op_stat_group(req_op(req));
799
800 part_stat_lock();
801 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
802 part_stat_unlock();
803 }
804}
805
Christoph Hellwig0d7a29a2021-11-17 07:14:03 +0100806static void blk_print_req_error(struct request *req, blk_status_t status)
807{
808 printk_ratelimited(KERN_ERR
809 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
810 "phys_seg %u prio class %u\n",
811 blk_status_to_str(status),
Christoph Hellwigf3fa33a2021-11-26 13:18:00 +0100812 req->q->disk ? req->q->disk->disk_name : "?",
Bart Van Assche16458cf2022-07-14 11:06:32 -0700813 blk_rq_pos(req), (__force u32)req_op(req),
814 blk_op_str(req_op(req)),
815 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
Christoph Hellwig0d7a29a2021-11-17 07:14:03 +0100816 req->nr_phys_segments,
817 IOPRIO_PRIO_CLASS(req->ioprio));
818}
819
Jens Axboe5581a5d2021-12-01 15:01:51 -0700820/*
821 * Fully end IO on a request. Does not support partial completions, or
822 * errors.
823 */
824static void blk_complete_request(struct request *req)
825{
826 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
827 int total_bytes = blk_rq_bytes(req);
828 struct bio *bio = req->bio;
829
830 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
831
832 if (!bio)
833 return;
834
835#ifdef CONFIG_BLK_DEV_INTEGRITY
836 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
837 req->q->integrity.profile->complete_fn(req, total_bytes);
838#endif
839
Eric Biggers9cd1e562023-03-15 11:39:02 -0700840 /*
841 * Upper layers may call blk_crypto_evict_key() anytime after the last
842 * bio_endio(). Therefore, the keyslot must be released before that.
843 */
844 blk_crypto_rq_put_keyslot(req);
845
Jens Axboe5581a5d2021-12-01 15:01:51 -0700846 blk_account_io_completion(req, total_bytes);
847
848 do {
849 struct bio *next = bio->bi_next;
850
851 /* Completion has already been traced */
852 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
Pankaj Raghava12821d2022-02-11 10:34:25 +0100853
854 if (req_op(req) == REQ_OP_ZONE_APPEND)
855 bio->bi_iter.bi_sector = req->__sector;
856
Jens Axboe5581a5d2021-12-01 15:01:51 -0700857 if (!is_flush)
858 bio_endio(bio);
859 bio = next;
860 } while (bio);
861
862 /*
863 * Reset counters so that the request stacking driver
864 * can find how many bytes remain in the request
865 * later.
866 */
Jens Axboeab3e1d32022-09-21 08:24:16 -0600867 if (!req->end_io) {
868 req->bio = NULL;
869 req->__data_len = 0;
870 }
Jens Axboe5581a5d2021-12-01 15:01:51 -0700871}
872
Jens Axboe9be3e062021-10-14 09:17:01 -0600873/**
874 * blk_update_request - Complete multiple bytes without completing the request
875 * @req: the request being processed
876 * @error: block status code
877 * @nr_bytes: number of bytes to complete for @req
878 *
879 * Description:
880 * Ends I/O on a number of bytes attached to @req, but doesn't complete
881 * the request structure even if @req doesn't have leftover.
882 * If @req has leftover, sets it up for the next range of segments.
883 *
884 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
885 * %false return from this function.
886 *
887 * Note:
888 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
889 * except in the consistency check at the end of this function.
890 *
891 * Return:
892 * %false - this request doesn't have any more data
893 * %true - this request has more data
894 **/
895bool blk_update_request(struct request *req, blk_status_t error,
896 unsigned int nr_bytes)
897{
898 int total_bytes;
899
Christoph Hellwig8a7d2672021-10-18 10:45:18 +0200900 trace_block_rq_complete(req, error, nr_bytes);
Jens Axboe9be3e062021-10-14 09:17:01 -0600901
902 if (!req->bio)
903 return false;
904
905#ifdef CONFIG_BLK_DEV_INTEGRITY
906 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
907 error == BLK_STS_OK)
908 req->q->integrity.profile->complete_fn(req, nr_bytes);
909#endif
910
Eric Biggers9cd1e562023-03-15 11:39:02 -0700911 /*
912 * Upper layers may call blk_crypto_evict_key() anytime after the last
913 * bio_endio(). Therefore, the keyslot must be released before that.
914 */
915 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
916 __blk_crypto_rq_put_keyslot(req);
917
Jens Axboe9be3e062021-10-14 09:17:01 -0600918 if (unlikely(error && !blk_rq_is_passthrough(req) &&
Christoph Hellwig3d973a72022-03-23 17:38:15 +0100919 !(req->rq_flags & RQF_QUIET)) &&
920 !test_bit(GD_DEAD, &req->q->disk->state)) {
Jens Axboe9be3e062021-10-14 09:17:01 -0600921 blk_print_req_error(req, error);
Yang Shid5869fd2022-02-10 14:52:22 -0800922 trace_block_rq_error(req, error, nr_bytes);
923 }
Jens Axboe9be3e062021-10-14 09:17:01 -0600924
925 blk_account_io_completion(req, nr_bytes);
926
927 total_bytes = 0;
928 while (req->bio) {
929 struct bio *bio = req->bio;
930 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
931
932 if (bio_bytes == bio->bi_iter.bi_size)
933 req->bio = bio->bi_next;
934
935 /* Completion has already been traced */
936 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
937 req_bio_endio(req, bio, bio_bytes, error);
938
939 total_bytes += bio_bytes;
940 nr_bytes -= bio_bytes;
941
942 if (!nr_bytes)
943 break;
944 }
945
946 /*
947 * completely done
948 */
949 if (!req->bio) {
950 /*
951 * Reset counters so that the request stacking driver
952 * can find how many bytes remain in the request
953 * later.
954 */
955 req->__data_len = 0;
956 return false;
957 }
958
959 req->__data_len -= total_bytes;
960
961 /* update sector only for requests with clear definition of sector */
962 if (!blk_rq_is_passthrough(req))
963 req->__sector += total_bytes >> 9;
964
965 /* mixed attributes always follow the first bio */
966 if (req->rq_flags & RQF_MIXED_MERGE) {
967 req->cmd_flags &= ~REQ_FAILFAST_MASK;
968 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
969 }
970
971 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
972 /*
973 * If total number of sectors is less than the first segment
974 * size, something has gone terribly wrong.
975 */
976 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
977 blk_dump_rq_flags(req, "request botched");
978 req->__data_len = blk_rq_cur_bytes(req);
979 }
980
981 /* recalculate the number of segments */
982 req->nr_phys_segments = blk_recalc_rq_segments(req);
983 }
984
985 return true;
986}
987EXPORT_SYMBOL_GPL(blk_update_request);
988
Christoph Hellwig450b7872021-11-17 07:14:01 +0100989static inline void blk_account_io_done(struct request *req, u64 now)
990{
Hengqi Chen5a80bd02023-05-20 08:40:57 +0000991 trace_block_io_done(req);
992
Christoph Hellwig450b7872021-11-17 07:14:01 +0100993 /*
994 * Account IO completion. flush_rq isn't accounted as a
995 * normal IO on queueing nor completion. Accounting the
996 * containing request is enough.
997 */
998 if (blk_do_io_stat(req) && req->part &&
Chaitanya Kulkarni06965032023-03-27 00:34:27 -0700999 !(req->rq_flags & RQF_FLUSH_SEQ)) {
1000 const int sgrp = op_stat_group(req_op(req));
Christoph Hellwig450b7872021-11-17 07:14:01 +01001001
Chaitanya Kulkarni06965032023-03-27 00:34:27 -07001002 part_stat_lock();
1003 update_io_ticks(req->part, jiffies, true);
1004 part_stat_inc(req->part, ios[sgrp]);
1005 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
1006 part_stat_unlock();
1007 }
Christoph Hellwig450b7872021-11-17 07:14:01 +01001008}
1009
1010static inline void blk_account_io_start(struct request *req)
1011{
Hengqi Chen5a80bd02023-05-20 08:40:57 +00001012 trace_block_io_start(req);
1013
Chaitanya Kulkarnie165fb42023-03-27 00:34:26 -07001014 if (blk_do_io_stat(req)) {
1015 /*
1016 * All non-passthrough requests are created from a bio with one
1017 * exception: when a flush command that is part of a flush sequence
1018 * generated by the state machine in blk-flush.c is cloned onto the
1019 * lower device by dm-multipath we can get here without a bio.
1020 */
1021 if (req->bio)
1022 req->part = req->bio->bi_bdev;
1023 else
1024 req->part = req->q->disk->part0;
1025
1026 part_stat_lock();
1027 update_io_ticks(req->part, jiffies, false);
1028 part_stat_unlock();
1029 }
Christoph Hellwig450b7872021-11-17 07:14:01 +01001030}
1031
Jens Axboef794f332021-10-08 05:50:46 -06001032static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1033{
Keith Busch54bdd672023-03-20 12:49:26 -07001034 if (rq->rq_flags & RQF_STATS)
Omar Sandoval522a7772018-05-09 02:08:53 -07001035 blk_stat_add(rq, now);
Omar Sandoval4bc63392018-05-09 02:08:52 -07001036
Baolin Wang87890092020-07-04 15:28:21 +08001037 blk_mq_sched_completed_request(rq, now);
Omar Sandoval522a7772018-05-09 02:08:53 -07001038 blk_account_io_done(rq, now);
Jens Axboef794f332021-10-08 05:50:46 -06001039}
1040
Ming Lei0d11e6a2013-12-05 10:50:39 -07001041inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
Christoph Hellwig91b63632014-04-16 09:44:53 +02001042{
Jens Axboef794f332021-10-08 05:50:46 -06001043 if (blk_mq_need_time_stamp(rq))
1044 __blk_mq_end_request_acct(rq, ktime_get_ns());
Jens Axboe320ae512013-10-24 09:20:05 +01001045
Chengming Zhoue5c0ca12023-08-13 23:23:25 +08001046 blk_mq_finish_request(rq);
1047
Jens Axboe320ae512013-10-24 09:20:05 +01001048 if (rq->end_io) {
Josef Bacika7905042018-07-03 09:32:35 -06001049 rq_qos_done(rq->q, rq);
Jens Axboede671d62022-09-21 15:19:54 -06001050 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1051 blk_mq_free_request(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +02001052 } else {
Jens Axboe320ae512013-10-24 09:20:05 +01001053 blk_mq_free_request(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +02001054 }
Jens Axboe320ae512013-10-24 09:20:05 +01001055}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -07001056EXPORT_SYMBOL(__blk_mq_end_request);
Christoph Hellwig63151a42014-04-16 09:44:52 +02001057
Christoph Hellwig2a842ac2017-06-03 09:38:04 +02001058void blk_mq_end_request(struct request *rq, blk_status_t error)
Christoph Hellwig63151a42014-04-16 09:44:52 +02001059{
1060 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1061 BUG();
Christoph Hellwigc8a446a2014-09-13 16:40:10 -07001062 __blk_mq_end_request(rq, error);
Christoph Hellwig63151a42014-04-16 09:44:52 +02001063}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -07001064EXPORT_SYMBOL(blk_mq_end_request);
Jens Axboe320ae512013-10-24 09:20:05 +01001065
Jens Axboef794f332021-10-08 05:50:46 -06001066#define TAG_COMP_BATCH 32
1067
1068static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1069 int *tag_array, int nr_tags)
1070{
1071 struct request_queue *q = hctx->queue;
1072
Chengming Zhoub8643d62023-09-13 15:16:12 +00001073 blk_mq_sub_active_requests(hctx, nr_tags);
Ming Lei3b87c6e2021-11-02 23:36:19 +08001074
Jens Axboef794f332021-10-08 05:50:46 -06001075 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1076 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1077}
1078
1079void blk_mq_end_request_batch(struct io_comp_batch *iob)
1080{
1081 int tags[TAG_COMP_BATCH], nr_tags = 0;
Jens Axboe02f7eab2021-10-28 12:08:34 -06001082 struct blk_mq_hw_ctx *cur_hctx = NULL;
Jens Axboef794f332021-10-08 05:50:46 -06001083 struct request *rq;
1084 u64 now = 0;
1085
1086 if (iob->need_ts)
1087 now = ktime_get_ns();
1088
1089 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1090 prefetch(rq->bio);
1091 prefetch(rq->rq_next);
1092
Jens Axboe5581a5d2021-12-01 15:01:51 -07001093 blk_complete_request(rq);
Jens Axboef794f332021-10-08 05:50:46 -06001094 if (iob->need_ts)
1095 __blk_mq_end_request_acct(rq, now);
1096
Chengming Zhoue5c0ca12023-08-13 23:23:25 +08001097 blk_mq_finish_request(rq);
1098
Jens Axboe98b26a02021-11-26 09:53:23 -07001099 rq_qos_done(rq->q, rq);
1100
Jens Axboeab3e1d32022-09-21 08:24:16 -06001101 /*
1102 * If end_io handler returns NONE, then it still has
1103 * ownership of the request.
1104 */
1105 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1106 continue;
1107
Jens Axboef794f332021-10-08 05:50:46 -06001108 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Jens Axboe0a467d02021-10-14 14:39:59 -06001109 if (!req_ref_put_and_test(rq))
Jens Axboef794f332021-10-08 05:50:46 -06001110 continue;
1111
1112 blk_crypto_free_request(rq);
1113 blk_pm_mark_last_busy(rq);
Jens Axboef794f332021-10-08 05:50:46 -06001114
Jens Axboe02f7eab2021-10-28 12:08:34 -06001115 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1116 if (cur_hctx)
1117 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
Jens Axboef794f332021-10-08 05:50:46 -06001118 nr_tags = 0;
Jens Axboe02f7eab2021-10-28 12:08:34 -06001119 cur_hctx = rq->mq_hctx;
Jens Axboef794f332021-10-08 05:50:46 -06001120 }
1121 tags[nr_tags++] = rq->tag;
Jens Axboef794f332021-10-08 05:50:46 -06001122 }
1123
1124 if (nr_tags)
Jens Axboe02f7eab2021-10-28 12:08:34 -06001125 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
Jens Axboef794f332021-10-08 05:50:46 -06001126}
1127EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1128
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001129static void blk_complete_reqs(struct llist_head *list)
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001130{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001131 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1132 struct request *rq, *next;
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001133
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001134 llist_for_each_entry_safe(rq, next, entry, ipi_list)
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001135 rq->q->mq_ops->complete(rq);
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001136}
1137
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001138static __latent_entropy void blk_done_softirq(struct softirq_action *h)
Christoph Hellwig115243f2020-06-11 08:44:42 +02001139{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001140 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001141}
1142
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001143static int blk_softirq_cpu_dead(unsigned int cpu)
1144{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001145 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001146 return 0;
1147}
1148
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08001149static void __blk_mq_complete_request_remote(void *data)
Jens Axboe320ae512013-10-24 09:20:05 +01001150{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001151 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
Jens Axboe320ae512013-10-24 09:20:05 +01001152}
1153
Christoph Hellwig963395262020-06-11 08:44:49 +02001154static inline bool blk_mq_complete_need_ipi(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001155{
Christoph Hellwig963395262020-06-11 08:44:49 +02001156 int cpu = raw_smp_processor_id();
Jens Axboe320ae512013-10-24 09:20:05 +01001157
Christoph Hellwig963395262020-06-11 08:44:49 +02001158 if (!IS_ENABLED(CONFIG_SMP) ||
1159 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1160 return false;
Sebastian Andrzej Siewior71425182020-12-04 20:13:54 +01001161 /*
1162 * With force threaded interrupts enabled, raising softirq from an SMP
1163 * function call will always result in waking the ksoftirqd thread.
1164 * This is probably worse than completing the request on a different
1165 * cache domain.
1166 */
Tanner Love91cc4702021-06-02 14:03:38 -04001167 if (force_irqthreads())
Sebastian Andrzej Siewior71425182020-12-04 20:13:54 +01001168 return false;
Christoph Hellwig963395262020-06-11 08:44:49 +02001169
1170 /* same CPU or cache domain? Complete locally */
1171 if (cpu == rq->mq_ctx->cpu ||
1172 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1173 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1174 return false;
1175
1176 /* don't try to IPI to an offline CPU */
1177 return cpu_online(rq->mq_ctx->cpu);
1178}
1179
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001180static void blk_mq_complete_send_ipi(struct request *rq)
1181{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001182 unsigned int cpu;
1183
1184 cpu = rq->mq_ctx->cpu;
Chengming Zhou660e8022023-07-17 12:00:55 +08001185 if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))
1186 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001187}
1188
1189static void blk_mq_raise_softirq(struct request *rq)
1190{
1191 struct llist_head *list;
1192
1193 preempt_disable();
1194 list = this_cpu_ptr(&blk_cpu_done);
1195 if (llist_add(&rq->ipi_list, list))
1196 raise_softirq(BLOCK_SOFTIRQ);
1197 preempt_enable();
1198}
1199
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001200bool blk_mq_complete_request_remote(struct request *rq)
1201{
Keith Buschaf78ff72018-11-26 09:54:30 -07001202 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
Ming Lei36e76532018-09-28 16:42:20 +08001203
Jens Axboe4ab32bf2018-11-18 16:15:35 -07001204 /*
Liu Songf1684202022-09-21 11:32:03 +08001205 * For request which hctx has only one ctx mapping,
1206 * or a polled request, always complete locally,
1207 * it's pointless to redirect the completion.
Jens Axboe4ab32bf2018-11-18 16:15:35 -07001208 */
Ed Tsai30654612023-06-14 08:25:29 +08001209 if ((rq->mq_hctx->nr_ctx == 1 &&
1210 rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1211 rq->cmd_flags & REQ_POLLED)
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001212 return false;
Jens Axboe320ae512013-10-24 09:20:05 +01001213
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001214 if (blk_mq_complete_need_ipi(rq)) {
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001215 blk_mq_complete_send_ipi(rq);
1216 return true;
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -08001217 }
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001218
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001219 if (rq->q->nr_hw_queues == 1) {
1220 blk_mq_raise_softirq(rq);
1221 return true;
1222 }
1223 return false;
Jens Axboe320ae512013-10-24 09:20:05 +01001224}
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001225EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1226
Jens Axboe320ae512013-10-24 09:20:05 +01001227/**
Christoph Hellwig15f73f52020-06-11 08:44:47 +02001228 * blk_mq_complete_request - end I/O on a request
1229 * @rq: the request being processed
Jens Axboe320ae512013-10-24 09:20:05 +01001230 *
Christoph Hellwig15f73f52020-06-11 08:44:47 +02001231 * Description:
1232 * Complete a request by scheduling the ->complete_rq operation.
1233 **/
1234void blk_mq_complete_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001235{
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001236 if (!blk_mq_complete_request_remote(rq))
Christoph Hellwig963395262020-06-11 08:44:49 +02001237 rq->q->mq_ops->complete(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01001238}
Christoph Hellwig15f73f52020-06-11 08:44:47 +02001239EXPORT_SYMBOL(blk_mq_complete_request);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08001240
1241/**
André Almeida105663f2020-01-06 15:08:18 -03001242 * blk_mq_start_request - Start processing a request
1243 * @rq: Pointer to request to be started
1244 *
1245 * Function used by device drivers to notify the block layer that a request
1246 * is going to be processed now, so blk layer can do proper initializations
1247 * such as starting the timeout timer.
1248 */
Christoph Hellwige2490072014-09-13 16:40:09 -07001249void blk_mq_start_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001250{
1251 struct request_queue *q = rq->q;
1252
Christoph Hellwiga54895f2020-12-03 17:21:39 +01001253 trace_block_rq_issue(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01001254
Kundan Kumar847c5bc2023-11-24 00:33:31 +05301255 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) &&
1256 !blk_rq_is_passthrough(rq)) {
Tejun Heo4cddeac2022-04-27 09:49:12 -10001257 rq->io_start_time_ns = ktime_get_ns();
Hou Tao3d244302019-05-21 15:59:03 +08001258 rq->stats_sectors = blk_rq_sectors(rq);
Jens Axboecf43e6b2016-11-07 21:32:37 -07001259 rq->rq_flags |= RQF_STATS;
Josef Bacika7905042018-07-03 09:32:35 -06001260 rq_qos_issue(q, rq);
Jens Axboecf43e6b2016-11-07 21:32:37 -07001261 }
1262
Tejun Heo1d9bd512018-01-09 08:29:48 -08001263 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
Jens Axboe538b7532014-09-16 10:37:37 -06001264
Tejun Heo1d9bd512018-01-09 08:29:48 -08001265 blk_add_timer(rq);
Keith Busch12f5b932018-05-29 15:52:28 +02001266 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
Chengming Zhou217b6132023-09-13 15:16:15 +00001267 rq->mq_hctx->tags->rqs[rq->tag] = rq;
Christoph Hellwig49f5baa2014-02-11 08:27:14 -08001268
Max Gurtovoy54d4e6a2019-09-16 18:44:29 +03001269#ifdef CONFIG_BLK_DEV_INTEGRITY
1270 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1271 q->integrity.profile->prepare_fn(rq);
1272#endif
Christoph Hellwig3e087732021-10-12 13:12:24 +02001273 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
Keith Buschf6c80cf2023-06-12 12:03:42 -07001274 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
Jens Axboe320ae512013-10-24 09:20:05 +01001275}
Christoph Hellwige2490072014-09-13 16:40:09 -07001276EXPORT_SYMBOL(blk_mq_start_request);
Jens Axboe320ae512013-10-24 09:20:05 +01001277
Ming Leia327c342022-05-12 22:00:10 +08001278/*
1279 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1280 * queues. This is important for md arrays to benefit from merging
1281 * requests.
1282 */
1283static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1284{
1285 if (plug->multiple_queues)
1286 return BLK_MAX_REQUEST_COUNT * 2;
1287 return BLK_MAX_REQUEST_COUNT;
1288}
1289
1290static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1291{
1292 struct request *last = rq_list_peek(&plug->mq_list);
1293
1294 if (!plug->rq_count) {
1295 trace_block_plug(rq->q);
1296 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1297 (!blk_queue_nomerges(rq->q) &&
1298 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1299 blk_mq_flush_plug_list(plug, false);
Al Viro878eb6e2022-11-01 00:54:13 +00001300 last = NULL;
Ming Leia327c342022-05-12 22:00:10 +08001301 trace_block_plug(rq->q);
1302 }
1303
1304 if (!plug->multiple_queues && last && last->q != rq->q)
1305 plug->multiple_queues = true;
Ming Leic6b7a3a2023-06-24 21:01:05 +08001306 /*
1307 * Any request allocated from sched tags can't be issued to
1308 * ->queue_rqs() directly
1309 */
1310 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
Ming Leia327c342022-05-12 22:00:10 +08001311 plug->has_elevator = true;
1312 rq->rq_next = NULL;
1313 rq_list_add(&plug->mq_list, rq);
1314 plug->rq_count++;
1315}
1316
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001317/**
1318 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001319 * @rq: request to insert
1320 * @at_head: insert request at head or tail of queue
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001321 *
1322 * Description:
1323 * Insert a fully prepared request at the back of the I/O scheduler queue
1324 * for execution. Don't wait for completion.
1325 *
1326 * Note:
1327 * This function will invoke @done directly if the queue is dead.
1328 */
Christoph Hellwige2e530862022-05-24 14:15:30 +02001329void blk_execute_rq_nowait(struct request *rq, bool at_head)
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001330{
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001331 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1332
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001333 WARN_ON(irqs_disabled());
1334 WARN_ON(!blk_rq_is_passthrough(rq));
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001335
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001336 blk_account_io_start(rq);
Pankaj Raghav110fdb42022-09-29 16:41:41 +02001337
1338 /*
1339 * As plugging can be enabled for passthrough requests on a zoned
1340 * device, directly accessing the plug instead of using blk_mq_plug()
1341 * should not have any consequences.
1342 */
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001343 if (current->plug && !at_head) {
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001344 blk_add_rq_to_plug(current->plug, rq);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001345 return;
1346 }
1347
Christoph Hellwig710fa372023-04-13 08:40:54 +02001348 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
Bart Van Assche65a558f2023-07-21 10:27:30 -07001349 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001350}
1351EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1352
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001353struct blk_rq_wait {
1354 struct completion done;
1355 blk_status_t ret;
1356};
1357
Jens Axboede671d62022-09-21 15:19:54 -06001358static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001359{
1360 struct blk_rq_wait *wait = rq->end_io_data;
1361
1362 wait->ret = ret;
1363 complete(&wait->done);
Jens Axboede671d62022-09-21 15:19:54 -06001364 return RQ_END_IO_NONE;
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001365}
1366
Kanchan Joshic6e99ea2022-08-23 21:44:42 +05301367bool blk_rq_is_poll(struct request *rq)
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001368{
1369 if (!rq->mq_hctx)
1370 return false;
1371 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1372 return false;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001373 return true;
1374}
Kanchan Joshic6e99ea2022-08-23 21:44:42 +05301375EXPORT_SYMBOL_GPL(blk_rq_is_poll);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001376
1377static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1378{
1379 do {
Keith Buschf6c80cf2023-06-12 12:03:42 -07001380 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001381 cond_resched();
1382 } while (!completion_done(wait));
1383}
1384
1385/**
1386 * blk_execute_rq - insert a request into queue for execution
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001387 * @rq: request to insert
1388 * @at_head: insert request at head or tail of queue
1389 *
1390 * Description:
1391 * Insert a fully prepared request at the back of the I/O scheduler queue
1392 * for execution and wait for completion.
