blob: e3c3c0c21b55367b5710fcd9e02f325d4169bca6 [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>
Ingo Molnar105ab3d82017-02-01 16:36:40 +010024#include <linux/sched/topology.h>
Ingo Molnar174cd4b2017-02-02 19:15:33 +010025#include <linux/sched/signal.h>
Jens Axboe320ae512013-10-24 09:20:05 +010026#include <linux/delay.h>
Jens Axboeaedcd722014-09-17 08:27:03 -060027#include <linux/crash_dump.h>
Jens Axboe88c7b2b2016-08-25 08:07:30 -060028#include <linux/prefetch.h>
Satya Tangiralaa892c8d2020-05-14 00:37:18 +000029#include <linux/blk-crypto.h>
Christoph Hellwig82d981d2021-11-23 19:53:12 +010030#include <linux/part_stat.h>
Ming Leia46c2702024-03-22 10:12:44 +080031#include <linux/sched/isolation.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) &&
Al Viro3f9b8fb2024-04-12 00:54:19 -040096 (!bdev_is_partition(mi->part) || 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;
Jens Axboe08420cf2024-01-15 14:45:07 -0700325 rq->start_time_ns = blk_time_get_ns();
Christoph Hellwig52fdbbc2021-11-17 07:13:59 +0100326 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))
Jens Axboe08420cf2024-01-15 14:45:07 -0700335 rq->start_time_ns = blk_time_get_ns();
Chengming Zhou5c17f452023-07-10 18:55:16 +0800336 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))
Jens Axboe08420cf2024-01-15 14:45:07 -0700446 alloc_time_ns = blk_time_get_ns();
Tejun Heo6f816b42019-08-28 15:05:57 -0700447
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
Bart Van Assche62591512024-05-09 10:01:48 -0700451retry:
452 data->ctx = blk_mq_get_ctx(q);
453 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
454
Jens Axboe781dd832021-11-02 08:34:09 -0600455 if (q->elevator) {
Christoph Hellwigdd6216b2023-05-18 07:31:01 +0200456 /*
457 * All requests use scheduler tags when an I/O scheduler is
458 * enabled for the queue.
459 */
460 data->rq_flags |= RQF_SCHED_TAGS;
Jens Axboe781dd832021-11-02 08:34:09 -0600461
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200462 /*
Lin Feng8d663f32021-04-15 11:39:20 +0800463 * Flush/passthrough requests are special and go directly to the
Christoph Hellwigdd6216b2023-05-18 07:31:01 +0200464 * dispatch list.
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200465 */
Bart Van Asschebe4c4272023-05-19 06:40:47 +0200466 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
Christoph Hellwigdd6216b2023-05-18 07:31:01 +0200467 !blk_op_is_passthrough(data->cmd_flags)) {
468 struct elevator_mq_ops *ops = &q->elevator->type->ops;
469
470 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
471
472 data->rq_flags |= RQF_USE_SCHED;
473 if (ops->limit_depth)
474 ops->limit_depth(data->cmd_flags, data);
475 }
Bart Van Assche62591512024-05-09 10:01:48 -0700476 } else {
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200477 blk_mq_tag_busy(data->hctx);
Bart Van Assche62591512024-05-09 10:01:48 -0700478 }
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200479
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 Hellwig5bbf4e52017-06-16 18:15:26 +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))
Jens Axboe08420cf2024-01-15 14:45:07 -0700631 alloc_time_ns = blk_time_get_ns();
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200632
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
Damien Le Moal347bde92024-05-01 20:09:04 +0900694 blk_zone_finish_request(rq);
695
Chengming Zhoue5c0ca12023-08-13 23:23:25 +0800696 if (rq->rq_flags & RQF_USE_SCHED) {
697 q->elevator->type->ops.finish_request(rq);
698 /*
699 * For postflush request that may need to be
700 * completed twice, we should clear this flag
701 * to avoid double finish_request() on the rq.
702 */
703 rq->rq_flags &= ~RQF_USE_SCHED;
704 }
705}
706
Keith Busch12f5b932018-05-29 15:52:28 +0200707static void __blk_mq_free_request(struct request *rq)
708{
709 struct request_queue *q = rq->q;
710 struct blk_mq_ctx *ctx = rq->mq_ctx;
Jens Axboeea4f9952018-10-29 15:06:13 -0600711 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Keith Busch12f5b932018-05-29 15:52:28 +0200712 const int sched_tag = rq->internal_tag;
713
Satya Tangiralaa892c8d2020-05-14 00:37:18 +0000714 blk_crypto_free_request(rq);
Bart Van Assche986d4132018-09-26 14:01:10 -0700715 blk_pm_mark_last_busy(rq);
Jens Axboeea4f9952018-10-29 15:06:13 -0600716 rq->mq_hctx = NULL;
Tian Landdad59332023-05-13 18:12:27 -0400717
Chengming Zhoub8643d62023-09-13 15:16:12 +0000718 if (rq->tag != BLK_MQ_NO_TAG) {
719 blk_mq_dec_active_requests(hctx);
John Garrycae740a2020-02-26 20:10:15 +0800720 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
Chengming Zhoub8643d62023-09-13 15:16:12 +0000721 }
Christoph Hellwig766473682020-05-29 15:53:12 +0200722 if (sched_tag != BLK_MQ_NO_TAG)
John Garrycae740a2020-02-26 20:10:15 +0800723 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
Keith Busch12f5b932018-05-29 15:52:28 +0200724 blk_mq_sched_restart(hctx);
725 blk_queue_exit(q);
726}
727
Christoph Hellwig6af54052017-06-16 18:15:22 +0200728void blk_mq_free_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100729{
Jens Axboe320ae512013-10-24 09:20:05 +0100730 struct request_queue *q = rq->q;
731
Chengming Zhoue5c0ca12023-08-13 23:23:25 +0800732 blk_mq_finish_request(rq);
Christoph Hellwig6af54052017-06-16 18:15:22 +0200733
Jens Axboe7beb2f82017-09-30 02:08:24 -0600734 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
Christoph Hellwigd152c682021-08-16 15:46:24 +0200735 laptop_io_completion(q->disk->bdi);
Jens Axboe7beb2f82017-09-30 02:08:24 -0600736
Josef Bacika7905042018-07-03 09:32:35 -0600737 rq_qos_done(q, rq);
Jens Axboe0d2602c2014-05-13 15:10:52 -0600738
Keith Busch12f5b932018-05-29 15:52:28 +0200739 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Jens Axboe0a467d02021-10-14 14:39:59 -0600740 if (req_ref_put_and_test(rq))
Keith Busch12f5b932018-05-29 15:52:28 +0200741 __blk_mq_free_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100742}
Jens Axboe1a3b5952014-11-17 10:40:48 -0700743EXPORT_SYMBOL_GPL(blk_mq_free_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100744
Jens Axboe47c122e2021-10-06 06:34:11 -0600745void blk_mq_free_plug_rqs(struct blk_plug *plug)
Jens Axboe320ae512013-10-24 09:20:05 +0100746{
Jens Axboe013a7f92021-10-13 07:58:52 -0600747 struct request *rq;
Jens Axboefe1f4522018-11-28 10:50:07 -0700748
Jens Axboec5fc7b92021-11-03 05:49:07 -0600749 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
Jens Axboe47c122e2021-10-06 06:34:11 -0600750 blk_mq_free_request(rq);
Jens Axboe47c122e2021-10-06 06:34:11 -0600751}
Omar Sandoval522a7772018-05-09 02:08:53 -0700752
Christoph Hellwig22350ad2021-11-17 07:14:02 +0100753void blk_dump_rq_flags(struct request *rq, char *msg)
754{
755 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
Christoph Hellwigf3fa33a2021-11-26 13:18:00 +0100756 rq->q->disk ? rq->q->disk->disk_name : "?",
Bart Van Assche16458cf2022-07-14 11:06:32 -0700757 (__force unsigned long long) rq->cmd_flags);
Christoph Hellwig22350ad2021-11-17 07:14:02 +0100758
759 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
760 (unsigned long long)blk_rq_pos(rq),
761 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
762 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
763 rq->bio, rq->biotail, blk_rq_bytes(rq));
764}
765EXPORT_SYMBOL(blk_dump_rq_flags);
766
Jens Axboe9be3e062021-10-14 09:17:01 -0600767static void blk_account_io_completion(struct request *req, unsigned int bytes)
768{
769 if (req->part && blk_do_io_stat(req)) {
770 const int sgrp = op_stat_group(req_op(req));
771
772 part_stat_lock();
773 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
774 part_stat_unlock();
775 }
776}
777
Christoph Hellwig0d7a29a2021-11-17 07:14:03 +0100778static void blk_print_req_error(struct request *req, blk_status_t status)
779{
780 printk_ratelimited(KERN_ERR
781 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
782 "phys_seg %u prio class %u\n",
783 blk_status_to_str(status),
Christoph Hellwigf3fa33a2021-11-26 13:18:00 +0100784 req->q->disk ? req->q->disk->disk_name : "?",
Bart Van Assche16458cf2022-07-14 11:06:32 -0700785 blk_rq_pos(req), (__force u32)req_op(req),
786 blk_op_str(req_op(req)),
787 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
Christoph Hellwig0d7a29a2021-11-17 07:14:03 +0100788 req->nr_phys_segments,
789 IOPRIO_PRIO_CLASS(req->ioprio));
790}
791
Jens Axboe5581a5d2021-12-01 15:01:51 -0700792/*
793 * Fully end IO on a request. Does not support partial completions, or
794 * errors.
795 */
796static void blk_complete_request(struct request *req)
797{
798 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
799 int total_bytes = blk_rq_bytes(req);
800 struct bio *bio = req->bio;
801
802 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
803
804 if (!bio)
805 return;
806
Jens Axboe5581a5d2021-12-01 15:01:51 -0700807 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
Christoph Hellwige9f5f442024-06-13 10:48:15 +0200808 blk_integrity_complete(req, total_bytes);
Jens Axboe5581a5d2021-12-01 15:01:51 -0700809
Eric Biggers9cd1e562023-03-15 11:39:02 -0700810 /*
811 * Upper layers may call blk_crypto_evict_key() anytime after the last
812 * bio_endio(). Therefore, the keyslot must be released before that.
813 */
814 blk_crypto_rq_put_keyslot(req);
815
Jens Axboe5581a5d2021-12-01 15:01:51 -0700816 blk_account_io_completion(req, total_bytes);
817
818 do {
819 struct bio *next = bio->bi_next;
820
821 /* Completion has already been traced */
822 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
Pankaj Raghava12821d2022-02-11 10:34:25 +0100823
Damien Le Moala0508c32024-04-08 10:41:03 +0900824 blk_zone_update_request_bio(req, bio);
Pankaj Raghava12821d2022-02-11 10:34:25 +0100825
Jens Axboe5581a5d2021-12-01 15:01:51 -0700826 if (!is_flush)
827 bio_endio(bio);
828 bio = next;
829 } while (bio);
830
831 /*
832 * Reset counters so that the request stacking driver
833 * can find how many bytes remain in the request
834 * later.
835 */
Jens Axboeab3e1d32022-09-21 08:24:16 -0600836 if (!req->end_io) {
837 req->bio = NULL;
838 req->__data_len = 0;
839 }
Jens Axboe5581a5d2021-12-01 15:01:51 -0700840}
841
Jens Axboe9be3e062021-10-14 09:17:01 -0600842/**
843 * blk_update_request - Complete multiple bytes without completing the request
844 * @req: the request being processed
845 * @error: block status code
846 * @nr_bytes: number of bytes to complete for @req
847 *
848 * Description:
849 * Ends I/O on a number of bytes attached to @req, but doesn't complete
850 * the request structure even if @req doesn't have leftover.
851 * If @req has leftover, sets it up for the next range of segments.
852 *
853 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
854 * %false return from this function.
855 *
856 * Note:
857 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
858 * except in the consistency check at the end of this function.
859 *
860 * Return:
861 * %false - this request doesn't have any more data
862 * %true - this request has more data
863 **/
864bool blk_update_request(struct request *req, blk_status_t error,
865 unsigned int nr_bytes)
866{
Damien Le Moalc0da26f2024-04-08 10:41:02 +0900867 bool is_flush = req->rq_flags & RQF_FLUSH_SEQ;
868 bool quiet = req->rq_flags & RQF_QUIET;
Jens Axboe9be3e062021-10-14 09:17:01 -0600869 int total_bytes;
870
Christoph Hellwig8a7d2672021-10-18 10:45:18 +0200871 trace_block_rq_complete(req, error, nr_bytes);
Jens Axboe9be3e062021-10-14 09:17:01 -0600872
873 if (!req->bio)
874 return false;
875
Jens Axboe9be3e062021-10-14 09:17:01 -0600876 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
877 error == BLK_STS_OK)
Christoph Hellwige9f5f442024-06-13 10:48:15 +0200878 blk_integrity_complete(req, nr_bytes);
Jens Axboe9be3e062021-10-14 09:17:01 -0600879
Eric Biggers9cd1e562023-03-15 11:39:02 -0700880 /*
881 * Upper layers may call blk_crypto_evict_key() anytime after the last
882 * bio_endio(). Therefore, the keyslot must be released before that.
883 */
884 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
885 __blk_crypto_rq_put_keyslot(req);
886
Damien Le Moalc0da26f2024-04-08 10:41:02 +0900887 if (unlikely(error && !blk_rq_is_passthrough(req) && !quiet) &&
888 !test_bit(GD_DEAD, &req->q->disk->state)) {
Jens Axboe9be3e062021-10-14 09:17:01 -0600889 blk_print_req_error(req, error);
Yang Shid5869fd2022-02-10 14:52:22 -0800890 trace_block_rq_error(req, error, nr_bytes);
891 }
Jens Axboe9be3e062021-10-14 09:17:01 -0600892
893 blk_account_io_completion(req, nr_bytes);
894
895 total_bytes = 0;
896 while (req->bio) {
897 struct bio *bio = req->bio;
898 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
899
Damien Le Moalc0da26f2024-04-08 10:41:02 +0900900 if (unlikely(error))
901 bio->bi_status = error;
902
903 if (bio_bytes == bio->bi_iter.bi_size) {
Jens Axboe9be3e062021-10-14 09:17:01 -0600904 req->bio = bio->bi_next;
Damien Le Moal9b1ce7f2024-04-08 10:41:10 +0900905 } else if (bio_is_zone_append(bio) && error == BLK_STS_OK) {
Damien Le Moalc0da26f2024-04-08 10:41:02 +0900906 /*
907 * Partial zone append completions cannot be supported
908 * as the BIO fragments may end up not being written
909 * sequentially.
910 */
911 bio->bi_status = BLK_STS_IOERR;
912 }
Jens Axboe9be3e062021-10-14 09:17:01 -0600913
914 /* Completion has already been traced */
915 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
Damien Le Moalc0da26f2024-04-08 10:41:02 +0900916 if (unlikely(quiet))
917 bio_set_flag(bio, BIO_QUIET);
918
919 bio_advance(bio, bio_bytes);
920
921 /* Don't actually finish bio if it's part of flush sequence */
Damien Le Moala0508c32024-04-08 10:41:03 +0900922 if (!bio->bi_iter.bi_size) {
923 blk_zone_update_request_bio(req, bio);
924 if (!is_flush)
925 bio_endio(bio);
Damien Le Moalc0da26f2024-04-08 10:41:02 +0900926 }
Jens Axboe9be3e062021-10-14 09:17:01 -0600927
928 total_bytes += bio_bytes;
929 nr_bytes -= bio_bytes;
930
931 if (!nr_bytes)
932 break;
933 }
934
935 /*
936 * completely done
937 */
938 if (!req->bio) {
939 /*
940 * Reset counters so that the request stacking driver
941 * can find how many bytes remain in the request
942 * later.
943 */
944 req->__data_len = 0;
945 return false;
946 }
947
948 req->__data_len -= total_bytes;
949
950 /* update sector only for requests with clear definition of sector */
951 if (!blk_rq_is_passthrough(req))
952 req->__sector += total_bytes >> 9;
953
954 /* mixed attributes always follow the first bio */
955 if (req->rq_flags & RQF_MIXED_MERGE) {
956 req->cmd_flags &= ~REQ_FAILFAST_MASK;
957 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
958 }
959
960 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
961 /*
962 * If total number of sectors is less than the first segment
963 * size, something has gone terribly wrong.
964 */
965 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
966 blk_dump_rq_flags(req, "request botched");
967 req->__data_len = blk_rq_cur_bytes(req);
968 }
969
970 /* recalculate the number of segments */
971 req->nr_phys_segments = blk_recalc_rq_segments(req);
972 }
973
974 return true;
975}
976EXPORT_SYMBOL_GPL(blk_update_request);
977
Christoph Hellwig450b7872021-11-17 07:14:01 +0100978static inline void blk_account_io_done(struct request *req, u64 now)
979{
Hengqi Chen5a80bd02023-05-20 08:40:57 +0000980 trace_block_io_done(req);
981
Christoph Hellwig450b7872021-11-17 07:14:01 +0100982 /*
983 * Account IO completion. flush_rq isn't accounted as a
984 * normal IO on queueing nor completion. Accounting the
985 * containing request is enough.
986 */
987 if (blk_do_io_stat(req) && req->part &&
Chaitanya Kulkarni06965032023-03-27 00:34:27 -0700988 !(req->rq_flags & RQF_FLUSH_SEQ)) {
989 const int sgrp = op_stat_group(req_op(req));
Christoph Hellwig450b7872021-11-17 07:14:01 +0100990
Chaitanya Kulkarni06965032023-03-27 00:34:27 -0700991 part_stat_lock();
992 update_io_ticks(req->part, jiffies, true);
993 part_stat_inc(req->part, ios[sgrp]);
994 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
Yu Kuai99dc4222024-05-09 20:37:16 +0800995 part_stat_local_dec(req->part,
996 in_flight[op_is_write(req_op(req))]);
Chaitanya Kulkarni06965032023-03-27 00:34:27 -0700997 part_stat_unlock();
998 }
Christoph Hellwig450b7872021-11-17 07:14:01 +0100999}
1000
1001static inline void blk_account_io_start(struct request *req)
1002{
Hengqi Chen5a80bd02023-05-20 08:40:57 +00001003 trace_block_io_start(req);
1004
Chaitanya Kulkarnie165fb42023-03-27 00:34:26 -07001005 if (blk_do_io_stat(req)) {
1006 /*
1007 * All non-passthrough requests are created from a bio with one
1008 * exception: when a flush command that is part of a flush sequence
1009 * generated by the state machine in blk-flush.c is cloned onto the
1010 * lower device by dm-multipath we can get here without a bio.