1393 * Return: The blk_status_t result provided to blk_mq_end_request().
1394 */
Christoph Hellwigb84ba302021-11-26 13:18:01 +01001395blk_status_t blk_execute_rq(struct request *rq, bool at_head)
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001396{
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001397 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001398 struct blk_rq_wait wait = {
1399 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1400 };
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001401
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001402 WARN_ON(irqs_disabled());
1403 WARN_ON(!blk_rq_is_passthrough(rq));
1404
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001405 rq->end_io_data = &wait;
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001406 rq->end_io = blk_end_sync_rq;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001407
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001408 blk_account_io_start(rq);
Christoph Hellwig710fa372023-04-13 08:40:54 +02001409 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001410 blk_mq_run_hw_queue(hctx, false);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001411
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001412 if (blk_rq_is_poll(rq)) {
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001413 blk_rq_poll_completion(rq, &wait.done);
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001414 } else {
1415 /*
1416 * Prevent hang_check timer from firing at us during very long
1417 * I/O
1418 */
1419 unsigned long hang_check = sysctl_hung_task_timeout_secs;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001420
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001421 if (hang_check)
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001422 while (!wait_for_completion_io_timeout(&wait.done,
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001423 hang_check * (HZ/2)))
1424 ;
1425 else
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001426 wait_for_completion_io(&wait.done);
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001427 }
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001428
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001429 return wait.ret;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001430}
1431EXPORT_SYMBOL(blk_execute_rq);
1432
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001433static void __blk_mq_requeue_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001434{
1435 struct request_queue *q = rq->q;
1436
Ming Lei923218f2017-11-02 23:24:38 +08001437 blk_mq_put_driver_tag(rq);
1438
Christoph Hellwiga54895f2020-12-03 17:21:39 +01001439 trace_block_rq_requeue(rq);
Josef Bacika7905042018-07-03 09:32:35 -06001440 rq_qos_requeue(q, rq);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -08001441
Keith Busch12f5b932018-05-29 15:52:28 +02001442 if (blk_mq_request_started(rq)) {
1443 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Christoph Hellwigda661262018-06-14 13:58:45 +02001444 rq->rq_flags &= ~RQF_TIMED_OUT;
Christoph Hellwige2490072014-09-13 16:40:09 -07001445 }
Jens Axboe320ae512013-10-24 09:20:05 +01001446}
1447
Bart Van Assche2b053ac2016-10-28 17:21:41 -07001448void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001449{
Christoph Hellwig214a4412023-04-13 08:40:53 +02001450 struct request_queue *q = rq->q;
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001451 unsigned long flags;
Christoph Hellwig214a4412023-04-13 08:40:53 +02001452
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001453 __blk_mq_requeue_request(rq);
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001454
Ming Lei105976f2018-02-23 23:36:56 +08001455 /* this request will be re-inserted to io scheduler queue */
1456 blk_mq_sched_requeue_request(rq);
1457
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001458 spin_lock_irqsave(&q->requeue_lock, flags);
1459 list_add_tail(&rq->queuelist, &q->requeue_list);
1460 spin_unlock_irqrestore(&q->requeue_lock, flags);
Christoph Hellwig214a4412023-04-13 08:40:53 +02001461
1462 if (kick_requeue_list)
1463 blk_mq_kick_requeue_list(q);
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001464}
1465EXPORT_SYMBOL(blk_mq_requeue_request);
1466
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001467static void blk_mq_requeue_work(struct work_struct *work)
1468{
1469 struct request_queue *q =
Mike Snitzer28494502016-09-14 13:28:30 -04001470 container_of(work, struct request_queue, requeue_work.work);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001471 LIST_HEAD(rq_list);
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001472 LIST_HEAD(flush_list);
1473 struct request *rq;
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001474
Jens Axboe18e97812017-07-27 08:03:57 -06001475 spin_lock_irq(&q->requeue_lock);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001476 list_splice_init(&q->requeue_list, &rq_list);
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001477 list_splice_init(&q->flush_list, &flush_list);
Jens Axboe18e97812017-07-27 08:03:57 -06001478 spin_unlock_irq(&q->requeue_lock);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001479
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001480 while (!list_empty(&rq_list)) {
1481 rq = list_entry(rq_list.next, struct request, queuelist);
Jianchao Wangaef18972019-02-12 09:56:25 +08001482 /*
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001483 * If RQF_DONTPREP ist set, the request has been started by the
1484 * driver already and might have driver-specific data allocated
1485 * already. Insert it into the hctx dispatch list to avoid
1486 * block layer merges for the request.
Jianchao Wangaef18972019-02-12 09:56:25 +08001487 */
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001488 if (rq->rq_flags & RQF_DONTPREP) {
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001489 list_del_init(&rq->queuelist);
Christoph Hellwig2b597612023-04-13 08:40:55 +02001490 blk_mq_request_bypass_insert(rq, 0);
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001491 } else {
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001492 list_del_init(&rq->queuelist);
Christoph Hellwig710fa372023-04-13 08:40:54 +02001493 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001494 }
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001495 }
1496
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001497 while (!list_empty(&flush_list)) {
1498 rq = list_entry(flush_list.next, struct request, queuelist);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001499 list_del_init(&rq->queuelist);
Christoph Hellwig710fa372023-04-13 08:40:54 +02001500 blk_mq_insert_request(rq, 0);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001501 }
1502
Bart Van Assche52d7f1b2016-10-28 17:20:32 -07001503 blk_mq_run_hw_queues(q, false);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001504}
1505
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001506void blk_mq_kick_requeue_list(struct request_queue *q)
1507{
Bart Van Asscheae943d22018-01-19 08:58:55 -08001508 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001509}
1510EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1511
Mike Snitzer28494502016-09-14 13:28:30 -04001512void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1513 unsigned long msecs)
1514{
Bart Van Assched4acf362017-08-09 11:28:06 -07001515 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1516 msecs_to_jiffies(msecs));
Mike Snitzer28494502016-09-14 13:28:30 -04001517}
1518EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1519
Ming Lei0e4237a2023-12-01 16:56:05 +08001520static bool blk_is_flush_data_rq(struct request *rq)
1521{
1522 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);
1523}
1524
John Garry2dd65322022-07-06 20:03:53 +08001525static bool blk_mq_rq_inflight(struct request *rq, void *priv)
Jens Axboeae879912018-11-08 09:03:51 -07001526{
1527 /*
John Garry8ab30a32021-12-06 20:49:48 +08001528 * If we find a request that isn't idle we know the queue is busy
1529 * as it's checked in the iter.
1530 * Return false to stop the iteration.
Ming Lei0e4237a2023-12-01 16:56:05 +08001531 *
1532 * In case of queue quiesce, if one flush data request is completed,
1533 * don't count it as inflight given the flush sequence is suspended,
1534 * and the original flush data request is invisible to driver, just
1535 * like other pending requests because of quiesce
Jens Axboeae879912018-11-08 09:03:51 -07001536 */
Ming Lei0e4237a2023-12-01 16:56:05 +08001537 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
1538 blk_is_flush_data_rq(rq) &&
1539 blk_mq_request_completed(rq))) {
Jens Axboeae879912018-11-08 09:03:51 -07001540 bool *busy = priv;
1541
1542 *busy = true;
1543 return false;
1544 }
1545
1546 return true;
1547}
1548
Jens Axboe3c94d832018-12-17 21:11:17 -07001549bool blk_mq_queue_inflight(struct request_queue *q)
Jens Axboeae879912018-11-08 09:03:51 -07001550{
1551 bool busy = false;
1552
Jens Axboe3c94d832018-12-17 21:11:17 -07001553 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
Jens Axboeae879912018-11-08 09:03:51 -07001554 return busy;
1555}
Jens Axboe3c94d832018-12-17 21:11:17 -07001556EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
Jens Axboeae879912018-11-08 09:03:51 -07001557
John Garry9bdb4832022-07-06 20:03:51 +08001558static void blk_mq_rq_timed_out(struct request *req)
Jens Axboe320ae512013-10-24 09:20:05 +01001559{
Christoph Hellwigda661262018-06-14 13:58:45 +02001560 req->rq_flags |= RQF_TIMED_OUT;
Christoph Hellwigd1210d52018-05-29 15:52:39 +02001561 if (req->q->mq_ops->timeout) {
1562 enum blk_eh_timer_return ret;
Jens Axboe87ee7b12014-04-24 08:51:47 -06001563
John Garry9bdb4832022-07-06 20:03:51 +08001564 ret = req->q->mq_ops->timeout(req);
Christoph Hellwigd1210d52018-05-29 15:52:39 +02001565 if (ret == BLK_EH_DONE)
1566 return;
1567 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
Christoph Hellwig46f92d42014-09-13 16:40:12 -07001568 }
Christoph Hellwigd1210d52018-05-29 15:52:39 +02001569
1570 blk_add_timer(req);
Jens Axboe87ee7b12014-04-24 08:51:47 -06001571}
Keith Busch5b3f25f2015-01-07 18:55:46 -07001572
David Jeffery82c22942022-10-26 13:19:57 +08001573struct blk_expired_data {
1574 bool has_timedout_rq;
1575 unsigned long next;
1576 unsigned long timeout_start;
1577};
1578
1579static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
Keith Busch12f5b932018-05-29 15:52:28 +02001580{
1581 unsigned long deadline;
1582
1583 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1584 return false;
Christoph Hellwigda661262018-06-14 13:58:45 +02001585 if (rq->rq_flags & RQF_TIMED_OUT)
1586 return false;
Keith Busch12f5b932018-05-29 15:52:28 +02001587
Christoph Hellwig079076b2018-11-14 17:02:05 +01001588 deadline = READ_ONCE(rq->deadline);
David Jeffery82c22942022-10-26 13:19:57 +08001589 if (time_after_eq(expired->timeout_start, deadline))
Keith Busch12f5b932018-05-29 15:52:28 +02001590 return true;
1591
David Jeffery82c22942022-10-26 13:19:57 +08001592 if (expired->next == 0)
1593 expired->next = deadline;
1594 else if (time_after(expired->next, deadline))
1595 expired->next = deadline;
Keith Busch12f5b932018-05-29 15:52:28 +02001596 return false;
1597}
1598
Ming Lei2e315dc2021-05-11 23:22:34 +08001599void blk_mq_put_rq_ref(struct request *rq)
1600{
Jens Axboede671d62022-09-21 15:19:54 -06001601 if (is_flush_rq(rq)) {
1602 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1603 blk_mq_free_request(rq);
1604 } else if (req_ref_put_and_test(rq)) {
Ming Lei2e315dc2021-05-11 23:22:34 +08001605 __blk_mq_free_request(rq);
Jens Axboede671d62022-09-21 15:19:54 -06001606 }
Ming Lei2e315dc2021-05-11 23:22:34 +08001607}
1608
John Garry2dd65322022-07-06 20:03:53 +08001609static bool blk_mq_check_expired(struct request *rq, void *priv)
Jens Axboe320ae512013-10-24 09:20:05 +01001610{
David Jeffery82c22942022-10-26 13:19:57 +08001611 struct blk_expired_data *expired = priv;
Christoph Hellwig81481eb2014-09-13 16:40:11 -07001612
Keith Busch12f5b932018-05-29 15:52:28 +02001613 /*
Ming Leic797b402021-08-11 23:52:02 +08001614 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1615 * be reallocated underneath the timeout handler's processing, then
1616 * the expire check is reliable. If the request is not expired, then
1617 * it was completed and reallocated as a new request after returning
1618 * from blk_mq_check_expired().
Keith Busch12f5b932018-05-29 15:52:28 +02001619 */
David Jeffery82c22942022-10-26 13:19:57 +08001620 if (blk_mq_req_expired(rq, expired)) {
1621 expired->has_timedout_rq = true;
1622 return false;
1623 }
1624 return true;
1625}
1626
1627static bool blk_mq_handle_expired(struct request *rq, void *priv)
1628{
1629 struct blk_expired_data *expired = priv;
1630
1631 if (blk_mq_req_expired(rq, expired))
John Garry9bdb4832022-07-06 20:03:51 +08001632 blk_mq_rq_timed_out(rq);
Jens Axboe7baa8572018-11-08 10:24:07 -07001633 return true;
Tejun Heo1d9bd512018-01-09 08:29:48 -08001634}
1635
Christoph Hellwig287922e2015-10-30 20:57:30 +08001636static void blk_mq_timeout_work(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +01001637{
Christoph Hellwig287922e2015-10-30 20:57:30 +08001638 struct request_queue *q =
1639 container_of(work, struct request_queue, timeout_work);
David Jeffery82c22942022-10-26 13:19:57 +08001640 struct blk_expired_data expired = {
1641 .timeout_start = jiffies,
1642 };
Tejun Heo1d9bd512018-01-09 08:29:48 -08001643 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08001644 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01001645
Gabriel Krisman Bertazi71f79fb2016-08-01 08:23:39 -06001646 /* A deadlock might occur if a request is stuck requiring a
1647 * timeout at the same time a queue freeze is waiting
1648 * completion, since the timeout code would not be able to
1649 * acquire the queue reference here.
1650 *
1651 * That's why we don't use blk_queue_enter here; instead, we use
1652 * percpu_ref_tryget directly, because we need to be able to
1653 * obtain a reference even in the short window between the queue
1654 * starting to freeze, by dropping the first reference in
Ming Lei1671d522017-03-27 20:06:57 +08001655 * blk_freeze_queue_start, and the moment the last request is
Gabriel Krisman Bertazi71f79fb2016-08-01 08:23:39 -06001656 * consumed, marked by the instant q_usage_counter reaches
1657 * zero.
1658 */
1659 if (!percpu_ref_tryget(&q->q_usage_counter))
Christoph Hellwig287922e2015-10-30 20:57:30 +08001660 return;
1661
David Jeffery82c22942022-10-26 13:19:57 +08001662 /* check if there is any timed-out request */
1663 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1664 if (expired.has_timedout_rq) {
1665 /*
1666 * Before walking tags, we must ensure any submit started
1667 * before the current time has finished. Since the submit
1668 * uses srcu or rcu, wait for a synchronization point to
1669 * ensure all running submits have finished
1670 */
Christoph Hellwig483239c2022-11-01 16:00:48 +01001671 blk_mq_wait_quiesce_done(q->tag_set);
Jens Axboe320ae512013-10-24 09:20:05 +01001672
David Jeffery82c22942022-10-26 13:19:57 +08001673 expired.next = 0;
1674 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1675 }
1676
1677 if (expired.next != 0) {
1678 mod_timer(&q->timeout, expired.next);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001679 } else {
Bart Van Asschefcd36c32018-01-10 08:33:33 -08001680 /*
1681 * Request timeouts are handled as a forward rolling timer. If
1682 * we end up here it means that no requests are pending and
1683 * also that no request has been pending for a while. Mark
1684 * each hctx as idle.
1685 */
Ming Leif054b562015-04-21 10:00:19 +08001686 queue_for_each_hw_ctx(q, hctx, i) {
1687 /* the hctx may be unmapped, so check it here */
1688 if (blk_mq_hw_queue_mapped(hctx))
1689 blk_mq_tag_idle(hctx);
1690 }
Jens Axboe0d2602c2014-05-13 15:10:52 -06001691 }
Christoph Hellwig287922e2015-10-30 20:57:30 +08001692 blk_queue_exit(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001693}
1694
Omar Sandoval88459642016-09-17 08:38:44 -06001695struct flush_busy_ctx_data {
1696 struct blk_mq_hw_ctx *hctx;
1697 struct list_head *list;
1698};
1699
1700static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1701{
1702 struct flush_busy_ctx_data *flush_data = data;
1703 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1704 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
Ming Leic16d6b52018-12-17 08:44:05 -07001705 enum hctx_type type = hctx->type;
Omar Sandoval88459642016-09-17 08:38:44 -06001706
Omar Sandoval88459642016-09-17 08:38:44 -06001707 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07001708 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
Omar Sandovale9a99a62018-02-27 16:56:42 -08001709 sbitmap_clear_bit(sb, bitnr);
Omar Sandoval88459642016-09-17 08:38:44 -06001710 spin_unlock(&ctx->lock);
1711 return true;
1712}
1713
Jens Axboe320ae512013-10-24 09:20:05 +01001714/*
Jens Axboe1429d7c2014-05-19 09:23:55 -06001715 * Process software queues that have been marked busy, splicing them
1716 * to the for-dispatch
1717 */
Jens Axboe2c3ad662016-12-14 14:34:47 -07001718void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
Jens Axboe1429d7c2014-05-19 09:23:55 -06001719{
Omar Sandoval88459642016-09-17 08:38:44 -06001720 struct flush_busy_ctx_data data = {
1721 .hctx = hctx,
1722 .list = list,
1723 };
Jens Axboe1429d7c2014-05-19 09:23:55 -06001724
Omar Sandoval88459642016-09-17 08:38:44 -06001725 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
Jens Axboe1429d7c2014-05-19 09:23:55 -06001726}
Jens Axboe2c3ad662016-12-14 14:34:47 -07001727EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
Jens Axboe1429d7c2014-05-19 09:23:55 -06001728
Ming Leib3476892017-10-14 17:22:30 +08001729struct dispatch_rq_data {
1730 struct blk_mq_hw_ctx *hctx;
1731 struct request *rq;
1732};
1733
1734static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1735 void *data)
1736{
1737 struct dispatch_rq_data *dispatch_data = data;
1738 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1739 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
Ming Leic16d6b52018-12-17 08:44:05 -07001740 enum hctx_type type = hctx->type;
Ming Leib3476892017-10-14 17:22:30 +08001741
1742 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07001743 if (!list_empty(&ctx->rq_lists[type])) {
1744 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
Ming Leib3476892017-10-14 17:22:30 +08001745 list_del_init(&dispatch_data->rq->queuelist);
Ming Leic16d6b52018-12-17 08:44:05 -07001746 if (list_empty(&ctx->rq_lists[type]))
Ming Leib3476892017-10-14 17:22:30 +08001747 sbitmap_clear_bit(sb, bitnr);
1748 }
1749 spin_unlock(&ctx->lock);
1750
1751 return !dispatch_data->rq;
1752}
1753
1754struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1755 struct blk_mq_ctx *start)
1756{
Jens Axboef31967f2018-10-29 13:13:29 -06001757 unsigned off = start ? start->index_hw[hctx->type] : 0;
Ming Leib3476892017-10-14 17:22:30 +08001758 struct dispatch_rq_data data = {
1759 .hctx = hctx,
1760 .rq = NULL,
1761 };
1762
1763 __sbitmap_for_each_set(&hctx->ctx_map, off,
1764 dispatch_rq_from_ctx, &data);
1765
1766 return data.rq;
1767}
1768
Chengming Zhoub8643d62023-09-13 15:16:12 +00001769bool __blk_mq_alloc_driver_tag(struct request *rq)
Jens Axboe703fd1c2016-09-16 13:59:14 -06001770{
John Garryae0f1a72021-10-05 18:23:38 +08001771 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
Ming Lei570e9b72020-06-30 22:03:55 +08001772 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
Ming Lei570e9b72020-06-30 22:03:55 +08001773 int tag;
1774
Ming Lei568f2702020-07-06 22:41:11 +08001775 blk_mq_tag_busy(rq->mq_hctx);
1776
Ming Lei570e9b72020-06-30 22:03:55 +08001777 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
John Garryae0f1a72021-10-05 18:23:38 +08001778 bt = &rq->mq_hctx->tags->breserved_tags;
Ming Lei570e9b72020-06-30 22:03:55 +08001779 tag_offset = 0;
Ming Lei28500852020-09-11 18:41:14 +08001780 } else {
1781 if (!hctx_may_queue(rq->mq_hctx, bt))
1782 return false;
Ming Lei570e9b72020-06-30 22:03:55 +08001783 }
1784
Ming Lei570e9b72020-06-30 22:03:55 +08001785 tag = __sbitmap_queue_get(bt);
1786 if (tag == BLK_MQ_NO_TAG)
1787 return false;
1788
1789 rq->tag = tag + tag_offset;
Chengming Zhoub8643d62023-09-13 15:16:12 +00001790 blk_mq_inc_active_requests(rq->mq_hctx);
Ming Lei568f2702020-07-06 22:41:11 +08001791 return true;
Ming Lei570e9b72020-06-30 22:03:55 +08001792}
1793
Jens Axboeeb619fd2017-11-09 08:32:43 -07001794static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1795 int flags, void *key)
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001796{
1797 struct blk_mq_hw_ctx *hctx;
1798
1799 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1800
Ming Lei5815839b2018-06-25 19:31:47 +08001801 spin_lock(&hctx->dispatch_wait_lock);
Jens Axboee8618572019-03-25 12:34:10 -06001802 if (!list_empty(&wait->entry)) {
1803 struct sbitmap_queue *sbq;
1804
1805 list_del_init(&wait->entry);
John Garryae0f1a72021-10-05 18:23:38 +08001806 sbq = &hctx->tags->bitmap_tags;
Jens Axboee8618572019-03-25 12:34:10 -06001807 atomic_dec(&sbq->ws_active);
1808 }
Ming Lei5815839b2018-06-25 19:31:47 +08001809 spin_unlock(&hctx->dispatch_wait_lock);
1810
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001811 blk_mq_run_hw_queue(hctx, true);
1812 return 1;
1813}
1814
Jens Axboef906a6a2017-11-09 16:10:13 -07001815/*
1816 * Mark us waiting for a tag. For shared tags, this involves hooking us into
Bart Van Asscheee3e4de2018-01-09 10:09:15 -08001817 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1818 * restart. For both cases, take care to check the condition again after
Jens Axboef906a6a2017-11-09 16:10:13 -07001819 * marking us as waiting.
1820 */
Ming Lei2278d692018-06-25 19:31:46 +08001821static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
Jens Axboef906a6a2017-11-09 16:10:13 -07001822 struct request *rq)
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001823{
Kemeng Shi98b99e92023-01-18 17:37:15 +08001824 struct sbitmap_queue *sbq;
Ming Lei5815839b2018-06-25 19:31:47 +08001825 struct wait_queue_head *wq;
Jens Axboef906a6a2017-11-09 16:10:13 -07001826 wait_queue_entry_t *wait;
1827 bool ret;
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001828
Kemeng Shi47df9ce2023-01-18 17:37:16 +08001829 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1830 !(blk_mq_is_shared_tags(hctx->flags))) {
Yufen Yu684b7322019-03-15 11:05:10 +08001831 blk_mq_sched_mark_restart_hctx(hctx);
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001832
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001833 /*
1834 * It's possible that a tag was freed in the window between the
1835 * allocation failure and adding the hardware queue to the wait
1836 * queue.
1837 *
1838 * Don't clear RESTART here, someone else could have set it.
1839 * At most this will cost an extra queue run.
1840 */
Ming Lei8ab6bb9e2018-06-25 19:31:45 +08001841 return blk_mq_get_driver_tag(rq);
Jens Axboeeb619fd2017-11-09 08:32:43 -07001842 }
1843
Ming Lei2278d692018-06-25 19:31:46 +08001844 wait = &hctx->dispatch_wait;
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001845 if (!list_empty_careful(&wait->entry))
1846 return false;
1847
Kemeng Shi98b99e92023-01-18 17:37:15 +08001848 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1849 sbq = &hctx->tags->breserved_tags;
1850 else
1851 sbq = &hctx->tags->bitmap_tags;
Jens Axboee8618572019-03-25 12:34:10 -06001852 wq = &bt_wait_ptr(sbq, hctx)->wait;
Ming Lei5815839b2018-06-25 19:31:47 +08001853
1854 spin_lock_irq(&wq->lock);
1855 spin_lock(&hctx->dispatch_wait_lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001856 if (!list_empty(&wait->entry)) {
Ming Lei5815839b2018-06-25 19:31:47 +08001857 spin_unlock(&hctx->dispatch_wait_lock);
1858 spin_unlock_irq(&wq->lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001859 return false;
1860 }
1861
Jens Axboee8618572019-03-25 12:34:10 -06001862 atomic_inc(&sbq->ws_active);
Ming Lei5815839b2018-06-25 19:31:47 +08001863 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1864 __add_wait_queue(wq, wait);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001865
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001866 /*
Ming Lei5266caa2024-01-12 20:26:26 +08001867 * Add one explicit barrier since blk_mq_get_driver_tag() may
1868 * not imply barrier in case of failure.
1869 *
1870 * Order adding us to wait queue and allocating driver tag.
1871 *
1872 * The pair is the one implied in sbitmap_queue_wake_up() which
1873 * orders clearing sbitmap tag bits and waitqueue_active() in
1874 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1875 *
1876 * Otherwise, re-order of adding wait queue and getting driver tag
1877 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1878 * the waitqueue_active() may not observe us in wait queue.
1879 */
1880 smp_mb();
1881
1882 /*
Jens Axboeeb619fd2017-11-09 08:32:43 -07001883 * It's possible that a tag was freed in the window between the
1884 * allocation failure and adding the hardware queue to the wait
1885 * queue.
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001886 */
Ming Lei8ab6bb9e2018-06-25 19:31:45 +08001887 ret = blk_mq_get_driver_tag(rq);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001888 if (!ret) {
Ming Lei5815839b2018-06-25 19:31:47 +08001889 spin_unlock(&hctx->dispatch_wait_lock);
1890 spin_unlock_irq(&wq->lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001891 return false;
Jens Axboef906a6a2017-11-09 16:10:13 -07001892 }
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001893
1894 /*
1895 * We got a tag, remove ourselves from the wait queue to ensure
1896 * someone else gets the wakeup.
1897 */
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001898 list_del_init(&wait->entry);
Jens Axboee8618572019-03-25 12:34:10 -06001899 atomic_dec(&sbq->ws_active);
Ming Lei5815839b2018-06-25 19:31:47 +08001900 spin_unlock(&hctx->dispatch_wait_lock);
1901 spin_unlock_irq(&wq->lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001902
1903 return true;
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001904}
1905
Ming Lei6e7687172018-07-03 09:03:16 -06001906#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1907#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1908/*
1909 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1910 * - EWMA is one simple way to compute running average value
1911 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1912 * - take 4 as factor for avoiding to get too small(0) result, and this
1913 * factor doesn't matter because EWMA decreases exponentially
1914 */
1915static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1916{
1917 unsigned int ewma;
1918
Ming Lei6e7687172018-07-03 09:03:16 -06001919 ewma = hctx->dispatch_busy;
1920
1921 if (!ewma && !busy)
1922 return;
1923
1924 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1925 if (busy)
1926 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1927 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1928
1929 hctx->dispatch_busy = ewma;
1930}
1931
Ming Lei86ff7c22018-01-30 22:04:57 -05001932#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1933
Johannes Thumshirnc92a4102020-03-25 00:24:44 +09001934static void blk_mq_handle_dev_resource(struct request *rq,
1935 struct list_head *list)
1936{
Johannes Thumshirnc92a4102020-03-25 00:24:44 +09001937 list_add(&rq->queuelist, list);
1938 __blk_mq_requeue_request(rq);
1939}
1940
Keith Busch0512a752020-05-12 17:55:47 +09001941static void blk_mq_handle_zone_resource(struct request *rq,
1942 struct list_head *zone_list)
1943{
1944 /*
1945 * If we end up here it is because we cannot dispatch a request to a
1946 * specific zone due to LLD level zone-write locking or other zone
1947 * related resource not being available. In this case, set the request
1948 * aside in zone_list for retrying it later.
1949 */
1950 list_add(&rq->queuelist, zone_list);
1951 __blk_mq_requeue_request(rq);
1952}
1953
Ming Lei75383522020-06-30 18:24:58 +08001954enum prep_dispatch {
1955 PREP_DISPATCH_OK,
1956 PREP_DISPATCH_NO_TAG,
1957 PREP_DISPATCH_NO_BUDGET,
1958};
1959
1960static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1961 bool need_budget)
1962{
1963 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Ming Lei2a5a24a2021-01-22 10:33:12 +08001964 int budget_token = -1;
Ming Lei75383522020-06-30 18:24:58 +08001965
Ming Lei2a5a24a2021-01-22 10:33:12 +08001966 if (need_budget) {
1967 budget_token = blk_mq_get_dispatch_budget(rq->q);
1968 if (budget_token < 0) {
1969 blk_mq_put_driver_tag(rq);
1970 return PREP_DISPATCH_NO_BUDGET;
1971 }
1972 blk_mq_set_rq_budget_token(rq, budget_token);
Ming Lei75383522020-06-30 18:24:58 +08001973 }
1974
1975 if (!blk_mq_get_driver_tag(rq)) {
1976 /*
1977 * The initial allocation attempt failed, so we need to
1978 * rerun the hardware queue when a tag is freed. The
1979 * waitqueue takes care of that. If the queue is run
1980 * before we add this entry back on the dispatch list,
1981 * we'll re-run it below.
1982 */
1983 if (!blk_mq_mark_tag_wait(hctx, rq)) {
Ming Lei1fd40b52020-06-30 18:25:00 +08001984 /*
1985 * All budgets not got from this function will be put
1986 * together during handling partial dispatch
1987 */
1988 if (need_budget)
Ming Lei2a5a24a2021-01-22 10:33:12 +08001989 blk_mq_put_dispatch_budget(rq->q, budget_token);
Ming Lei75383522020-06-30 18:24:58 +08001990 return PREP_DISPATCH_NO_TAG;
1991 }
1992 }
1993
1994 return PREP_DISPATCH_OK;
1995}
1996
Ming Lei1fd40b52020-06-30 18:25:00 +08001997/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1998static void blk_mq_release_budgets(struct request_queue *q,
Ming Lei2a5a24a2021-01-22 10:33:12 +08001999 struct list_head *list)
Ming Lei1fd40b52020-06-30 18:25:00 +08002000{
Ming Lei2a5a24a2021-01-22 10:33:12 +08002001 struct request *rq;
Ming Lei1fd40b52020-06-30 18:25:00 +08002002
Ming Lei2a5a24a2021-01-22 10:33:12 +08002003 list_for_each_entry(rq, list, queuelist) {
2004 int budget_token = blk_mq_get_rq_budget_token(rq);
2005
2006 if (budget_token >= 0)
2007 blk_mq_put_dispatch_budget(q, budget_token);
2008 }
Ming Lei1fd40b52020-06-30 18:25:00 +08002009}
2010
Jens Axboe1f57f8d2018-06-28 11:54:01 -06002011/*
Kemeng Shi34c9f5472023-01-18 17:37:19 +08002012 * blk_mq_commit_rqs will notify driver using bd->last that there is no
2013 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
2014 * details)
2015 * Attention, we should explicitly call this in unusual cases:
2016 * 1) did not queue everything initially scheduled to queue
2017 * 2) the last attempt to queue a request failed
2018 */
2019static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2020 bool from_schedule)
2021{
2022 if (hctx->queue->mq_ops->commit_rqs && queued) {
2023 trace_block_unplug(hctx->queue, queued, !from_schedule);
2024 hctx->queue->mq_ops->commit_rqs(hctx);
2025 }
2026}
2027
2028/*
Jens Axboe1f57f8d2018-06-28 11:54:01 -06002029 * Returns true if we did some work AND can potentially do more.
2030 */
Ming Lei445874e2020-06-30 18:24:57 +08002031bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
Ming Lei1fd40b52020-06-30 18:25:00 +08002032 unsigned int nr_budgets)
Jens Axboef04c3df2016-12-07 08:41:17 -07002033{
Ming Lei75383522020-06-30 18:24:58 +08002034 enum prep_dispatch prep;
Ming Lei445874e2020-06-30 18:24:57 +08002035 struct request_queue *q = hctx->queue;
Kemeng Shif1ce99f2023-01-18 17:37:24 +08002036 struct request *rq;
Kemeng Shi4ea58fe2023-01-18 17:37:23 +08002037 int queued;
Ming Lei86ff7c22018-01-30 22:04:57 -05002038 blk_status_t ret = BLK_STS_OK;
Keith Busch0512a752020-05-12 17:55:47 +09002039 LIST_HEAD(zone_list);
Naohiro Aota9586e672021-10-27 01:51:27 +09002040 bool needs_resource = false;
Jens Axboef04c3df2016-12-07 08:41:17 -07002041
Omar Sandoval81380ca2017-04-07 08:56:26 -06002042 if (list_empty(list))
2043 return false;
2044
Jens Axboef04c3df2016-12-07 08:41:17 -07002045 /*
Jens Axboef04c3df2016-12-07 08:41:17 -07002046 * Now process all the entries, sending them to the driver.
2047 */
Kemeng Shi4ea58fe2023-01-18 17:37:23 +08002048 queued = 0;
Omar Sandoval81380ca2017-04-07 08:56:26 -06002049 do {
Jens Axboef04c3df2016-12-07 08:41:17 -07002050 struct blk_mq_queue_data bd;
2051
2052 rq = list_first_entry(list, struct request, queuelist);
Ming Lei0bca7992018-04-05 00:35:21 +08002053
Ming Lei445874e2020-06-30 18:24:57 +08002054 WARN_ON_ONCE(hctx != rq->mq_hctx);
Ming Lei1fd40b52020-06-30 18:25:00 +08002055 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
Ming Lei75383522020-06-30 18:24:58 +08002056 if (prep != PREP_DISPATCH_OK)
Ming Lei0bca7992018-04-05 00:35:21 +08002057 break;
Ming Leide148292017-10-14 17:22:29 +08002058
Jens Axboef04c3df2016-12-07 08:41:17 -07002059 list_del_init(&rq->queuelist);
2060
2061 bd.rq = rq;
Kemeng Shif1ce99f2023-01-18 17:37:24 +08002062 bd.last = list_empty(list);
Jens Axboef04c3df2016-12-07 08:41:17 -07002063
Ming Lei1fd40b52020-06-30 18:25:00 +08002064 /*
2065 * once the request is queued to lld, no need to cover the
2066 * budget any more
2067 */
2068 if (nr_budgets)
2069 nr_budgets--;
Jens Axboef04c3df2016-12-07 08:41:17 -07002070 ret = q->mq_ops->queue_rq(hctx, &bd);
Ming Lei7bf13722020-07-01 21:58:57 +08002071 switch (ret) {
2072 case BLK_STS_OK:
2073 queued++;
Jens Axboef04c3df2016-12-07 08:41:17 -07002074 break;
Ming Lei7bf13722020-07-01 21:58:57 +08002075 case BLK_STS_RESOURCE:
Naohiro Aota9586e672021-10-27 01:51:27 +09002076 needs_resource = true;
2077 fallthrough;
Ming Lei7bf13722020-07-01 21:58:57 +08002078 case BLK_STS_DEV_RESOURCE:
2079 blk_mq_handle_dev_resource(rq, list);
2080 goto out;
2081 case BLK_STS_ZONE_RESOURCE:
Keith Busch0512a752020-05-12 17:55:47 +09002082 /*
2083 * Move the request to zone_list and keep going through
2084 * the dispatch list to find more requests the drive can
2085 * accept.
2086 */
2087 blk_mq_handle_zone_resource(rq, &zone_list);
Naohiro Aota9586e672021-10-27 01:51:27 +09002088 needs_resource = true;
Ming Lei7bf13722020-07-01 21:58:57 +08002089 break;
2090 default:
Hannes Reineckee21ee5a2020-09-30 10:02:53 +02002091 blk_mq_end_request(rq, ret);
Christoph Hellwigfc17b652017-06-03 09:38:05 +02002092 }
Omar Sandoval81380ca2017-04-07 08:56:26 -06002093 } while (!list_empty(list));
Ming Lei7bf13722020-07-01 21:58:57 +08002094out:
Keith Busch0512a752020-05-12 17:55:47 +09002095 if (!list_empty(&zone_list))
2096 list_splice_tail_init(&zone_list, list);
2097
yangerkun632bfb62020-09-05 19:25:56 +08002098 /* If we didn't flush the entire list, we could have told the driver
2099 * there was more coming, but that turned out to be a lie.
2100 */
Kemeng Shie4ef2e02023-01-18 17:37:22 +08002101 if (!list_empty(list) || ret != BLK_STS_OK)
2102 blk_mq_commit_rqs(hctx, queued, false);
2103
Jens Axboef04c3df2016-12-07 08:41:17 -07002104 /*
2105 * Any items that need requeuing? Stuff them into hctx->dispatch,
2106 * that is where we will continue on next queue run.
2107 */
2108 if (!list_empty(list)) {
Ming Lei86ff7c22018-01-30 22:04:57 -05002109 bool needs_restart;
Ming Lei75383522020-06-30 18:24:58 +08002110 /* For non-shared tags, the RESTART check will suffice */
2111 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
Kemeng Shi47df9ce2023-01-18 17:37:16 +08002112 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2113 blk_mq_is_shared_tags(hctx->flags));
Ming Lei86ff7c22018-01-30 22:04:57 -05002114
Ming Lei2a5a24a2021-01-22 10:33:12 +08002115 if (nr_budgets)
2116 blk_mq_release_budgets(q, list);
Jens Axboef04c3df2016-12-07 08:41:17 -07002117
2118 spin_lock(&hctx->lock);
Ming Lei01e99ae2020-02-25 09:04:32 +08002119 list_splice_tail_init(list, &hctx->dispatch);
Jens Axboef04c3df2016-12-07 08:41:17 -07002120 spin_unlock(&hctx->lock);
2121
2122 /*
Ming Leid7d85352020-08-17 18:01:15 +08002123 * Order adding requests to hctx->dispatch and checking
2124 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2125 * in blk_mq_sched_restart(). Avoid restart code path to
2126 * miss the new added requests to hctx->dispatch, meantime
2127 * SCHED_RESTART is observed here.
2128 */
2129 smp_mb();
2130
2131 /*
Bart Van Assche710c7852017-04-07 11:16:51 -07002132 * If SCHED_RESTART was set by the caller of this function and
2133 * it is no longer set that means that it was cleared by another
2134 * thread and hence that a queue rerun is needed.
Jens Axboef04c3df2016-12-07 08:41:17 -07002135 *
Jens Axboeeb619fd2017-11-09 08:32:43 -07002136 * If 'no_tag' is set, that means that we failed getting
2137 * a driver tag with an I/O scheduler attached. If our dispatch
2138 * waitqueue is no longer active, ensure that we run the queue
2139 * AFTER adding our entries back to the list.
Jens Axboebd166ef2017-01-17 06:03:22 -07002140 *
Bart Van Assche710c7852017-04-07 11:16:51 -07002141 * If no I/O scheduler has been configured it is possible that
2142 * the hardware queue got stopped and restarted before requests
2143 * were pushed back onto the dispatch list. Rerun the queue to
2144 * avoid starvation. Notes:
2145 * - blk_mq_run_hw_queue() checks whether or not a queue has
2146 * been stopped before rerunning a queue.
2147 * - Some but not all block drivers stop a queue before
Christoph Hellwigfc17b652017-06-03 09:38:05 +02002148 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
Bart Van Assche710c7852017-04-07 11:16:51 -07002149 * and dm-rq.
Ming Lei86ff7c22018-01-30 22:04:57 -05002150 *
2151 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2152 * bit is set, run queue after a delay to avoid IO stalls
Douglas Andersonab3cee32020-04-20 09:24:51 -07002153 * that could otherwise occur if the queue is idle. We'll do
Naohiro Aota9586e672021-10-27 01:51:27 +09002154 * similar if we couldn't get budget or couldn't lock a zone
2155 * and SCHED_RESTART is set.
Jens Axboebd166ef2017-01-17 06:03:22 -07002156 */
Ming Lei86ff7c22018-01-30 22:04:57 -05002157 needs_restart = blk_mq_sched_needs_restart(hctx);
Naohiro Aota9586e672021-10-27 01:51:27 +09002158 if (prep == PREP_DISPATCH_NO_BUDGET)
2159 needs_resource = true;
Ming Lei86ff7c22018-01-30 22:04:57 -05002160 if (!needs_restart ||
Jens Axboeeb619fd2017-11-09 08:32:43 -07002161 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
Jens Axboebd166ef2017-01-17 06:03:22 -07002162 blk_mq_run_hw_queue(hctx, true);
Miaohe Lin6d5e8d22022-09-05 18:19:50 +08002163 else if (needs_resource)
Ming Lei86ff7c22018-01-30 22:04:57 -05002164 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
Jens Axboe1f57f8d2018-06-28 11:54:01 -06002165
Ming Lei6e7687172018-07-03 09:03:16 -06002166 blk_mq_update_dispatch_busy(hctx, true);
Jens Axboe1f57f8d2018-06-28 11:54:01 -06002167 return false;
Kemeng Shi4ea58fe2023-01-18 17:37:23 +08002168 }
Jens Axboef04c3df2016-12-07 08:41:17 -07002169
Kemeng Shi4ea58fe2023-01-18 17:37:23 +08002170 blk_mq_update_dispatch_busy(hctx, false);
2171 return true;
Jens Axboef04c3df2016-12-07 08:41:17 -07002172}
2173
Ming Leif82ddf12018-04-08 17:48:10 +08002174static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2175{
2176 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2177
2178 if (cpu >= nr_cpu_ids)
2179 cpu = cpumask_first(hctx->cpumask);
2180 return cpu;
2181}
2182
Jens Axboe506e9312014-05-07 10:26:44 -06002183/*
2184 * It'd be great if the workqueue API had a way to pass
2185 * in a mask and had some smarts for more clever placement.
2186 * For now we just round-robin here, switching for every
2187 * BLK_MQ_CPU_WORK_BATCH queued items.
2188 */
2189static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2190{
Ming Lei7bed4592018-01-18 00:41:51 +08002191 bool tried = false;
Ming Lei476f8c92018-04-08 17:48:09 +08002192 int next_cpu = hctx->next_cpu;
Ming Lei7bed4592018-01-18 00:41:51 +08002193
Christoph Hellwigb657d7e2014-11-24 09:27:23 +01002194 if (hctx->queue->nr_hw_queues == 1)
2195 return WORK_CPU_UNBOUND;
Jens Axboe506e9312014-05-07 10:26:44 -06002196
2197 if (--hctx->next_cpu_batch <= 0) {
Ming Lei7bed4592018-01-18 00:41:51 +08002198select_cpu:
Ming Lei476f8c92018-04-08 17:48:09 +08002199 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
Christoph Hellwig20e4d8132018-01-12 10:53:06 +08002200 cpu_online_mask);
Jens Axboe506e9312014-05-07 10:26:44 -06002201 if (next_cpu >= nr_cpu_ids)
Ming Leif82ddf12018-04-08 17:48:10 +08002202 next_cpu = blk_mq_first_mapped_cpu(hctx);
Jens Axboe506e9312014-05-07 10:26:44 -06002203 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2204 }
2205
Ming Lei7bed4592018-01-18 00:41:51 +08002206 /*
2207 * Do unbound schedule if we can't find a online CPU for this hctx,
2208 * and it should only happen in the path of handling CPU DEAD.
2209 */
Ming Lei476f8c92018-04-08 17:48:09 +08002210 if (!cpu_online(next_cpu)) {
Ming Lei7bed4592018-01-18 00:41:51 +08002211 if (!tried) {
2212 tried = true;
2213 goto select_cpu;
2214 }
2215
2216 /*
2217 * Make sure to re-select CPU next time once after CPUs
2218 * in hctx->cpumask become online again.
2219 */
Ming Lei476f8c92018-04-08 17:48:09 +08002220 hctx->next_cpu = next_cpu;
Ming Lei7bed4592018-01-18 00:41:51 +08002221 hctx->next_cpu_batch = 1;
2222 return WORK_CPU_UNBOUND;
2223 }
Ming Lei476f8c92018-04-08 17:48:09 +08002224
2225 hctx->next_cpu = next_cpu;
2226 return next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -06002227}
2228
André Almeida105663f2020-01-06 15:08:18 -03002229/**
André Almeida105663f2020-01-06 15:08:18 -03002230 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2231 * @hctx: Pointer to the hardware queue to run.
Minwoo Imfa94ba82020-12-05 00:20:55 +09002232 * @msecs: Milliseconds of delay to wait before running the queue.
André Almeida105663f2020-01-06 15:08:18 -03002233 *
2234 * Run a hardware queue asynchronously with a delay of @msecs.
2235 */
Bart Van Assche7587a5a2017-04-07 11:16:52 -07002236void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2237{
Christoph Hellwig1aa8d872023-04-13 08:06:50 +02002238 if (unlikely(blk_mq_hctx_stopped(hctx)))
2239 return;
2240 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2241 msecs_to_jiffies(msecs));
Bart Van Assche7587a5a2017-04-07 11:16:52 -07002242}
2243EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2244
André Almeida105663f2020-01-06 15:08:18 -03002245/**
2246 * blk_mq_run_hw_queue - Start to run a hardware queue.
2247 * @hctx: Pointer to the hardware queue to run.
2248 * @async: If we want to run the queue asynchronously.
2249 *
2250 * Check if the request queue is not in a quiesced state and if there are
2251 * pending requests to be sent. If this is true, run the queue to send requests
2252 * to hardware.
2253 */
John Garry626fb732019-10-30 00:59:30 +08002254void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
Bart Van Assche7587a5a2017-04-07 11:16:52 -07002255{
Ming Lei24f5a902018-01-06 16:27:38 +08002256 bool need_run;
2257
2258 /*
Christoph Hellwig4d5bba52023-04-13 08:06:51 +02002259 * We can't run the queue inline with interrupts disabled.
2260 */
2261 WARN_ON_ONCE(!async && in_interrupt());
2262
Bart Van Assche65a558f2023-07-21 10:27:30 -07002263 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2264
Christoph Hellwig4d5bba52023-04-13 08:06:51 +02002265 /*
Ming Lei24f5a902018-01-06 16:27:38 +08002266 * When queue is quiesced, we may be switching io scheduler, or
2267 * updating nr_hw_queues, or other things, and we can't run queue
2268 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2269 *
2270 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2271 * quiesced.
2272 */
Ming Lei41adf532021-12-06 19:12:13 +08002273 __blk_mq_run_dispatch_ops(hctx->queue, false,
Ming Lei2a904d02021-12-03 21:15:31 +08002274 need_run = !blk_queue_quiesced(hctx->queue) &&
2275 blk_mq_hctx_has_pending(hctx));
Ming Lei24f5a902018-01-06 16:27:38 +08002276
Christoph Hellwig1aa8d872023-04-13 08:06:50 +02002277 if (!need_run)
2278 return;
2279
Bart Van Assche65a558f2023-07-21 10:27:30 -07002280 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
Christoph Hellwig1aa8d872023-04-13 08:06:50 +02002281 blk_mq_delay_run_hw_queue(hctx, 0);
2282 return;
2283 }
2284
Christoph Hellwig4d5bba52023-04-13 08:06:51 +02002285 blk_mq_run_dispatch_ops(hctx->queue,
2286 blk_mq_sched_dispatch_requests(hctx));
Jens Axboe320ae512013-10-24 09:20:05 +01002287}
Omar Sandoval5b727272017-04-14 01:00:00 -07002288EXPORT_SYMBOL(blk_mq_run_hw_queue);
Jens Axboe320ae512013-10-24 09:20:05 +01002289
Jan Karab6e68ee2021-01-11 17:47:17 +01002290/*
Jan Karab6e68ee2021-01-11 17:47:17 +01002291 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2292 * scheduler.
2293 */
2294static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2295{
Ming Lei5d054262022-05-22 20:23:50 +08002296 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
Jan Karab6e68ee2021-01-11 17:47:17 +01002297 /*
2298 * If the IO scheduler does not respect hardware queues when
2299 * dispatching, we just don't bother with multiple HW queues and
2300 * dispatch from hctx for the current CPU since running multiple queues
2301 * just causes lock contention inside the scheduler and pointless cache
2302 * bouncing.
2303 */
Bart Van Assche51ab80f2022-06-15 15:55:49 -07002304 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
Ming Lei5d054262022-05-22 20:23:50 +08002305
Jan Karab6e68ee2021-01-11 17:47:17 +01002306 if (!blk_mq_hctx_stopped(hctx))
2307 return hctx;
2308 return NULL;
2309}
2310
André Almeida105663f2020-01-06 15:08:18 -03002311/**
Mauro Carvalho Chehab24f7bb82020-10-23 18:32:54 +02002312 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
André Almeida105663f2020-01-06 15:08:18 -03002313 * @q: Pointer to the request queue to run.
2314 * @async: If we want to run the queue asynchronously.
2315 */
Mike Snitzerb94ec292015-03-11 23:56:38 -04002316void blk_mq_run_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01002317{
Jan Karab6e68ee2021-01-11 17:47:17 +01002318 struct blk_mq_hw_ctx *hctx, *sq_hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002319 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01002320
Jan Karab6e68ee2021-01-11 17:47:17 +01002321 sq_hctx = NULL;
Ming Lei4d337ceb2022-06-16 09:44:00 +08002322 if (blk_queue_sq_sched(q))
Jan Karab6e68ee2021-01-11 17:47:17 +01002323 sq_hctx = blk_mq_get_sq_hctx(q);
Jens Axboe320ae512013-10-24 09:20:05 +01002324 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboe79f720a2017-11-10 09:13:21 -07002325 if (blk_mq_hctx_stopped(hctx))
Jens Axboe320ae512013-10-24 09:20:05 +01002326 continue;
Jan Karab6e68ee2021-01-11 17:47:17 +01002327 /*
2328 * 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_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +01002335 }
2336}
Mike Snitzerb94ec292015-03-11 23:56:38 -04002337EXPORT_SYMBOL(blk_mq_run_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +01002338
Bart Van Asschefd001442016-10-28 17:19:37 -07002339/**
Douglas Andersonb9151e72020-04-20 09:24:52 -07002340 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2341 * @q: Pointer to the request queue to run.
Minwoo Imfa94ba82020-12-05 00:20:55 +09002342 * @msecs: Milliseconds of delay to wait before running the queues.
Douglas Andersonb9151e72020-04-20 09:24:52 -07002343 */
2344void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2345{
Jan Karab6e68ee2021-01-11 17:47:17 +01002346 struct blk_mq_hw_ctx *hctx, *sq_hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002347 unsigned long i;
Douglas Andersonb9151e72020-04-20 09:24:52 -07002348
Jan Karab6e68ee2021-01-11 17:47:17 +01002349 sq_hctx = NULL;
Ming Lei4d337ceb2022-06-16 09:44:00 +08002350 if (blk_queue_sq_sched(q))
Jan Karab6e68ee2021-01-11 17:47:17 +01002351 sq_hctx = blk_mq_get_sq_hctx(q);
Douglas Andersonb9151e72020-04-20 09:24:52 -07002352 queue_for_each_hw_ctx(q, hctx, i) {
2353 if (blk_mq_hctx_stopped(hctx))
2354 continue;
Jan Karab6e68ee2021-01-11 17:47:17 +01002355 /*
David Jeffery8f5fea62022-01-31 15:33:37 -05002356 * If there is already a run_work pending, leave the
2357 * pending delay untouched. Otherwise, a hctx can stall
2358 * if another hctx is re-delaying the other's work
2359 * before the work executes.