1011 */
1012 if (req->bio)
1013 req->part = req->bio->bi_bdev;
1014 else
1015 req->part = req->q->disk->part0;
1016
1017 part_stat_lock();
1018 update_io_ticks(req->part, jiffies, false);
Yu Kuai99dc4222024-05-09 20:37:16 +08001019 part_stat_local_inc(req->part,
1020 in_flight[op_is_write(req_op(req))]);
Chaitanya Kulkarnie165fb42023-03-27 00:34:26 -07001021 part_stat_unlock();
1022 }
Christoph Hellwig450b7872021-11-17 07:14:01 +01001023}
1024
Jens Axboef794f332021-10-08 05:50:46 -06001025static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1026{
Keith Busch54bdd672023-03-20 12:49:26 -07001027 if (rq->rq_flags & RQF_STATS)
Omar Sandoval522a7772018-05-09 02:08:53 -07001028 blk_stat_add(rq, now);
Omar Sandoval4bc63392018-05-09 02:08:52 -07001029
Baolin Wang87890092020-07-04 15:28:21 +08001030 blk_mq_sched_completed_request(rq, now);
Omar Sandoval522a7772018-05-09 02:08:53 -07001031 blk_account_io_done(rq, now);
Jens Axboef794f332021-10-08 05:50:46 -06001032}
1033
Ming Lei0d11e6a2013-12-05 10:50:39 -07001034inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
Christoph Hellwig91b63632014-04-16 09:44:53 +02001035{
Jens Axboef794f332021-10-08 05:50:46 -06001036 if (blk_mq_need_time_stamp(rq))
Jens Axboe08420cf2024-01-15 14:45:07 -07001037 __blk_mq_end_request_acct(rq, blk_time_get_ns());
Jens Axboe320ae512013-10-24 09:20:05 +01001038
Chengming Zhoue5c0ca12023-08-13 23:23:25 +08001039 blk_mq_finish_request(rq);
1040
Jens Axboe320ae512013-10-24 09:20:05 +01001041 if (rq->end_io) {
Josef Bacika7905042018-07-03 09:32:35 -06001042 rq_qos_done(rq->q, rq);
Jens Axboede671d62022-09-21 15:19:54 -06001043 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1044 blk_mq_free_request(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +02001045 } else {
Jens Axboe320ae512013-10-24 09:20:05 +01001046 blk_mq_free_request(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +02001047 }
Jens Axboe320ae512013-10-24 09:20:05 +01001048}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -07001049EXPORT_SYMBOL(__blk_mq_end_request);
Christoph Hellwig63151a42014-04-16 09:44:52 +02001050
Christoph Hellwig2a842ac2017-06-03 09:38:04 +02001051void blk_mq_end_request(struct request *rq, blk_status_t error)
Christoph Hellwig63151a42014-04-16 09:44:52 +02001052{
1053 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1054 BUG();
Christoph Hellwigc8a446a2014-09-13 16:40:10 -07001055 __blk_mq_end_request(rq, error);
Christoph Hellwig63151a42014-04-16 09:44:52 +02001056}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -07001057EXPORT_SYMBOL(blk_mq_end_request);
Jens Axboe320ae512013-10-24 09:20:05 +01001058
Jens Axboef794f332021-10-08 05:50:46 -06001059#define TAG_COMP_BATCH 32
1060
1061static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1062 int *tag_array, int nr_tags)
1063{
1064 struct request_queue *q = hctx->queue;
1065
Chengming Zhoub8643d62023-09-13 15:16:12 +00001066 blk_mq_sub_active_requests(hctx, nr_tags);
Ming Lei3b87c6e2021-11-02 23:36:19 +08001067
Jens Axboef794f332021-10-08 05:50:46 -06001068 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1069 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1070}
1071
1072void blk_mq_end_request_batch(struct io_comp_batch *iob)
1073{
1074 int tags[TAG_COMP_BATCH], nr_tags = 0;
Jens Axboe02f7eab2021-10-28 12:08:34 -06001075 struct blk_mq_hw_ctx *cur_hctx = NULL;
Jens Axboef794f332021-10-08 05:50:46 -06001076 struct request *rq;
1077 u64 now = 0;
1078
1079 if (iob->need_ts)
Jens Axboe08420cf2024-01-15 14:45:07 -07001080 now = blk_time_get_ns();
Jens Axboef794f332021-10-08 05:50:46 -06001081
1082 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1083 prefetch(rq->bio);
1084 prefetch(rq->rq_next);
1085
Jens Axboe5581a5d2021-12-01 15:01:51 -07001086 blk_complete_request(rq);
Jens Axboef794f332021-10-08 05:50:46 -06001087 if (iob->need_ts)
1088 __blk_mq_end_request_acct(rq, now);
1089
Chengming Zhoue5c0ca12023-08-13 23:23:25 +08001090 blk_mq_finish_request(rq);
1091
Jens Axboe98b26a02021-11-26 09:53:23 -07001092 rq_qos_done(rq->q, rq);
1093
Jens Axboeab3e1d32022-09-21 08:24:16 -06001094 /*
1095 * If end_io handler returns NONE, then it still has
1096 * ownership of the request.
1097 */
1098 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1099 continue;
1100
Jens Axboef794f332021-10-08 05:50:46 -06001101 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Jens Axboe0a467d02021-10-14 14:39:59 -06001102 if (!req_ref_put_and_test(rq))
Jens Axboef794f332021-10-08 05:50:46 -06001103 continue;
1104
1105 blk_crypto_free_request(rq);
1106 blk_pm_mark_last_busy(rq);
Jens Axboef794f332021-10-08 05:50:46 -06001107
Jens Axboe02f7eab2021-10-28 12:08:34 -06001108 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1109 if (cur_hctx)
1110 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
Jens Axboef794f332021-10-08 05:50:46 -06001111 nr_tags = 0;
Jens Axboe02f7eab2021-10-28 12:08:34 -06001112 cur_hctx = rq->mq_hctx;
Jens Axboef794f332021-10-08 05:50:46 -06001113 }
1114 tags[nr_tags++] = rq->tag;
Jens Axboef794f332021-10-08 05:50:46 -06001115 }
1116
1117 if (nr_tags)
Jens Axboe02f7eab2021-10-28 12:08:34 -06001118 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
Jens Axboef794f332021-10-08 05:50:46 -06001119}
1120EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1121
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001122static void blk_complete_reqs(struct llist_head *list)
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001123{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001124 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1125 struct request *rq, *next;
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001126
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001127 llist_for_each_entry_safe(rq, next, entry, ipi_list)
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001128 rq->q->mq_ops->complete(rq);
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001129}
1130
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001131static __latent_entropy void blk_done_softirq(struct softirq_action *h)
Christoph Hellwig115243f2020-06-11 08:44:42 +02001132{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001133 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001134}
1135
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001136static int blk_softirq_cpu_dead(unsigned int cpu)
1137{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001138 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001139 return 0;
1140}
1141
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08001142static void __blk_mq_complete_request_remote(void *data)
Jens Axboe320ae512013-10-24 09:20:05 +01001143{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001144 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
Jens Axboe320ae512013-10-24 09:20:05 +01001145}
1146
Christoph Hellwig963395262020-06-11 08:44:49 +02001147static inline bool blk_mq_complete_need_ipi(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001148{
Christoph Hellwig963395262020-06-11 08:44:49 +02001149 int cpu = raw_smp_processor_id();
Jens Axboe320ae512013-10-24 09:20:05 +01001150
Christoph Hellwig963395262020-06-11 08:44:49 +02001151 if (!IS_ENABLED(CONFIG_SMP) ||
1152 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1153 return false;
Sebastian Andrzej Siewior71425182020-12-04 20:13:54 +01001154 /*
1155 * With force threaded interrupts enabled, raising softirq from an SMP
1156 * function call will always result in waking the ksoftirqd thread.
1157 * This is probably worse than completing the request on a different
1158 * cache domain.
1159 */
Tanner Love91cc4702021-06-02 14:03:38 -04001160 if (force_irqthreads())
Sebastian Andrzej Siewior71425182020-12-04 20:13:54 +01001161 return false;
Christoph Hellwig963395262020-06-11 08:44:49 +02001162
Qais Yousefaf550e42024-02-23 15:57:49 +00001163 /* same CPU or cache domain and capacity? Complete locally */
Christoph Hellwig963395262020-06-11 08:44:49 +02001164 if (cpu == rq->mq_ctx->cpu ||
1165 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
Qais Yousefaf550e42024-02-23 15:57:49 +00001166 cpus_share_cache(cpu, rq->mq_ctx->cpu) &&
1167 cpus_equal_capacity(cpu, rq->mq_ctx->cpu)))
Christoph Hellwig963395262020-06-11 08:44:49 +02001168 return false;
1169
1170 /* don't try to IPI to an offline CPU */
1171 return cpu_online(rq->mq_ctx->cpu);
1172}
1173
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001174static void blk_mq_complete_send_ipi(struct request *rq)
1175{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001176 unsigned int cpu;
1177
1178 cpu = rq->mq_ctx->cpu;
Chengming Zhou660e8022023-07-17 12:00:55 +08001179 if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))
1180 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001181}
1182
1183static void blk_mq_raise_softirq(struct request *rq)
1184{
1185 struct llist_head *list;
1186
1187 preempt_disable();
1188 list = this_cpu_ptr(&blk_cpu_done);
1189 if (llist_add(&rq->ipi_list, list))
1190 raise_softirq(BLOCK_SOFTIRQ);
1191 preempt_enable();
1192}
1193
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001194bool blk_mq_complete_request_remote(struct request *rq)
1195{
Keith Buschaf78ff72018-11-26 09:54:30 -07001196 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
Ming Lei36e76532018-09-28 16:42:20 +08001197
Jens Axboe4ab32bf2018-11-18 16:15:35 -07001198 /*
Liu Songf1684202022-09-21 11:32:03 +08001199 * For request which hctx has only one ctx mapping,
1200 * or a polled request, always complete locally,
1201 * it's pointless to redirect the completion.
Jens Axboe4ab32bf2018-11-18 16:15:35 -07001202 */
Ed Tsai30654612023-06-14 08:25:29 +08001203 if ((rq->mq_hctx->nr_ctx == 1 &&
1204 rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1205 rq->cmd_flags & REQ_POLLED)
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001206 return false;
Jens Axboe320ae512013-10-24 09:20:05 +01001207
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001208 if (blk_mq_complete_need_ipi(rq)) {
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001209 blk_mq_complete_send_ipi(rq);
1210 return true;
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -08001211 }
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001212
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001213 if (rq->q->nr_hw_queues == 1) {
1214 blk_mq_raise_softirq(rq);
1215 return true;
1216 }
1217 return false;
Jens Axboe320ae512013-10-24 09:20:05 +01001218}
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001219EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1220
Jens Axboe320ae512013-10-24 09:20:05 +01001221/**
Christoph Hellwig15f73f52020-06-11 08:44:47 +02001222 * blk_mq_complete_request - end I/O on a request
1223 * @rq: the request being processed
Jens Axboe320ae512013-10-24 09:20:05 +01001224 *
Christoph Hellwig15f73f52020-06-11 08:44:47 +02001225 * Description:
1226 * Complete a request by scheduling the ->complete_rq operation.
1227 **/
1228void blk_mq_complete_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001229{
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001230 if (!blk_mq_complete_request_remote(rq))
Christoph Hellwig963395262020-06-11 08:44:49 +02001231 rq->q->mq_ops->complete(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01001232}
Christoph Hellwig15f73f52020-06-11 08:44:47 +02001233EXPORT_SYMBOL(blk_mq_complete_request);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08001234
1235/**
André Almeida105663f2020-01-06 15:08:18 -03001236 * blk_mq_start_request - Start processing a request
1237 * @rq: Pointer to request to be started
1238 *
1239 * Function used by device drivers to notify the block layer that a request
1240 * is going to be processed now, so blk layer can do proper initializations
1241 * such as starting the timeout timer.
1242 */
Christoph Hellwige2490072014-09-13 16:40:09 -07001243void blk_mq_start_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001244{
1245 struct request_queue *q = rq->q;
1246
Christoph Hellwiga54895f2020-12-03 17:21:39 +01001247 trace_block_rq_issue(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01001248
Kundan Kumar847c5bc2023-11-24 00:33:31 +05301249 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) &&
1250 !blk_rq_is_passthrough(rq)) {
Jens Axboe08420cf2024-01-15 14:45:07 -07001251 rq->io_start_time_ns = blk_time_get_ns();
Hou Tao3d244302019-05-21 15:59:03 +08001252 rq->stats_sectors = blk_rq_sectors(rq);
Jens Axboecf43e6b2016-11-07 21:32:37 -07001253 rq->rq_flags |= RQF_STATS;
Josef Bacika7905042018-07-03 09:32:35 -06001254 rq_qos_issue(q, rq);
Jens Axboecf43e6b2016-11-07 21:32:37 -07001255 }
1256
Tejun Heo1d9bd512018-01-09 08:29:48 -08001257 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
Jens Axboe538b7532014-09-16 10:37:37 -06001258
Tejun Heo1d9bd512018-01-09 08:29:48 -08001259 blk_add_timer(rq);
Keith Busch12f5b932018-05-29 15:52:28 +02001260 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
Chengming Zhou217b6132023-09-13 15:16:15 +00001261 rq->mq_hctx->tags->rqs[rq->tag] = rq;
Christoph Hellwig49f5baa2014-02-11 08:27:14 -08001262
Max Gurtovoy54d4e6a2019-09-16 18:44:29 +03001263 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
Christoph Hellwige9f5f442024-06-13 10:48:15 +02001264 blk_integrity_prepare(rq);
1265
Christoph Hellwig3e087732021-10-12 13:12:24 +02001266 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
Keith Buschf6c80cf2023-06-12 12:03:42 -07001267 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
Jens Axboe320ae512013-10-24 09:20:05 +01001268}
Christoph Hellwige2490072014-09-13 16:40:09 -07001269EXPORT_SYMBOL(blk_mq_start_request);
Jens Axboe320ae512013-10-24 09:20:05 +01001270
Ming Leia327c342022-05-12 22:00:10 +08001271/*
1272 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1273 * queues. This is important for md arrays to benefit from merging
1274 * requests.
1275 */
1276static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1277{
1278 if (plug->multiple_queues)
1279 return BLK_MAX_REQUEST_COUNT * 2;
1280 return BLK_MAX_REQUEST_COUNT;
1281}
1282
1283static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1284{
1285 struct request *last = rq_list_peek(&plug->mq_list);
1286
1287 if (!plug->rq_count) {
1288 trace_block_plug(rq->q);
1289 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1290 (!blk_queue_nomerges(rq->q) &&
1291 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1292 blk_mq_flush_plug_list(plug, false);
Al Viro878eb6e2022-11-01 00:54:13 +00001293 last = NULL;
Ming Leia327c342022-05-12 22:00:10 +08001294 trace_block_plug(rq->q);
1295 }
1296
1297 if (!plug->multiple_queues && last && last->q != rq->q)
1298 plug->multiple_queues = true;
Ming Leic6b7a3a2023-06-24 21:01:05 +08001299 /*
1300 * Any request allocated from sched tags can't be issued to
1301 * ->queue_rqs() directly
1302 */
1303 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
Ming Leia327c342022-05-12 22:00:10 +08001304 plug->has_elevator = true;
1305 rq->rq_next = NULL;
1306 rq_list_add(&plug->mq_list, rq);
1307 plug->rq_count++;
1308}
1309
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001310/**
1311 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001312 * @rq: request to insert
1313 * @at_head: insert request at head or tail of queue
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001314 *
1315 * Description:
1316 * Insert a fully prepared request at the back of the I/O scheduler queue
1317 * for execution. Don't wait for completion.
1318 *
1319 * Note:
1320 * This function will invoke @done directly if the queue is dead.
1321 */
Christoph Hellwige2e530862022-05-24 14:15:30 +02001322void blk_execute_rq_nowait(struct request *rq, bool at_head)
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001323{
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001324 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1325
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001326 WARN_ON(irqs_disabled());
1327 WARN_ON(!blk_rq_is_passthrough(rq));
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001328
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001329 blk_account_io_start(rq);
Pankaj Raghav110fdb42022-09-29 16:41:41 +02001330
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001331 if (current->plug && !at_head) {
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001332 blk_add_rq_to_plug(current->plug, rq);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001333 return;
1334 }
1335
Christoph Hellwig710fa372023-04-13 08:40:54 +02001336 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
Bart Van Assche65a558f2023-07-21 10:27:30 -07001337 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001338}
1339EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1340
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001341struct blk_rq_wait {
1342 struct completion done;
1343 blk_status_t ret;
1344};
1345
Jens Axboede671d62022-09-21 15:19:54 -06001346static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001347{
1348 struct blk_rq_wait *wait = rq->end_io_data;
1349
1350 wait->ret = ret;
1351 complete(&wait->done);
Jens Axboede671d62022-09-21 15:19:54 -06001352 return RQ_END_IO_NONE;
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001353}
1354
Kanchan Joshic6e99ea2022-08-23 21:44:42 +05301355bool blk_rq_is_poll(struct request *rq)
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001356{
1357 if (!rq->mq_hctx)
1358 return false;
1359 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1360 return false;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001361 return true;
1362}
Kanchan Joshic6e99ea2022-08-23 21:44:42 +05301363EXPORT_SYMBOL_GPL(blk_rq_is_poll);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001364
1365static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1366{
1367 do {
Keith Buschf6c80cf2023-06-12 12:03:42 -07001368 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001369 cond_resched();
1370 } while (!completion_done(wait));
1371}
1372
1373/**
1374 * blk_execute_rq - insert a request into queue for execution
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001375 * @rq: request to insert
1376 * @at_head: insert request at head or tail of queue
1377 *
1378 * Description:
1379 * Insert a fully prepared request at the back of the I/O scheduler queue
1380 * for execution and wait for completion.
1381 * Return: The blk_status_t result provided to blk_mq_end_request().