2360 */
2361 if (delayed_work_pending(&hctx->run_work))
2362 continue;
2363 /*
Jan Karab6e68ee2021-01-11 17:47:17 +01002364 * Dispatch from this hctx either if there's no hctx preferred
2365 * by IO scheduler or if it has requests that bypass the
2366 * scheduler.
2367 */
2368 if (!sq_hctx || sq_hctx == hctx ||
2369 !list_empty_careful(&hctx->dispatch))
2370 blk_mq_delay_run_hw_queue(hctx, msecs);
Douglas Andersonb9151e72020-04-20 09:24:52 -07002371 }
2372}
2373EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2374
Ming Lei39a70c72017-06-06 23:22:09 +08002375/*
2376 * This function is often used for pausing .queue_rq() by driver when
2377 * there isn't enough resource or some conditions aren't satisfied, and
Bart Van Assche4d606212017-08-17 16:23:00 -07002378 * BLK_STS_RESOURCE is usually returned.
Ming Lei39a70c72017-06-06 23:22:09 +08002379 *
2380 * We do not guarantee that dispatch can be drained or blocked
2381 * after blk_mq_stop_hw_queue() returns. Please use
2382 * blk_mq_quiesce_queue() for that requirement.
2383 */
Jens Axboe320ae512013-10-24 09:20:05 +01002384void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2385{
Ming Lei641a9ed2017-06-06 23:22:10 +08002386 cancel_delayed_work(&hctx->run_work);
2387
2388 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboe320ae512013-10-24 09:20:05 +01002389}
2390EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2391
Ming Lei39a70c72017-06-06 23:22:09 +08002392/*
2393 * This function is often used for pausing .queue_rq() by driver when
2394 * there isn't enough resource or some conditions aren't satisfied, and
Bart Van Assche4d606212017-08-17 16:23:00 -07002395 * BLK_STS_RESOURCE is usually returned.
Ming Lei39a70c72017-06-06 23:22:09 +08002396 *
2397 * We do not guarantee that dispatch can be drained or blocked
2398 * after blk_mq_stop_hw_queues() returns. Please use
2399 * blk_mq_quiesce_queue() for that requirement.
2400 */
Jens Axboe2719aa22017-05-03 11:08:14 -06002401void blk_mq_stop_hw_queues(struct request_queue *q)
2402{
Ming Lei641a9ed2017-06-06 23:22:10 +08002403 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002404 unsigned long i;
Ming Lei641a9ed2017-06-06 23:22:10 +08002405
2406 queue_for_each_hw_ctx(q, hctx, i)
2407 blk_mq_stop_hw_queue(hctx);
Christoph Hellwig280d45f2013-10-25 14:45:58 +01002408}
2409EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2410
Jens Axboe320ae512013-10-24 09:20:05 +01002411void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2412{
2413 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboee4043dc2014-04-09 10:18:23 -06002414
Bart Van Assche65a558f2023-07-21 10:27:30 -07002415 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
Jens Axboe320ae512013-10-24 09:20:05 +01002416}
2417EXPORT_SYMBOL(blk_mq_start_hw_queue);
2418
Christoph Hellwig2f268552014-04-16 09:44:56 +02002419void blk_mq_start_hw_queues(struct request_queue *q)
2420{
2421 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002422 unsigned long i;
Christoph Hellwig2f268552014-04-16 09:44:56 +02002423
2424 queue_for_each_hw_ctx(q, hctx, i)
2425 blk_mq_start_hw_queue(hctx);
2426}
2427EXPORT_SYMBOL(blk_mq_start_hw_queues);
2428
Jens Axboeae911c52016-12-08 13:19:30 -07002429void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2430{
2431 if (!blk_mq_hctx_stopped(hctx))
2432 return;
2433
2434 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2435 blk_mq_run_hw_queue(hctx, async);
2436}
2437EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2438
Christoph Hellwig1b4a3252014-04-16 09:44:54 +02002439void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01002440{
2441 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002442 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01002443
Jens Axboeae911c52016-12-08 13:19:30 -07002444 queue_for_each_hw_ctx(q, hctx, i)
Bart Van Assche65a558f2023-07-21 10:27:30 -07002445 blk_mq_start_stopped_hw_queue(hctx, async ||
2446 (hctx->flags & BLK_MQ_F_BLOCKING));
Jens Axboe320ae512013-10-24 09:20:05 +01002447}
2448EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2449
Christoph Hellwig70f4db62014-04-16 10:48:08 -06002450static void blk_mq_run_work_fn(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +01002451{
Christoph Hellwigc20a1a22023-04-13 08:06:48 +02002452 struct blk_mq_hw_ctx *hctx =
2453 container_of(work, struct blk_mq_hw_ctx, run_work.work);
Jens Axboe320ae512013-10-24 09:20:05 +01002454
Christoph Hellwig4d5bba52023-04-13 08:06:51 +02002455 blk_mq_run_dispatch_ops(hctx->queue,
2456 blk_mq_sched_dispatch_requests(hctx));
Jens Axboe320ae512013-10-24 09:20:05 +01002457}
2458
André Almeida105663f2020-01-06 15:08:18 -03002459/**
2460 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2461 * @rq: Pointer to request to be inserted.
Christoph Hellwig2b597612023-04-13 08:40:55 +02002462 * @flags: BLK_MQ_INSERT_*
André Almeida105663f2020-01-06 15:08:18 -03002463 *
Jens Axboe157f3772017-09-11 16:43:57 -06002464 * Should only be used carefully, when the caller knows we want to
2465 * bypass a potential IO scheduler on the target device.
2466 */
Christoph Hellwig360f2642023-05-19 06:40:46 +02002467static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
Jens Axboe157f3772017-09-11 16:43:57 -06002468{
Jens Axboeea4f9952018-10-29 15:06:13 -06002469 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Jens Axboe157f3772017-09-11 16:43:57 -06002470
2471 spin_lock(&hctx->lock);
Christoph Hellwig2b597612023-04-13 08:40:55 +02002472 if (flags & BLK_MQ_INSERT_AT_HEAD)
Ming Lei01e99ae2020-02-25 09:04:32 +08002473 list_add(&rq->queuelist, &hctx->dispatch);
2474 else
2475 list_add_tail(&rq->queuelist, &hctx->dispatch);
Jens Axboe157f3772017-09-11 16:43:57 -06002476 spin_unlock(&hctx->lock);
Jens Axboe157f3772017-09-11 16:43:57 -06002477}
2478
Christoph Hellwig05a93112023-04-13 08:40:42 +02002479static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2480 struct blk_mq_ctx *ctx, struct list_head *list,
2481 bool run_queue_async)
Jens Axboe320ae512013-10-24 09:20:05 +01002482{
Ming Lei3f0cedc2018-07-02 17:35:58 +08002483 struct request *rq;
Ming Leic16d6b52018-12-17 08:44:05 -07002484 enum hctx_type type = hctx->type;
Ming Lei3f0cedc2018-07-02 17:35:58 +08002485
Jens Axboe320ae512013-10-24 09:20:05 +01002486 /*
Christoph Hellwig94aa2282023-04-13 08:40:41 +02002487 * Try to issue requests directly if the hw queue isn't busy to save an
2488 * extra enqueue & dequeue to the sw queue.
2489 */
2490 if (!hctx->dispatch_busy && !run_queue_async) {
2491 blk_mq_run_dispatch_ops(hctx->queue,
2492 blk_mq_try_issue_list_directly(hctx, list));
2493 if (list_empty(list))
2494 goto out;
2495 }
2496
2497 /*
Jens Axboe320ae512013-10-24 09:20:05 +01002498 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2499 * offline now
2500 */
Ming Lei3f0cedc2018-07-02 17:35:58 +08002501 list_for_each_entry(rq, list, queuelist) {
Jens Axboee57690f2016-08-24 15:34:35 -06002502 BUG_ON(rq->mq_ctx != ctx);
Christoph Hellwiga54895f2020-12-03 17:21:39 +01002503 trace_block_rq_insert(rq);
Bart Van Assche65a558f2023-07-21 10:27:30 -07002504 if (rq->cmd_flags & REQ_NOWAIT)
2505 run_queue_async = true;
Jens Axboe320ae512013-10-24 09:20:05 +01002506 }
Ming Lei3f0cedc2018-07-02 17:35:58 +08002507
2508 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07002509 list_splice_tail_init(list, &ctx->rq_lists[type]);
Ming Leicfd0c552015-10-20 23:13:57 +08002510 blk_mq_hctx_mark_pending(hctx, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01002511 spin_unlock(&ctx->lock);
Christoph Hellwig94aa2282023-04-13 08:40:41 +02002512out:
2513 blk_mq_run_hw_queue(hctx, run_queue_async);
Jens Axboe320ae512013-10-24 09:20:05 +01002514}
2515
Christoph Hellwig710fa372023-04-13 08:40:54 +02002516static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002517{
2518 struct request_queue *q = rq->q;
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002519 struct blk_mq_ctx *ctx = rq->mq_ctx;
2520 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2521
Christoph Hellwig53548d22023-04-13 08:40:47 +02002522 if (blk_rq_is_passthrough(rq)) {
2523 /*
2524 * Passthrough request have to be added to hctx->dispatch
2525 * directly. The device may be in a situation where it can't
2526 * handle FS request, and always returns BLK_STS_RESOURCE for
2527 * them, which gets them added to hctx->dispatch.
2528 *
2529 * If a passthrough request is required to unblock the queues,
2530 * and it is added to the scheduler queue, there is no chance to
2531 * dispatch it given we prioritize requests in hctx->dispatch.
2532 */
Christoph Hellwig2b597612023-04-13 08:40:55 +02002533 blk_mq_request_bypass_insert(rq, flags);
Bart Van Asschebe4c4272023-05-19 06:40:47 +02002534 } else if (req_op(rq) == REQ_OP_FLUSH) {
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002535 /*
2536 * Firstly normal IO request is inserted to scheduler queue or
2537 * sw queue, meantime we add flush request to dispatch queue(
2538 * hctx->dispatch) directly and there is at most one in-flight
2539 * flush request for each hw queue, so it doesn't matter to add
2540 * flush request to tail or front of the dispatch queue.
2541 *
2542 * Secondly in case of NCQ, flush request belongs to non-NCQ
2543 * command, and queueing it will fail when there is any
2544 * in-flight normal IO request(NCQ command). When adding flush
2545 * rq to the front of hctx->dispatch, it is easier to introduce
2546 * extra time to flush rq's latency because of S_SCHED_RESTART
2547 * compared with adding to the tail of dispatch queue, then
2548 * chance of flush merge is increased, and less flush requests
2549 * will be issued to controller. It is observed that ~10% time
2550 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2551 * drive when adding flush rq to the front of hctx->dispatch.
2552 *
2553 * Simply queue flush rq to the front of hctx->dispatch so that
2554 * intensive flush workloads can benefit in case of NCQ HW.
2555 */
Christoph Hellwig2b597612023-04-13 08:40:55 +02002556 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
Christoph Hellwig53548d22023-04-13 08:40:47 +02002557 } else if (q->elevator) {
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002558 LIST_HEAD(list);
2559
Christoph Hellwig53548d22023-04-13 08:40:47 +02002560 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2561
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002562 list_add(&rq->queuelist, &list);
Christoph Hellwig93fffe12023-04-13 08:40:56 +02002563 q->elevator->type->ops.insert_requests(hctx, &list, flags);
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002564 } else {
Christoph Hellwig4ec5c052023-04-13 08:40:45 +02002565 trace_block_rq_insert(rq);
2566
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002567 spin_lock(&ctx->lock);
Christoph Hellwig710fa372023-04-13 08:40:54 +02002568 if (flags & BLK_MQ_INSERT_AT_HEAD)
Christoph Hellwig4ec5c052023-04-13 08:40:45 +02002569 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2570 else
2571 list_add_tail(&rq->queuelist,
2572 &ctx->rq_lists[hctx->type]);
Christoph Hellwiga88db1e2023-04-13 08:40:44 +02002573 blk_mq_hctx_mark_pending(hctx, ctx);
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002574 spin_unlock(&ctx->lock);
2575 }
Jens Axboe320ae512013-10-24 09:20:05 +01002576}
2577
Christoph Hellwig14ccb662019-06-06 12:29:01 +02002578static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2579 unsigned int nr_segs)
Jens Axboe320ae512013-10-24 09:20:05 +01002580{
Eric Biggers93f221a2020-09-15 20:53:14 -07002581 int err;
2582
Christoph Hellwigf924cdd2019-06-06 12:29:00 +02002583 if (bio->bi_opf & REQ_RAHEAD)
2584 rq->cmd_flags |= REQ_FAILFAST_MASK;
2585
2586 rq->__sector = bio->bi_iter.bi_sector;
Christoph Hellwig14ccb662019-06-06 12:29:01 +02002587 blk_rq_bio_prep(rq, bio, nr_segs);
Eric Biggers93f221a2020-09-15 20:53:14 -07002588
2589 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2590 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2591 WARN_ON_ONCE(err);
Jens Axboe4b570522014-05-29 11:00:11 -06002592
Konstantin Khlebnikovb5af37a2020-05-27 07:24:16 +02002593 blk_account_io_start(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01002594}
2595
Mike Snitzer0f955492018-01-17 11:25:56 -05002596static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig3e087732021-10-12 13:12:24 +02002597 struct request *rq, bool last)
Shaohua Lif984df12015-05-08 10:51:32 -07002598{
Shaohua Lif984df12015-05-08 10:51:32 -07002599 struct request_queue *q = rq->q;
Shaohua Lif984df12015-05-08 10:51:32 -07002600 struct blk_mq_queue_data bd = {
2601 .rq = rq,
Jens Axboebe94f052018-11-24 10:15:46 -07002602 .last = last,
Shaohua Lif984df12015-05-08 10:51:32 -07002603 };
Jens Axboef06345a2017-06-12 11:22:46 -06002604 blk_status_t ret;
Mike Snitzer0f955492018-01-17 11:25:56 -05002605
Mike Snitzer0f955492018-01-17 11:25:56 -05002606 /*
2607 * For OK queue, we are done. For error, caller may kill it.
2608 * Any other error (busy), just add it to our list as we
2609 * previously would have done.
2610 */
2611 ret = q->mq_ops->queue_rq(hctx, &bd);
2612 switch (ret) {
2613 case BLK_STS_OK:
Ming Lei6ce3dd62018-07-10 09:03:31 +08002614 blk_mq_update_dispatch_busy(hctx, false);
Mike Snitzer0f955492018-01-17 11:25:56 -05002615 break;
2616 case BLK_STS_RESOURCE:
Ming Lei86ff7c22018-01-30 22:04:57 -05002617 case BLK_STS_DEV_RESOURCE:
Ming Lei6ce3dd62018-07-10 09:03:31 +08002618 blk_mq_update_dispatch_busy(hctx, true);
Mike Snitzer0f955492018-01-17 11:25:56 -05002619 __blk_mq_requeue_request(rq);
2620 break;
2621 default:
Ming Lei6ce3dd62018-07-10 09:03:31 +08002622 blk_mq_update_dispatch_busy(hctx, false);
Mike Snitzer0f955492018-01-17 11:25:56 -05002623 break;
2624 }
2625
2626 return ret;
2627}
2628
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002629static bool blk_mq_get_budget_and_tag(struct request *rq)
Mike Snitzer0f955492018-01-17 11:25:56 -05002630{
Ming Lei2a5a24a2021-01-22 10:33:12 +08002631 int budget_token;
Ming Leid964f042017-06-06 23:22:00 +08002632
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002633 budget_token = blk_mq_get_dispatch_budget(rq->q);
Ming Lei2a5a24a2021-01-22 10:33:12 +08002634 if (budget_token < 0)
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002635 return false;
Ming Lei2a5a24a2021-01-22 10:33:12 +08002636 blk_mq_set_rq_budget_token(rq, budget_token);
Ming Lei8ab6bb9e2018-06-25 19:31:45 +08002637 if (!blk_mq_get_driver_tag(rq)) {
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002638 blk_mq_put_dispatch_budget(rq->q, budget_token);
2639 return false;
Ming Lei88022d72017-11-05 02:21:12 +08002640 }
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002641 return true;
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002642}
2643
André Almeida105663f2020-01-06 15:08:18 -03002644/**
2645 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2646 * @hctx: Pointer of the associated hardware queue.
2647 * @rq: Pointer to request to be sent.
André Almeida105663f2020-01-06 15:08:18 -03002648 *
2649 * If the device has enough resources to accept a new request now, send the
2650 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2651 * we can try send it another time in the future. Requests inserted at this
2652 * queue have higher priority.
2653 */
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002654static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig3e087732021-10-12 13:12:24 +02002655 struct request *rq)
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002656{
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002657 blk_status_t ret;
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002658
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002659 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
Christoph Hellwig710fa372023-04-13 08:40:54 +02002660 blk_mq_insert_request(rq, 0);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002661 return;
2662 }
2663
Christoph Hellwigdd6216b2023-05-18 07:31:01 +02002664 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
Christoph Hellwig710fa372023-04-13 08:40:54 +02002665 blk_mq_insert_request(rq, 0);
Bart Van Assche65a558f2023-07-21 10:27:30 -07002666 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002667 return;
2668 }
2669
2670 ret = __blk_mq_issue_directly(hctx, rq, true);
2671 switch (ret) {
2672 case BLK_STS_OK:
2673 break;
2674 case BLK_STS_RESOURCE:
2675 case BLK_STS_DEV_RESOURCE:
Christoph Hellwig2b597612023-04-13 08:40:55 +02002676 blk_mq_request_bypass_insert(rq, 0);
Christoph Hellwig23943952023-04-13 08:40:52 +02002677 blk_mq_run_hw_queue(hctx, false);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002678 break;
2679 default:
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002680 blk_mq_end_request(rq, ret);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002681 break;
2682 }
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002683}
2684
Christoph Hellwig06c8c692021-11-17 07:13:58 +01002685static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002686{
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002687 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2688
2689 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
Christoph Hellwig710fa372023-04-13 08:40:54 +02002690 blk_mq_insert_request(rq, 0);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002691 return BLK_STS_OK;
2692 }
2693
2694 if (!blk_mq_get_budget_and_tag(rq))
2695 return BLK_STS_RESOURCE;
2696 return __blk_mq_issue_directly(hctx, rq, last);
Christoph Hellwig5eb61262017-03-22 15:01:51 -04002697}
2698
Kemeng Shi3e368fb2023-01-18 17:37:18 +08002699static void blk_mq_plug_issue_direct(struct blk_plug *plug)
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002700{
2701 struct blk_mq_hw_ctx *hctx = NULL;
2702 struct request *rq;
2703 int queued = 0;
Kemeng Shi0d617a82023-01-18 17:37:20 +08002704 blk_status_t ret = BLK_STS_OK;
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002705
2706 while ((rq = rq_list_pop(&plug->mq_list))) {
2707 bool last = rq_list_empty(plug->mq_list);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002708
2709 if (hctx != rq->mq_hctx) {
Kemeng Shi34c9f5472023-01-18 17:37:19 +08002710 if (hctx) {
2711 blk_mq_commit_rqs(hctx, queued, false);
2712 queued = 0;
2713 }
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002714 hctx = rq->mq_hctx;
2715 }
2716
2717 ret = blk_mq_request_issue_directly(rq, last);
2718 switch (ret) {
2719 case BLK_STS_OK:
2720 queued++;
2721 break;
2722 case BLK_STS_RESOURCE:
2723 case BLK_STS_DEV_RESOURCE:
Christoph Hellwig2b597612023-04-13 08:40:55 +02002724 blk_mq_request_bypass_insert(rq, 0);
Christoph Hellwig23943952023-04-13 08:40:52 +02002725 blk_mq_run_hw_queue(hctx, false);
Kemeng Shi0d617a82023-01-18 17:37:20 +08002726 goto out;
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002727 default:
2728 blk_mq_end_request(rq, ret);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002729 break;
2730 }
2731 }
2732
Kemeng Shi0d617a82023-01-18 17:37:20 +08002733out:
2734 if (ret != BLK_STS_OK)
Kemeng Shi34c9f5472023-01-18 17:37:19 +08002735 blk_mq_commit_rqs(hctx, queued, false);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002736}
2737
Keith Busch518579a2021-12-20 12:59:19 -08002738static void __blk_mq_flush_plug_list(struct request_queue *q,
2739 struct blk_plug *plug)
2740{
2741 if (blk_queue_quiesced(q))
2742 return;
2743 q->mq_ops->queue_rqs(&plug->mq_list);
2744}
2745
Jens Axboe26fed4a2022-03-11 10:24:17 -07002746static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2747{
2748 struct blk_mq_hw_ctx *this_hctx = NULL;
2749 struct blk_mq_ctx *this_ctx = NULL;
2750 struct request *requeue_list = NULL;
Jan Kara34e0a272023-03-13 10:30:02 +01002751 struct request **requeue_lastp = &requeue_list;
Jens Axboe26fed4a2022-03-11 10:24:17 -07002752 unsigned int depth = 0;
Ming Leid97217e2023-05-18 07:30:59 +02002753 bool is_passthrough = false;
Jens Axboe26fed4a2022-03-11 10:24:17 -07002754 LIST_HEAD(list);
2755
2756 do {
2757 struct request *rq = rq_list_pop(&plug->mq_list);
2758
2759 if (!this_hctx) {
2760 this_hctx = rq->mq_hctx;
2761 this_ctx = rq->mq_ctx;
Ming Leid97217e2023-05-18 07:30:59 +02002762 is_passthrough = blk_rq_is_passthrough(rq);
2763 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2764 is_passthrough != blk_rq_is_passthrough(rq)) {
Jan Kara34e0a272023-03-13 10:30:02 +01002765 rq_list_add_tail(&requeue_lastp, rq);
Jens Axboe26fed4a2022-03-11 10:24:17 -07002766 continue;
2767 }
Jan Kara34e0a272023-03-13 10:30:02 +01002768 list_add(&rq->queuelist, &list);
Jens Axboe26fed4a2022-03-11 10:24:17 -07002769 depth++;
2770 } while (!rq_list_empty(plug->mq_list));
2771
2772 plug->mq_list = requeue_list;
2773 trace_block_unplug(this_hctx->queue, depth, !from_sched);
Christoph Hellwig05a93112023-04-13 08:40:42 +02002774
2775 percpu_ref_get(&this_hctx->queue->q_usage_counter);
Ming Leid97217e2023-05-18 07:30:59 +02002776 /* passthrough requests should never be issued to the I/O scheduler */
Ming Lei2293cae2023-06-21 21:22:08 +08002777 if (is_passthrough) {
2778 spin_lock(&this_hctx->lock);
2779 list_splice_tail_init(&list, &this_hctx->dispatch);
2780 spin_unlock(&this_hctx->lock);
2781 blk_mq_run_hw_queue(this_hctx, from_sched);
2782 } else if (this_hctx->queue->elevator) {
Christoph Hellwig05a93112023-04-13 08:40:42 +02002783 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
Christoph Hellwig93fffe12023-04-13 08:40:56 +02002784 &list, 0);
Christoph Hellwig05a93112023-04-13 08:40:42 +02002785 blk_mq_run_hw_queue(this_hctx, from_sched);
2786 } else {
2787 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2788 }
2789 percpu_ref_put(&this_hctx->queue->q_usage_counter);
Jens Axboe26fed4a2022-03-11 10:24:17 -07002790}
2791
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002792void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2793{
Jens Axboe3c67d442021-12-03 06:48:53 -07002794 struct request *rq;
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002795
Ross Lagerwall70904262023-07-14 11:11:06 +01002796 /*
2797 * We may have been called recursively midway through handling
2798 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2799 * To avoid mq_list changing under our feet, clear rq_count early and
2800 * bail out specifically if rq_count is 0 rather than checking
2801 * whether the mq_list is empty.
2802 */
2803 if (plug->rq_count == 0)
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002804 return;
2805 plug->rq_count = 0;
2806
2807 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
Jens Axboe3c67d442021-12-03 06:48:53 -07002808 struct request_queue *q;
2809
2810 rq = rq_list_peek(&plug->mq_list);
2811 q = rq->q;
2812
2813 /*
2814 * Peek first request and see if we have a ->queue_rqs() hook.
2815 * If we do, we can dispatch the whole plug list in one go. We
2816 * already know at this point that all requests belong to the
2817 * same queue, caller must ensure that's the case.