1382 */
Christoph Hellwigb84ba302021-11-26 13:18:01 +01001383blk_status_t blk_execute_rq(struct request *rq, bool at_head)
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001384{
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001385 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001386 struct blk_rq_wait wait = {
1387 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1388 };
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001389
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001390 WARN_ON(irqs_disabled());
1391 WARN_ON(!blk_rq_is_passthrough(rq));
1392
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001393 rq->end_io_data = &wait;
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001394 rq->end_io = blk_end_sync_rq;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001395
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001396 blk_account_io_start(rq);
Christoph Hellwig710fa372023-04-13 08:40:54 +02001397 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02001398 blk_mq_run_hw_queue(hctx, false);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001399
Keith Busch0eb4db42024-02-23 07:59:09 -08001400 if (blk_rq_is_poll(rq))
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001401 blk_rq_poll_completion(rq, &wait.done);
Keith Busch0eb4db42024-02-23 07:59:09 -08001402 else
1403 blk_wait_io(&wait.done);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001404
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001405 return wait.ret;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001406}
1407EXPORT_SYMBOL(blk_execute_rq);
1408
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001409static void __blk_mq_requeue_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001410{
1411 struct request_queue *q = rq->q;
1412
Ming Lei923218f2017-11-02 23:24:38 +08001413 blk_mq_put_driver_tag(rq);
1414
Christoph Hellwiga54895f2020-12-03 17:21:39 +01001415 trace_block_rq_requeue(rq);
Josef Bacika7905042018-07-03 09:32:35 -06001416 rq_qos_requeue(q, rq);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -08001417
Keith Busch12f5b932018-05-29 15:52:28 +02001418 if (blk_mq_request_started(rq)) {
1419 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Christoph Hellwigda661262018-06-14 13:58:45 +02001420 rq->rq_flags &= ~RQF_TIMED_OUT;
Christoph Hellwige2490072014-09-13 16:40:09 -07001421 }
Jens Axboe320ae512013-10-24 09:20:05 +01001422}
1423
Bart Van Assche2b053ac2016-10-28 17:21:41 -07001424void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001425{
Christoph Hellwig214a4412023-04-13 08:40:53 +02001426 struct request_queue *q = rq->q;
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001427 unsigned long flags;
Christoph Hellwig214a4412023-04-13 08:40:53 +02001428
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001429 __blk_mq_requeue_request(rq);
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001430
Ming Lei105976f2018-02-23 23:36:56 +08001431 /* this request will be re-inserted to io scheduler queue */
1432 blk_mq_sched_requeue_request(rq);
1433
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001434 spin_lock_irqsave(&q->requeue_lock, flags);
1435 list_add_tail(&rq->queuelist, &q->requeue_list);
1436 spin_unlock_irqrestore(&q->requeue_lock, flags);
Christoph Hellwig214a4412023-04-13 08:40:53 +02001437
1438 if (kick_requeue_list)
1439 blk_mq_kick_requeue_list(q);
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001440}
1441EXPORT_SYMBOL(blk_mq_requeue_request);
1442
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001443static void blk_mq_requeue_work(struct work_struct *work)
1444{
1445 struct request_queue *q =
Mike Snitzer28494502016-09-14 13:28:30 -04001446 container_of(work, struct request_queue, requeue_work.work);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001447 LIST_HEAD(rq_list);
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001448 LIST_HEAD(flush_list);
1449 struct request *rq;
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001450
Jens Axboe18e97812017-07-27 08:03:57 -06001451 spin_lock_irq(&q->requeue_lock);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001452 list_splice_init(&q->requeue_list, &rq_list);
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001453 list_splice_init(&q->flush_list, &flush_list);
Jens Axboe18e97812017-07-27 08:03:57 -06001454 spin_unlock_irq(&q->requeue_lock);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001455
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001456 while (!list_empty(&rq_list)) {
1457 rq = list_entry(rq_list.next, struct request, queuelist);
Jianchao Wangaef18972019-02-12 09:56:25 +08001458 /*
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001459 * If RQF_DONTPREP ist set, the request has been started by the
1460 * driver already and might have driver-specific data allocated
1461 * already. Insert it into the hctx dispatch list to avoid
1462 * block layer merges for the request.
Jianchao Wangaef18972019-02-12 09:56:25 +08001463 */
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001464 if (rq->rq_flags & RQF_DONTPREP) {
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001465 list_del_init(&rq->queuelist);
Christoph Hellwig2b597612023-04-13 08:40:55 +02001466 blk_mq_request_bypass_insert(rq, 0);
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001467 } else {
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001468 list_del_init(&rq->queuelist);
Christoph Hellwig710fa372023-04-13 08:40:54 +02001469 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
Christoph Hellwiga1e948b2023-04-13 08:40:48 +02001470 }
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001471 }
1472
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02001473 while (!list_empty(&flush_list)) {
1474 rq = list_entry(flush_list.next, struct request, queuelist);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001475 list_del_init(&rq->queuelist);
Christoph Hellwig710fa372023-04-13 08:40:54 +02001476 blk_mq_insert_request(rq, 0);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001477 }
1478
Bart Van Assche52d7f1b2016-10-28 17:20:32 -07001479 blk_mq_run_hw_queues(q, false);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001480}
1481
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001482void blk_mq_kick_requeue_list(struct request_queue *q)
1483{
Bart Van Asscheae943d22018-01-19 08:58:55 -08001484 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001485}
1486EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1487
Mike Snitzer28494502016-09-14 13:28:30 -04001488void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1489 unsigned long msecs)
1490{
Bart Van Assched4acf362017-08-09 11:28:06 -07001491 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1492 msecs_to_jiffies(msecs));
Mike Snitzer28494502016-09-14 13:28:30 -04001493}
1494EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1495
Ming Lei0e4237a2023-12-01 16:56:05 +08001496static bool blk_is_flush_data_rq(struct request *rq)
1497{
1498 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);
1499}
1500
John Garry2dd65322022-07-06 20:03:53 +08001501static bool blk_mq_rq_inflight(struct request *rq, void *priv)
Jens Axboeae879912018-11-08 09:03:51 -07001502{
1503 /*
John Garry8ab30a32021-12-06 20:49:48 +08001504 * If we find a request that isn't idle we know the queue is busy
1505 * as it's checked in the iter.
1506 * Return false to stop the iteration.
Ming Lei0e4237a2023-12-01 16:56:05 +08001507 *
1508 * In case of queue quiesce, if one flush data request is completed,
1509 * don't count it as inflight given the flush sequence is suspended,
1510 * and the original flush data request is invisible to driver, just
1511 * like other pending requests because of quiesce
Jens Axboeae879912018-11-08 09:03:51 -07001512 */
Ming Lei0e4237a2023-12-01 16:56:05 +08001513 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
1514 blk_is_flush_data_rq(rq) &&
1515 blk_mq_request_completed(rq))) {
Jens Axboeae879912018-11-08 09:03:51 -07001516 bool *busy = priv;
1517
1518 *busy = true;
1519 return false;
1520 }
1521
1522 return true;
1523}
1524
Jens Axboe3c94d832018-12-17 21:11:17 -07001525bool blk_mq_queue_inflight(struct request_queue *q)
Jens Axboeae879912018-11-08 09:03:51 -07001526{
1527 bool busy = false;
1528
Jens Axboe3c94d832018-12-17 21:11:17 -07001529 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
Jens Axboeae879912018-11-08 09:03:51 -07001530 return busy;
1531}
Jens Axboe3c94d832018-12-17 21:11:17 -07001532EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
Jens Axboeae879912018-11-08 09:03:51 -07001533
John Garry9bdb4832022-07-06 20:03:51 +08001534static void blk_mq_rq_timed_out(struct request *req)
Jens Axboe320ae512013-10-24 09:20:05 +01001535{
Christoph Hellwigda661262018-06-14 13:58:45 +02001536 req->rq_flags |= RQF_TIMED_OUT;
Christoph Hellwigd1210d52018-05-29 15:52:39 +02001537 if (req->q->mq_ops->timeout) {
1538 enum blk_eh_timer_return ret;
Jens Axboe87ee7b12014-04-24 08:51:47 -06001539
John Garry9bdb4832022-07-06 20:03:51 +08001540 ret = req->q->mq_ops->timeout(req);
Christoph Hellwigd1210d52018-05-29 15:52:39 +02001541 if (ret == BLK_EH_DONE)
1542 return;
1543 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
Christoph Hellwig46f92d42014-09-13 16:40:12 -07001544 }
Christoph Hellwigd1210d52018-05-29 15:52:39 +02001545
1546 blk_add_timer(req);
Jens Axboe87ee7b12014-04-24 08:51:47 -06001547}
Keith Busch5b3f25f2015-01-07 18:55:46 -07001548
David Jeffery82c22942022-10-26 13:19:57 +08001549struct blk_expired_data {
1550 bool has_timedout_rq;
1551 unsigned long next;
1552 unsigned long timeout_start;
1553};
1554
1555static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
Keith Busch12f5b932018-05-29 15:52:28 +02001556{
1557 unsigned long deadline;
1558
1559 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1560 return false;
Christoph Hellwigda661262018-06-14 13:58:45 +02001561 if (rq->rq_flags & RQF_TIMED_OUT)
1562 return false;
Keith Busch12f5b932018-05-29 15:52:28 +02001563
Christoph Hellwig079076b2018-11-14 17:02:05 +01001564 deadline = READ_ONCE(rq->deadline);
David Jeffery82c22942022-10-26 13:19:57 +08001565 if (time_after_eq(expired->timeout_start, deadline))
Keith Busch12f5b932018-05-29 15:52:28 +02001566 return true;
1567
David Jeffery82c22942022-10-26 13:19:57 +08001568 if (expired->next == 0)
1569 expired->next = deadline;
1570 else if (time_after(expired->next, deadline))
1571 expired->next = deadline;
Keith Busch12f5b932018-05-29 15:52:28 +02001572 return false;
1573}
1574
Ming Lei2e315dc2021-05-11 23:22:34 +08001575void blk_mq_put_rq_ref(struct request *rq)
1576{
Jens Axboede671d62022-09-21 15:19:54 -06001577 if (is_flush_rq(rq)) {
1578 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1579 blk_mq_free_request(rq);
1580 } else if (req_ref_put_and_test(rq)) {
Ming Lei2e315dc2021-05-11 23:22:34 +08001581 __blk_mq_free_request(rq);
Jens Axboede671d62022-09-21 15:19:54 -06001582 }
Ming Lei2e315dc2021-05-11 23:22:34 +08001583}
1584
John Garry2dd65322022-07-06 20:03:53 +08001585static bool blk_mq_check_expired(struct request *rq, void *priv)
Jens Axboe320ae512013-10-24 09:20:05 +01001586{
David Jeffery82c22942022-10-26 13:19:57 +08001587 struct blk_expired_data *expired = priv;
Christoph Hellwig81481eb2014-09-13 16:40:11 -07001588
Keith Busch12f5b932018-05-29 15:52:28 +02001589 /*
Ming Leic797b402021-08-11 23:52:02 +08001590 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1591 * be reallocated underneath the timeout handler's processing, then
1592 * the expire check is reliable. If the request is not expired, then
1593 * it was completed and reallocated as a new request after returning
1594 * from blk_mq_check_expired().
Keith Busch12f5b932018-05-29 15:52:28 +02001595 */
David Jeffery82c22942022-10-26 13:19:57 +08001596 if (blk_mq_req_expired(rq, expired)) {
1597 expired->has_timedout_rq = true;
1598 return false;
1599 }
1600 return true;
1601}
1602
1603static bool blk_mq_handle_expired(struct request *rq, void *priv)
1604{
1605 struct blk_expired_data *expired = priv;
1606
1607 if (blk_mq_req_expired(rq, expired))
John Garry9bdb4832022-07-06 20:03:51 +08001608 blk_mq_rq_timed_out(rq);
Jens Axboe7baa8572018-11-08 10:24:07 -07001609 return true;
Tejun Heo1d9bd512018-01-09 08:29:48 -08001610}
1611
Christoph Hellwig287922e2015-10-30 20:57:30 +08001612static void blk_mq_timeout_work(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +01001613{
Christoph Hellwig287922e2015-10-30 20:57:30 +08001614 struct request_queue *q =
1615 container_of(work, struct request_queue, timeout_work);
David Jeffery82c22942022-10-26 13:19:57 +08001616 struct blk_expired_data expired = {
1617 .timeout_start = jiffies,
1618 };
Tejun Heo1d9bd512018-01-09 08:29:48 -08001619 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08001620 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01001621
Gabriel Krisman Bertazi71f79fb2016-08-01 08:23:39 -06001622 /* A deadlock might occur if a request is stuck requiring a
1623 * timeout at the same time a queue freeze is waiting
1624 * completion, since the timeout code would not be able to
1625 * acquire the queue reference here.
1626 *
1627 * That's why we don't use blk_queue_enter here; instead, we use
1628 * percpu_ref_tryget directly, because we need to be able to
1629 * obtain a reference even in the short window between the queue
1630 * starting to freeze, by dropping the first reference in
Ming Lei1671d522017-03-27 20:06:57 +08001631 * blk_freeze_queue_start, and the moment the last request is
Gabriel Krisman Bertazi71f79fb2016-08-01 08:23:39 -06001632 * consumed, marked by the instant q_usage_counter reaches
1633 * zero.
1634 */
1635 if (!percpu_ref_tryget(&q->q_usage_counter))
Christoph Hellwig287922e2015-10-30 20:57:30 +08001636 return;
1637
David Jeffery82c22942022-10-26 13:19:57 +08001638 /* check if there is any timed-out request */
1639 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1640 if (expired.has_timedout_rq) {
1641 /*
1642 * Before walking tags, we must ensure any submit started
1643 * before the current time has finished. Since the submit
1644 * uses srcu or rcu, wait for a synchronization point to
1645 * ensure all running submits have finished
1646 */
Christoph Hellwig483239c2022-11-01 16:00:48 +01001647 blk_mq_wait_quiesce_done(q->tag_set);
Jens Axboe320ae512013-10-24 09:20:05 +01001648
David Jeffery82c22942022-10-26 13:19:57 +08001649 expired.next = 0;
1650 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1651 }
1652
1653 if (expired.next != 0) {
1654 mod_timer(&q->timeout, expired.next);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001655 } else {
Bart Van Asschefcd36c32018-01-10 08:33:33 -08001656 /*
1657 * Request timeouts are handled as a forward rolling timer. If
1658 * we end up here it means that no requests are pending and
1659 * also that no request has been pending for a while. Mark
1660 * each hctx as idle.
1661 */
Ming Leif054b562015-04-21 10:00:19 +08001662 queue_for_each_hw_ctx(q, hctx, i) {
1663 /* the hctx may be unmapped, so check it here */
1664 if (blk_mq_hw_queue_mapped(hctx))
1665 blk_mq_tag_idle(hctx);
1666 }
Jens Axboe0d2602c2014-05-13 15:10:52 -06001667 }
Christoph Hellwig287922e2015-10-30 20:57:30 +08001668 blk_queue_exit(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001669}
1670
Omar Sandoval88459642016-09-17 08:38:44 -06001671struct flush_busy_ctx_data {
1672 struct blk_mq_hw_ctx *hctx;
1673 struct list_head *list;
1674};
1675
1676static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1677{
1678 struct flush_busy_ctx_data *flush_data = data;
1679 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1680 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
Ming Leic16d6b52018-12-17 08:44:05 -07001681 enum hctx_type type = hctx->type;
Omar Sandoval88459642016-09-17 08:38:44 -06001682
Omar Sandoval88459642016-09-17 08:38:44 -06001683 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07001684 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
Omar Sandovale9a99a62018-02-27 16:56:42 -08001685 sbitmap_clear_bit(sb, bitnr);
Omar Sandoval88459642016-09-17 08:38:44 -06001686 spin_unlock(&ctx->lock);
1687 return true;
1688}
1689
Jens Axboe320ae512013-10-24 09:20:05 +01001690/*
Jens Axboe1429d7c2014-05-19 09:23:55 -06001691 * Process software queues that have been marked busy, splicing them
1692 * to the for-dispatch
1693 */
Jens Axboe2c3ad662016-12-14 14:34:47 -07001694void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
Jens Axboe1429d7c2014-05-19 09:23:55 -06001695{
Omar Sandoval88459642016-09-17 08:38:44 -06001696 struct flush_busy_ctx_data data = {
1697 .hctx = hctx,
1698 .list = list,
1699 };
Jens Axboe1429d7c2014-05-19 09:23:55 -06001700
Omar Sandoval88459642016-09-17 08:38:44 -06001701 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
Jens Axboe1429d7c2014-05-19 09:23:55 -06001702}
Jens Axboe2c3ad662016-12-14 14:34:47 -07001703EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
Jens Axboe1429d7c2014-05-19 09:23:55 -06001704
Ming Leib3476892017-10-14 17:22:30 +08001705struct dispatch_rq_data {
1706 struct blk_mq_hw_ctx *hctx;
1707 struct request *rq;
1708};
1709
1710static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1711 void *data)
1712{
1713 struct dispatch_rq_data *dispatch_data = data;
1714 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1715 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
Ming Leic16d6b52018-12-17 08:44:05 -07001716 enum hctx_type type = hctx->type;
Ming Leib3476892017-10-14 17:22:30 +08001717
1718 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07001719 if (!list_empty(&ctx->rq_lists[type])) {
1720 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
Ming Leib3476892017-10-14 17:22:30 +08001721 list_del_init(&dispatch_data->rq->queuelist);
Ming Leic16d6b52018-12-17 08:44:05 -07001722 if (list_empty(&ctx->rq_lists[type]))
Ming Leib3476892017-10-14 17:22:30 +08001723 sbitmap_clear_bit(sb, bitnr);
1724 }
1725 spin_unlock(&ctx->lock);
1726
1727 return !dispatch_data->rq;
1728}
1729
1730struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1731 struct blk_mq_ctx *start)
1732{
Jens Axboef31967f2018-10-29 13:13:29 -06001733 unsigned off = start ? start->index_hw[hctx->type] : 0;
Ming Leib3476892017-10-14 17:22:30 +08001734 struct dispatch_rq_data data = {
1735 .hctx = hctx,
1736 .rq = NULL,
1737 };
1738
1739 __sbitmap_for_each_set(&hctx->ctx_map, off,
1740 dispatch_rq_from_ctx, &data);
1741
1742 return data.rq;
1743}
1744
Chengming Zhoub8643d62023-09-13 15:16:12 +00001745bool __blk_mq_alloc_driver_tag(struct request *rq)
Jens Axboe703fd1c2016-09-16 13:59:14 -06001746{
John Garryae0f1a72021-10-05 18:23:38 +08001747 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
Ming Lei570e9b72020-06-30 22:03:55 +08001748 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
Ming Lei570e9b72020-06-30 22:03:55 +08001749 int tag;
1750
Ming Lei568f2702020-07-06 22:41:11 +08001751 blk_mq_tag_busy(rq->mq_hctx);
1752
Ming Lei570e9b72020-06-30 22:03:55 +08001753 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
John Garryae0f1a72021-10-05 18:23:38 +08001754 bt = &rq->mq_hctx->tags->breserved_tags;
Ming Lei570e9b72020-06-30 22:03:55 +08001755 tag_offset = 0;
Ming Lei28500852020-09-11 18:41:14 +08001756 } else {
1757 if (!hctx_may_queue(rq->mq_hctx, bt))
1758 return false;
Ming Lei570e9b72020-06-30 22:03:55 +08001759 }
1760
Ming Lei570e9b72020-06-30 22:03:55 +08001761 tag = __sbitmap_queue_get(bt);
1762 if (tag == BLK_MQ_NO_TAG)
1763 return false;
1764
1765 rq->tag = tag + tag_offset;
Chengming Zhoub8643d62023-09-13 15:16:12 +00001766 blk_mq_inc_active_requests(rq->mq_hctx);
Ming Lei568f2702020-07-06 22:41:11 +08001767 return true;
Ming Lei570e9b72020-06-30 22:03:55 +08001768}
1769
Jens Axboeeb619fd2017-11-09 08:32:43 -07001770static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1771 int flags, void *key)
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001772{
1773 struct blk_mq_hw_ctx *hctx;
1774
1775 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1776
Ming Lei5815839b2018-06-25 19:31:47 +08001777 spin_lock(&hctx->dispatch_wait_lock);
Jens Axboee8618572019-03-25 12:34:10 -06001778 if (!list_empty(&wait->entry)) {
1779 struct sbitmap_queue *sbq;
1780
1781 list_del_init(&wait->entry);
John Garryae0f1a72021-10-05 18:23:38 +08001782 sbq = &hctx->tags->bitmap_tags;
Jens Axboee8618572019-03-25 12:34:10 -06001783 atomic_dec(&sbq->ws_active);
1784 }
Ming Lei5815839b2018-06-25 19:31:47 +08001785 spin_unlock(&hctx->dispatch_wait_lock);
1786
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001787 blk_mq_run_hw_queue(hctx, true);
1788 return 1;
1789}
1790
Jens Axboef906a6a2017-11-09 16:10:13 -07001791/*
1792 * Mark us waiting for a tag. For shared tags, this involves hooking us into
Bart Van Asscheee3e4de2018-01-09 10:09:15 -08001793 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1794 * restart. For both cases, take care to check the condition again after
Jens Axboef906a6a2017-11-09 16:10:13 -07001795 * marking us as waiting.