Jens Axboe3c67d442021-12-03 06:48:53 -07002818 */
Chengming Zhou434097ee2023-09-13 15:16:14 +00002819 if (q->mq_ops->queue_rqs) {
Jens Axboe3c67d442021-12-03 06:48:53 -07002820 blk_mq_run_dispatch_ops(q,
Keith Busch518579a2021-12-20 12:59:19 -08002821 __blk_mq_flush_plug_list(q, plug));
Jens Axboe3c67d442021-12-03 06:48:53 -07002822 if (rq_list_empty(plug->mq_list))
2823 return;
2824 }
Ming Lei73f37602021-12-06 11:33:50 +08002825
2826 blk_mq_run_dispatch_ops(q,
Kemeng Shi3e368fb2023-01-18 17:37:18 +08002827 blk_mq_plug_issue_direct(plug));
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002828 if (rq_list_empty(plug->mq_list))
2829 return;
2830 }
2831
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002832 do {
Jens Axboe26fed4a2022-03-11 10:24:17 -07002833 blk_mq_dispatch_plug_list(plug, from_schedule);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002834 } while (!rq_list_empty(plug->mq_list));
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002835}
2836
Christoph Hellwig94aa2282023-04-13 08:40:41 +02002837static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
Ming Lei6ce3dd62018-07-10 09:03:31 +08002838 struct list_head *list)
2839{
Keith Busch536167d42020-04-07 03:13:48 +09002840 int queued = 0;
Kemeng Shi984ce0a2023-01-18 17:37:21 +08002841 blk_status_t ret = BLK_STS_OK;
Keith Busch536167d42020-04-07 03:13:48 +09002842
Ming Lei6ce3dd62018-07-10 09:03:31 +08002843 while (!list_empty(list)) {
Ming Lei6ce3dd62018-07-10 09:03:31 +08002844 struct request *rq = list_first_entry(list, struct request,
2845 queuelist);
2846
2847 list_del_init(&rq->queuelist);
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002848 ret = blk_mq_request_issue_directly(rq, list_empty(list));
Kemeng Shi27e8b2b2023-01-18 17:37:25 +08002849 switch (ret) {
2850 case BLK_STS_OK:
Keith Busch536167d42020-04-07 03:13:48 +09002851 queued++;
Kemeng Shi27e8b2b2023-01-18 17:37:25 +08002852 break;
2853 case BLK_STS_RESOURCE:
2854 case BLK_STS_DEV_RESOURCE:
Christoph Hellwig2b597612023-04-13 08:40:55 +02002855 blk_mq_request_bypass_insert(rq, 0);
Christoph Hellwig23943952023-04-13 08:40:52 +02002856 if (list_empty(list))
2857 blk_mq_run_hw_queue(hctx, false);
Kemeng Shi27e8b2b2023-01-18 17:37:25 +08002858 goto out;
2859 default:
2860 blk_mq_end_request(rq, ret);
2861 break;
2862 }
Ming Lei6ce3dd62018-07-10 09:03:31 +08002863 }
Jens Axboed666ba92018-11-27 17:02:25 -07002864
Kemeng Shi27e8b2b2023-01-18 17:37:25 +08002865out:
Kemeng Shi984ce0a2023-01-18 17:37:21 +08002866 if (ret != BLK_STS_OK)
2867 blk_mq_commit_rqs(hctx, queued, false);
Ming Lei6ce3dd62018-07-10 09:03:31 +08002868}
2869
Ming Leib131f202021-11-11 16:51:34 +08002870static bool blk_mq_attempt_bio_merge(struct request_queue *q,
Christoph Hellwig0c5bcc92021-11-23 17:04:41 +01002871 struct bio *bio, unsigned int nr_segs)
Jens Axboe900e0802021-11-03 05:47:09 -06002872{
2873 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
Christoph Hellwig0c5bcc92021-11-23 17:04:41 +01002874 if (blk_attempt_plug_merge(q, bio, nr_segs))
Jens Axboe900e0802021-11-03 05:47:09 -06002875 return true;
2876 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2877 return true;
2878 }
2879 return false;
2880}
2881
Jens Axboe71539712021-11-03 05:52:45 -06002882static struct request *blk_mq_get_new_requests(struct request_queue *q,
2883 struct blk_plug *plug,
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002884 struct bio *bio,
2885 unsigned int nsegs)
Jens Axboe71539712021-11-03 05:52:45 -06002886{
2887 struct blk_mq_alloc_data data = {
2888 .q = q,
2889 .nr_tags = 1,
Ming Lei9d497e22022-01-04 21:42:23 +08002890 .cmd_flags = bio->bi_opf,
Jens Axboe71539712021-11-03 05:52:45 -06002891 };
2892 struct request *rq;
2893
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002894 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002895 return NULL;
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002896
2897 rq_qos_throttle(q, bio);
2898
Jens Axboe71539712021-11-03 05:52:45 -06002899 if (plug) {
2900 data.nr_tags = plug->nr_ios;
2901 plug->nr_ios = 1;
2902 data.cached_rq = &plug->cached_rq;
2903 }
2904
2905 rq = __blk_mq_alloc_requests(&data);
Jens Axboe373b5412021-12-02 12:42:58 -07002906 if (rq)
2907 return rq;
Jens Axboe71539712021-11-03 05:52:45 -06002908 rq_qos_cleanup(q, bio);
2909 if (bio->bi_opf & REQ_NOWAIT)
2910 bio_wouldblock_error(bio);
Jens Axboe95febeb2021-11-15 14:23:08 -07002911 return NULL;
Jens Axboe71539712021-11-03 05:52:45 -06002912}
2913
Christoph Hellwig309ce672024-01-11 14:57:04 +01002914/*
2915 * Check if we can use the passed on request for submitting the passed in bio,
2916 * and remove it from the request list if it can be used.
2917 */
2918static bool blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug,
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002919 struct bio *bio)
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002920{
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002921 enum hctx_type type = blk_mq_get_hctx_type(bio->bi_opf);
2922 enum hctx_type hctx_type = rq->mq_hctx->type;
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002923
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002924 WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq);
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002925
Pavel Begunkov77465642023-01-17 11:42:15 +00002926 if (type != hctx_type &&
2927 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002928 return false;
2929 if (op_is_flush(rq->cmd_flags) != op_is_flush(bio->bi_opf))
2930 return false;
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002931
Jens Axboe26456722022-06-21 10:03:57 -06002932 /*
2933 * If any qos ->throttle() end up blocking, we will have flushed the
2934 * plug and hence killed the cached_rq list as well. Pop this entry
2935 * before we throttle.
2936 */
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002937 plug->cached_rq = rq_list_next(rq);
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002938 rq_qos_throttle(rq->q, bio);
Jens Axboe26456722022-06-21 10:03:57 -06002939
Chengming Zhou5c17f452023-07-10 18:55:16 +08002940 blk_mq_rq_time_init(rq, 0);
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002941 rq->cmd_flags = bio->bi_opf;
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002942 INIT_LIST_HEAD(&rq->queuelist);
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002943 return true;
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002944}
2945
André Almeida105663f2020-01-06 15:08:18 -03002946/**
Christoph Hellwigc62b37d2020-07-01 10:59:43 +02002947 * blk_mq_submit_bio - Create and send a request to block device.
André Almeida105663f2020-01-06 15:08:18 -03002948 * @bio: Bio pointer.
2949 *
2950 * Builds up a request structure from @q and @bio and send to the device. The
2951 * request may not be queued directly to hardware if:
2952 * * This request can be merged with another one
2953 * * We want to place request at plug queue for possible future merging
2954 * * There is an IO scheduler active at this queue
2955 *
2956 * It will not queue the request if there is an error with the bio, or at the
2957 * request creation.
André Almeida105663f2020-01-06 15:08:18 -03002958 */
Christoph Hellwig3e087732021-10-12 13:12:24 +02002959void blk_mq_submit_bio(struct bio *bio)
Jens Axboe07068d52014-05-22 10:40:51 -06002960{
Pavel Begunkoved6cdde2021-10-14 15:03:30 +01002961 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
Christoph Hellwig6deacb32022-07-06 09:03:38 +02002962 struct blk_plug *plug = blk_mq_plug(bio);
Christoph Hellwigef295ec2016-10-28 08:48:16 -06002963 const int is_sync = op_is_sync(bio->bi_opf);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02002964 struct blk_mq_hw_ctx *hctx;
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002965 struct request *rq = NULL;
Jens Axboeabd45c12021-10-13 12:43:41 -06002966 unsigned int nr_segs = 1;
Satya Tangiralaa892c8d2020-05-14 00:37:18 +00002967 blk_status_t ret;
Jens Axboe07068d52014-05-22 10:40:51 -06002968
Christoph Hellwig51d798c2022-07-27 12:22:56 -04002969 bio = blk_queue_bounce(bio, q);
Jan Kara9c6227e2022-06-23 09:48:33 +02002970
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002971 if (plug) {
2972 rq = rq_list_peek(&plug->cached_rq);
2973 if (rq && rq->q != q)
2974 rq = NULL;
2975 }
2976 if (rq) {
Jens Axboe7b4f36c2024-01-12 09:12:20 -07002977 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2978 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2979 if (!bio)
2980 return;
2981 }
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002982 if (!bio_integrity_prep(bio))
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002983 return;
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002984 if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002985 return;
Christoph Hellwig309ce672024-01-11 14:57:04 +01002986 if (blk_mq_use_cached_rq(rq, plug, bio))
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002987 goto done;
2988 percpu_ref_get(&q->q_usage_counter);
2989 } else {
2990 if (unlikely(bio_queue_enter(bio)))
2991 return;
Jens Axboe7b4f36c2024-01-12 09:12:20 -07002992 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2993 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2994 if (!bio)
2995 goto fail;
2996 }
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002997 if (!bio_integrity_prep(bio))
2998 goto fail;
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002999 }
Jens Axboe87760e52016-11-09 12:38:14 -07003000
Christoph Hellwigb0077e22023-11-13 11:52:31 +08003001 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
3002 if (unlikely(!rq)) {
3003fail:
3004 blk_queue_exit(q);
3005 return;
3006 }
3007
3008done:
Christoph Hellwige8a676d2020-12-03 17:21:36 +01003009 trace_block_getrq(bio);
Xiaoguang Wangd6f1dda2018-10-23 22:30:50 +08003010
Josef Bacikc1c80382018-07-03 11:14:59 -04003011 rq_qos_track(q, rq, bio);
Jens Axboe07068d52014-05-22 10:40:51 -06003012
Bart Van Assche970d1682019-07-01 08:47:30 -07003013 blk_mq_bio_to_request(rq, bio, nr_segs);
3014
Eric Biggers9cd1e562023-03-15 11:39:02 -07003015 ret = blk_crypto_rq_get_keyslot(rq);
Satya Tangiralaa892c8d2020-05-14 00:37:18 +00003016 if (ret != BLK_STS_OK) {
3017 bio->bi_status = ret;
3018 bio_endio(bio);
3019 blk_mq_free_request(rq);
Christoph Hellwig3e087732021-10-12 13:12:24 +02003020 return;
Satya Tangiralaa892c8d2020-05-14 00:37:18 +00003021 }
3022
Christoph Hellwig360f2642023-05-19 06:40:46 +02003023 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
Christoph Hellwigd92ca9d82021-10-19 14:25:53 +02003024 return;
3025
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02003026 if (plug) {
Jens Axboece5b0092018-11-27 17:13:56 -07003027 blk_add_rq_to_plug(plug, rq);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02003028 return;
3029 }
3030
3031 hctx = rq->mq_hctx;
Christoph Hellwigdd6216b2023-05-18 07:31:01 +02003032 if ((rq->rq_flags & RQF_USE_SCHED) ||
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02003033 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
Christoph Hellwig710fa372023-04-13 08:40:54 +02003034 blk_mq_insert_request(rq, 0);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02003035 blk_mq_run_hw_queue(hctx, true);
3036 } else {
3037 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3038 }
Jens Axboe320ae512013-10-24 09:20:05 +01003039}
3040
Christoph Hellwig248c7932022-02-15 11:05:36 +01003041#ifdef CONFIG_BLK_MQ_STACKING
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003042/**
Christoph Hellwiga5efda32022-02-15 11:05:37 +01003043 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
Christoph Hellwiga5efda32022-02-15 11:05:37 +01003044 * @rq: the request being queued
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003045 */
Christoph Hellwig28db4712022-02-15 11:05:38 +01003046blk_status_t blk_insert_cloned_request(struct request *rq)
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003047{
Christoph Hellwig28db4712022-02-15 11:05:38 +01003048 struct request_queue *q = rq->q;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003049 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
Uday Shankar49d24392023-02-28 17:06:55 -07003050 unsigned int max_segments = blk_rq_get_max_segments(rq);
Christoph Hellwiga5efda32022-02-15 11:05:37 +01003051 blk_status_t ret;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003052
3053 if (blk_rq_sectors(rq) > max_sectors) {
3054 /*
3055 * SCSI device does not have a good way to return if
3056 * Write Same/Zero is actually supported. If a device rejects
3057 * a non-read/write command (discard, write same,etc.) the
3058 * low-level device driver will set the relevant queue limit to
3059 * 0 to prevent blk-lib from issuing more of the offending
3060 * operations. Commands queued prior to the queue limit being
3061 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3062 * errors being propagated to upper layers.
3063 */
3064 if (max_sectors == 0)
3065 return BLK_STS_NOTSUPP;
3066
3067 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3068 __func__, blk_rq_sectors(rq), max_sectors);
3069 return BLK_STS_IOERR;
3070 }
3071
3072 /*
3073 * The queue settings related to segment counting may differ from the
3074 * original queue.
3075 */
3076 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
Uday Shankar49d24392023-02-28 17:06:55 -07003077 if (rq->nr_phys_segments > max_segments) {
3078 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3079 __func__, rq->nr_phys_segments, max_segments);
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003080 return BLK_STS_IOERR;
3081 }
3082
Christoph Hellwig28db4712022-02-15 11:05:38 +01003083 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003084 return BLK_STS_IOERR;
3085
Eric Biggers5b8562f2023-03-15 11:39:06 -07003086 ret = blk_crypto_rq_get_keyslot(rq);
3087 if (ret != BLK_STS_OK)
3088 return ret;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003089
3090 blk_account_io_start(rq);
3091
3092 /*
3093 * Since we have a scheduler attached on the top device,
3094 * bypass a potential scheduler on the bottom device for
3095 * insert.
3096 */
Christoph Hellwig28db4712022-02-15 11:05:38 +01003097 blk_mq_run_dispatch_ops(q,
Ming Lei4cafe862021-12-03 21:15:34 +08003098 ret = blk_mq_request_issue_directly(rq, true));
Yu Kuai592ee112022-01-26 09:21:32 +08003099 if (ret)
3100 blk_account_io_done(rq, ktime_get_ns());
Ming Lei4cafe862021-12-03 21:15:34 +08003101 return ret;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003102}
3103EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3104
3105/**
3106 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3107 * @rq: the clone request to be cleaned up
3108 *
3109 * Description:
3110 * Free all bios in @rq for a cloned request.
3111 */
3112void blk_rq_unprep_clone(struct request *rq)
3113{
3114 struct bio *bio;
3115
3116 while ((bio = rq->bio) != NULL) {
3117 rq->bio = bio->bi_next;
3118
3119 bio_put(bio);
3120 }
3121}
3122EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3123
3124/**
3125 * blk_rq_prep_clone - Helper function to setup clone request
3126 * @rq: the request to be setup
3127 * @rq_src: original request to be cloned
3128 * @bs: bio_set that bios for clone are allocated from
3129 * @gfp_mask: memory allocation mask for bio
3130 * @bio_ctr: setup function to be called for each clone bio.
3131 * Returns %0 for success, non %0 for failure.
3132 * @data: private data to be passed to @bio_ctr
3133 *
3134 * Description:
3135 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3136 * Also, pages which the original bios are pointing to are not copied
3137 * and the cloned bios just point same pages.
3138 * So cloned bios must be completed before original bios, which means
3139 * the caller must complete @rq before @rq_src.
3140 */
3141int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3142 struct bio_set *bs, gfp_t gfp_mask,
3143 int (*bio_ctr)(struct bio *, struct bio *, void *),
3144 void *data)
3145{
3146 struct bio *bio, *bio_src;
3147
3148 if (!bs)
3149 bs = &fs_bio_set;
3150
3151 __rq_for_each_bio(bio_src, rq_src) {
Christoph Hellwigabfc4262022-02-02 17:01:09 +01003152 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3153 bs);
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003154 if (!bio)
3155 goto free_and_out;
3156
3157 if (bio_ctr && bio_ctr(bio, bio_src, data))
3158 goto free_and_out;
3159
3160 if (rq->bio) {
3161 rq->biotail->bi_next = bio;
3162 rq->biotail = bio;
3163 } else {
3164 rq->bio = rq->biotail = bio;
3165 }
3166 bio = NULL;
3167 }
3168
3169 /* Copy attributes of the original request to the clone request. */
3170 rq->__sector = blk_rq_pos(rq_src);
3171 rq->__data_len = blk_rq_bytes(rq_src);
3172 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3173 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3174 rq->special_vec = rq_src->special_vec;
3175 }
3176 rq->nr_phys_segments = rq_src->nr_phys_segments;
3177 rq->ioprio = rq_src->ioprio;
3178
3179 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3180 goto free_and_out;
3181
3182 return 0;
3183
3184free_and_out:
3185 if (bio)
3186 bio_put(bio);
3187 blk_rq_unprep_clone(rq);
3188
3189 return -ENOMEM;
3190}
3191EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
Christoph Hellwig248c7932022-02-15 11:05:36 +01003192#endif /* CONFIG_BLK_MQ_STACKING */
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003193
Christoph Hellwigf2b8f3c2021-11-17 07:14:00 +01003194/*
3195 * Steal bios from a request and add them to a bio list.
3196 * The request must not have been partially completed before.
3197 */
3198void blk_steal_bios(struct bio_list *list, struct request *rq)
3199{
3200 if (rq->bio) {
3201 if (list->tail)
3202 list->tail->bi_next = rq->bio;
3203 else
3204 list->head = rq->bio;
3205 list->tail = rq->biotail;
3206
3207 rq->bio = NULL;
3208 rq->biotail = NULL;
3209 }
3210
3211 rq->__data_len = 0;
3212}
3213EXPORT_SYMBOL_GPL(blk_steal_bios);
3214
Ming Leibd631412021-05-11 23:22:35 +08003215static size_t order_to_size(unsigned int order)
3216{
3217 return (size_t)PAGE_SIZE << order;
3218}
3219
3220/* called before freeing request pool in @tags */
John Garryf32e4ea2021-10-05 18:23:32 +08003221static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3222 struct blk_mq_tags *tags)
Ming Leibd631412021-05-11 23:22:35 +08003223{
Ming Leibd631412021-05-11 23:22:35 +08003224 struct page *page;
3225 unsigned long flags;
3226
Yu Kuai76dd2982022-10-11 22:22:53 +08003227 /*
3228 * There is no need to clear mapping if driver tags is not initialized
3229 * or the mapping belongs to the driver tags.
3230 */
3231 if (!drv_tags || drv_tags == tags)
John Garry4f245d52021-10-05 18:23:33 +08003232 return;
3233
Ming Leibd631412021-05-11 23:22:35 +08003234 list_for_each_entry(page, &tags->page_list, lru) {
3235 unsigned long start = (unsigned long)page_address(page);
3236 unsigned long end = start + order_to_size(page->private);
3237 int i;
3238
John Garryf32e4ea2021-10-05 18:23:32 +08003239 for (i = 0; i < drv_tags->nr_tags; i++) {
Ming Leibd631412021-05-11 23:22:35 +08003240 struct request *rq = drv_tags->rqs[i];
3241 unsigned long rq_addr = (unsigned long)rq;
3242
3243 if (rq_addr >= start && rq_addr < end) {
Jens Axboe0a467d02021-10-14 14:39:59 -06003244 WARN_ON_ONCE(req_ref_read(rq) != 0);
Ming Leibd631412021-05-11 23:22:35 +08003245 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3246 }
3247 }
3248 }
3249
3250 /*
3251 * Wait until all pending iteration is done.
3252 *
3253 * Request reference is cleared and it is guaranteed to be observed
3254 * after the ->lock is released.
3255 */
3256 spin_lock_irqsave(&drv_tags->lock, flags);
3257 spin_unlock_irqrestore(&drv_tags->lock, flags);
3258}
3259
Jens Axboecc71a6f2017-01-11 14:29:56 -07003260void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3261 unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01003262{
John Garryf32e4ea2021-10-05 18:23:32 +08003263 struct blk_mq_tags *drv_tags;
Jens Axboe320ae512013-10-24 09:20:05 +01003264 struct page *page;
3265
Ming Leie02657e2022-03-08 06:51:48 +01003266 if (list_empty(&tags->page_list))
3267 return;
3268
John Garry079a2e32021-10-05 18:23:39 +08003269 if (blk_mq_is_shared_tags(set->flags))
3270 drv_tags = set->shared_tags;
John Garrye155b0c2021-10-05 18:23:37 +08003271 else
3272 drv_tags = set->tags[hctx_idx];
John Garryf32e4ea2021-10-05 18:23:32 +08003273
John Garry65de57b2021-10-05 18:23:26 +08003274 if (tags->static_rqs && set->ops->exit_request) {
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003275 int i;
3276
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003277 for (i = 0; i < tags->nr_tags; i++) {
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003278 struct request *rq = tags->static_rqs[i];
3279
3280 if (!rq)
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003281 continue;
Christoph Hellwigd6296d392017-05-01 10:19:08 -06003282 set->ops->exit_request(set, rq, hctx_idx);
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003283 tags->static_rqs[i] = NULL;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003284 }
3285 }
3286
John Garryf32e4ea2021-10-05 18:23:32 +08003287 blk_mq_clear_rq_mapping(drv_tags, tags);
Ming Leibd631412021-05-11 23:22:35 +08003288
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003289 while (!list_empty(&tags->page_list)) {
3290 page = list_first_entry(&tags->page_list, struct page, lru);
Dave Hansen67534712014-01-08 20:17:46 -07003291 list_del_init(&page->lru);
Catalin Marinasf75782e2015-09-14 18:16:02 +01003292 /*
3293 * Remove kmemleak object previously allocated in
Raul E Rangel273938b2019-05-02 13:48:11 -06003294 * blk_mq_alloc_rqs().
Catalin Marinasf75782e2015-09-14 18:16:02 +01003295 */
3296 kmemleak_free(page_address(page));
Jens Axboe320ae512013-10-24 09:20:05 +01003297 __free_pages(page, page->private);
3298 }
Jens Axboecc71a6f2017-01-11 14:29:56 -07003299}
Jens Axboe320ae512013-10-24 09:20:05 +01003300
John Garrye155b0c2021-10-05 18:23:37 +08003301void blk_mq_free_rq_map(struct blk_mq_tags *tags)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003302{
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003303 kfree(tags->rqs);
Jens Axboecc71a6f2017-01-11 14:29:56 -07003304 tags->rqs = NULL;
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003305 kfree(tags->static_rqs);
3306 tags->static_rqs = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003307
John Garrye155b0c2021-10-05 18:23:37 +08003308 blk_mq_free_tags(tags);
Jens Axboe320ae512013-10-24 09:20:05 +01003309}
3310
Ming Lei4d805132022-03-08 15:32:14 +08003311static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3312 unsigned int hctx_idx)
3313{
3314 int i;
3315
3316 for (i = 0; i < set->nr_maps; i++) {
3317 unsigned int start = set->map[i].queue_offset;
3318 unsigned int end = start + set->map[i].nr_queues;
3319
3320 if (hctx_idx >= start && hctx_idx < end)
3321 break;
3322 }
3323
3324 if (i >= set->nr_maps)
3325 i = HCTX_TYPE_DEFAULT;
3326
3327 return i;
3328}
3329
3330static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3331 unsigned int hctx_idx)
3332{
3333 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3334
3335 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3336}
3337
John Garry63064be2021-10-05 18:23:35 +08003338static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3339 unsigned int hctx_idx,
3340 unsigned int nr_tags,
John Garrye155b0c2021-10-05 18:23:37 +08003341 unsigned int reserved_tags)
Jens Axboe320ae512013-10-24 09:20:05 +01003342{
Ming Lei4d805132022-03-08 15:32:14 +08003343 int node = blk_mq_get_hctx_node(set, hctx_idx);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003344 struct blk_mq_tags *tags;
Jens Axboe320ae512013-10-24 09:20:05 +01003345
Shaohua Li59f082e2017-02-01 09:53:14 -08003346 if (node == NUMA_NO_NODE)
3347 node = set->numa_node;
3348
John Garrye155b0c2021-10-05 18:23:37 +08003349 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3350 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003351 if (!tags)
3352 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003353
Kees Cook590b5b72018-06-12 14:04:20 -07003354 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
Gabriel Krisman Bertazi36e1f3d12016-12-06 13:31:44 -02003355 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
Shaohua Li59f082e2017-02-01 09:53:14 -08003356 node);
Jinlong Chen7edfd682022-11-02 10:52:29 +08003357 if (!tags->rqs)
3358 goto err_free_tags;
Jens Axboe320ae512013-10-24 09:20:05 +01003359
Kees Cook590b5b72018-06-12 14:04:20 -07003360 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3361 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3362 node);
Jinlong Chen7edfd682022-11-02 10:52:29 +08003363 if (!tags->static_rqs)
3364 goto err_free_rqs;
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003365
Jens Axboecc71a6f2017-01-11 14:29:56 -07003366 return tags;
Jinlong Chen7edfd682022-11-02 10:52:29 +08003367
3368err_free_rqs:
3369 kfree(tags->rqs);
3370err_free_tags:
3371 blk_mq_free_tags(tags);
3372 return NULL;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003373}
3374
Tejun Heo1d9bd512018-01-09 08:29:48 -08003375static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3376 unsigned int hctx_idx, int node)
3377{
3378 int ret;
3379
3380 if (set->ops->init_request) {
3381 ret = set->ops->init_request(set, rq, hctx_idx, node);
3382 if (ret)
3383 return ret;
3384 }
3385
Keith Busch12f5b932018-05-29 15:52:28 +02003386 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Tejun Heo1d9bd512018-01-09 08:29:48 -08003387 return 0;
3388}
3389
John Garry63064be2021-10-05 18:23:35 +08003390static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3391 struct blk_mq_tags *tags,
3392 unsigned int hctx_idx, unsigned int depth)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003393{
3394 unsigned int i, j, entries_per_page, max_order = 4;
Ming Lei4d805132022-03-08 15:32:14 +08003395 int node = blk_mq_get_hctx_node(set, hctx_idx);
Jens Axboecc71a6f2017-01-11 14:29:56 -07003396 size_t rq_size, left;
Shaohua Li59f082e2017-02-01 09:53:14 -08003397
Shaohua Li59f082e2017-02-01 09:53:14 -08003398 if (node == NUMA_NO_NODE)
3399 node = set->numa_node;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003400
3401 INIT_LIST_HEAD(&tags->page_list);
3402
Jens Axboe320ae512013-10-24 09:20:05 +01003403 /*
3404 * rq_size is the size of the request plus driver payload, rounded
3405 * to the cacheline size
3406 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003407 rq_size = round_up(sizeof(struct request) + set->cmd_size,
Jens Axboe320ae512013-10-24 09:20:05 +01003408 cache_line_size());
Jens Axboecc71a6f2017-01-11 14:29:56 -07003409 left = rq_size * depth;
Jens Axboe320ae512013-10-24 09:20:05 +01003410
Jens Axboecc71a6f2017-01-11 14:29:56 -07003411 for (i = 0; i < depth; ) {
Jens Axboe320ae512013-10-24 09:20:05 +01003412 int this_order = max_order;
3413 struct page *page;
3414 int to_do;
3415 void *p;
3416
Bartlomiej Zolnierkiewiczb3a834b2016-05-16 09:54:47 -06003417 while (this_order && left < order_to_size(this_order - 1))
Jens Axboe320ae512013-10-24 09:20:05 +01003418 this_order--;
3419
3420 do {
Shaohua Li59f082e2017-02-01 09:53:14 -08003421 page = alloc_pages_node(node,
Gabriel Krisman Bertazi36e1f3d12016-12-06 13:31:44 -02003422 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
Jens Axboea5164402014-09-10 09:02:03 -06003423 this_order);
Jens Axboe320ae512013-10-24 09:20:05 +01003424 if (page)
3425 break;
3426 if (!this_order--)
3427 break;
3428 if (order_to_size(this_order) < rq_size)
3429 break;
3430 } while (1);
3431
3432 if (!page)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003433 goto fail;
Jens Axboe320ae512013-10-24 09:20:05 +01003434
3435 page->private = this_order;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003436 list_add_tail(&page->lru, &tags->page_list);
Jens Axboe320ae512013-10-24 09:20:05 +01003437
3438 p = page_address(page);
Catalin Marinasf75782e2015-09-14 18:16:02 +01003439 /*
3440 * Allow kmemleak to scan these pages as they contain pointers
3441 * to additional allocations like via ops->init_request().