1796 */
Ming Lei2278d692018-06-25 19:31:46 +08001797static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
Jens Axboef906a6a2017-11-09 16:10:13 -07001798 struct request *rq)
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001799{
Kemeng Shi98b99e92023-01-18 17:37:15 +08001800 struct sbitmap_queue *sbq;
Ming Lei5815839b2018-06-25 19:31:47 +08001801 struct wait_queue_head *wq;
Jens Axboef906a6a2017-11-09 16:10:13 -07001802 wait_queue_entry_t *wait;
1803 bool ret;
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001804
Kemeng Shi47df9ce2023-01-18 17:37:16 +08001805 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1806 !(blk_mq_is_shared_tags(hctx->flags))) {
Yufen Yu684b7322019-03-15 11:05:10 +08001807 blk_mq_sched_mark_restart_hctx(hctx);
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001808
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001809 /*
1810 * It's possible that a tag was freed in the window between the
1811 * allocation failure and adding the hardware queue to the wait
1812 * queue.
1813 *
1814 * Don't clear RESTART here, someone else could have set it.
1815 * At most this will cost an extra queue run.
1816 */
Ming Lei8ab6bb9e2018-06-25 19:31:45 +08001817 return blk_mq_get_driver_tag(rq);
Jens Axboeeb619fd2017-11-09 08:32:43 -07001818 }
1819
Ming Lei2278d692018-06-25 19:31:46 +08001820 wait = &hctx->dispatch_wait;
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001821 if (!list_empty_careful(&wait->entry))
1822 return false;
1823
Kemeng Shi98b99e92023-01-18 17:37:15 +08001824 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1825 sbq = &hctx->tags->breserved_tags;
1826 else
1827 sbq = &hctx->tags->bitmap_tags;
Jens Axboee8618572019-03-25 12:34:10 -06001828 wq = &bt_wait_ptr(sbq, hctx)->wait;
Ming Lei5815839b2018-06-25 19:31:47 +08001829
1830 spin_lock_irq(&wq->lock);
1831 spin_lock(&hctx->dispatch_wait_lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001832 if (!list_empty(&wait->entry)) {
Ming Lei5815839b2018-06-25 19:31:47 +08001833 spin_unlock(&hctx->dispatch_wait_lock);
1834 spin_unlock_irq(&wq->lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001835 return false;
1836 }
1837
Jens Axboee8618572019-03-25 12:34:10 -06001838 atomic_inc(&sbq->ws_active);
Ming Lei5815839b2018-06-25 19:31:47 +08001839 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1840 __add_wait_queue(wq, wait);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001841
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001842 /*
Ming Lei5266caa2024-01-12 20:26:26 +08001843 * Add one explicit barrier since blk_mq_get_driver_tag() may
1844 * not imply barrier in case of failure.
1845 *
1846 * Order adding us to wait queue and allocating driver tag.
1847 *
1848 * The pair is the one implied in sbitmap_queue_wake_up() which
1849 * orders clearing sbitmap tag bits and waitqueue_active() in
1850 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1851 *
1852 * Otherwise, re-order of adding wait queue and getting driver tag
1853 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1854 * the waitqueue_active() may not observe us in wait queue.
1855 */
1856 smp_mb();
1857
1858 /*
Jens Axboeeb619fd2017-11-09 08:32:43 -07001859 * It's possible that a tag was freed in the window between the
1860 * allocation failure and adding the hardware queue to the wait
1861 * queue.
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001862 */
Ming Lei8ab6bb9e2018-06-25 19:31:45 +08001863 ret = blk_mq_get_driver_tag(rq);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001864 if (!ret) {
Ming Lei5815839b2018-06-25 19:31:47 +08001865 spin_unlock(&hctx->dispatch_wait_lock);
1866 spin_unlock_irq(&wq->lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001867 return false;
Jens Axboef906a6a2017-11-09 16:10:13 -07001868 }
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001869
1870 /*
1871 * We got a tag, remove ourselves from the wait queue to ensure
1872 * someone else gets the wakeup.
1873 */
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001874 list_del_init(&wait->entry);
Jens Axboee8618572019-03-25 12:34:10 -06001875 atomic_dec(&sbq->ws_active);
Ming Lei5815839b2018-06-25 19:31:47 +08001876 spin_unlock(&hctx->dispatch_wait_lock);
1877 spin_unlock_irq(&wq->lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001878
1879 return true;
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001880}
1881
Ming Lei6e7687172018-07-03 09:03:16 -06001882#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1883#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1884/*
1885 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1886 * - EWMA is one simple way to compute running average value
1887 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1888 * - take 4 as factor for avoiding to get too small(0) result, and this
1889 * factor doesn't matter because EWMA decreases exponentially
1890 */
1891static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1892{
1893 unsigned int ewma;
1894
Ming Lei6e7687172018-07-03 09:03:16 -06001895 ewma = hctx->dispatch_busy;
1896
1897 if (!ewma && !busy)
1898 return;
1899
1900 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1901 if (busy)
1902 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1903 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1904
1905 hctx->dispatch_busy = ewma;
1906}
1907
Ming Lei86ff7c22018-01-30 22:04:57 -05001908#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1909
Johannes Thumshirnc92a4102020-03-25 00:24:44 +09001910static void blk_mq_handle_dev_resource(struct request *rq,
1911 struct list_head *list)
1912{
Johannes Thumshirnc92a4102020-03-25 00:24:44 +09001913 list_add(&rq->queuelist, list);
1914 __blk_mq_requeue_request(rq);
1915}
1916
Ming Lei75383522020-06-30 18:24:58 +08001917enum prep_dispatch {
1918 PREP_DISPATCH_OK,
1919 PREP_DISPATCH_NO_TAG,
1920 PREP_DISPATCH_NO_BUDGET,
1921};
1922
1923static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1924 bool need_budget)
1925{
1926 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Ming Lei2a5a24a2021-01-22 10:33:12 +08001927 int budget_token = -1;
Ming Lei75383522020-06-30 18:24:58 +08001928
Ming Lei2a5a24a2021-01-22 10:33:12 +08001929 if (need_budget) {
1930 budget_token = blk_mq_get_dispatch_budget(rq->q);
1931 if (budget_token < 0) {
1932 blk_mq_put_driver_tag(rq);
1933 return PREP_DISPATCH_NO_BUDGET;
1934 }
1935 blk_mq_set_rq_budget_token(rq, budget_token);
Ming Lei75383522020-06-30 18:24:58 +08001936 }
1937
1938 if (!blk_mq_get_driver_tag(rq)) {
1939 /*
1940 * The initial allocation attempt failed, so we need to
1941 * rerun the hardware queue when a tag is freed. The
1942 * waitqueue takes care of that. If the queue is run
1943 * before we add this entry back on the dispatch list,
1944 * we'll re-run it below.
1945 */
1946 if (!blk_mq_mark_tag_wait(hctx, rq)) {
Ming Lei1fd40b52020-06-30 18:25:00 +08001947 /*
1948 * All budgets not got from this function will be put
1949 * together during handling partial dispatch
1950 */
1951 if (need_budget)
Ming Lei2a5a24a2021-01-22 10:33:12 +08001952 blk_mq_put_dispatch_budget(rq->q, budget_token);
Ming Lei75383522020-06-30 18:24:58 +08001953 return PREP_DISPATCH_NO_TAG;
1954 }
1955 }
1956
1957 return PREP_DISPATCH_OK;
1958}
1959
Ming Lei1fd40b52020-06-30 18:25:00 +08001960/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1961static void blk_mq_release_budgets(struct request_queue *q,
Ming Lei2a5a24a2021-01-22 10:33:12 +08001962 struct list_head *list)
Ming Lei1fd40b52020-06-30 18:25:00 +08001963{
Ming Lei2a5a24a2021-01-22 10:33:12 +08001964 struct request *rq;
Ming Lei1fd40b52020-06-30 18:25:00 +08001965
Ming Lei2a5a24a2021-01-22 10:33:12 +08001966 list_for_each_entry(rq, list, queuelist) {
1967 int budget_token = blk_mq_get_rq_budget_token(rq);
1968
1969 if (budget_token >= 0)
1970 blk_mq_put_dispatch_budget(q, budget_token);
1971 }
Ming Lei1fd40b52020-06-30 18:25:00 +08001972}
1973
Jens Axboe1f57f8d2018-06-28 11:54:01 -06001974/*
Kemeng Shi34c9f5472023-01-18 17:37:19 +08001975 * blk_mq_commit_rqs will notify driver using bd->last that there is no
1976 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1977 * details)
1978 * Attention, we should explicitly call this in unusual cases:
1979 * 1) did not queue everything initially scheduled to queue
1980 * 2) the last attempt to queue a request failed
1981 */
1982static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
1983 bool from_schedule)
1984{
1985 if (hctx->queue->mq_ops->commit_rqs && queued) {
1986 trace_block_unplug(hctx->queue, queued, !from_schedule);
1987 hctx->queue->mq_ops->commit_rqs(hctx);
1988 }
1989}
1990
1991/*
Jens Axboe1f57f8d2018-06-28 11:54:01 -06001992 * Returns true if we did some work AND can potentially do more.
1993 */
Ming Lei445874e2020-06-30 18:24:57 +08001994bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
Ming Lei1fd40b52020-06-30 18:25:00 +08001995 unsigned int nr_budgets)
Jens Axboef04c3df2016-12-07 08:41:17 -07001996{
Ming Lei75383522020-06-30 18:24:58 +08001997 enum prep_dispatch prep;
Ming Lei445874e2020-06-30 18:24:57 +08001998 struct request_queue *q = hctx->queue;
Kemeng Shif1ce99f2023-01-18 17:37:24 +08001999 struct request *rq;
Kemeng Shi4ea58fe2023-01-18 17:37:23 +08002000 int queued;
Ming Lei86ff7c22018-01-30 22:04:57 -05002001 blk_status_t ret = BLK_STS_OK;
Naohiro Aota9586e672021-10-27 01:51:27 +09002002 bool needs_resource = false;
Jens Axboef04c3df2016-12-07 08:41:17 -07002003
Omar Sandoval81380ca2017-04-07 08:56:26 -06002004 if (list_empty(list))
2005 return false;
2006
Jens Axboef04c3df2016-12-07 08:41:17 -07002007 /*
Jens Axboef04c3df2016-12-07 08:41:17 -07002008 * Now process all the entries, sending them to the driver.
2009 */
Kemeng Shi4ea58fe2023-01-18 17:37:23 +08002010 queued = 0;
Omar Sandoval81380ca2017-04-07 08:56:26 -06002011 do {
Jens Axboef04c3df2016-12-07 08:41:17 -07002012 struct blk_mq_queue_data bd;
2013
2014 rq = list_first_entry(list, struct request, queuelist);
Ming Lei0bca7992018-04-05 00:35:21 +08002015
Ming Lei445874e2020-06-30 18:24:57 +08002016 WARN_ON_ONCE(hctx != rq->mq_hctx);
Ming Lei1fd40b52020-06-30 18:25:00 +08002017 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
Ming Lei75383522020-06-30 18:24:58 +08002018 if (prep != PREP_DISPATCH_OK)
Ming Lei0bca7992018-04-05 00:35:21 +08002019 break;
Ming Leide148292017-10-14 17:22:29 +08002020
Jens Axboef04c3df2016-12-07 08:41:17 -07002021 list_del_init(&rq->queuelist);
2022
2023 bd.rq = rq;
Kemeng Shif1ce99f2023-01-18 17:37:24 +08002024 bd.last = list_empty(list);
Jens Axboef04c3df2016-12-07 08:41:17 -07002025
Ming Lei1fd40b52020-06-30 18:25:00 +08002026 /*
2027 * once the request is queued to lld, no need to cover the
2028 * budget any more
2029 */
2030 if (nr_budgets)
2031 nr_budgets--;
Jens Axboef04c3df2016-12-07 08:41:17 -07002032 ret = q->mq_ops->queue_rq(hctx, &bd);
Ming Lei7bf13722020-07-01 21:58:57 +08002033 switch (ret) {
2034 case BLK_STS_OK:
2035 queued++;
Jens Axboef04c3df2016-12-07 08:41:17 -07002036 break;
Ming Lei7bf13722020-07-01 21:58:57 +08002037 case BLK_STS_RESOURCE:
Naohiro Aota9586e672021-10-27 01:51:27 +09002038 needs_resource = true;
2039 fallthrough;
Ming Lei7bf13722020-07-01 21:58:57 +08002040 case BLK_STS_DEV_RESOURCE:
2041 blk_mq_handle_dev_resource(rq, list);
2042 goto out;
Ming Lei7bf13722020-07-01 21:58:57 +08002043 default:
Hannes Reineckee21ee5a2020-09-30 10:02:53 +02002044 blk_mq_end_request(rq, ret);
Christoph Hellwigfc17b652017-06-03 09:38:05 +02002045 }
Omar Sandoval81380ca2017-04-07 08:56:26 -06002046 } while (!list_empty(list));
Ming Lei7bf13722020-07-01 21:58:57 +08002047out:
yangerkun632bfb62020-09-05 19:25:56 +08002048 /* If we didn't flush the entire list, we could have told the driver
2049 * there was more coming, but that turned out to be a lie.
2050 */
Kemeng Shie4ef2e02023-01-18 17:37:22 +08002051 if (!list_empty(list) || ret != BLK_STS_OK)
2052 blk_mq_commit_rqs(hctx, queued, false);
2053
Jens Axboef04c3df2016-12-07 08:41:17 -07002054 /*
2055 * Any items that need requeuing? Stuff them into hctx->dispatch,
2056 * that is where we will continue on next queue run.
2057 */
2058 if (!list_empty(list)) {
Ming Lei86ff7c22018-01-30 22:04:57 -05002059 bool needs_restart;
Ming Lei75383522020-06-30 18:24:58 +08002060 /* For non-shared tags, the RESTART check will suffice */
2061 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
Kemeng Shi47df9ce2023-01-18 17:37:16 +08002062 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2063 blk_mq_is_shared_tags(hctx->flags));
Ming Lei86ff7c22018-01-30 22:04:57 -05002064
Ming Lei2a5a24a2021-01-22 10:33:12 +08002065 if (nr_budgets)
2066 blk_mq_release_budgets(q, list);
Jens Axboef04c3df2016-12-07 08:41:17 -07002067
2068 spin_lock(&hctx->lock);
Ming Lei01e99ae2020-02-25 09:04:32 +08002069 list_splice_tail_init(list, &hctx->dispatch);
Jens Axboef04c3df2016-12-07 08:41:17 -07002070 spin_unlock(&hctx->lock);
2071
2072 /*
Ming Leid7d85352020-08-17 18:01:15 +08002073 * Order adding requests to hctx->dispatch and checking
2074 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2075 * in blk_mq_sched_restart(). Avoid restart code path to
2076 * miss the new added requests to hctx->dispatch, meantime
2077 * SCHED_RESTART is observed here.
2078 */
2079 smp_mb();
2080
2081 /*
Bart Van Assche710c7852017-04-07 11:16:51 -07002082 * If SCHED_RESTART was set by the caller of this function and
2083 * it is no longer set that means that it was cleared by another
2084 * thread and hence that a queue rerun is needed.
Jens Axboef04c3df2016-12-07 08:41:17 -07002085 *
Jens Axboeeb619fd2017-11-09 08:32:43 -07002086 * If 'no_tag' is set, that means that we failed getting
2087 * a driver tag with an I/O scheduler attached. If our dispatch
2088 * waitqueue is no longer active, ensure that we run the queue
2089 * AFTER adding our entries back to the list.
Jens Axboebd166ef2017-01-17 06:03:22 -07002090 *
Bart Van Assche710c7852017-04-07 11:16:51 -07002091 * If no I/O scheduler has been configured it is possible that
2092 * the hardware queue got stopped and restarted before requests
2093 * were pushed back onto the dispatch list. Rerun the queue to
2094 * avoid starvation. Notes:
2095 * - blk_mq_run_hw_queue() checks whether or not a queue has
2096 * been stopped before rerunning a queue.
2097 * - Some but not all block drivers stop a queue before
Christoph Hellwigfc17b652017-06-03 09:38:05 +02002098 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
Bart Van Assche710c7852017-04-07 11:16:51 -07002099 * and dm-rq.
Ming Lei86ff7c22018-01-30 22:04:57 -05002100 *
2101 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2102 * bit is set, run queue after a delay to avoid IO stalls
Douglas Andersonab3cee32020-04-20 09:24:51 -07002103 * that could otherwise occur if the queue is idle. We'll do
Naohiro Aota9586e672021-10-27 01:51:27 +09002104 * similar if we couldn't get budget or couldn't lock a zone
2105 * and SCHED_RESTART is set.
Jens Axboebd166ef2017-01-17 06:03:22 -07002106 */
Ming Lei86ff7c22018-01-30 22:04:57 -05002107 needs_restart = blk_mq_sched_needs_restart(hctx);
Naohiro Aota9586e672021-10-27 01:51:27 +09002108 if (prep == PREP_DISPATCH_NO_BUDGET)
2109 needs_resource = true;
Ming Lei86ff7c22018-01-30 22:04:57 -05002110 if (!needs_restart ||
Jens Axboeeb619fd2017-11-09 08:32:43 -07002111 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
Jens Axboebd166ef2017-01-17 06:03:22 -07002112 blk_mq_run_hw_queue(hctx, true);
Miaohe Lin6d5e8d22022-09-05 18:19:50 +08002113 else if (needs_resource)
Ming Lei86ff7c22018-01-30 22:04:57 -05002114 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
Jens Axboe1f57f8d2018-06-28 11:54:01 -06002115
Ming Lei6e7687172018-07-03 09:03:16 -06002116 blk_mq_update_dispatch_busy(hctx, true);
Jens Axboe1f57f8d2018-06-28 11:54:01 -06002117 return false;
Kemeng Shi4ea58fe2023-01-18 17:37:23 +08002118 }
Jens Axboef04c3df2016-12-07 08:41:17 -07002119
Kemeng Shi4ea58fe2023-01-18 17:37:23 +08002120 blk_mq_update_dispatch_busy(hctx, false);
2121 return true;
Jens Axboef04c3df2016-12-07 08:41:17 -07002122}
2123
Ming Leif82ddf12018-04-08 17:48:10 +08002124static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2125{
2126 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2127
2128 if (cpu >= nr_cpu_ids)
2129 cpu = cpumask_first(hctx->cpumask);
2130 return cpu;
2131}
2132
Jens Axboe506e9312014-05-07 10:26:44 -06002133/*
Ming Leia46c2702024-03-22 10:12:44 +08002134 * ->next_cpu is always calculated from hctx->cpumask, so simply use
2135 * it for speeding up the check
2136 */
2137static bool blk_mq_hctx_empty_cpumask(struct blk_mq_hw_ctx *hctx)
2138{
2139 return hctx->next_cpu >= nr_cpu_ids;
2140}
2141
2142/*
Jens Axboe506e9312014-05-07 10:26:44 -06002143 * It'd be great if the workqueue API had a way to pass
2144 * in a mask and had some smarts for more clever placement.