3442 */
Gabriel Krisman Bertazi36e1f3d12016-12-06 13:31:44 -02003443 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
Jens Axboe320ae512013-10-24 09:20:05 +01003444 entries_per_page = order_to_size(this_order) / rq_size;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003445 to_do = min(entries_per_page, depth - i);
Jens Axboe320ae512013-10-24 09:20:05 +01003446 left -= to_do * rq_size;
3447 for (j = 0; j < to_do; j++) {
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003448 struct request *rq = p;
3449
3450 tags->static_rqs[i] = rq;
Tejun Heo1d9bd512018-01-09 08:29:48 -08003451 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3452 tags->static_rqs[i] = NULL;
3453 goto fail;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003454 }
3455
Jens Axboe320ae512013-10-24 09:20:05 +01003456 p += rq_size;
3457 i++;
3458 }
3459 }
Jens Axboecc71a6f2017-01-11 14:29:56 -07003460 return 0;
Jens Axboe320ae512013-10-24 09:20:05 +01003461
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003462fail:
Jens Axboecc71a6f2017-01-11 14:29:56 -07003463 blk_mq_free_rqs(set, tags, hctx_idx);
3464 return -ENOMEM;
Jens Axboe320ae512013-10-24 09:20:05 +01003465}
3466
Ming Leibf0beec2020-05-29 15:53:15 +02003467struct rq_iter_data {
3468 struct blk_mq_hw_ctx *hctx;
3469 bool has_rq;
3470};
3471
John Garry2dd65322022-07-06 20:03:53 +08003472static bool blk_mq_has_request(struct request *rq, void *data)
Ming Leibf0beec2020-05-29 15:53:15 +02003473{
3474 struct rq_iter_data *iter_data = data;
3475
3476 if (rq->mq_hctx != iter_data->hctx)
3477 return true;
3478 iter_data->has_rq = true;
3479 return false;
3480}
3481
3482static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3483{
3484 struct blk_mq_tags *tags = hctx->sched_tags ?
3485 hctx->sched_tags : hctx->tags;
3486 struct rq_iter_data data = {
3487 .hctx = hctx,
3488 };
3489
3490 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3491 return data.has_rq;
3492}
3493
3494static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3495 struct blk_mq_hw_ctx *hctx)
3496{
Yury Norov9b51d9d2021-08-14 14:17:05 -07003497 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
Ming Leibf0beec2020-05-29 15:53:15 +02003498 return false;
3499 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3500 return false;
3501 return true;
3502}
3503
3504static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3505{
3506 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3507 struct blk_mq_hw_ctx, cpuhp_online);
3508
3509 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3510 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3511 return 0;
3512
3513 /*
3514 * Prevent new request from being allocated on the current hctx.
3515 *
3516 * The smp_mb__after_atomic() Pairs with the implied barrier in
3517 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3518 * seen once we return from the tag allocator.
3519 */
3520 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3521 smp_mb__after_atomic();
3522
3523 /*
3524 * Try to grab a reference to the queue and wait for any outstanding
3525 * requests. If we could not grab a reference the queue has been
3526 * frozen and there are no requests.
3527 */
3528 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3529 while (blk_mq_hctx_has_requests(hctx))
3530 msleep(5);
3531 percpu_ref_put(&hctx->queue->q_usage_counter);
3532 }
3533
3534 return 0;
3535}
3536
3537static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3538{
3539 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3540 struct blk_mq_hw_ctx, cpuhp_online);
3541
3542 if (cpumask_test_cpu(cpu, hctx->cpumask))
3543 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3544 return 0;
3545}
3546
Jens Axboee57690f2016-08-24 15:34:35 -06003547/*
3548 * 'cpu' is going away. splice any existing rq_list entries from this
3549 * software queue to the hw queue dispatch list, and ensure that it
3550 * gets run.
3551 */
Thomas Gleixner9467f852016-09-22 08:05:17 -06003552static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
Jens Axboe484b4062014-05-21 14:01:15 -06003553{
Thomas Gleixner9467f852016-09-22 08:05:17 -06003554 struct blk_mq_hw_ctx *hctx;
Jens Axboe484b4062014-05-21 14:01:15 -06003555 struct blk_mq_ctx *ctx;
3556 LIST_HEAD(tmp);
Ming Leic16d6b52018-12-17 08:44:05 -07003557 enum hctx_type type;
Jens Axboe484b4062014-05-21 14:01:15 -06003558
Thomas Gleixner9467f852016-09-22 08:05:17 -06003559 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
Ming Leibf0beec2020-05-29 15:53:15 +02003560 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3561 return 0;
3562
Jens Axboee57690f2016-08-24 15:34:35 -06003563 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
Ming Leic16d6b52018-12-17 08:44:05 -07003564 type = hctx->type;
Jens Axboe484b4062014-05-21 14:01:15 -06003565
3566 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07003567 if (!list_empty(&ctx->rq_lists[type])) {
3568 list_splice_init(&ctx->rq_lists[type], &tmp);
Jens Axboe484b4062014-05-21 14:01:15 -06003569 blk_mq_hctx_clear_pending(hctx, ctx);
3570 }
3571 spin_unlock(&ctx->lock);
3572
3573 if (list_empty(&tmp))
Thomas Gleixner9467f852016-09-22 08:05:17 -06003574 return 0;
Jens Axboe484b4062014-05-21 14:01:15 -06003575
Jens Axboee57690f2016-08-24 15:34:35 -06003576 spin_lock(&hctx->lock);
3577 list_splice_tail_init(&tmp, &hctx->dispatch);
3578 spin_unlock(&hctx->lock);
Jens Axboe484b4062014-05-21 14:01:15 -06003579
3580 blk_mq_run_hw_queue(hctx, true);
Thomas Gleixner9467f852016-09-22 08:05:17 -06003581 return 0;
Jens Axboe484b4062014-05-21 14:01:15 -06003582}
3583
Thomas Gleixner9467f852016-09-22 08:05:17 -06003584static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
Jens Axboe484b4062014-05-21 14:01:15 -06003585{
Ming Leibf0beec2020-05-29 15:53:15 +02003586 if (!(hctx->flags & BLK_MQ_F_STACKING))
3587 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3588 &hctx->cpuhp_online);
Thomas Gleixner9467f852016-09-22 08:05:17 -06003589 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3590 &hctx->cpuhp_dead);
Jens Axboe484b4062014-05-21 14:01:15 -06003591}
3592
Ming Lei364b6182021-05-11 23:22:36 +08003593/*
3594 * Before freeing hw queue, clearing the flush request reference in
3595 * tags->rqs[] for avoiding potential UAF.
3596 */
3597static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3598 unsigned int queue_depth, struct request *flush_rq)
3599{
3600 int i;
3601 unsigned long flags;
3602
3603 /* The hw queue may not be mapped yet */
3604 if (!tags)
3605 return;
3606
Jens Axboe0a467d02021-10-14 14:39:59 -06003607 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
Ming Lei364b6182021-05-11 23:22:36 +08003608
3609 for (i = 0; i < queue_depth; i++)
3610 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3611
3612 /*
3613 * Wait until all pending iteration is done.
3614 *
3615 * Request reference is cleared and it is guaranteed to be observed
3616 * after the ->lock is released.
3617 */
3618 spin_lock_irqsave(&tags->lock, flags);
3619 spin_unlock_irqrestore(&tags->lock, flags);
3620}
3621
Ming Leic3b4afc2015-06-04 22:25:04 +08003622/* hctx->ctxs will be freed in queue's release handler */
Ming Lei08e98fc2014-09-25 23:23:38 +08003623static void blk_mq_exit_hctx(struct request_queue *q,
3624 struct blk_mq_tag_set *set,
3625 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3626{
Ming Lei364b6182021-05-11 23:22:36 +08003627 struct request *flush_rq = hctx->fq->flush_rq;
3628
Ming Lei8ab0b7d2018-01-09 21:28:29 +08003629 if (blk_mq_hw_queue_mapped(hctx))
3630 blk_mq_tag_idle(hctx);
Ming Lei08e98fc2014-09-25 23:23:38 +08003631
Ming Lei6cfeadb2022-06-16 09:44:01 +08003632 if (blk_queue_init_done(q))
3633 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3634 set->queue_depth, flush_rq);
Ming Leif70ced02014-09-25 23:23:47 +08003635 if (set->ops->exit_request)
Ming Lei364b6182021-05-11 23:22:36 +08003636 set->ops->exit_request(set, flush_rq, hctx_idx);
Ming Leif70ced02014-09-25 23:23:47 +08003637
Ming Lei08e98fc2014-09-25 23:23:38 +08003638 if (set->ops->exit_hctx)
3639 set->ops->exit_hctx(hctx, hctx_idx);
3640
Thomas Gleixner9467f852016-09-22 08:05:17 -06003641 blk_mq_remove_cpuhp(hctx);
Ming Lei2f8f1332019-04-30 09:52:27 +08003642
Ming Lei4e5cc992022-03-08 15:32:19 +08003643 xa_erase(&q->hctx_table, hctx_idx);
3644
Ming Lei2f8f1332019-04-30 09:52:27 +08003645 spin_lock(&q->unused_hctx_lock);
3646 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3647 spin_unlock(&q->unused_hctx_lock);
Ming Lei08e98fc2014-09-25 23:23:38 +08003648}
3649
Ming Lei624dbe42014-05-27 23:35:13 +08003650static void blk_mq_exit_hw_queues(struct request_queue *q,
3651 struct blk_mq_tag_set *set, int nr_queue)
3652{
3653 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08003654 unsigned long i;
Ming Lei624dbe42014-05-27 23:35:13 +08003655
3656 queue_for_each_hw_ctx(q, hctx, i) {
3657 if (i == nr_queue)
3658 break;
Ming Lei08e98fc2014-09-25 23:23:38 +08003659 blk_mq_exit_hctx(q, set, hctx, i);
Ming Lei624dbe42014-05-27 23:35:13 +08003660 }
Ming Lei624dbe42014-05-27 23:35:13 +08003661}
3662
Ming Lei08e98fc2014-09-25 23:23:38 +08003663static int blk_mq_init_hctx(struct request_queue *q,
3664 struct blk_mq_tag_set *set,
3665 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3666{
Ming Lei7c6c5b72019-04-30 09:52:26 +08003667 hctx->queue_num = hctx_idx;
Ming Lei08e98fc2014-09-25 23:23:38 +08003668
Ming Leibf0beec2020-05-29 15:53:15 +02003669 if (!(hctx->flags & BLK_MQ_F_STACKING))
3670 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3671 &hctx->cpuhp_online);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003672 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3673
3674 hctx->tags = set->tags[hctx_idx];
3675
3676 if (set->ops->init_hctx &&
3677 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3678 goto unregister_cpu_notifier;
3679
3680 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3681 hctx->numa_node))
3682 goto exit_hctx;
Ming Lei4e5cc992022-03-08 15:32:19 +08003683
3684 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3685 goto exit_flush_rq;
3686
Ming Lei7c6c5b72019-04-30 09:52:26 +08003687 return 0;
3688
Ming Lei4e5cc992022-03-08 15:32:19 +08003689 exit_flush_rq:
3690 if (set->ops->exit_request)
3691 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003692 exit_hctx:
3693 if (set->ops->exit_hctx)
3694 set->ops->exit_hctx(hctx, hctx_idx);
3695 unregister_cpu_notifier:
3696 blk_mq_remove_cpuhp(hctx);
3697 return -1;
3698}
3699
3700static struct blk_mq_hw_ctx *
3701blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3702 int node)
3703{
3704 struct blk_mq_hw_ctx *hctx;
3705 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3706
Ming Lei704b9142021-12-03 21:15:32 +08003707 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003708 if (!hctx)
3709 goto fail_alloc_hctx;
3710
3711 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3712 goto free_hctx;
3713
3714 atomic_set(&hctx->nr_active, 0);
Ming Lei08e98fc2014-09-25 23:23:38 +08003715 if (node == NUMA_NO_NODE)
Ming Lei7c6c5b72019-04-30 09:52:26 +08003716 node = set->numa_node;
3717 hctx->numa_node = node;
Ming Lei08e98fc2014-09-25 23:23:38 +08003718
Jens Axboe9f993732017-04-10 09:54:54 -06003719 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
Ming Lei08e98fc2014-09-25 23:23:38 +08003720 spin_lock_init(&hctx->lock);
3721 INIT_LIST_HEAD(&hctx->dispatch);
3722 hctx->queue = q;
Ming Lei51db1c32020-08-19 23:20:19 +08003723 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
Ming Lei08e98fc2014-09-25 23:23:38 +08003724
Ming Lei2f8f1332019-04-30 09:52:27 +08003725 INIT_LIST_HEAD(&hctx->hctx_list);
3726
Ming Lei08e98fc2014-09-25 23:23:38 +08003727 /*
3728 * Allocate space for all possible cpus to avoid allocation at
3729 * runtime
3730 */
Johannes Thumshirnd904bfa2017-11-15 17:32:33 -08003731 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
Ming Lei7c6c5b72019-04-30 09:52:26 +08003732 gfp, node);
Ming Lei08e98fc2014-09-25 23:23:38 +08003733 if (!hctx->ctxs)
Ming Lei7c6c5b72019-04-30 09:52:26 +08003734 goto free_cpumask;
Ming Lei08e98fc2014-09-25 23:23:38 +08003735
Jianchao Wang5b202852018-10-12 18:07:26 +08003736 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
Ming Leic548e622021-01-22 10:33:08 +08003737 gfp, node, false, false))
Ming Lei08e98fc2014-09-25 23:23:38 +08003738 goto free_ctxs;
Ming Lei08e98fc2014-09-25 23:23:38 +08003739 hctx->nr_ctx = 0;
3740
Ming Lei5815839b2018-06-25 19:31:47 +08003741 spin_lock_init(&hctx->dispatch_wait_lock);
Jens Axboeeb619fd2017-11-09 08:32:43 -07003742 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3743 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3744
Guoqing Jiang754a1572020-03-09 22:41:37 +01003745 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
Ming Leif70ced02014-09-25 23:23:47 +08003746 if (!hctx->fq)
Ming Lei7c6c5b72019-04-30 09:52:26 +08003747 goto free_bitmap;
Ming Leif70ced02014-09-25 23:23:47 +08003748
Ming Lei7c6c5b72019-04-30 09:52:26 +08003749 blk_mq_hctx_kobj_init(hctx);
Bart Van Assche6a83e742016-11-02 10:09:51 -06003750
Ming Lei7c6c5b72019-04-30 09:52:26 +08003751 return hctx;
Ming Lei08e98fc2014-09-25 23:23:38 +08003752
3753 free_bitmap:
Omar Sandoval88459642016-09-17 08:38:44 -06003754 sbitmap_free(&hctx->ctx_map);
Ming Lei08e98fc2014-09-25 23:23:38 +08003755 free_ctxs:
3756 kfree(hctx->ctxs);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003757 free_cpumask:
3758 free_cpumask_var(hctx->cpumask);
3759 free_hctx:
3760 kfree(hctx);
3761 fail_alloc_hctx:
3762 return NULL;
Ming Lei08e98fc2014-09-25 23:23:38 +08003763}
3764
Jens Axboe320ae512013-10-24 09:20:05 +01003765static void blk_mq_init_cpu_queues(struct request_queue *q,
3766 unsigned int nr_hw_queues)
3767{
Jens Axboeb3c661b2018-10-30 10:36:06 -06003768 struct blk_mq_tag_set *set = q->tag_set;
3769 unsigned int i, j;
Jens Axboe320ae512013-10-24 09:20:05 +01003770
3771 for_each_possible_cpu(i) {
3772 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3773 struct blk_mq_hw_ctx *hctx;
Ming Leic16d6b52018-12-17 08:44:05 -07003774 int k;
Jens Axboe320ae512013-10-24 09:20:05 +01003775
Jens Axboe320ae512013-10-24 09:20:05 +01003776 __ctx->cpu = i;
3777 spin_lock_init(&__ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07003778 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3779 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3780
Jens Axboe320ae512013-10-24 09:20:05 +01003781 __ctx->queue = q;
3782
Jens Axboe320ae512013-10-24 09:20:05 +01003783 /*
3784 * Set local node, IFF we have more than one hw queue. If
3785 * not, we remain on the home node of the device
3786 */
Jens Axboeb3c661b2018-10-30 10:36:06 -06003787 for (j = 0; j < set->nr_maps; j++) {
3788 hctx = blk_mq_map_queue_type(q, j, i);
3789 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
Xianting Tian576e85c2020-10-19 16:20:47 +08003790 hctx->numa_node = cpu_to_node(i);
Jens Axboeb3c661b2018-10-30 10:36:06 -06003791 }
Jens Axboe320ae512013-10-24 09:20:05 +01003792 }
3793}
3794
John Garry63064be2021-10-05 18:23:35 +08003795struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3796 unsigned int hctx_idx,
3797 unsigned int depth)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003798{
John Garry63064be2021-10-05 18:23:35 +08003799 struct blk_mq_tags *tags;
3800 int ret;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003801
John Garrye155b0c2021-10-05 18:23:37 +08003802 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
John Garry63064be2021-10-05 18:23:35 +08003803 if (!tags)
3804 return NULL;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003805
John Garry63064be2021-10-05 18:23:35 +08003806 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3807 if (ret) {
John Garrye155b0c2021-10-05 18:23:37 +08003808 blk_mq_free_rq_map(tags);
John Garry63064be2021-10-05 18:23:35 +08003809 return NULL;
3810 }
Jens Axboecc71a6f2017-01-11 14:29:56 -07003811
John Garry63064be2021-10-05 18:23:35 +08003812 return tags;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003813}
3814
John Garry63064be2021-10-05 18:23:35 +08003815static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3816 int hctx_idx)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003817{
John Garry079a2e32021-10-05 18:23:39 +08003818 if (blk_mq_is_shared_tags(set->flags)) {
3819 set->tags[hctx_idx] = set->shared_tags;
John Garry1c0706a2020-08-19 23:20:22 +08003820
John Garrye155b0c2021-10-05 18:23:37 +08003821 return true;
Jens Axboebd166ef2017-01-17 06:03:22 -07003822 }
John Garrye155b0c2021-10-05 18:23:37 +08003823
John Garry63064be2021-10-05 18:23:35 +08003824 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3825 set->queue_depth);
3826
3827 return set->tags[hctx_idx];
Jens Axboecc71a6f2017-01-11 14:29:56 -07003828}
3829
John Garry645db342021-10-05 18:23:36 +08003830void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3831 struct blk_mq_tags *tags,
3832 unsigned int hctx_idx)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003833{
John Garry645db342021-10-05 18:23:36 +08003834 if (tags) {
3835 blk_mq_free_rqs(set, tags, hctx_idx);
John Garrye155b0c2021-10-05 18:23:37 +08003836 blk_mq_free_rq_map(tags);
Jens Axboecc71a6f2017-01-11 14:29:56 -07003837 }
3838}
3839
John Garrye155b0c2021-10-05 18:23:37 +08003840static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3841 unsigned int hctx_idx)
3842{
John Garry079a2e32021-10-05 18:23:39 +08003843 if (!blk_mq_is_shared_tags(set->flags))
John Garrye155b0c2021-10-05 18:23:37 +08003844 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3845
3846 set->tags[hctx_idx] = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003847}
3848
Christoph Hellwig4b855ad2017-06-26 12:20:57 +02003849static void blk_mq_map_swqueue(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01003850{
Ming Lei4f481202022-03-08 15:32:18 +08003851 unsigned int j, hctx_idx;
3852 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01003853 struct blk_mq_hw_ctx *hctx;
3854 struct blk_mq_ctx *ctx;
Ming Lei2a34c082015-04-21 10:00:20 +08003855 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01003856
3857 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboee4043dc2014-04-09 10:18:23 -06003858 cpumask_clear(hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01003859 hctx->nr_ctx = 0;
huhaid416c922018-05-18 08:32:30 -06003860 hctx->dispatch_from = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003861 }
3862
3863 /*
Christoph Hellwig4b855ad2017-06-26 12:20:57 +02003864 * Map software to hardware queues.
Ming Lei4412efe2018-04-25 04:01:44 +08003865 *
3866 * If the cpu isn't present, the cpu is mapped to first hctx.
Jens Axboe320ae512013-10-24 09:20:05 +01003867 */
Christoph Hellwig20e4d8132018-01-12 10:53:06 +08003868 for_each_possible_cpu(i) {
Ming Lei4412efe2018-04-25 04:01:44 +08003869
Thomas Gleixner897bb0c2016-03-19 11:30:33 +01003870 ctx = per_cpu_ptr(q->queue_ctx, i);
Jens Axboeb3c661b2018-10-30 10:36:06 -06003871 for (j = 0; j < set->nr_maps; j++) {
Jianchao Wangbb94aea2019-01-24 18:25:33 +08003872 if (!set->map[j].nr_queues) {
3873 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3874 HCTX_TYPE_DEFAULT, i);
Ming Leie5edd5f2018-12-18 01:28:56 +08003875 continue;
Jianchao Wangbb94aea2019-01-24 18:25:33 +08003876 }
Ming Leifd689872020-05-07 21:04:08 +08003877 hctx_idx = set->map[j].mq_map[i];
3878 /* unmapped hw queue can be remapped after CPU topo changed */
3879 if (!set->tags[hctx_idx] &&
John Garry63064be2021-10-05 18:23:35 +08003880 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
Ming Leifd689872020-05-07 21:04:08 +08003881 /*
3882 * If tags initialization fail for some hctx,
3883 * that hctx won't be brought online. In this
3884 * case, remap the current ctx to hctx[0] which
3885 * is guaranteed to always have tags allocated
3886 */
3887 set->map[j].mq_map[i] = 0;
3888 }
Ming Leie5edd5f2018-12-18 01:28:56 +08003889
Jens Axboeb3c661b2018-10-30 10:36:06 -06003890 hctx = blk_mq_map_queue_type(q, j, i);
Jianchao Wang8ccdf4a2019-01-24 18:25:32 +08003891 ctx->hctxs[j] = hctx;
Jens Axboeb3c661b2018-10-30 10:36:06 -06003892 /*
3893 * If the CPU is already set in the mask, then we've
3894 * mapped this one already. This can happen if
3895 * devices share queues across queue maps.