2145 * For now we just round-robin here, switching for every
2146 * BLK_MQ_CPU_WORK_BATCH queued items.
2147 */
2148static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2149{
Ming Lei7bed4592018-01-18 00:41:51 +08002150 bool tried = false;
Ming Lei476f8c92018-04-08 17:48:09 +08002151 int next_cpu = hctx->next_cpu;
Ming Lei7bed4592018-01-18 00:41:51 +08002152
Ming Leia46c2702024-03-22 10:12:44 +08002153 /* Switch to unbound if no allowable CPUs in this hctx */
2154 if (hctx->queue->nr_hw_queues == 1 || blk_mq_hctx_empty_cpumask(hctx))
Christoph Hellwigb657d7e2014-11-24 09:27:23 +01002155 return WORK_CPU_UNBOUND;
Jens Axboe506e9312014-05-07 10:26:44 -06002156
2157 if (--hctx->next_cpu_batch <= 0) {
Ming Lei7bed4592018-01-18 00:41:51 +08002158select_cpu:
Ming Lei476f8c92018-04-08 17:48:09 +08002159 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
Christoph Hellwig20e4d8132018-01-12 10:53:06 +08002160 cpu_online_mask);
Jens Axboe506e9312014-05-07 10:26:44 -06002161 if (next_cpu >= nr_cpu_ids)
Ming Leif82ddf12018-04-08 17:48:10 +08002162 next_cpu = blk_mq_first_mapped_cpu(hctx);
Jens Axboe506e9312014-05-07 10:26:44 -06002163 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2164 }
2165
Ming Lei7bed4592018-01-18 00:41:51 +08002166 /*
2167 * Do unbound schedule if we can't find a online CPU for this hctx,
2168 * and it should only happen in the path of handling CPU DEAD.
2169 */
Ming Lei476f8c92018-04-08 17:48:09 +08002170 if (!cpu_online(next_cpu)) {
Ming Lei7bed4592018-01-18 00:41:51 +08002171 if (!tried) {
2172 tried = true;
2173 goto select_cpu;
2174 }
2175
2176 /*
2177 * Make sure to re-select CPU next time once after CPUs
2178 * in hctx->cpumask become online again.
2179 */
Ming Lei476f8c92018-04-08 17:48:09 +08002180 hctx->next_cpu = next_cpu;
Ming Lei7bed4592018-01-18 00:41:51 +08002181 hctx->next_cpu_batch = 1;
2182 return WORK_CPU_UNBOUND;
2183 }
Ming Lei476f8c92018-04-08 17:48:09 +08002184
2185 hctx->next_cpu = next_cpu;
2186 return next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -06002187}
2188
André Almeida105663f2020-01-06 15:08:18 -03002189/**
André Almeida105663f2020-01-06 15:08:18 -03002190 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2191 * @hctx: Pointer to the hardware queue to run.
Minwoo Imfa94ba82020-12-05 00:20:55 +09002192 * @msecs: Milliseconds of delay to wait before running the queue.
André Almeida105663f2020-01-06 15:08:18 -03002193 *
2194 * Run a hardware queue asynchronously with a delay of @msecs.
2195 */
Bart Van Assche7587a5a2017-04-07 11:16:52 -07002196void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2197{
Christoph Hellwig1aa8d872023-04-13 08:06:50 +02002198 if (unlikely(blk_mq_hctx_stopped(hctx)))
2199 return;
2200 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2201 msecs_to_jiffies(msecs));
Bart Van Assche7587a5a2017-04-07 11:16:52 -07002202}
2203EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2204
André Almeida105663f2020-01-06 15:08:18 -03002205/**
2206 * blk_mq_run_hw_queue - Start to run a hardware queue.
2207 * @hctx: Pointer to the hardware queue to run.
2208 * @async: If we want to run the queue asynchronously.
2209 *
2210 * Check if the request queue is not in a quiesced state and if there are
2211 * pending requests to be sent. If this is true, run the queue to send requests
2212 * to hardware.
2213 */
John Garry626fb732019-10-30 00:59:30 +08002214void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
Bart Van Assche7587a5a2017-04-07 11:16:52 -07002215{
Ming Lei24f5a902018-01-06 16:27:38 +08002216 bool need_run;
2217
2218 /*
Christoph Hellwig4d5bba52023-04-13 08:06:51 +02002219 * We can't run the queue inline with interrupts disabled.
2220 */
2221 WARN_ON_ONCE(!async && in_interrupt());
2222
Bart Van Assche65a558f2023-07-21 10:27:30 -07002223 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2224
Christoph Hellwig4d5bba52023-04-13 08:06:51 +02002225 /*
Ming Lei24f5a902018-01-06 16:27:38 +08002226 * When queue is quiesced, we may be switching io scheduler, or
2227 * updating nr_hw_queues, or other things, and we can't run queue
2228 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2229 *
2230 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2231 * quiesced.
2232 */
Ming Lei41adf532021-12-06 19:12:13 +08002233 __blk_mq_run_dispatch_ops(hctx->queue, false,
Ming Lei2a904d02021-12-03 21:15:31 +08002234 need_run = !blk_queue_quiesced(hctx->queue) &&
2235 blk_mq_hctx_has_pending(hctx));
Ming Lei24f5a902018-01-06 16:27:38 +08002236
Christoph Hellwig1aa8d872023-04-13 08:06:50 +02002237 if (!need_run)
2238 return;
2239
Bart Van Assche65a558f2023-07-21 10:27:30 -07002240 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
Christoph Hellwig1aa8d872023-04-13 08:06:50 +02002241 blk_mq_delay_run_hw_queue(hctx, 0);
2242 return;
2243 }
2244
Christoph Hellwig4d5bba52023-04-13 08:06:51 +02002245 blk_mq_run_dispatch_ops(hctx->queue,
2246 blk_mq_sched_dispatch_requests(hctx));
Jens Axboe320ae512013-10-24 09:20:05 +01002247}
Omar Sandoval5b727272017-04-14 01:00:00 -07002248EXPORT_SYMBOL(blk_mq_run_hw_queue);
Jens Axboe320ae512013-10-24 09:20:05 +01002249
Jan Karab6e68ee2021-01-11 17:47:17 +01002250/*
Jan Karab6e68ee2021-01-11 17:47:17 +01002251 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2252 * scheduler.
2253 */
2254static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2255{
Ming Lei5d054262022-05-22 20:23:50 +08002256 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
Jan Karab6e68ee2021-01-11 17:47:17 +01002257 /*
2258 * If the IO scheduler does not respect hardware queues when
2259 * dispatching, we just don't bother with multiple HW queues and
2260 * dispatch from hctx for the current CPU since running multiple queues
2261 * just causes lock contention inside the scheduler and pointless cache
2262 * bouncing.
2263 */
Bart Van Assche51ab80f2022-06-15 15:55:49 -07002264 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
Ming Lei5d054262022-05-22 20:23:50 +08002265
Jan Karab6e68ee2021-01-11 17:47:17 +01002266 if (!blk_mq_hctx_stopped(hctx))
2267 return hctx;
2268 return NULL;
2269}
2270
André Almeida105663f2020-01-06 15:08:18 -03002271/**
Mauro Carvalho Chehab24f7bb82020-10-23 18:32:54 +02002272 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
André Almeida105663f2020-01-06 15:08:18 -03002273 * @q: Pointer to the request queue to run.
2274 * @async: If we want to run the queue asynchronously.
2275 */
Mike Snitzerb94ec292015-03-11 23:56:38 -04002276void blk_mq_run_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01002277{
Jan Karab6e68ee2021-01-11 17:47:17 +01002278 struct blk_mq_hw_ctx *hctx, *sq_hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002279 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01002280
Jan Karab6e68ee2021-01-11 17:47:17 +01002281 sq_hctx = NULL;
Ming Lei4d337ceb2022-06-16 09:44:00 +08002282 if (blk_queue_sq_sched(q))
Jan Karab6e68ee2021-01-11 17:47:17 +01002283 sq_hctx = blk_mq_get_sq_hctx(q);
Jens Axboe320ae512013-10-24 09:20:05 +01002284 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboe79f720a2017-11-10 09:13:21 -07002285 if (blk_mq_hctx_stopped(hctx))
Jens Axboe320ae512013-10-24 09:20:05 +01002286 continue;
Jan Karab6e68ee2021-01-11 17:47:17 +01002287 /*
2288 * Dispatch from this hctx either if there's no hctx preferred
2289 * by IO scheduler or if it has requests that bypass the
2290 * scheduler.
2291 */
2292 if (!sq_hctx || sq_hctx == hctx ||
2293 !list_empty_careful(&hctx->dispatch))
2294 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +01002295 }
2296}
Mike Snitzerb94ec292015-03-11 23:56:38 -04002297EXPORT_SYMBOL(blk_mq_run_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +01002298
Bart Van Asschefd001442016-10-28 17:19:37 -07002299/**
Douglas Andersonb9151e72020-04-20 09:24:52 -07002300 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2301 * @q: Pointer to the request queue to run.
Minwoo Imfa94ba82020-12-05 00:20:55 +09002302 * @msecs: Milliseconds of delay to wait before running the queues.
Douglas Andersonb9151e72020-04-20 09:24:52 -07002303 */
2304void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2305{
Jan Karab6e68ee2021-01-11 17:47:17 +01002306 struct blk_mq_hw_ctx *hctx, *sq_hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002307 unsigned long i;
Douglas Andersonb9151e72020-04-20 09:24:52 -07002308
Jan Karab6e68ee2021-01-11 17:47:17 +01002309 sq_hctx = NULL;
Ming Lei4d337ceb2022-06-16 09:44:00 +08002310 if (blk_queue_sq_sched(q))
Jan Karab6e68ee2021-01-11 17:47:17 +01002311 sq_hctx = blk_mq_get_sq_hctx(q);
Douglas Andersonb9151e72020-04-20 09:24:52 -07002312 queue_for_each_hw_ctx(q, hctx, i) {
2313 if (blk_mq_hctx_stopped(hctx))
2314 continue;
Jan Karab6e68ee2021-01-11 17:47:17 +01002315 /*
David Jeffery8f5fea62022-01-31 15:33:37 -05002316 * If there is already a run_work pending, leave the
2317 * pending delay untouched. Otherwise, a hctx can stall
2318 * if another hctx is re-delaying the other's work
2319 * before the work executes.
2320 */
2321 if (delayed_work_pending(&hctx->run_work))
2322 continue;
2323 /*
Jan Karab6e68ee2021-01-11 17:47:17 +01002324 * Dispatch from this hctx either if there's no hctx preferred
2325 * by IO scheduler or if it has requests that bypass the
2326 * scheduler.
2327 */
2328 if (!sq_hctx || sq_hctx == hctx ||
2329 !list_empty_careful(&hctx->dispatch))
2330 blk_mq_delay_run_hw_queue(hctx, msecs);
Douglas Andersonb9151e72020-04-20 09:24:52 -07002331 }
2332}
2333EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2334
Ming Lei39a70c72017-06-06 23:22:09 +08002335/*
2336 * This function is often used for pausing .queue_rq() by driver when
2337 * there isn't enough resource or some conditions aren't satisfied, and
Bart Van Assche4d606212017-08-17 16:23:00 -07002338 * BLK_STS_RESOURCE is usually returned.
Ming Lei39a70c72017-06-06 23:22:09 +08002339 *
2340 * We do not guarantee that dispatch can be drained or blocked
2341 * after blk_mq_stop_hw_queue() returns. Please use
2342 * blk_mq_quiesce_queue() for that requirement.
2343 */
Jens Axboe320ae512013-10-24 09:20:05 +01002344void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2345{
Ming Lei641a9ed2017-06-06 23:22:10 +08002346 cancel_delayed_work(&hctx->run_work);
2347
2348 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboe320ae512013-10-24 09:20:05 +01002349}
2350EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2351
Ming Lei39a70c72017-06-06 23:22:09 +08002352/*
2353 * This function is often used for pausing .queue_rq() by driver when
2354 * there isn't enough resource or some conditions aren't satisfied, and
Bart Van Assche4d606212017-08-17 16:23:00 -07002355 * BLK_STS_RESOURCE is usually returned.
Ming Lei39a70c72017-06-06 23:22:09 +08002356 *
2357 * We do not guarantee that dispatch can be drained or blocked
2358 * after blk_mq_stop_hw_queues() returns. Please use
2359 * blk_mq_quiesce_queue() for that requirement.
2360 */
Jens Axboe2719aa22017-05-03 11:08:14 -06002361void blk_mq_stop_hw_queues(struct request_queue *q)
2362{
Ming Lei641a9ed2017-06-06 23:22:10 +08002363 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002364 unsigned long i;
Ming Lei641a9ed2017-06-06 23:22:10 +08002365
2366 queue_for_each_hw_ctx(q, hctx, i)
2367 blk_mq_stop_hw_queue(hctx);
Christoph Hellwig280d45f2013-10-25 14:45:58 +01002368}
2369EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2370
Jens Axboe320ae512013-10-24 09:20:05 +01002371void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2372{
2373 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboee4043dc2014-04-09 10:18:23 -06002374
Bart Van Assche65a558f2023-07-21 10:27:30 -07002375 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
Jens Axboe320ae512013-10-24 09:20:05 +01002376}
2377EXPORT_SYMBOL(blk_mq_start_hw_queue);
2378
Christoph Hellwig2f268552014-04-16 09:44:56 +02002379void blk_mq_start_hw_queues(struct request_queue *q)
2380{
2381 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002382 unsigned long i;
Christoph Hellwig2f268552014-04-16 09:44:56 +02002383
2384 queue_for_each_hw_ctx(q, hctx, i)
2385 blk_mq_start_hw_queue(hctx);
2386}
2387EXPORT_SYMBOL(blk_mq_start_hw_queues);
2388
Jens Axboeae911c52016-12-08 13:19:30 -07002389void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2390{
2391 if (!blk_mq_hctx_stopped(hctx))
2392 return;
2393
2394 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2395 blk_mq_run_hw_queue(hctx, async);
2396}
2397EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2398
Christoph Hellwig1b4a3252014-04-16 09:44:54 +02002399void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01002400{
2401 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002402 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01002403
Jens Axboeae911c52016-12-08 13:19:30 -07002404 queue_for_each_hw_ctx(q, hctx, i)
Bart Van Assche65a558f2023-07-21 10:27:30 -07002405 blk_mq_start_stopped_hw_queue(hctx, async ||
2406 (hctx->flags & BLK_MQ_F_BLOCKING));
Jens Axboe320ae512013-10-24 09:20:05 +01002407}
2408EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2409
Christoph Hellwig70f4db62014-04-16 10:48:08 -06002410static void blk_mq_run_work_fn(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +01002411{
Christoph Hellwigc20a1a22023-04-13 08:06:48 +02002412 struct blk_mq_hw_ctx *hctx =
2413 container_of(work, struct blk_mq_hw_ctx, run_work.work);
Jens Axboe320ae512013-10-24 09:20:05 +01002414
Christoph Hellwig4d5bba52023-04-13 08:06:51 +02002415 blk_mq_run_dispatch_ops(hctx->queue,
2416 blk_mq_sched_dispatch_requests(hctx));
Jens Axboe320ae512013-10-24 09:20:05 +01002417}
2418
André Almeida105663f2020-01-06 15:08:18 -03002419/**
2420 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2421 * @rq: Pointer to request to be inserted.
Christoph Hellwig2b597612023-04-13 08:40:55 +02002422 * @flags: BLK_MQ_INSERT_*
André Almeida105663f2020-01-06 15:08:18 -03002423 *
Jens Axboe157f3772017-09-11 16:43:57 -06002424 * Should only be used carefully, when the caller knows we want to
2425 * bypass a potential IO scheduler on the target device.
2426 */
Christoph Hellwig360f2642023-05-19 06:40:46 +02002427static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
Jens Axboe157f3772017-09-11 16:43:57 -06002428{
Jens Axboeea4f9952018-10-29 15:06:13 -06002429 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Jens Axboe157f3772017-09-11 16:43:57 -06002430
2431 spin_lock(&hctx->lock);
Christoph Hellwig2b597612023-04-13 08:40:55 +02002432 if (flags & BLK_MQ_INSERT_AT_HEAD)
Ming Lei01e99ae2020-02-25 09:04:32 +08002433 list_add(&rq->queuelist, &hctx->dispatch);
2434 else
2435 list_add_tail(&rq->queuelist, &hctx->dispatch);
Jens Axboe157f3772017-09-11 16:43:57 -06002436 spin_unlock(&hctx->lock);
Jens Axboe157f3772017-09-11 16:43:57 -06002437}
2438
Christoph Hellwig05a93112023-04-13 08:40:42 +02002439static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2440 struct blk_mq_ctx *ctx, struct list_head *list,
2441 bool run_queue_async)
Jens Axboe320ae512013-10-24 09:20:05 +01002442{
Ming Lei3f0cedc2018-07-02 17:35:58 +08002443 struct request *rq;
Ming Leic16d6b52018-12-17 08:44:05 -07002444 enum hctx_type type = hctx->type;
Ming Lei3f0cedc2018-07-02 17:35:58 +08002445
Jens Axboe320ae512013-10-24 09:20:05 +01002446 /*
Christoph Hellwig94aa2282023-04-13 08:40:41 +02002447 * Try to issue requests directly if the hw queue isn't busy to save an
2448 * extra enqueue & dequeue to the sw queue.
2449 */
2450 if (!hctx->dispatch_busy && !run_queue_async) {
2451 blk_mq_run_dispatch_ops(hctx->queue,
2452 blk_mq_try_issue_list_directly(hctx, list));
2453 if (list_empty(list))
2454 goto out;
2455 }
2456
2457 /*
Jens Axboe320ae512013-10-24 09:20:05 +01002458 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2459 * offline now
2460 */
Ming Lei3f0cedc2018-07-02 17:35:58 +08002461 list_for_each_entry(rq, list, queuelist) {
Jens Axboee57690f2016-08-24 15:34:35 -06002462 BUG_ON(rq->mq_ctx != ctx);
Christoph Hellwiga54895f2020-12-03 17:21:39 +01002463 trace_block_rq_insert(rq);
Bart Van Assche65a558f2023-07-21 10:27:30 -07002464 if (rq->cmd_flags & REQ_NOWAIT)
2465 run_queue_async = true;
Jens Axboe320ae512013-10-24 09:20:05 +01002466 }
Ming Lei3f0cedc2018-07-02 17:35:58 +08002467
2468 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07002469 list_splice_tail_init(list, &ctx->rq_lists[type]);
Ming Leicfd0c552015-10-20 23:13:57 +08002470 blk_mq_hctx_mark_pending(hctx, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01002471 spin_unlock(&ctx->lock);
Christoph Hellwig94aa2282023-04-13 08:40:41 +02002472out:
2473 blk_mq_run_hw_queue(hctx, run_queue_async);
Jens Axboe320ae512013-10-24 09:20:05 +01002474}
2475
Christoph Hellwig710fa372023-04-13 08:40:54 +02002476static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002477{
2478 struct request_queue *q = rq->q;
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002479 struct blk_mq_ctx *ctx = rq->mq_ctx;
2480 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2481
Christoph Hellwig53548d22023-04-13 08:40:47 +02002482 if (blk_rq_is_passthrough(rq)) {
2483 /*
2484 * Passthrough request have to be added to hctx->dispatch
2485 * directly. The device may be in a situation where it can't
2486 * handle FS request, and always returns BLK_STS_RESOURCE for
2487 * them, which gets them added to hctx->dispatch.