3896 */
3897 if (cpumask_test_cpu(i, hctx->cpumask))
3898 continue;
3899
3900 cpumask_set_cpu(i, hctx->cpumask);
3901 hctx->type = j;
3902 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3903 hctx->ctxs[hctx->nr_ctx++] = ctx;
3904
3905 /*
3906 * If the nr_ctx type overflows, we have exceeded the
3907 * amount of sw queues we can support.
3908 */
3909 BUG_ON(!hctx->nr_ctx);
3910 }
Jianchao Wangbb94aea2019-01-24 18:25:33 +08003911
3912 for (; j < HCTX_MAX_TYPES; j++)
3913 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3914 HCTX_TYPE_DEFAULT, i);
Jens Axboe320ae512013-10-24 09:20:05 +01003915 }
Jens Axboe506e9312014-05-07 10:26:44 -06003916
3917 queue_for_each_hw_ctx(q, hctx, i) {
Ming Lei4412efe2018-04-25 04:01:44 +08003918 /*
3919 * If no software queues are mapped to this hardware queue,
3920 * disable it and free the request entries.
3921 */
3922 if (!hctx->nr_ctx) {
3923 /* Never unmap queue 0. We need it as a
3924 * fallback in case of a new remap fails
3925 * allocation
3926 */
John Garrye155b0c2021-10-05 18:23:37 +08003927 if (i)
3928 __blk_mq_free_map_and_rqs(set, i);
Ming Lei4412efe2018-04-25 04:01:44 +08003929
3930 hctx->tags = NULL;
3931 continue;
3932 }
Jens Axboe484b4062014-05-21 14:01:15 -06003933
Ming Lei2a34c082015-04-21 10:00:20 +08003934 hctx->tags = set->tags[i];
3935 WARN_ON(!hctx->tags);
3936
Jens Axboe484b4062014-05-21 14:01:15 -06003937 /*
Chong Yuan889fa312015-04-15 11:39:29 -06003938 * Set the map size to the number of mapped software queues.
3939 * This is more accurate and more efficient than looping
3940 * over all possibly mapped software queues.
3941 */
Omar Sandoval88459642016-09-17 08:38:44 -06003942 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
Chong Yuan889fa312015-04-15 11:39:29 -06003943
3944 /*
Jens Axboe484b4062014-05-21 14:01:15 -06003945 * Initialize batch roundrobin counts
3946 */
Ming Leif82ddf12018-04-08 17:48:10 +08003947 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
Jens Axboe506e9312014-05-07 10:26:44 -06003948 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3949 }
Jens Axboe320ae512013-10-24 09:20:05 +01003950}
3951
Jens Axboe8e8320c2017-06-20 17:56:13 -06003952/*
3953 * Caller needs to ensure that we're either frozen/quiesced, or that
3954 * the queue isn't live yet.
3955 */
Jeff Moyer2404e602015-11-03 10:40:06 -05003956static void queue_set_hctx_shared(struct request_queue *q, bool shared)
Jens Axboe0d2602c2014-05-13 15:10:52 -06003957{
3958 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08003959 unsigned long i;
Jens Axboe0d2602c2014-05-13 15:10:52 -06003960
Jeff Moyer2404e602015-11-03 10:40:06 -05003961 queue_for_each_hw_ctx(q, hctx, i) {
Yu Kuai454bb672021-07-31 14:21:30 +08003962 if (shared) {
Ming Lei51db1c32020-08-19 23:20:19 +08003963 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
Yu Kuai454bb672021-07-31 14:21:30 +08003964 } else {
3965 blk_mq_tag_idle(hctx);
Ming Lei51db1c32020-08-19 23:20:19 +08003966 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
Yu Kuai454bb672021-07-31 14:21:30 +08003967 }
Jeff Moyer2404e602015-11-03 10:40:06 -05003968 }
3969}
3970
Hannes Reinecke655ac302020-08-19 23:20:20 +08003971static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3972 bool shared)
Jeff Moyer2404e602015-11-03 10:40:06 -05003973{
3974 struct request_queue *q;
Jens Axboe0d2602c2014-05-13 15:10:52 -06003975
Bart Van Assche705cda92017-04-07 11:16:49 -07003976 lockdep_assert_held(&set->tag_list_lock);
3977
Jens Axboe0d2602c2014-05-13 15:10:52 -06003978 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3979 blk_mq_freeze_queue(q);
Jeff Moyer2404e602015-11-03 10:40:06 -05003980 queue_set_hctx_shared(q, shared);
Jens Axboe0d2602c2014-05-13 15:10:52 -06003981 blk_mq_unfreeze_queue(q);
3982 }
3983}
3984
3985static void blk_mq_del_queue_tag_set(struct request_queue *q)
3986{
3987 struct blk_mq_tag_set *set = q->tag_set;
3988
Jens Axboe0d2602c2014-05-13 15:10:52 -06003989 mutex_lock(&set->tag_list_lock);
Daniel Wagner08c875c2020-07-28 15:29:51 +02003990 list_del(&q->tag_set_list);
Jeff Moyer2404e602015-11-03 10:40:06 -05003991 if (list_is_singular(&set->tag_list)) {
3992 /* just transitioned to unshared */
Ming Lei51db1c32020-08-19 23:20:19 +08003993 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
Jeff Moyer2404e602015-11-03 10:40:06 -05003994 /* update existing queue */
Hannes Reinecke655ac302020-08-19 23:20:20 +08003995 blk_mq_update_tag_set_shared(set, false);
Jeff Moyer2404e602015-11-03 10:40:06 -05003996 }
Jens Axboe0d2602c2014-05-13 15:10:52 -06003997 mutex_unlock(&set->tag_list_lock);
Roman Pena347c7a2018-06-10 22:38:24 +02003998 INIT_LIST_HEAD(&q->tag_set_list);
Jens Axboe0d2602c2014-05-13 15:10:52 -06003999}
4000
4001static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
4002 struct request_queue *q)
4003{
Jens Axboe0d2602c2014-05-13 15:10:52 -06004004 mutex_lock(&set->tag_list_lock);
Jeff Moyer2404e602015-11-03 10:40:06 -05004005
Jens Axboeff821d22017-11-10 22:05:12 -07004006 /*
4007 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4008 */
4009 if (!list_empty(&set->tag_list) &&
Ming Lei51db1c32020-08-19 23:20:19 +08004010 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
4011 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
Jeff Moyer2404e602015-11-03 10:40:06 -05004012 /* update existing queue */
Hannes Reinecke655ac302020-08-19 23:20:20 +08004013 blk_mq_update_tag_set_shared(set, true);
Jeff Moyer2404e602015-11-03 10:40:06 -05004014 }
Ming Lei51db1c32020-08-19 23:20:19 +08004015 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
Jeff Moyer2404e602015-11-03 10:40:06 -05004016 queue_set_hctx_shared(q, true);
Daniel Wagner08c875c2020-07-28 15:29:51 +02004017 list_add_tail(&q->tag_set_list, &set->tag_list);
Jeff Moyer2404e602015-11-03 10:40:06 -05004018
Jens Axboe0d2602c2014-05-13 15:10:52 -06004019 mutex_unlock(&set->tag_list_lock);
4020}
4021
Ming Lei1db49092018-11-20 09:44:35 +08004022/* All allocations will be freed in release handler of q->mq_kobj */
4023static int blk_mq_alloc_ctxs(struct request_queue *q)
4024{
4025 struct blk_mq_ctxs *ctxs;
4026 int cpu;
4027
4028 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4029 if (!ctxs)
4030 return -ENOMEM;
4031
4032 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4033 if (!ctxs->queue_ctx)
4034 goto fail;
4035
4036 for_each_possible_cpu(cpu) {
4037 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4038 ctx->ctxs = ctxs;
4039 }
4040
4041 q->mq_kobj = &ctxs->kobj;
4042 q->queue_ctx = ctxs->queue_ctx;
4043
4044 return 0;
4045 fail:
4046 kfree(ctxs);
4047 return -ENOMEM;
4048}
4049
Ming Leie09aae72015-01-29 20:17:27 +08004050/*
4051 * It is the actual release handler for mq, but we do it from
4052 * request queue's release handler for avoiding use-after-free
4053 * and headache because q->mq_kobj shouldn't have been introduced,
4054 * but we can't group ctx/kctx kobj without it.
4055 */
4056void blk_mq_release(struct request_queue *q)
4057{
Ming Lei2f8f1332019-04-30 09:52:27 +08004058 struct blk_mq_hw_ctx *hctx, *next;
Ming Lei4f481202022-03-08 15:32:18 +08004059 unsigned long i;
Ming Leie09aae72015-01-29 20:17:27 +08004060
Ming Lei2f8f1332019-04-30 09:52:27 +08004061 queue_for_each_hw_ctx(q, hctx, i)
4062 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4063
4064 /* all hctx are in .unused_hctx_list now */
4065 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4066 list_del_init(&hctx->hctx_list);
Ming Lei6c8b2322017-02-22 18:14:01 +08004067 kobject_put(&hctx->kobj);
Ming Leic3b4afc2015-06-04 22:25:04 +08004068 }
Ming Leie09aae72015-01-29 20:17:27 +08004069
Ming Lei4e5cc992022-03-08 15:32:19 +08004070 xa_destroy(&q->hctx_table);
Ming Leie09aae72015-01-29 20:17:27 +08004071
Ming Lei7ea5fe32017-02-22 18:14:00 +08004072 /*
4073 * release .mq_kobj and sw queue's kobject now because
4074 * both share lifetime with request queue.
4075 */
4076 blk_mq_sysfs_deinit(q);
Ming Leie09aae72015-01-29 20:17:27 +08004077}
4078
Christoph Hellwig5ec780a2021-06-24 10:10:12 +02004079static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
Christoph Hellwig2f227bb2020-03-27 09:30:08 +01004080 void *queuedata)
Jens Axboe320ae512013-10-24 09:20:05 +01004081{
Christoph Hellwig26a97502021-06-02 09:53:17 +03004082 struct request_queue *q;
4083 int ret;
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004084
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004085 q = blk_alloc_queue(set->numa_node);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004086 if (!q)
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004087 return ERR_PTR(-ENOMEM);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004088 q->queuedata = queuedata;
4089 ret = blk_mq_init_allocated_queue(set, q);
4090 if (ret) {
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004091 blk_put_queue(q);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004092 return ERR_PTR(ret);
4093 }
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004094 return q;
4095}
Christoph Hellwig2f227bb2020-03-27 09:30:08 +01004096
4097struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4098{
4099 return blk_mq_init_queue_data(set, NULL);
4100}
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004101EXPORT_SYMBOL(blk_mq_init_queue);
4102
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004103/**
4104 * blk_mq_destroy_queue - shutdown a request queue
4105 * @q: request queue to shutdown
4106 *
Bart Van Assche81ea42b2023-01-30 13:12:33 -08004107 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4108 * requests will be failed with -ENODEV. The caller is responsible for dropping
4109 * the reference from blk_mq_init_queue() by calling blk_put_queue().
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004110 *
4111 * Context: can sleep
4112 */
4113void blk_mq_destroy_queue(struct request_queue *q)
4114{
4115 WARN_ON_ONCE(!queue_is_mq(q));
4116 WARN_ON_ONCE(blk_queue_registered(q));
4117
4118 might_sleep();
4119
4120 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4121 blk_queue_start_drain(q);
Jinlong Chen56c1ee92022-10-30 16:32:12 +08004122 blk_mq_freeze_queue_wait(q);
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004123
4124 blk_sync_queue(q);
4125 blk_mq_cancel_work_sync(q);
4126 blk_mq_exit_queue(q);
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004127}
4128EXPORT_SYMBOL(blk_mq_destroy_queue);
4129
Christoph Hellwig4dcc48742021-08-16 15:19:05 +02004130struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4131 struct lock_class_key *lkclass)
Jens Axboe9316a9e2018-10-15 08:40:37 -06004132{
4133 struct request_queue *q;
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004134 struct gendisk *disk;
Jens Axboe9316a9e2018-10-15 08:40:37 -06004135
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004136 q = blk_mq_init_queue_data(set, queuedata);
4137 if (IS_ERR(q))
4138 return ERR_CAST(q);
Jens Axboe9316a9e2018-10-15 08:40:37 -06004139
Christoph Hellwig4a1fa412021-08-16 15:19:08 +02004140 disk = __alloc_disk_node(q, set->numa_node, lkclass);
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004141 if (!disk) {
Christoph Hellwig0a3e5cc2022-07-20 15:05:40 +02004142 blk_mq_destroy_queue(q);
Christoph Hellwig2b3f0562022-10-18 15:57:17 +02004143 blk_put_queue(q);
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004144 return ERR_PTR(-ENOMEM);
Jens Axboe9316a9e2018-10-15 08:40:37 -06004145 }
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004146 set_bit(GD_OWNS_QUEUE, &disk->state);
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004147 return disk;
Jens Axboe9316a9e2018-10-15 08:40:37 -06004148}
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004149EXPORT_SYMBOL(__blk_mq_alloc_disk);
Jens Axboe9316a9e2018-10-15 08:40:37 -06004150
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004151struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4152 struct lock_class_key *lkclass)
4153{
Christoph Hellwig22c17e22022-11-22 08:27:53 +01004154 struct gendisk *disk;
4155
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004156 if (!blk_get_queue(q))
4157 return NULL;
Christoph Hellwig22c17e22022-11-22 08:27:53 +01004158 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4159 if (!disk)
4160 blk_put_queue(q);
4161 return disk;
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004162}
4163EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4164
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004165static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4166 struct blk_mq_tag_set *set, struct request_queue *q,
4167 int hctx_idx, int node)
4168{
Ming Lei2f8f1332019-04-30 09:52:27 +08004169 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004170
Ming Lei2f8f1332019-04-30 09:52:27 +08004171 /* reuse dead hctx first */
4172 spin_lock(&q->unused_hctx_lock);
4173 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4174 if (tmp->numa_node == node) {
4175 hctx = tmp;
4176 break;
4177 }
4178 }
4179 if (hctx)
4180 list_del_init(&hctx->hctx_list);
4181 spin_unlock(&q->unused_hctx_lock);
4182
4183 if (!hctx)
4184 hctx = blk_mq_alloc_hctx(q, set, node);
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004185 if (!hctx)
Ming Lei7c6c5b72019-04-30 09:52:26 +08004186 goto fail;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004187
Ming Lei7c6c5b72019-04-30 09:52:26 +08004188 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4189 goto free_hctx;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004190
4191 return hctx;
Ming Lei7c6c5b72019-04-30 09:52:26 +08004192
4193 free_hctx:
4194 kobject_put(&hctx->kobj);
4195 fail:
4196 return NULL;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004197}
4198
Keith Busch868f2f02015-12-17 17:08:14 -07004199static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4200 struct request_queue *q)
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004201{
Ming Lei4e5cc992022-03-08 15:32:19 +08004202 struct blk_mq_hw_ctx *hctx;
4203 unsigned long i, j;
Bart Van Asscheac0d6b92019-10-25 09:50:09 -07004204
Ming Leifb350e02018-01-06 16:27:40 +08004205 /* protect against switching io scheduler */
4206 mutex_lock(&q->sysfs_lock);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004207 for (i = 0; i < set->nr_hw_queues; i++) {
Ming Lei306f13ee2022-03-08 15:32:15 +08004208 int old_node;
Ming Lei4d805132022-03-08 15:32:14 +08004209 int node = blk_mq_get_hctx_node(set, i);
Ming Lei4e5cc992022-03-08 15:32:19 +08004210 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
Keith Busch868f2f02015-12-17 17:08:14 -07004211
Ming Lei306f13ee2022-03-08 15:32:15 +08004212 if (old_hctx) {
4213 old_node = old_hctx->numa_node;
4214 blk_mq_exit_hctx(q, set, old_hctx, i);
4215 }
Jens Axboe320ae512013-10-24 09:20:05 +01004216
Ming Lei4e5cc992022-03-08 15:32:19 +08004217 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
Ming Lei306f13ee2022-03-08 15:32:15 +08004218 if (!old_hctx)
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004219 break;
Ming Lei306f13ee2022-03-08 15:32:15 +08004220 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4221 node, old_node);
Ming Lei4e5cc992022-03-08 15:32:19 +08004222 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4223 WARN_ON_ONCE(!hctx);
Keith Busch868f2f02015-12-17 17:08:14 -07004224 }
Jens Axboe320ae512013-10-24 09:20:05 +01004225 }
Jianchao Wange01ad462018-10-12 18:07:28 +08004226 /*
4227 * Increasing nr_hw_queues fails. Free the newly allocated
4228 * hctxs and keep the previous q->nr_hw_queues.
4229 */
4230 if (i != set->nr_hw_queues) {
4231 j = q->nr_hw_queues;
Jianchao Wange01ad462018-10-12 18:07:28 +08004232 } else {
4233 j = i;
Jianchao Wange01ad462018-10-12 18:07:28 +08004234 q->nr_hw_queues = set->nr_hw_queues;
4235 }
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004236
Ming Lei4e5cc992022-03-08 15:32:19 +08004237 xa_for_each_start(&q->hctx_table, j, hctx, j)
4238 blk_mq_exit_hctx(q, set, hctx, j);
Ming Leifb350e02018-01-06 16:27:40 +08004239 mutex_unlock(&q->sysfs_lock);
Keith Busch868f2f02015-12-17 17:08:14 -07004240}
4241
Ming Lei42ee3062022-03-08 15:32:16 +08004242static void blk_mq_update_poll_flag(struct request_queue *q)
4243{
4244 struct blk_mq_tag_set *set = q->tag_set;
4245
4246 if (set->nr_maps > HCTX_TYPE_POLL &&
4247 set->map[HCTX_TYPE_POLL].nr_queues)
4248 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4249 else
4250 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4251}
4252
Christoph Hellwig26a97502021-06-02 09:53:17 +03004253int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4254 struct request_queue *q)
Keith Busch868f2f02015-12-17 17:08:14 -07004255{
Ming Lei66841672016-02-12 15:27:00 +08004256 /* mark the queue as mq asap */
4257 q->mq_ops = set->ops;
4258
Ming Lei1db49092018-11-20 09:44:35 +08004259 if (blk_mq_alloc_ctxs(q))
Keith Busch54bdd672023-03-20 12:49:26 -07004260 goto err_exit;
Keith Busch868f2f02015-12-17 17:08:14 -07004261
Ming Lei737f98c2017-02-22 18:13:59 +08004262 /* init q->mq_kobj and sw queues' kobjects */
4263 blk_mq_sysfs_init(q);
4264
Ming Lei2f8f1332019-04-30 09:52:27 +08004265 INIT_LIST_HEAD(&q->unused_hctx_list);
4266 spin_lock_init(&q->unused_hctx_lock);
4267
Ming Lei4e5cc992022-03-08 15:32:19 +08004268 xa_init(&q->hctx_table);
4269
Keith Busch868f2f02015-12-17 17:08:14 -07004270 blk_mq_realloc_hw_ctxs(set, q);
4271 if (!q->nr_hw_queues)
4272 goto err_hctxs;
Jens Axboe320ae512013-10-24 09:20:05 +01004273
Christoph Hellwig287922e2015-10-30 20:57:30 +08004274 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
Ming Leie56f6982015-07-16 19:53:22 +08004275 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
Jens Axboe320ae512013-10-24 09:20:05 +01004276
Jens Axboea8908932018-10-16 14:23:06 -06004277 q->tag_set = set;
Jens Axboe320ae512013-10-24 09:20:05 +01004278
Jens Axboe94eddfb2013-11-19 09:25:07 -07004279 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
Ming Lei42ee3062022-03-08 15:32:16 +08004280 blk_mq_update_poll_flag(q);
Jens Axboe320ae512013-10-24 09:20:05 +01004281
Mike Snitzer28494502016-09-14 13:28:30 -04004282 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02004283 INIT_LIST_HEAD(&q->flush_list);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06004284 INIT_LIST_HEAD(&q->requeue_list);
4285 spin_lock_init(&q->requeue_lock);
4286
Jens Axboeeba71762014-05-20 15:17:27 -06004287 q->nr_requests = set->queue_depth;
4288
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004289 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
Jens Axboe0d2602c2014-05-13 15:10:52 -06004290 blk_mq_add_queue_tag_set(set, q);
Christoph Hellwig4b855ad2017-06-26 12:20:57 +02004291 blk_mq_map_swqueue(q);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004292 return 0;
Christoph Hellwig18741982014-02-10 09:29:00 -07004293
Jens Axboe320ae512013-10-24 09:20:05 +01004294err_hctxs:
Chen Jun943f45b2022-10-31 03:12:42 +00004295 blk_mq_release(q);
Ming Linc7de5722016-05-25 23:23:27 -07004296err_exit:
4297 q->mq_ops = NULL;
Christoph Hellwig26a97502021-06-02 09:53:17 +03004298 return -ENOMEM;
Jens Axboe320ae512013-10-24 09:20:05 +01004299}
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004300EXPORT_SYMBOL(blk_mq_init_allocated_queue);
Jens Axboe320ae512013-10-24 09:20:05 +01004301
Ming Leic7e2d942019-04-30 09:52:25 +08004302/* tags can _not_ be used after returning from blk_mq_exit_queue */
4303void blk_mq_exit_queue(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01004304{
Bart Van Assche630ef622021-05-13 10:15:29 -07004305 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01004306
Bart Van Assche630ef622021-05-13 10:15:29 -07004307 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
Ming Lei624dbe42014-05-27 23:35:13 +08004308 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
Bart Van Assche630ef622021-05-13 10:15:29 -07004309 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4310 blk_mq_del_queue_tag_set(q);
Jens Axboe320ae512013-10-24 09:20:05 +01004311}
Jens Axboe320ae512013-10-24 09:20:05 +01004312
Jens Axboea5164402014-09-10 09:02:03 -06004313static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4314{
4315 int i;
4316
John Garry079a2e32021-10-05 18:23:39 +08004317 if (blk_mq_is_shared_tags(set->flags)) {
4318 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
John Garrye155b0c2021-10-05 18:23:37 +08004319 BLK_MQ_NO_HCTX_IDX,
4320 set->queue_depth);
John Garry079a2e32021-10-05 18:23:39 +08004321 if (!set->shared_tags)
John Garrye155b0c2021-10-05 18:23:37 +08004322 return -ENOMEM;
4323 }
4324
Xianting Tian8229cca2020-09-26 10:39:47 +08004325 for (i = 0; i < set->nr_hw_queues; i++) {
John Garry63064be2021-10-05 18:23:35 +08004326 if (!__blk_mq_alloc_map_and_rqs(set, i))
Jens Axboea5164402014-09-10 09:02:03 -06004327 goto out_unwind;
Xianting Tian8229cca2020-09-26 10:39:47 +08004328 cond_resched();
4329 }
Jens Axboea5164402014-09-10 09:02:03 -06004330
4331 return 0;
4332
4333out_unwind:
4334 while (--i >= 0)
John Garrye155b0c2021-10-05 18:23:37 +08004335 __blk_mq_free_map_and_rqs(set, i);
4336
John Garry079a2e32021-10-05 18:23:39 +08004337 if (blk_mq_is_shared_tags(set->flags)) {
4338 blk_mq_free_map_and_rqs(set, set->shared_tags,
John Garrye155b0c2021-10-05 18:23:37 +08004339 BLK_MQ_NO_HCTX_IDX);
John Garry645db342021-10-05 18:23:36 +08004340 }
Jens Axboea5164402014-09-10 09:02:03 -06004341
Jens Axboea5164402014-09-10 09:02:03 -06004342 return -ENOMEM;
4343}
4344
4345/*
4346 * Allocate the request maps associated with this tag_set. Note that this
4347 * may reduce the depth asked for, if memory is tight. set->queue_depth
4348 * will be updated to reflect the allocated depth.
4349 */
John Garry63064be2021-10-05 18:23:35 +08004350static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
Jens Axboea5164402014-09-10 09:02:03 -06004351{
4352 unsigned int depth;
4353 int err;
4354
4355 depth = set->queue_depth;
4356 do {
4357 err = __blk_mq_alloc_rq_maps(set);
4358 if (!err)
4359 break;
4360
4361 set->queue_depth >>= 1;
4362 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4363 err = -ENOMEM;
4364 break;
4365 }
4366 } while (set->queue_depth);
4367
4368 if (!set->queue_depth || err) {
4369 pr_err("blk-mq: failed to allocate request map\n");
4370 return -ENOMEM;
4371 }
4372
4373 if (depth != set->queue_depth)
4374 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4375 depth, set->queue_depth);
4376
4377 return 0;
4378}
4379
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004380static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
Omar Sandovalebe8bdd2017-04-07 08:53:11 -06004381{
Bart Van Assche6e66b492020-03-09 21:26:17 -07004382 /*
4383 * blk_mq_map_queues() and multiple .map_queues() implementations
4384 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4385 * number of hardware queues.
4386 */
4387 if (set->nr_maps == 1)
4388 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4389
Ming Lei59388702018-12-07 11:03:53 +08004390 if (set->ops->map_queues && !is_kdump_kernel()) {
Jens Axboeb3c661b2018-10-30 10:36:06 -06004391 int i;
4392
Ming Lei7d4901a2018-01-06 16:27:39 +08004393 /*
4394 * transport .map_queues is usually done in the following
4395 * way:
4396 *
4397 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4398 * mask = get_cpu_mask(queue)
4399 * for_each_cpu(cpu, mask)
Jens Axboeb3c661b2018-10-30 10:36:06 -06004400 * set->map[x].mq_map[cpu] = queue;
Ming Lei7d4901a2018-01-06 16:27:39 +08004401 * }
4402 *
4403 * When we need to remap, the table has to be cleared for
4404 * killing stale mapping since one CPU may not be mapped
4405 * to any hw queue.