2488 *
2489 * If a passthrough request is required to unblock the queues,
2490 * and it is added to the scheduler queue, there is no chance to
2491 * dispatch it given we prioritize requests in hctx->dispatch.
2492 */
Christoph Hellwig2b597612023-04-13 08:40:55 +02002493 blk_mq_request_bypass_insert(rq, flags);
Bart Van Asschebe4c4272023-05-19 06:40:47 +02002494 } else if (req_op(rq) == REQ_OP_FLUSH) {
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002495 /*
2496 * Firstly normal IO request is inserted to scheduler queue or
2497 * sw queue, meantime we add flush request to dispatch queue(
2498 * hctx->dispatch) directly and there is at most one in-flight
2499 * flush request for each hw queue, so it doesn't matter to add
2500 * flush request to tail or front of the dispatch queue.
2501 *
2502 * Secondly in case of NCQ, flush request belongs to non-NCQ
2503 * command, and queueing it will fail when there is any
2504 * in-flight normal IO request(NCQ command). When adding flush
2505 * rq to the front of hctx->dispatch, it is easier to introduce
2506 * extra time to flush rq's latency because of S_SCHED_RESTART
2507 * compared with adding to the tail of dispatch queue, then
2508 * chance of flush merge is increased, and less flush requests
2509 * will be issued to controller. It is observed that ~10% time
2510 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2511 * drive when adding flush rq to the front of hctx->dispatch.
2512 *
2513 * Simply queue flush rq to the front of hctx->dispatch so that
2514 * intensive flush workloads can benefit in case of NCQ HW.
2515 */
Christoph Hellwig2b597612023-04-13 08:40:55 +02002516 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
Christoph Hellwig53548d22023-04-13 08:40:47 +02002517 } else if (q->elevator) {
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002518 LIST_HEAD(list);
2519
Christoph Hellwig53548d22023-04-13 08:40:47 +02002520 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2521
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002522 list_add(&rq->queuelist, &list);
Christoph Hellwig93fffe12023-04-13 08:40:56 +02002523 q->elevator->type->ops.insert_requests(hctx, &list, flags);
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002524 } else {
Christoph Hellwig4ec5c052023-04-13 08:40:45 +02002525 trace_block_rq_insert(rq);
2526
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002527 spin_lock(&ctx->lock);
Christoph Hellwig710fa372023-04-13 08:40:54 +02002528 if (flags & BLK_MQ_INSERT_AT_HEAD)
Christoph Hellwig4ec5c052023-04-13 08:40:45 +02002529 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2530 else
2531 list_add_tail(&rq->queuelist,
2532 &ctx->rq_lists[hctx->type]);
Christoph Hellwiga88db1e2023-04-13 08:40:44 +02002533 blk_mq_hctx_mark_pending(hctx, ctx);
Christoph Hellwig2bd215d2023-04-13 08:40:43 +02002534 spin_unlock(&ctx->lock);
2535 }
Jens Axboe320ae512013-10-24 09:20:05 +01002536}
2537
Christoph Hellwig14ccb662019-06-06 12:29:01 +02002538static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2539 unsigned int nr_segs)
Jens Axboe320ae512013-10-24 09:20:05 +01002540{
Eric Biggers93f221a2020-09-15 20:53:14 -07002541 int err;
2542
Christoph Hellwigf924cdd2019-06-06 12:29:00 +02002543 if (bio->bi_opf & REQ_RAHEAD)
2544 rq->cmd_flags |= REQ_FAILFAST_MASK;
2545
2546 rq->__sector = bio->bi_iter.bi_sector;
Bart Van Assche44981352024-02-02 12:39:25 -08002547 rq->write_hint = bio->bi_write_hint;
Christoph Hellwig14ccb662019-06-06 12:29:01 +02002548 blk_rq_bio_prep(rq, bio, nr_segs);
Eric Biggers93f221a2020-09-15 20:53:14 -07002549
2550 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2551 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2552 WARN_ON_ONCE(err);
Jens Axboe4b570522014-05-29 11:00:11 -06002553
Konstantin Khlebnikovb5af37a2020-05-27 07:24:16 +02002554 blk_account_io_start(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01002555}
2556
Mike Snitzer0f955492018-01-17 11:25:56 -05002557static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig3e087732021-10-12 13:12:24 +02002558 struct request *rq, bool last)
Shaohua Lif984df12015-05-08 10:51:32 -07002559{
Shaohua Lif984df12015-05-08 10:51:32 -07002560 struct request_queue *q = rq->q;
Shaohua Lif984df12015-05-08 10:51:32 -07002561 struct blk_mq_queue_data bd = {
2562 .rq = rq,
Jens Axboebe94f052018-11-24 10:15:46 -07002563 .last = last,
Shaohua Lif984df12015-05-08 10:51:32 -07002564 };
Jens Axboef06345a2017-06-12 11:22:46 -06002565 blk_status_t ret;
Mike Snitzer0f955492018-01-17 11:25:56 -05002566
Mike Snitzer0f955492018-01-17 11:25:56 -05002567 /*
2568 * For OK queue, we are done. For error, caller may kill it.
2569 * Any other error (busy), just add it to our list as we
2570 * previously would have done.
2571 */
2572 ret = q->mq_ops->queue_rq(hctx, &bd);
2573 switch (ret) {
2574 case BLK_STS_OK:
Ming Lei6ce3dd62018-07-10 09:03:31 +08002575 blk_mq_update_dispatch_busy(hctx, false);
Mike Snitzer0f955492018-01-17 11:25:56 -05002576 break;
2577 case BLK_STS_RESOURCE:
Ming Lei86ff7c22018-01-30 22:04:57 -05002578 case BLK_STS_DEV_RESOURCE:
Ming Lei6ce3dd62018-07-10 09:03:31 +08002579 blk_mq_update_dispatch_busy(hctx, true);
Mike Snitzer0f955492018-01-17 11:25:56 -05002580 __blk_mq_requeue_request(rq);
2581 break;
2582 default:
Ming Lei6ce3dd62018-07-10 09:03:31 +08002583 blk_mq_update_dispatch_busy(hctx, false);
Mike Snitzer0f955492018-01-17 11:25:56 -05002584 break;
2585 }
2586
2587 return ret;
2588}
2589
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002590static bool blk_mq_get_budget_and_tag(struct request *rq)
Mike Snitzer0f955492018-01-17 11:25:56 -05002591{
Ming Lei2a5a24a2021-01-22 10:33:12 +08002592 int budget_token;
Ming Leid964f042017-06-06 23:22:00 +08002593
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002594 budget_token = blk_mq_get_dispatch_budget(rq->q);
Ming Lei2a5a24a2021-01-22 10:33:12 +08002595 if (budget_token < 0)
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002596 return false;
Ming Lei2a5a24a2021-01-22 10:33:12 +08002597 blk_mq_set_rq_budget_token(rq, budget_token);
Ming Lei8ab6bb9e2018-06-25 19:31:45 +08002598 if (!blk_mq_get_driver_tag(rq)) {
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002599 blk_mq_put_dispatch_budget(rq->q, budget_token);
2600 return false;
Ming Lei88022d72017-11-05 02:21:12 +08002601 }
Christoph Hellwig2b71b872023-04-13 08:40:49 +02002602 return true;
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002603}
2604
André Almeida105663f2020-01-06 15:08:18 -03002605/**
2606 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2607 * @hctx: Pointer of the associated hardware queue.
2608 * @rq: Pointer to request to be sent.
André Almeida105663f2020-01-06 15:08:18 -03002609 *
2610 * If the device has enough resources to accept a new request now, send the
2611 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2612 * we can try send it another time in the future. Requests inserted at this
2613 * queue have higher priority.
2614 */
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002615static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig3e087732021-10-12 13:12:24 +02002616 struct request *rq)
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002617{
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002618 blk_status_t ret;
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002619
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002620 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
Christoph Hellwig710fa372023-04-13 08:40:54 +02002621 blk_mq_insert_request(rq, 0);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002622 return;
2623 }
2624
Christoph Hellwigdd6216b2023-05-18 07:31:01 +02002625 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
Christoph Hellwig710fa372023-04-13 08:40:54 +02002626 blk_mq_insert_request(rq, 0);
Bart Van Assche65a558f2023-07-21 10:27:30 -07002627 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002628 return;
2629 }
2630
2631 ret = __blk_mq_issue_directly(hctx, rq, true);
2632 switch (ret) {
2633 case BLK_STS_OK:
2634 break;
2635 case BLK_STS_RESOURCE:
2636 case BLK_STS_DEV_RESOURCE:
Christoph Hellwig2b597612023-04-13 08:40:55 +02002637 blk_mq_request_bypass_insert(rq, 0);
Christoph Hellwig23943952023-04-13 08:40:52 +02002638 blk_mq_run_hw_queue(hctx, false);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002639 break;
2640 default:
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002641 blk_mq_end_request(rq, ret);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002642 break;
2643 }
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002644}
2645
Christoph Hellwig06c8c692021-11-17 07:13:58 +01002646static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002647{
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002648 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2649
2650 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
Christoph Hellwig710fa372023-04-13 08:40:54 +02002651 blk_mq_insert_request(rq, 0);
Christoph Hellwige1f44ac2023-04-13 08:40:50 +02002652 return BLK_STS_OK;
2653 }
2654
2655 if (!blk_mq_get_budget_and_tag(rq))
2656 return BLK_STS_RESOURCE;
2657 return __blk_mq_issue_directly(hctx, rq, last);
Christoph Hellwig5eb61262017-03-22 15:01:51 -04002658}
2659
Kemeng Shi3e368fb2023-01-18 17:37:18 +08002660static void blk_mq_plug_issue_direct(struct blk_plug *plug)
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002661{
2662 struct blk_mq_hw_ctx *hctx = NULL;
2663 struct request *rq;
2664 int queued = 0;
Kemeng Shi0d617a82023-01-18 17:37:20 +08002665 blk_status_t ret = BLK_STS_OK;
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002666
2667 while ((rq = rq_list_pop(&plug->mq_list))) {
2668 bool last = rq_list_empty(plug->mq_list);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002669
2670 if (hctx != rq->mq_hctx) {
Kemeng Shi34c9f5472023-01-18 17:37:19 +08002671 if (hctx) {
2672 blk_mq_commit_rqs(hctx, queued, false);
2673 queued = 0;
2674 }
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002675 hctx = rq->mq_hctx;
2676 }
2677
2678 ret = blk_mq_request_issue_directly(rq, last);
2679 switch (ret) {
2680 case BLK_STS_OK:
2681 queued++;
2682 break;
2683 case BLK_STS_RESOURCE:
2684 case BLK_STS_DEV_RESOURCE:
Christoph Hellwig2b597612023-04-13 08:40:55 +02002685 blk_mq_request_bypass_insert(rq, 0);
Christoph Hellwig23943952023-04-13 08:40:52 +02002686 blk_mq_run_hw_queue(hctx, false);
Kemeng Shi0d617a82023-01-18 17:37:20 +08002687 goto out;
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002688 default:
2689 blk_mq_end_request(rq, ret);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002690 break;
2691 }
2692 }
2693
Kemeng Shi0d617a82023-01-18 17:37:20 +08002694out:
2695 if (ret != BLK_STS_OK)
Kemeng Shi34c9f5472023-01-18 17:37:19 +08002696 blk_mq_commit_rqs(hctx, queued, false);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002697}
2698
Keith Busch518579a2021-12-20 12:59:19 -08002699static void __blk_mq_flush_plug_list(struct request_queue *q,
2700 struct blk_plug *plug)
2701{
2702 if (blk_queue_quiesced(q))
2703 return;
2704 q->mq_ops->queue_rqs(&plug->mq_list);
2705}
2706
Jens Axboe26fed4a2022-03-11 10:24:17 -07002707static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2708{
2709 struct blk_mq_hw_ctx *this_hctx = NULL;
2710 struct blk_mq_ctx *this_ctx = NULL;
2711 struct request *requeue_list = NULL;
Jan Kara34e0a272023-03-13 10:30:02 +01002712 struct request **requeue_lastp = &requeue_list;
Jens Axboe26fed4a2022-03-11 10:24:17 -07002713 unsigned int depth = 0;
Ming Leid97217e2023-05-18 07:30:59 +02002714 bool is_passthrough = false;
Jens Axboe26fed4a2022-03-11 10:24:17 -07002715 LIST_HEAD(list);
2716
2717 do {
2718 struct request *rq = rq_list_pop(&plug->mq_list);
2719
2720 if (!this_hctx) {
2721 this_hctx = rq->mq_hctx;
2722 this_ctx = rq->mq_ctx;
Ming Leid97217e2023-05-18 07:30:59 +02002723 is_passthrough = blk_rq_is_passthrough(rq);
2724 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2725 is_passthrough != blk_rq_is_passthrough(rq)) {
Jan Kara34e0a272023-03-13 10:30:02 +01002726 rq_list_add_tail(&requeue_lastp, rq);
Jens Axboe26fed4a2022-03-11 10:24:17 -07002727 continue;
2728 }
Jan Kara34e0a272023-03-13 10:30:02 +01002729 list_add(&rq->queuelist, &list);
Jens Axboe26fed4a2022-03-11 10:24:17 -07002730 depth++;
2731 } while (!rq_list_empty(plug->mq_list));
2732
2733 plug->mq_list = requeue_list;
2734 trace_block_unplug(this_hctx->queue, depth, !from_sched);
Christoph Hellwig05a93112023-04-13 08:40:42 +02002735
2736 percpu_ref_get(&this_hctx->queue->q_usage_counter);
Ming Leid97217e2023-05-18 07:30:59 +02002737 /* passthrough requests should never be issued to the I/O scheduler */
Ming Lei2293cae2023-06-21 21:22:08 +08002738 if (is_passthrough) {
2739 spin_lock(&this_hctx->lock);
2740 list_splice_tail_init(&list, &this_hctx->dispatch);
2741 spin_unlock(&this_hctx->lock);
2742 blk_mq_run_hw_queue(this_hctx, from_sched);
2743 } else if (this_hctx->queue->elevator) {
Christoph Hellwig05a93112023-04-13 08:40:42 +02002744 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
Christoph Hellwig93fffe12023-04-13 08:40:56 +02002745 &list, 0);
Christoph Hellwig05a93112023-04-13 08:40:42 +02002746 blk_mq_run_hw_queue(this_hctx, from_sched);
2747 } else {
2748 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2749 }
2750 percpu_ref_put(&this_hctx->queue->q_usage_counter);
Jens Axboe26fed4a2022-03-11 10:24:17 -07002751}
2752
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002753void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2754{
Jens Axboe3c67d442021-12-03 06:48:53 -07002755 struct request *rq;
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002756
Ross Lagerwall70904262023-07-14 11:11:06 +01002757 /*
2758 * We may have been called recursively midway through handling
2759 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2760 * To avoid mq_list changing under our feet, clear rq_count early and
2761 * bail out specifically if rq_count is 0 rather than checking
2762 * whether the mq_list is empty.
2763 */
2764 if (plug->rq_count == 0)
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002765 return;
2766 plug->rq_count = 0;
2767
2768 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
Jens Axboe3c67d442021-12-03 06:48:53 -07002769 struct request_queue *q;
2770
2771 rq = rq_list_peek(&plug->mq_list);
2772 q = rq->q;
2773
2774 /*
2775 * Peek first request and see if we have a ->queue_rqs() hook.
2776 * If we do, we can dispatch the whole plug list in one go. We
2777 * already know at this point that all requests belong to the
2778 * same queue, caller must ensure that's the case.
Jens Axboe3c67d442021-12-03 06:48:53 -07002779 */
Chengming Zhou434097ee2023-09-13 15:16:14 +00002780 if (q->mq_ops->queue_rqs) {
Jens Axboe3c67d442021-12-03 06:48:53 -07002781 blk_mq_run_dispatch_ops(q,
Keith Busch518579a2021-12-20 12:59:19 -08002782 __blk_mq_flush_plug_list(q, plug));
Jens Axboe3c67d442021-12-03 06:48:53 -07002783 if (rq_list_empty(plug->mq_list))
2784 return;
2785 }
Ming Lei73f37602021-12-06 11:33:50 +08002786
2787 blk_mq_run_dispatch_ops(q,
Kemeng Shi3e368fb2023-01-18 17:37:18 +08002788 blk_mq_plug_issue_direct(plug));
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002789 if (rq_list_empty(plug->mq_list))
2790 return;
2791 }
2792
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002793 do {
Jens Axboe26fed4a2022-03-11 10:24:17 -07002794 blk_mq_dispatch_plug_list(plug, from_schedule);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002795 } while (!rq_list_empty(plug->mq_list));
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002796}
2797
Christoph Hellwig94aa2282023-04-13 08:40:41 +02002798static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
Ming Lei6ce3dd62018-07-10 09:03:31 +08002799 struct list_head *list)
2800{
Keith Busch536167d42020-04-07 03:13:48 +09002801 int queued = 0;
Kemeng Shi984ce0a2023-01-18 17:37:21 +08002802 blk_status_t ret = BLK_STS_OK;
Keith Busch536167d42020-04-07 03:13:48 +09002803
Ming Lei6ce3dd62018-07-10 09:03:31 +08002804 while (!list_empty(list)) {
Ming Lei6ce3dd62018-07-10 09:03:31 +08002805 struct request *rq = list_first_entry(list, struct request,
2806 queuelist);
2807
2808 list_del_init(&rq->queuelist);
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002809 ret = blk_mq_request_issue_directly(rq, list_empty(list));
Kemeng Shi27e8b2b2023-01-18 17:37:25 +08002810 switch (ret) {
2811 case BLK_STS_OK:
Keith Busch536167d42020-04-07 03:13:48 +09002812 queued++;
Kemeng Shi27e8b2b2023-01-18 17:37:25 +08002813 break;
2814 case BLK_STS_RESOURCE:
2815 case BLK_STS_DEV_RESOURCE:
Christoph Hellwig2b597612023-04-13 08:40:55 +02002816 blk_mq_request_bypass_insert(rq, 0);
Christoph Hellwig23943952023-04-13 08:40:52 +02002817 if (list_empty(list))
2818 blk_mq_run_hw_queue(hctx, false);
Kemeng Shi27e8b2b2023-01-18 17:37:25 +08002819 goto out;
2820 default:
2821 blk_mq_end_request(rq, ret);
2822 break;
2823 }
Ming Lei6ce3dd62018-07-10 09:03:31 +08002824 }
Jens Axboed666ba92018-11-27 17:02:25 -07002825
Kemeng Shi27e8b2b2023-01-18 17:37:25 +08002826out:
Kemeng Shi984ce0a2023-01-18 17:37:21 +08002827 if (ret != BLK_STS_OK)
2828 blk_mq_commit_rqs(hctx, queued, false);
Ming Lei6ce3dd62018-07-10 09:03:31 +08002829}
2830
Ming Leib131f202021-11-11 16:51:34 +08002831static bool blk_mq_attempt_bio_merge(struct request_queue *q,
Christoph Hellwig0c5bcc92021-11-23 17:04:41 +01002832 struct bio *bio, unsigned int nr_segs)
Jens Axboe900e0802021-11-03 05:47:09 -06002833{
2834 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
Christoph Hellwig0c5bcc92021-11-23 17:04:41 +01002835 if (blk_attempt_plug_merge(q, bio, nr_segs))
Jens Axboe900e0802021-11-03 05:47:09 -06002836 return true;
2837 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2838 return true;
2839 }
2840 return false;
2841}
2842
Jens Axboe71539712021-11-03 05:52:45 -06002843static struct request *blk_mq_get_new_requests(struct request_queue *q,
2844 struct blk_plug *plug,
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002845 struct bio *bio,
2846 unsigned int nsegs)
Jens Axboe71539712021-11-03 05:52:45 -06002847{
2848 struct blk_mq_alloc_data data = {
2849 .q = q,
2850 .nr_tags = 1,
Ming Lei9d497e22022-01-04 21:42:23 +08002851 .cmd_flags = bio->bi_opf,
Jens Axboe71539712021-11-03 05:52:45 -06002852 };
2853 struct request *rq;
2854
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002855 rq_qos_throttle(q, bio);
2856
Jens Axboe71539712021-11-03 05:52:45 -06002857 if (plug) {
2858 data.nr_tags = plug->nr_ios;
2859 plug->nr_ios = 1;
2860 data.cached_rq = &plug->cached_rq;
2861 }
2862
2863 rq = __blk_mq_alloc_requests(&data);
Jens Axboe373b5412021-12-02 12:42:58 -07002864 if (rq)
2865 return rq;
Jens Axboe71539712021-11-03 05:52:45 -06002866 rq_qos_cleanup(q, bio);
2867 if (bio->bi_opf & REQ_NOWAIT)
2868 bio_wouldblock_error(bio);
Jens Axboe95febeb2021-11-15 14:23:08 -07002869 return NULL;
Jens Axboe71539712021-11-03 05:52:45 -06002870}
2871
Christoph Hellwig309ce672024-01-11 14:57:04 +01002872/*
Christoph Hellwig337e89f2024-01-24 10:26:57 +01002873 * Check if there is a suitable cached request and return it.
Christoph Hellwig309ce672024-01-11 14:57:04 +01002874 */
Christoph Hellwig337e89f2024-01-24 10:26:57 +01002875static struct request *blk_mq_peek_cached_request(struct blk_plug *plug,
2876 struct request_queue *q, blk_opf_t opf)
2877{
2878 enum hctx_type type = blk_mq_get_hctx_type(opf);
2879 struct request *rq;
2880
2881 if (!plug)
2882 return NULL;
2883 rq = rq_list_peek(&plug->cached_rq);
2884 if (!rq || rq->q != q)
2885 return NULL;
2886 if (type != rq->mq_hctx->type &&
2887 (type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT))
2888 return NULL;
2889 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
2890 return NULL;
2891 return rq;
2892}
2893
2894static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug,
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002895 struct bio *bio)
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002896{
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002897 WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq);
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002898
Jens Axboe26456722022-06-21 10:03:57 -06002899 /*
2900 * If any qos ->throttle() end up blocking, we will have flushed the
2901 * plug and hence killed the cached_rq list as well. Pop this entry
2902 * before we throttle.
2903 */
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002904 plug->cached_rq = rq_list_next(rq);
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002905 rq_qos_throttle(rq->q, bio);
Jens Axboe26456722022-06-21 10:03:57 -06002906
Chengming Zhou5c17f452023-07-10 18:55:16 +08002907 blk_mq_rq_time_init(rq, 0);
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002908 rq->cmd_flags = bio->bi_opf;
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002909 INIT_LIST_HEAD(&rq->queuelist);
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002910}
2911
Ming Lei0676c432024-06-20 11:06:31 +08002912static bool bio_unaligned(const struct bio *bio, struct request_queue *q)
2913{
2914 unsigned int bs_mask = queue_logical_block_size(q) - 1;
2915
2916 /* .bi_sector of any zero sized bio need to be initialized */
2917 if ((bio->bi_iter.bi_size & bs_mask) ||
2918 ((bio->bi_iter.bi_sector << SECTOR_SHIFT) & bs_mask))
2919 return true;
2920 return false;
2921}
2922
André Almeida105663f2020-01-06 15:08:18 -03002923/**
Christoph Hellwigc62b37d2020-07-01 10:59:43 +02002924 * blk_mq_submit_bio - Create and send a request to block device.
André Almeida105663f2020-01-06 15:08:18 -03002925 * @bio: Bio pointer.
2926 *
2927 * Builds up a request structure from @q and @bio and send to the device. The
2928 * request may not be queued directly to hardware if:
2929 * * This request can be merged with another one
2930 * * We want to place request at plug queue for possible future merging
2931 * * There is an IO scheduler active at this queue
2932 *
2933 * It will not queue the request if there is an error with the bio, or at the
2934 * request creation.
André Almeida105663f2020-01-06 15:08:18 -03002935 */
Christoph Hellwig3e087732021-10-12 13:12:24 +02002936void blk_mq_submit_bio(struct bio *bio)
Jens Axboe07068d52014-05-22 10:40:51 -06002937{
Pavel Begunkoved6cdde2021-10-14 15:03:30 +01002938 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
Damien Le Moal99a94762024-04-08 10:41:28 +09002939 struct blk_plug *plug = current->plug;
Christoph Hellwigef295ec2016-10-28 08:48:16 -06002940 const int is_sync = op_is_sync(bio->bi_opf);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02002941 struct blk_mq_hw_ctx *hctx;
Jens Axboeabd45c12021-10-13 12:43:41 -06002942 unsigned int nr_segs = 1;
Christoph Hellwig72e84e92024-01-24 10:26:58 +01002943 struct request *rq;
Satya Tangiralaa892c8d2020-05-14 00:37:18 +00002944 blk_status_t ret;
Jens Axboe07068d52014-05-22 10:40:51 -06002945
Damien Le Moaldd291d72024-04-08 10:41:07 +09002946 /*
2947 * If the plug has a cached request for this queue, try to use it.
2948 */
2949 rq = blk_mq_peek_cached_request(plug, q, bio->bi_opf);
2950
2951 /*
2952 * A BIO that was released from a zone write plug has already been
2953 * through the preparation in this function, already holds a reference
2954 * on the queue usage counter, and is the only write BIO in-flight for
2955 * the target zone. Go straight to preparing a request for it.
2956 */
2957 if (bio_zone_write_plugging(bio)) {
2958 nr_segs = bio->__bi_nr_segments;
2959 if (rq)
2960 blk_queue_exit(q);
2961 goto new_request;
2962 }
2963
Christoph Hellwig51d798c2022-07-27 12:22:56 -04002964 bio = blk_queue_bounce(bio, q);
Jan Kara9c6227e2022-06-23 09:48:33 +02002965
Christoph Hellwig72e84e92024-01-24 10:26:58 +01002966 /*
Christoph Hellwig72e84e92024-01-24 10:26:58 +01002967 * The cached request already holds a q_usage_counter reference and we
2968 * don't have to acquire a new one if we use it.
2969 */
Christoph Hellwig72e84e92024-01-24 10:26:58 +01002970 if (!rq) {
Christoph Hellwigb0077e22023-11-13 11:52:31 +08002971 if (unlikely(bio_queue_enter(bio)))
2972 return;
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002973 }
Jens Axboe87760e52016-11-09 12:38:14 -07002974
Ming Lei0676c432024-06-20 11:06:31 +08002975 /*
2976 * Device reconfiguration may change logical block size, so alignment
2977 * check has to be done with queue usage counter held
2978 */
2979 if (unlikely(bio_unaligned(bio, q))) {
2980 bio_io_error(bio);
2981 goto queue_exit;
2982 }
2983
Christoph Hellwig337e89f2024-01-24 10:26:57 +01002984 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2985 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2986 if (!bio)
Christoph Hellwig0f299da2024-01-24 10:26:56 +01002987 goto queue_exit;
Jens Axboedece1632015-11-05 10:41:16 -07002988 }
Christoph Hellwig337e89f2024-01-24 10:26:57 +01002989 if (!bio_integrity_prep(bio))
2990 goto queue_exit;
Jens Axboe87760e52016-11-09 12:38:14 -07002991
Christoph Hellwig0f299da2024-01-24 10:26:56 +01002992 if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
2993 goto queue_exit;
2994
Damien Le Moaldd291d72024-04-08 10:41:07 +09002995 if (blk_queue_is_zoned(q) && blk_zone_plug_bio(bio, nr_segs))
2996 goto queue_exit;
2997
2998new_request:
Christoph Hellwig72e84e92024-01-24 10:26:58 +01002999 if (!rq) {
3000 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
3001 if (unlikely(!rq))
3002 goto queue_exit;
3003 } else {
3004 blk_mq_use_cached_rq(rq, plug, bio);
Christoph Hellwigb0077e22023-11-13 11:52:31 +08003005 }
3006
Christoph Hellwige8a676d2020-12-03 17:21:36 +01003007 trace_block_getrq(bio);
Xiaoguang Wangd6f1dda2018-10-23 22:30:50 +08003008
Josef Bacikc1c80382018-07-03 11:14:59 -04003009 rq_qos_track(q, rq, bio);
Jens Axboe07068d52014-05-22 10:40:51 -06003010
Bart Van Assche970d1682019-07-01 08:47:30 -07003011 blk_mq_bio_to_request(rq, bio, nr_segs);
3012
Eric Biggers9cd1e562023-03-15 11:39:02 -07003013 ret = blk_crypto_rq_get_keyslot(rq);
Satya Tangiralaa892c8d2020-05-14 00:37:18 +00003014 if (ret != BLK_STS_OK) {
3015 bio->bi_status = ret;
3016 bio_endio(bio);
3017 blk_mq_free_request(rq);
Christoph Hellwig3e087732021-10-12 13:12:24 +02003018 return;
Satya Tangiralaa892c8d2020-05-14 00:37:18 +00003019 }
3020
Damien Le Moal096bc7e2024-05-01 20:09:02 +09003021 if (bio_zone_write_plugging(bio))
3022 blk_zone_write_plug_init_request(rq);
3023
Christoph Hellwig360f2642023-05-19 06:40:46 +02003024 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
Christoph Hellwigd92ca9d82021-10-19 14:25:53 +02003025 return;
3026
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02003027 if (plug) {
Jens Axboece5b0092018-11-27 17:13:56 -07003028 blk_add_rq_to_plug(plug, rq);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02003029 return;
3030 }
3031
3032 hctx = rq->mq_hctx;
Christoph Hellwigdd6216b2023-05-18 07:31:01 +02003033 if ((rq->rq_flags & RQF_USE_SCHED) ||
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02003034 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
Christoph Hellwig710fa372023-04-13 08:40:54 +02003035 blk_mq_insert_request(rq, 0);
Christoph Hellwigf0dbe6e2023-04-13 08:40:51 +02003036 blk_mq_run_hw_queue(hctx, true);
3037 } else {
3038 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3039 }
Christoph Hellwig0f299da2024-01-24 10:26:56 +01003040 return;
3041
3042queue_exit:
Christoph Hellwig72e84e92024-01-24 10:26:58 +01003043 /*
3044 * Don't drop the queue reference if we were trying to use a cached
3045 * request and thus didn't acquire one.
3046 */
3047 if (!rq)
3048 blk_queue_exit(q);
Jens Axboe320ae512013-10-24 09:20:05 +01003049}
3050
Christoph Hellwig248c7932022-02-15 11:05:36 +01003051#ifdef CONFIG_BLK_MQ_STACKING
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003052/**
Christoph Hellwiga5efda32022-02-15 11:05:37 +01003053 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
Christoph Hellwiga5efda32022-02-15 11:05:37 +01003054 * @rq: the request being queued
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003055 */
Christoph Hellwig28db4712022-02-15 11:05:38 +01003056blk_status_t blk_insert_cloned_request(struct request *rq)
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003057{
Christoph Hellwig28db4712022-02-15 11:05:38 +01003058 struct request_queue *q = rq->q;
John Garry8d1dfd52024-06-20 12:53:50 +00003059 unsigned int max_sectors = blk_queue_get_max_sectors(rq);
Uday Shankar49d24392023-02-28 17:06:55 -07003060 unsigned int max_segments = blk_rq_get_max_segments(rq);
Christoph Hellwiga5efda32022-02-15 11:05:37 +01003061 blk_status_t ret;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003062
3063 if (blk_rq_sectors(rq) > max_sectors) {
3064 /*
3065 * SCSI device does not have a good way to return if
3066 * Write Same/Zero is actually supported. If a device rejects
3067 * a non-read/write command (discard, write same,etc.) the
3068 * low-level device driver will set the relevant queue limit to
3069 * 0 to prevent blk-lib from issuing more of the offending
3070 * operations. Commands queued prior to the queue limit being
3071 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3072 * errors being propagated to upper layers.
3073 */
3074 if (max_sectors == 0)
3075 return BLK_STS_NOTSUPP;
3076
3077 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3078 __func__, blk_rq_sectors(rq), max_sectors);
3079 return BLK_STS_IOERR;
3080 }
3081
3082 /*
3083 * The queue settings related to segment counting may differ from the
3084 * original queue.
3085 */
3086 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
Uday Shankar49d24392023-02-28 17:06:55 -07003087 if (rq->nr_phys_segments > max_segments) {
3088 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3089 __func__, rq->nr_phys_segments, max_segments);
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003090 return BLK_STS_IOERR;
3091 }
3092
Christoph Hellwig28db4712022-02-15 11:05:38 +01003093 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003094 return BLK_STS_IOERR;
3095
Eric Biggers5b8562f2023-03-15 11:39:06 -07003096 ret = blk_crypto_rq_get_keyslot(rq);
3097 if (ret != BLK_STS_OK)
3098 return ret;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003099
3100 blk_account_io_start(rq);
3101
3102 /*
3103 * Since we have a scheduler attached on the top device,
3104 * bypass a potential scheduler on the bottom device for
3105 * insert.
3106 */
Christoph Hellwig28db4712022-02-15 11:05:38 +01003107 blk_mq_run_dispatch_ops(q,
Ming Lei4cafe862021-12-03 21:15:34 +08003108 ret = blk_mq_request_issue_directly(rq, true));
Yu Kuai592ee112022-01-26 09:21:32 +08003109 if (ret)
Jens Axboe08420cf2024-01-15 14:45:07 -07003110 blk_account_io_done(rq, blk_time_get_ns());
Ming Lei4cafe862021-12-03 21:15:34 +08003111 return ret;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003112}
3113EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3114
3115/**
3116 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3117 * @rq: the clone request to be cleaned up
3118 *
3119 * Description:
3120 * Free all bios in @rq for a cloned request.
3121 */
3122void blk_rq_unprep_clone(struct request *rq)
3123{
3124 struct bio *bio;
3125
3126 while ((bio = rq->bio) != NULL) {
3127 rq->bio = bio->bi_next;
3128
3129 bio_put(bio);
3130 }
3131}
3132EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3133
3134/**
3135 * blk_rq_prep_clone - Helper function to setup clone request
3136 * @rq: the request to be setup
3137 * @rq_src: original request to be cloned
3138 * @bs: bio_set that bios for clone are allocated from
3139 * @gfp_mask: memory allocation mask for bio
3140 * @bio_ctr: setup function to be called for each clone bio.
3141 * Returns %0 for success, non %0 for failure.
3142 * @data: private data to be passed to @bio_ctr
3143 *
3144 * Description:
3145 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3146 * Also, pages which the original bios are pointing to are not copied
3147 * and the cloned bios just point same pages.
3148 * So cloned bios must be completed before original bios, which means
3149 * the caller must complete @rq before @rq_src.
3150 */
3151int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3152 struct bio_set *bs, gfp_t gfp_mask,
3153 int (*bio_ctr)(struct bio *, struct bio *, void *),
3154 void *data)
3155{
3156 struct bio *bio, *bio_src;
3157
3158 if (!bs)
3159 bs = &fs_bio_set;
3160
3161 __rq_for_each_bio(bio_src, rq_src) {
Christoph Hellwigabfc4262022-02-02 17:01:09 +01003162 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3163 bs);
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003164 if (!bio)
3165 goto free_and_out;
3166
3167 if (bio_ctr && bio_ctr(bio, bio_src, data))
3168 goto free_and_out;
3169
3170 if (rq->bio) {
3171 rq->biotail->bi_next = bio;
3172 rq->biotail = bio;
3173 } else {
3174 rq->bio = rq->biotail = bio;
3175 }
3176 bio = NULL;
3177 }
3178
3179 /* Copy attributes of the original request to the clone request. */
3180 rq->__sector = blk_rq_pos(rq_src);
3181 rq->__data_len = blk_rq_bytes(rq_src);
3182 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3183 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3184 rq->special_vec = rq_src->special_vec;
3185 }
3186 rq->nr_phys_segments = rq_src->nr_phys_segments;
3187 rq->ioprio = rq_src->ioprio;
Bart Van Assche44981352024-02-02 12:39:25 -08003188 rq->write_hint = rq_src->write_hint;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003189
3190 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3191 goto free_and_out;
3192
3193 return 0;
3194
3195free_and_out:
3196 if (bio)
3197 bio_put(bio);
3198 blk_rq_unprep_clone(rq);
3199
3200 return -ENOMEM;
3201}
3202EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
Christoph Hellwig248c7932022-02-15 11:05:36 +01003203#endif /* CONFIG_BLK_MQ_STACKING */
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003204
Christoph Hellwigf2b8f3c2021-11-17 07:14:00 +01003205/*
3206 * Steal bios from a request and add them to a bio list.
3207 * The request must not have been partially completed before.
3208 */
3209void blk_steal_bios(struct bio_list *list, struct request *rq)
3210{
3211 if (rq->bio) {
3212 if (list->tail)
3213 list->tail->bi_next = rq->bio;
3214 else
3215 list->head = rq->bio;
3216 list->tail = rq->biotail;
3217
3218 rq->bio = NULL;
3219 rq->biotail = NULL;
3220 }
3221
3222 rq->__data_len = 0;
3223}
3224EXPORT_SYMBOL_GPL(blk_steal_bios);
3225
Ming Leibd631412021-05-11 23:22:35 +08003226static size_t order_to_size(unsigned int order)
3227{
3228 return (size_t)PAGE_SIZE << order;
3229}
3230
3231/* called before freeing request pool in @tags */
John Garryf32e4ea2021-10-05 18:23:32 +08003232static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3233 struct blk_mq_tags *tags)
Ming Leibd631412021-05-11 23:22:35 +08003234{
Ming Leibd631412021-05-11 23:22:35 +08003235 struct page *page;
3236 unsigned long flags;
3237
Yu Kuai76dd2982022-10-11 22:22:53 +08003238 /*
3239 * There is no need to clear mapping if driver tags is not initialized
3240 * or the mapping belongs to the driver tags.