4406 */
Jens Axboeb3c661b2018-10-30 10:36:06 -06004407 for (i = 0; i < set->nr_maps; i++)
4408 blk_mq_clear_mq_map(&set->map[i]);
Ming Lei7d4901a2018-01-06 16:27:39 +08004409
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004410 set->ops->map_queues(set);
Jens Axboeb3c661b2018-10-30 10:36:06 -06004411 } else {
4412 BUG_ON(set->nr_maps > 1);
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004413 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
Jens Axboeb3c661b2018-10-30 10:36:06 -06004414 }
Omar Sandovalebe8bdd2017-04-07 08:53:11 -06004415}
4416
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004417static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004418 int new_nr_hw_queues)
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004419{
4420 struct blk_mq_tags **new_tags;
Chengming Zhoue1dd7bc92023-08-21 17:56:00 +08004421 int i;
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004422
Chengming Zhou6be6d112023-09-08 08:57:02 +08004423 if (set->nr_hw_queues >= new_nr_hw_queues)
Shin'ichiro Kawasakid4b2e0d2022-11-22 17:49:17 +09004424 goto done;
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004425
4426 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4427 GFP_KERNEL, set->numa_node);
4428 if (!new_tags)
4429 return -ENOMEM;
4430
4431 if (set->tags)
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004432 memcpy(new_tags, set->tags, set->nr_hw_queues *
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004433 sizeof(*set->tags));
4434 kfree(set->tags);
4435 set->tags = new_tags;
Chengming Zhou72226572023-08-21 17:56:02 +08004436
4437 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4438 if (!__blk_mq_alloc_map_and_rqs(set, i)) {
4439 while (--i >= set->nr_hw_queues)
4440 __blk_mq_free_map_and_rqs(set, i);
4441 return -ENOMEM;
4442 }
4443 cond_resched();
4444 }
4445
Shin'ichiro Kawasakid4b2e0d2022-11-22 17:49:17 +09004446done:
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004447 set->nr_hw_queues = new_nr_hw_queues;
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004448 return 0;
4449}
4450
Jens Axboea4391c62014-06-05 15:21:56 -06004451/*
4452 * Alloc a tag set to be associated with one or more request queues.
4453 * May fail with EINVAL for various error conditions. May adjust the
Minwoo Imc018c842018-06-30 22:12:41 +09004454 * requested depth down, if it's too large. In that case, the set
Jens Axboea4391c62014-06-05 15:21:56 -06004455 * value will be stored in set->queue_depth.
4456 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004457int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4458{
Jens Axboeb3c661b2018-10-30 10:36:06 -06004459 int i, ret;
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004460
Bart Van Assche205fb5f52014-10-30 14:45:11 +01004461 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4462
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004463 if (!set->nr_hw_queues)
4464 return -EINVAL;
Jens Axboea4391c62014-06-05 15:21:56 -06004465 if (!set->queue_depth)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004466 return -EINVAL;
4467 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4468 return -EINVAL;
4469
Christoph Hellwig7d7e0f92016-09-14 16:18:54 +02004470 if (!set->ops->queue_rq)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004471 return -EINVAL;
4472
Ming Leide148292017-10-14 17:22:29 +08004473 if (!set->ops->get_budget ^ !set->ops->put_budget)
4474 return -EINVAL;
4475
Jens Axboea4391c62014-06-05 15:21:56 -06004476 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4477 pr_info("blk-mq: reduced tag depth to %u\n",
4478 BLK_MQ_MAX_DEPTH);
4479 set->queue_depth = BLK_MQ_MAX_DEPTH;
4480 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004481
Jens Axboeb3c661b2018-10-30 10:36:06 -06004482 if (!set->nr_maps)
4483 set->nr_maps = 1;
4484 else if (set->nr_maps > HCTX_MAX_TYPES)
4485 return -EINVAL;
4486
Shaohua Li6637fad2014-11-30 16:00:58 -08004487 /*
4488 * If a crashdump is active, then we are potentially in a very
4489 * memory constrained environment. Limit us to 1 queue and
4490 * 64 tags to prevent using too much memory.
4491 */
4492 if (is_kdump_kernel()) {
4493 set->nr_hw_queues = 1;
Ming Lei59388702018-12-07 11:03:53 +08004494 set->nr_maps = 1;
Shaohua Li6637fad2014-11-30 16:00:58 -08004495 set->queue_depth = min(64U, set->queue_depth);
4496 }
Keith Busch868f2f02015-12-17 17:08:14 -07004497 /*
Jens Axboe392546a2018-10-29 13:25:27 -06004498 * There is no use for more h/w queues than cpus if we just have
4499 * a single map
Keith Busch868f2f02015-12-17 17:08:14 -07004500 */
Jens Axboe392546a2018-10-29 13:25:27 -06004501 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
Keith Busch868f2f02015-12-17 17:08:14 -07004502 set->nr_hw_queues = nr_cpu_ids;
Shaohua Li6637fad2014-11-30 16:00:58 -08004503
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004504 if (set->flags & BLK_MQ_F_BLOCKING) {
4505 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4506 if (!set->srcu)
4507 return -ENOMEM;
4508 ret = init_srcu_struct(set->srcu);
4509 if (ret)
4510 goto out_free_srcu;
4511 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004512
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004513 ret = -ENOMEM;
Christoph Hellwig5ee20292022-11-09 11:08:10 +01004514 set->tags = kcalloc_node(set->nr_hw_queues,
4515 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4516 set->numa_node);
4517 if (!set->tags)
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004518 goto out_cleanup_srcu;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004519
Jens Axboeb3c661b2018-10-30 10:36:06 -06004520 for (i = 0; i < set->nr_maps; i++) {
4521 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
Ming Lei07b35eb2018-12-17 18:42:45 +08004522 sizeof(set->map[i].mq_map[0]),
Jens Axboeb3c661b2018-10-30 10:36:06 -06004523 GFP_KERNEL, set->numa_node);
4524 if (!set->map[i].mq_map)
4525 goto out_free_mq_map;
Ming Lei59388702018-12-07 11:03:53 +08004526 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
Jens Axboeb3c661b2018-10-30 10:36:06 -06004527 }
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004528
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004529 blk_mq_update_queue_map(set);
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004530
John Garry63064be2021-10-05 18:23:35 +08004531 ret = blk_mq_alloc_set_map_and_rqs(set);
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004532 if (ret)
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004533 goto out_free_mq_map;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004534
Jens Axboe0d2602c2014-05-13 15:10:52 -06004535 mutex_init(&set->tag_list_lock);
4536 INIT_LIST_HEAD(&set->tag_list);
4537
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004538 return 0;
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004539
4540out_free_mq_map:
Jens Axboeb3c661b2018-10-30 10:36:06 -06004541 for (i = 0; i < set->nr_maps; i++) {
4542 kfree(set->map[i].mq_map);
4543 set->map[i].mq_map = NULL;
4544 }
Robert Elliott5676e7b2014-09-02 11:38:44 -05004545 kfree(set->tags);
4546 set->tags = NULL;
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004547out_cleanup_srcu:
4548 if (set->flags & BLK_MQ_F_BLOCKING)
4549 cleanup_srcu_struct(set->srcu);
4550out_free_srcu:
4551 if (set->flags & BLK_MQ_F_BLOCKING)
4552 kfree(set->srcu);
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004553 return ret;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004554}
4555EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4556
Christoph Hellwigcdb14e02021-06-02 09:53:16 +03004557/* allocate and initialize a tagset for a simple single-queue device */
4558int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4559 const struct blk_mq_ops *ops, unsigned int queue_depth,
4560 unsigned int set_flags)
4561{
4562 memset(set, 0, sizeof(*set));
4563 set->ops = ops;
4564 set->nr_hw_queues = 1;
4565 set->nr_maps = 1;
4566 set->queue_depth = queue_depth;
4567 set->numa_node = NUMA_NO_NODE;
4568 set->flags = set_flags;
4569 return blk_mq_alloc_tag_set(set);
4570}
4571EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4572
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004573void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4574{
Jens Axboeb3c661b2018-10-30 10:36:06 -06004575 int i, j;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004576
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004577 for (i = 0; i < set->nr_hw_queues; i++)
John Garrye155b0c2021-10-05 18:23:37 +08004578 __blk_mq_free_map_and_rqs(set, i);
Jens Axboe484b4062014-05-21 14:01:15 -06004579
John Garry079a2e32021-10-05 18:23:39 +08004580 if (blk_mq_is_shared_tags(set->flags)) {
4581 blk_mq_free_map_and_rqs(set, set->shared_tags,
John Garrye155b0c2021-10-05 18:23:37 +08004582 BLK_MQ_NO_HCTX_IDX);
4583 }
John Garry32bc15a2020-08-19 23:20:24 +08004584
Jens Axboeb3c661b2018-10-30 10:36:06 -06004585 for (j = 0; j < set->nr_maps; j++) {
4586 kfree(set->map[j].mq_map);
4587 set->map[j].mq_map = NULL;
4588 }
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004589
Ming Lei981bd182014-04-24 00:07:34 +08004590 kfree(set->tags);
Robert Elliott5676e7b2014-09-02 11:38:44 -05004591 set->tags = NULL;
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004592 if (set->flags & BLK_MQ_F_BLOCKING) {
4593 cleanup_srcu_struct(set->srcu);
4594 kfree(set->srcu);
4595 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004596}
4597EXPORT_SYMBOL(blk_mq_free_tag_set);
4598
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004599int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4600{
4601 struct blk_mq_tag_set *set = q->tag_set;
4602 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08004603 int ret;
4604 unsigned long i;
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004605
Jens Axboebd166ef2017-01-17 06:03:22 -07004606 if (!set)
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004607 return -EINVAL;
4608
Aleksei Zakharove5fa8142019-02-08 19:14:05 +03004609 if (q->nr_requests == nr)
4610 return 0;
4611
Jens Axboe70f36b62017-01-19 10:59:07 -07004612 blk_mq_freeze_queue(q);
Ming Lei24f5a902018-01-06 16:27:38 +08004613 blk_mq_quiesce_queue(q);
Jens Axboe70f36b62017-01-19 10:59:07 -07004614
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004615 ret = 0;
4616 queue_for_each_hw_ctx(q, hctx, i) {
Keith Busche9137d42016-02-18 14:56:35 -07004617 if (!hctx->tags)
4618 continue;
Jens Axboebd166ef2017-01-17 06:03:22 -07004619 /*
4620 * If we're using an MQ scheduler, just update the scheduler
4621 * queue depth. This is similar to what the old code would do.
4622 */
John Garryf6adcef2021-10-05 18:23:29 +08004623 if (hctx->sched_tags) {
Jens Axboe70f36b62017-01-19 10:59:07 -07004624 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
John Garryf6adcef2021-10-05 18:23:29 +08004625 nr, true);
John Garryf6adcef2021-10-05 18:23:29 +08004626 } else {
4627 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4628 false);
Jens Axboe70f36b62017-01-19 10:59:07 -07004629 }
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004630 if (ret)
4631 break;
Jens Axboe77f1e0a2019-01-18 10:34:16 -07004632 if (q->elevator && q->elevator->type->ops.depth_updated)
4633 q->elevator->type->ops.depth_updated(hctx);
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004634 }
John Garryd97e5942021-05-13 20:00:58 +08004635 if (!ret) {
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004636 q->nr_requests = nr;
John Garry079a2e32021-10-05 18:23:39 +08004637 if (blk_mq_is_shared_tags(set->flags)) {
John Garry8fa04462021-10-05 18:23:28 +08004638 if (q->elevator)
John Garry079a2e32021-10-05 18:23:39 +08004639 blk_mq_tag_update_sched_shared_tags(q);
John Garry8fa04462021-10-05 18:23:28 +08004640 else
John Garry079a2e32021-10-05 18:23:39 +08004641 blk_mq_tag_resize_shared_tags(set, nr);
John Garry8fa04462021-10-05 18:23:28 +08004642 }
John Garryd97e5942021-05-13 20:00:58 +08004643 }
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004644
Ming Lei24f5a902018-01-06 16:27:38 +08004645 blk_mq_unquiesce_queue(q);
Jens Axboe70f36b62017-01-19 10:59:07 -07004646 blk_mq_unfreeze_queue(q);
Jens Axboe70f36b62017-01-19 10:59:07 -07004647
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004648 return ret;
4649}
4650
Jianchao Wangd48ece22018-08-21 15:15:03 +08004651/*
4652 * request_queue and elevator_type pair.
4653 * It is just used by __blk_mq_update_nr_hw_queues to cache
4654 * the elevator_type associated with a request_queue.
4655 */
4656struct blk_mq_qe_pair {
4657 struct list_head node;
4658 struct request_queue *q;
4659 struct elevator_type *type;
4660};
4661
4662/*
4663 * Cache the elevator_type in qe pair list and switch the
4664 * io scheduler to 'none'
4665 */
4666static bool blk_mq_elv_switch_none(struct list_head *head,
4667 struct request_queue *q)
4668{
4669 struct blk_mq_qe_pair *qe;
4670
Jianchao Wangd48ece22018-08-21 15:15:03 +08004671 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4672 if (!qe)
4673 return false;
4674
Ming Lei5fd7a842022-06-16 09:43:59 +08004675 /* q->elevator needs protection from ->sysfs_lock */
4676 mutex_lock(&q->sysfs_lock);
4677
Ming Lei24516562023-06-16 21:23:54 +08004678 /* the check has to be done with holding sysfs_lock */
4679 if (!q->elevator) {
4680 kfree(qe);
4681 goto unlock;
4682 }
4683
Jianchao Wangd48ece22018-08-21 15:15:03 +08004684 INIT_LIST_HEAD(&qe->node);
4685 qe->q = q;
4686 qe->type = q->elevator->type;
Christoph Hellwigdd6f7f12022-10-20 08:48:16 +02004687 /* keep a reference to the elevator module as we'll switch back */
4688 __elevator_get(qe->type);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004689 list_add(&qe->node, head);
Christoph Hellwig64b36072022-10-30 11:07:14 +01004690 elevator_disable(q);
Ming Lei24516562023-06-16 21:23:54 +08004691unlock:
Jianchao Wangd48ece22018-08-21 15:15:03 +08004692 mutex_unlock(&q->sysfs_lock);
4693
4694 return true;
4695}
4696
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004697static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4698 struct request_queue *q)
Jianchao Wangd48ece22018-08-21 15:15:03 +08004699{
4700 struct blk_mq_qe_pair *qe;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004701
4702 list_for_each_entry(qe, head, node)
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004703 if (qe->q == q)
4704 return qe;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004705
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004706 return NULL;
4707}
4708
4709static void blk_mq_elv_switch_back(struct list_head *head,
4710 struct request_queue *q)
4711{
4712 struct blk_mq_qe_pair *qe;
4713 struct elevator_type *t;
4714
4715 qe = blk_lookup_qe_pair(head, q);
4716 if (!qe)
Jianchao Wangd48ece22018-08-21 15:15:03 +08004717 return;
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004718 t = qe->type;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004719 list_del(&qe->node);
4720 kfree(qe);
4721
4722 mutex_lock(&q->sysfs_lock);
Keith Busch8237c012022-09-27 08:56:52 -07004723 elevator_switch(q, t);
Jinlong Chen8ed40ee2022-10-20 08:48:19 +02004724 /* drop the reference acquired in blk_mq_elv_switch_none */
4725 elevator_put(t);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004726 mutex_unlock(&q->sysfs_lock);
4727}
4728
Keith Busche4dc2b32017-05-30 14:39:11 -04004729static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4730 int nr_hw_queues)
Keith Busch868f2f02015-12-17 17:08:14 -07004731{
4732 struct request_queue *q;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004733 LIST_HEAD(head);
Chengming Zhou6be6d112023-09-08 08:57:02 +08004734 int prev_nr_hw_queues = set->nr_hw_queues;
4735 int i;
Keith Busch868f2f02015-12-17 17:08:14 -07004736
Bart Van Assche705cda92017-04-07 11:16:49 -07004737 lockdep_assert_held(&set->tag_list_lock);
4738
Jens Axboe392546a2018-10-29 13:25:27 -06004739 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
Keith Busch868f2f02015-12-17 17:08:14 -07004740 nr_hw_queues = nr_cpu_ids;
Weiping Zhangfe35ec52020-06-17 14:18:37 +08004741 if (nr_hw_queues < 1)
4742 return;
4743 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
Keith Busch868f2f02015-12-17 17:08:14 -07004744 return;
4745
4746 list_for_each_entry(q, &set->tag_list, tag_set_list)
4747 blk_mq_freeze_queue(q);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004748 /*
4749 * Switch IO scheduler to 'none', cleaning up the data associated
4750 * with the previous scheduler. We will switch back once we are done
4751 * updating the new sw to hw queue mappings.
4752 */
4753 list_for_each_entry(q, &set->tag_list, tag_set_list)
4754 if (!blk_mq_elv_switch_none(&head, q))
4755 goto switch_back;
Keith Busch868f2f02015-12-17 17:08:14 -07004756
Jianchao Wang477e19d2018-10-12 18:07:25 +08004757 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4758 blk_mq_debugfs_unregister_hctxs(q);
Christoph Hellwigeaa870f2022-06-28 19:18:49 +02004759 blk_mq_sysfs_unregister_hctxs(q);
Jianchao Wang477e19d2018-10-12 18:07:25 +08004760 }
4761
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004762 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004763 goto reregister;
4764
Jianchao Wange01ad462018-10-12 18:07:28 +08004765fallback:
Weiping Zhangaa880ad2020-05-13 08:44:05 +08004766 blk_mq_update_queue_map(set);
Keith Busch868f2f02015-12-17 17:08:14 -07004767 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4768 blk_mq_realloc_hw_ctxs(set, q);
Ming Lei42ee3062022-03-08 15:32:16 +08004769 blk_mq_update_poll_flag(q);
Jianchao Wange01ad462018-10-12 18:07:28 +08004770 if (q->nr_hw_queues != set->nr_hw_queues) {
Ye Bina846a8e2021-11-08 15:40:19 +08004771 int i = prev_nr_hw_queues;
4772
Jianchao Wange01ad462018-10-12 18:07:28 +08004773 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4774 nr_hw_queues, prev_nr_hw_queues);
Ye Bina846a8e2021-11-08 15:40:19 +08004775 for (; i < set->nr_hw_queues; i++)
4776 __blk_mq_free_map_and_rqs(set, i);
4777
Jianchao Wange01ad462018-10-12 18:07:28 +08004778 set->nr_hw_queues = prev_nr_hw_queues;
Jianchao Wange01ad462018-10-12 18:07:28 +08004779 goto fallback;
4780 }
Jianchao Wang477e19d2018-10-12 18:07:25 +08004781 blk_mq_map_swqueue(q);
4782 }
4783
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004784reregister:
Jianchao Wang477e19d2018-10-12 18:07:25 +08004785 list_for_each_entry(q, &set->tag_list, tag_set_list) {
Christoph Hellwigeaa870f2022-06-28 19:18:49 +02004786 blk_mq_sysfs_register_hctxs(q);
Jianchao Wang477e19d2018-10-12 18:07:25 +08004787 blk_mq_debugfs_register_hctxs(q);
Keith Busch868f2f02015-12-17 17:08:14 -07004788 }
4789
Jianchao Wangd48ece22018-08-21 15:15:03 +08004790switch_back:
4791 list_for_each_entry(q, &set->tag_list, tag_set_list)
4792 blk_mq_elv_switch_back(&head, q);
4793
Keith Busch868f2f02015-12-17 17:08:14 -07004794 list_for_each_entry(q, &set->tag_list, tag_set_list)
4795 blk_mq_unfreeze_queue(q);
Chengming Zhou6be6d112023-09-08 08:57:02 +08004796
4797 /* Free the excess tags when nr_hw_queues shrink. */
4798 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
4799 __blk_mq_free_map_and_rqs(set, i);
Keith Busch868f2f02015-12-17 17:08:14 -07004800}
Keith Busche4dc2b32017-05-30 14:39:11 -04004801
4802void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4803{
4804 mutex_lock(&set->tag_list_lock);
4805 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4806 mutex_unlock(&set->tag_list_lock);
4807}
Keith Busch868f2f02015-12-17 17:08:14 -07004808EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4809
Keith Buschf6c80cf2023-06-12 12:03:42 -07004810static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4811 struct io_comp_batch *iob, unsigned int flags)
Jens Axboebbd7bb72016-11-04 09:34:34 -06004812{
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004813 long state = get_current_state();
4814 int ret;
Jens Axboe1052b8a2018-11-26 08:21:49 -07004815
Jens Axboeaa61bec2018-11-13 21:32:10 -07004816 do {
Jens Axboe5a72e892021-10-12 09:24:29 -06004817 ret = q->mq_ops->poll(hctx, iob);
Jens Axboebbd7bb72016-11-04 09:34:34 -06004818 if (ret > 0) {
Jens Axboe849a3702018-11-16 08:37:34 -07004819 __set_current_state(TASK_RUNNING);
Jens Axboe85f4d4b2018-11-06 13:30:55 -07004820 return ret;
Jens Axboebbd7bb72016-11-04 09:34:34 -06004821 }
4822
4823 if (signal_pending_state(state, current))
Jens Axboe849a3702018-11-16 08:37:34 -07004824 __set_current_state(TASK_RUNNING);
Peter Zijlstrab03fbd42021-06-11 10:28:12 +02004825 if (task_is_running(current))
Jens Axboe85f4d4b2018-11-06 13:30:55 -07004826 return 1;
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004827
Christoph Hellwigef99b2d2021-10-12 13:12:19 +02004828 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
Jens Axboebbd7bb72016-11-04 09:34:34 -06004829 break;
4830 cpu_relax();
Jens Axboeaa61bec2018-11-13 21:32:10 -07004831 } while (!need_resched());
Jens Axboebbd7bb72016-11-04 09:34:34 -06004832
Nitesh Shetty67b41102018-02-13 21:18:12 +05304833 __set_current_state(TASK_RUNNING);
Jens Axboe85f4d4b2018-11-06 13:30:55 -07004834 return 0;
Jens Axboebbd7bb72016-11-04 09:34:34 -06004835}
Jens Axboebbd7bb72016-11-04 09:34:34 -06004836
Keith Buschf6c80cf2023-06-12 12:03:42 -07004837int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4838 struct io_comp_batch *iob, unsigned int flags)
4839{
4840 struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4841
4842 return blk_hctx_poll(q, hctx, iob, flags);
4843}
4844
4845int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4846 unsigned int poll_flags)
4847{
4848 struct request_queue *q = rq->q;
4849 int ret;
4850
4851 if (!blk_rq_is_poll(rq))
4852 return 0;
4853 if (!percpu_ref_tryget(&q->q_usage_counter))
4854 return 0;
4855
4856 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4857 blk_queue_exit(q);
4858
4859 return ret;
4860}
4861EXPORT_SYMBOL_GPL(blk_rq_poll);
4862
Jens Axboe9cf2bab2018-10-31 17:01:22 -06004863unsigned int blk_mq_rq_cpu(struct request *rq)
4864{
4865 return rq->mq_ctx->cpu;
4866}
4867EXPORT_SYMBOL(blk_mq_rq_cpu);
4868
Ming Lei2a19b282021-11-16 09:43:43 +08004869void blk_mq_cancel_work_sync(struct request_queue *q)
4870{
Jinlong Chen219cf432022-10-30 17:47:30 +08004871 struct blk_mq_hw_ctx *hctx;
4872 unsigned long i;
Ming Lei2a19b282021-11-16 09:43:43 +08004873
Jinlong Chen219cf432022-10-30 17:47:30 +08004874 cancel_delayed_work_sync(&q->requeue_work);
Ming Lei2a19b282021-11-16 09:43:43 +08004875
Jinlong Chen219cf432022-10-30 17:47:30 +08004876 queue_for_each_hw_ctx(q, hctx, i)
4877 cancel_delayed_work_sync(&hctx->run_work);
Ming Lei2a19b282021-11-16 09:43:43 +08004878}
4879
Jens Axboe320ae512013-10-24 09:20:05 +01004880static int __init blk_mq_init(void)
4881{
Christoph Hellwigc3077b52020-06-11 08:44:41 +02004882 int i;
4883
4884 for_each_possible_cpu(i)
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01004885 init_llist_head(&per_cpu(blk_cpu_done, i));
Chengming Zhou660e8022023-07-17 12:00:55 +08004886 for_each_possible_cpu(i)
4887 INIT_CSD(&per_cpu(blk_cpu_csd, i),
4888 __blk_mq_complete_request_remote, NULL);
Christoph Hellwigc3077b52020-06-11 08:44:41 +02004889 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4890
4891 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4892 "block/softirq:dead", NULL,
4893 blk_softirq_cpu_dead);
Thomas Gleixner9467f852016-09-22 08:05:17 -06004894 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4895 blk_mq_hctx_notify_dead);
Ming Leibf0beec2020-05-29 15:53:15 +02004896 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4897 blk_mq_hctx_notify_online,
4898 blk_mq_hctx_notify_offline);
Jens Axboe320ae512013-10-24 09:20:05 +01004899 return 0;
4900}
4901subsys_initcall(blk_mq_init);