3241 */
3242 if (!drv_tags || drv_tags == tags)
John Garry4f245d52021-10-05 18:23:33 +08003243 return;
3244
Ming Leibd631412021-05-11 23:22:35 +08003245 list_for_each_entry(page, &tags->page_list, lru) {
3246 unsigned long start = (unsigned long)page_address(page);
3247 unsigned long end = start + order_to_size(page->private);
3248 int i;
3249
John Garryf32e4ea2021-10-05 18:23:32 +08003250 for (i = 0; i < drv_tags->nr_tags; i++) {
Ming Leibd631412021-05-11 23:22:35 +08003251 struct request *rq = drv_tags->rqs[i];
3252 unsigned long rq_addr = (unsigned long)rq;
3253
3254 if (rq_addr >= start && rq_addr < end) {
Jens Axboe0a467d02021-10-14 14:39:59 -06003255 WARN_ON_ONCE(req_ref_read(rq) != 0);
Ming Leibd631412021-05-11 23:22:35 +08003256 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3257 }
3258 }
3259 }
3260
3261 /*
3262 * Wait until all pending iteration is done.
3263 *
3264 * Request reference is cleared and it is guaranteed to be observed
3265 * after the ->lock is released.
3266 */
3267 spin_lock_irqsave(&drv_tags->lock, flags);
3268 spin_unlock_irqrestore(&drv_tags->lock, flags);
3269}
3270
Jens Axboecc71a6f2017-01-11 14:29:56 -07003271void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3272 unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01003273{
John Garryf32e4ea2021-10-05 18:23:32 +08003274 struct blk_mq_tags *drv_tags;
Jens Axboe320ae512013-10-24 09:20:05 +01003275 struct page *page;
3276
Ming Leie02657e2022-03-08 06:51:48 +01003277 if (list_empty(&tags->page_list))
3278 return;
3279
John Garry079a2e32021-10-05 18:23:39 +08003280 if (blk_mq_is_shared_tags(set->flags))
3281 drv_tags = set->shared_tags;
John Garrye155b0c2021-10-05 18:23:37 +08003282 else
3283 drv_tags = set->tags[hctx_idx];
John Garryf32e4ea2021-10-05 18:23:32 +08003284
John Garry65de57b2021-10-05 18:23:26 +08003285 if (tags->static_rqs && set->ops->exit_request) {
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003286 int i;
3287
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003288 for (i = 0; i < tags->nr_tags; i++) {
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003289 struct request *rq = tags->static_rqs[i];
3290
3291 if (!rq)
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003292 continue;
Christoph Hellwigd6296d392017-05-01 10:19:08 -06003293 set->ops->exit_request(set, rq, hctx_idx);
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003294 tags->static_rqs[i] = NULL;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003295 }
3296 }
3297
John Garryf32e4ea2021-10-05 18:23:32 +08003298 blk_mq_clear_rq_mapping(drv_tags, tags);
Ming Leibd631412021-05-11 23:22:35 +08003299
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003300 while (!list_empty(&tags->page_list)) {
3301 page = list_first_entry(&tags->page_list, struct page, lru);
Dave Hansen67534712014-01-08 20:17:46 -07003302 list_del_init(&page->lru);
Catalin Marinasf75782e2015-09-14 18:16:02 +01003303 /*
3304 * Remove kmemleak object previously allocated in
Raul E Rangel273938b2019-05-02 13:48:11 -06003305 * blk_mq_alloc_rqs().
Catalin Marinasf75782e2015-09-14 18:16:02 +01003306 */
3307 kmemleak_free(page_address(page));
Jens Axboe320ae512013-10-24 09:20:05 +01003308 __free_pages(page, page->private);
3309 }
Jens Axboecc71a6f2017-01-11 14:29:56 -07003310}
Jens Axboe320ae512013-10-24 09:20:05 +01003311
John Garrye155b0c2021-10-05 18:23:37 +08003312void blk_mq_free_rq_map(struct blk_mq_tags *tags)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003313{
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003314 kfree(tags->rqs);
Jens Axboecc71a6f2017-01-11 14:29:56 -07003315 tags->rqs = NULL;
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003316 kfree(tags->static_rqs);
3317 tags->static_rqs = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003318
John Garrye155b0c2021-10-05 18:23:37 +08003319 blk_mq_free_tags(tags);
Jens Axboe320ae512013-10-24 09:20:05 +01003320}
3321
Ming Lei4d805132022-03-08 15:32:14 +08003322static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3323 unsigned int hctx_idx)
3324{
3325 int i;
3326
3327 for (i = 0; i < set->nr_maps; i++) {
3328 unsigned int start = set->map[i].queue_offset;
3329 unsigned int end = start + set->map[i].nr_queues;
3330
3331 if (hctx_idx >= start && hctx_idx < end)
3332 break;
3333 }
3334
3335 if (i >= set->nr_maps)
3336 i = HCTX_TYPE_DEFAULT;
3337
3338 return i;
3339}
3340
3341static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3342 unsigned int hctx_idx)
3343{
3344 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3345
3346 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3347}
3348
John Garry63064be2021-10-05 18:23:35 +08003349static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3350 unsigned int hctx_idx,
3351 unsigned int nr_tags,
John Garrye155b0c2021-10-05 18:23:37 +08003352 unsigned int reserved_tags)
Jens Axboe320ae512013-10-24 09:20:05 +01003353{
Ming Lei4d805132022-03-08 15:32:14 +08003354 int node = blk_mq_get_hctx_node(set, hctx_idx);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003355 struct blk_mq_tags *tags;
Jens Axboe320ae512013-10-24 09:20:05 +01003356
Shaohua Li59f082e2017-02-01 09:53:14 -08003357 if (node == NUMA_NO_NODE)
3358 node = set->numa_node;
3359
John Garrye155b0c2021-10-05 18:23:37 +08003360 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3361 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003362 if (!tags)
3363 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003364
Kees Cook590b5b72018-06-12 14:04:20 -07003365 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
Gabriel Krisman Bertazi36e1f3d12016-12-06 13:31:44 -02003366 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
Shaohua Li59f082e2017-02-01 09:53:14 -08003367 node);
Jinlong Chen7edfd682022-11-02 10:52:29 +08003368 if (!tags->rqs)
3369 goto err_free_tags;
Jens Axboe320ae512013-10-24 09:20:05 +01003370
Kees Cook590b5b72018-06-12 14:04:20 -07003371 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3372 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3373 node);
Jinlong Chen7edfd682022-11-02 10:52:29 +08003374 if (!tags->static_rqs)
3375 goto err_free_rqs;
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003376
Jens Axboecc71a6f2017-01-11 14:29:56 -07003377 return tags;
Jinlong Chen7edfd682022-11-02 10:52:29 +08003378
3379err_free_rqs:
3380 kfree(tags->rqs);
3381err_free_tags:
3382 blk_mq_free_tags(tags);
3383 return NULL;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003384}
3385
Tejun Heo1d9bd512018-01-09 08:29:48 -08003386static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3387 unsigned int hctx_idx, int node)
3388{
3389 int ret;
3390
3391 if (set->ops->init_request) {
3392 ret = set->ops->init_request(set, rq, hctx_idx, node);
3393 if (ret)
3394 return ret;
3395 }
3396
Keith Busch12f5b932018-05-29 15:52:28 +02003397 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Tejun Heo1d9bd512018-01-09 08:29:48 -08003398 return 0;
3399}
3400
John Garry63064be2021-10-05 18:23:35 +08003401static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3402 struct blk_mq_tags *tags,
3403 unsigned int hctx_idx, unsigned int depth)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003404{
3405 unsigned int i, j, entries_per_page, max_order = 4;
Ming Lei4d805132022-03-08 15:32:14 +08003406 int node = blk_mq_get_hctx_node(set, hctx_idx);
Jens Axboecc71a6f2017-01-11 14:29:56 -07003407 size_t rq_size, left;
Shaohua Li59f082e2017-02-01 09:53:14 -08003408
Shaohua Li59f082e2017-02-01 09:53:14 -08003409 if (node == NUMA_NO_NODE)
3410 node = set->numa_node;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003411
3412 INIT_LIST_HEAD(&tags->page_list);
3413
Jens Axboe320ae512013-10-24 09:20:05 +01003414 /*
3415 * rq_size is the size of the request plus driver payload, rounded
3416 * to the cacheline size
3417 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003418 rq_size = round_up(sizeof(struct request) + set->cmd_size,
Jens Axboe320ae512013-10-24 09:20:05 +01003419 cache_line_size());
Jens Axboecc71a6f2017-01-11 14:29:56 -07003420 left = rq_size * depth;
Jens Axboe320ae512013-10-24 09:20:05 +01003421
Jens Axboecc71a6f2017-01-11 14:29:56 -07003422 for (i = 0; i < depth; ) {
Jens Axboe320ae512013-10-24 09:20:05 +01003423 int this_order = max_order;
3424 struct page *page;
3425 int to_do;
3426 void *p;
3427
Bartlomiej Zolnierkiewiczb3a834b2016-05-16 09:54:47 -06003428 while (this_order && left < order_to_size(this_order - 1))
Jens Axboe320ae512013-10-24 09:20:05 +01003429 this_order--;
3430
3431 do {
Shaohua Li59f082e2017-02-01 09:53:14 -08003432 page = alloc_pages_node(node,
Gabriel Krisman Bertazi36e1f3d12016-12-06 13:31:44 -02003433 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
Jens Axboea5164402014-09-10 09:02:03 -06003434 this_order);
Jens Axboe320ae512013-10-24 09:20:05 +01003435 if (page)
3436 break;
3437 if (!this_order--)
3438 break;
3439 if (order_to_size(this_order) < rq_size)
3440 break;
3441 } while (1);
3442
3443 if (!page)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003444 goto fail;
Jens Axboe320ae512013-10-24 09:20:05 +01003445
3446 page->private = this_order;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003447 list_add_tail(&page->lru, &tags->page_list);
Jens Axboe320ae512013-10-24 09:20:05 +01003448
3449 p = page_address(page);
Catalin Marinasf75782e2015-09-14 18:16:02 +01003450 /*
3451 * Allow kmemleak to scan these pages as they contain pointers
3452 * to additional allocations like via ops->init_request().
3453 */
Gabriel Krisman Bertazi36e1f3d12016-12-06 13:31:44 -02003454 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
Jens Axboe320ae512013-10-24 09:20:05 +01003455 entries_per_page = order_to_size(this_order) / rq_size;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003456 to_do = min(entries_per_page, depth - i);
Jens Axboe320ae512013-10-24 09:20:05 +01003457 left -= to_do * rq_size;
3458 for (j = 0; j < to_do; j++) {
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003459 struct request *rq = p;
3460
3461 tags->static_rqs[i] = rq;
Tejun Heo1d9bd512018-01-09 08:29:48 -08003462 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3463 tags->static_rqs[i] = NULL;
3464 goto fail;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003465 }
3466
Jens Axboe320ae512013-10-24 09:20:05 +01003467 p += rq_size;
3468 i++;
3469 }
3470 }
Jens Axboecc71a6f2017-01-11 14:29:56 -07003471 return 0;
Jens Axboe320ae512013-10-24 09:20:05 +01003472
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003473fail:
Jens Axboecc71a6f2017-01-11 14:29:56 -07003474 blk_mq_free_rqs(set, tags, hctx_idx);
3475 return -ENOMEM;
Jens Axboe320ae512013-10-24 09:20:05 +01003476}
3477
Ming Leibf0beec2020-05-29 15:53:15 +02003478struct rq_iter_data {
3479 struct blk_mq_hw_ctx *hctx;
3480 bool has_rq;
3481};
3482
John Garry2dd65322022-07-06 20:03:53 +08003483static bool blk_mq_has_request(struct request *rq, void *data)
Ming Leibf0beec2020-05-29 15:53:15 +02003484{
3485 struct rq_iter_data *iter_data = data;
3486
3487 if (rq->mq_hctx != iter_data->hctx)
3488 return true;
3489 iter_data->has_rq = true;
3490 return false;
3491}
3492
3493static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3494{
3495 struct blk_mq_tags *tags = hctx->sched_tags ?
3496 hctx->sched_tags : hctx->tags;
3497 struct rq_iter_data data = {
3498 .hctx = hctx,
3499 };
3500
3501 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3502 return data.has_rq;
3503}
3504
Ming Leia46c2702024-03-22 10:12:44 +08003505static bool blk_mq_hctx_has_online_cpu(struct blk_mq_hw_ctx *hctx,
3506 unsigned int this_cpu)
Ming Leibf0beec2020-05-29 15:53:15 +02003507{
Ming Leia46c2702024-03-22 10:12:44 +08003508 enum hctx_type type = hctx->type;
3509 int cpu;
3510
3511 /*
3512 * hctx->cpumask has to rule out isolated CPUs, but userspace still
3513 * might submit IOs on these isolated CPUs, so use the queue map to
3514 * check if all CPUs mapped to this hctx are offline
3515 */
3516 for_each_online_cpu(cpu) {
3517 struct blk_mq_hw_ctx *h = blk_mq_map_queue_type(hctx->queue,
3518 type, cpu);
3519
3520 if (h != hctx)
3521 continue;
3522
3523 /* this hctx has at least one online CPU */
3524 if (this_cpu != cpu)
3525 return true;
3526 }
3527
3528 return false;
Ming Leibf0beec2020-05-29 15:53:15 +02003529}
3530
3531static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3532{
3533 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3534 struct blk_mq_hw_ctx, cpuhp_online);
3535
Ming Leia46c2702024-03-22 10:12:44 +08003536 if (blk_mq_hctx_has_online_cpu(hctx, cpu))
Ming Leibf0beec2020-05-29 15:53:15 +02003537 return 0;
3538
3539 /*
3540 * Prevent new request from being allocated on the current hctx.
3541 *
3542 * The smp_mb__after_atomic() Pairs with the implied barrier in
3543 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3544 * seen once we return from the tag allocator.
3545 */
3546 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3547 smp_mb__after_atomic();
3548
3549 /*
3550 * Try to grab a reference to the queue and wait for any outstanding
3551 * requests. If we could not grab a reference the queue has been
3552 * frozen and there are no requests.
3553 */
3554 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3555 while (blk_mq_hctx_has_requests(hctx))
3556 msleep(5);
3557 percpu_ref_put(&hctx->queue->q_usage_counter);
3558 }
3559
3560 return 0;
3561}
3562
Ming Lei7b815812024-05-17 10:05:14 +08003563/*
3564 * Check if one CPU is mapped to the specified hctx
3565 *
3566 * Isolated CPUs have been ruled out from hctx->cpumask, which is supposed
3567 * to be used for scheduling kworker only. For other usage, please call this
3568 * helper for checking if one CPU belongs to the specified hctx
3569 */
3570static bool blk_mq_cpu_mapped_to_hctx(unsigned int cpu,
3571 const struct blk_mq_hw_ctx *hctx)
3572{
3573 struct blk_mq_hw_ctx *mapped_hctx = blk_mq_map_queue_type(hctx->queue,
3574 hctx->type, cpu);
3575
3576 return mapped_hctx == hctx;
3577}
3578
Ming Leibf0beec2020-05-29 15:53:15 +02003579static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3580{
3581 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3582 struct blk_mq_hw_ctx, cpuhp_online);
3583
Ming Lei7b815812024-05-17 10:05:14 +08003584 if (blk_mq_cpu_mapped_to_hctx(cpu, hctx))
Ming Leibf0beec2020-05-29 15:53:15 +02003585 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3586 return 0;
3587}
3588
Jens Axboee57690f2016-08-24 15:34:35 -06003589/*
3590 * 'cpu' is going away. splice any existing rq_list entries from this
3591 * software queue to the hw queue dispatch list, and ensure that it
3592 * gets run.
3593 */
Thomas Gleixner9467f852016-09-22 08:05:17 -06003594static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
Jens Axboe484b4062014-05-21 14:01:15 -06003595{
Thomas Gleixner9467f852016-09-22 08:05:17 -06003596 struct blk_mq_hw_ctx *hctx;
Jens Axboe484b4062014-05-21 14:01:15 -06003597 struct blk_mq_ctx *ctx;
3598 LIST_HEAD(tmp);
Ming Leic16d6b52018-12-17 08:44:05 -07003599 enum hctx_type type;
Jens Axboe484b4062014-05-21 14:01:15 -06003600
Thomas Gleixner9467f852016-09-22 08:05:17 -06003601 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
Ming Lei7b815812024-05-17 10:05:14 +08003602 if (!blk_mq_cpu_mapped_to_hctx(cpu, hctx))
Ming Leibf0beec2020-05-29 15:53:15 +02003603 return 0;
3604
Jens Axboee57690f2016-08-24 15:34:35 -06003605 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
Ming Leic16d6b52018-12-17 08:44:05 -07003606 type = hctx->type;
Jens Axboe484b4062014-05-21 14:01:15 -06003607
3608 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07003609 if (!list_empty(&ctx->rq_lists[type])) {
3610 list_splice_init(&ctx->rq_lists[type], &tmp);
Jens Axboe484b4062014-05-21 14:01:15 -06003611 blk_mq_hctx_clear_pending(hctx, ctx);
3612 }
3613 spin_unlock(&ctx->lock);
3614
3615 if (list_empty(&tmp))
Thomas Gleixner9467f852016-09-22 08:05:17 -06003616 return 0;
Jens Axboe484b4062014-05-21 14:01:15 -06003617
Jens Axboee57690f2016-08-24 15:34:35 -06003618 spin_lock(&hctx->lock);
3619 list_splice_tail_init(&tmp, &hctx->dispatch);
3620 spin_unlock(&hctx->lock);
Jens Axboe484b4062014-05-21 14:01:15 -06003621
3622 blk_mq_run_hw_queue(hctx, true);
Thomas Gleixner9467f852016-09-22 08:05:17 -06003623 return 0;
Jens Axboe484b4062014-05-21 14:01:15 -06003624}
3625
Thomas Gleixner9467f852016-09-22 08:05:17 -06003626static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
Jens Axboe484b4062014-05-21 14:01:15 -06003627{
Ming Leibf0beec2020-05-29 15:53:15 +02003628 if (!(hctx->flags & BLK_MQ_F_STACKING))
3629 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3630 &hctx->cpuhp_online);
Thomas Gleixner9467f852016-09-22 08:05:17 -06003631 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3632 &hctx->cpuhp_dead);
Jens Axboe484b4062014-05-21 14:01:15 -06003633}
3634
Ming Lei364b6182021-05-11 23:22:36 +08003635/*
3636 * Before freeing hw queue, clearing the flush request reference in
3637 * tags->rqs[] for avoiding potential UAF.
3638 */
3639static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3640 unsigned int queue_depth, struct request *flush_rq)
3641{
3642 int i;
3643 unsigned long flags;
3644
3645 /* The hw queue may not be mapped yet */
3646 if (!tags)
3647 return;
3648
Jens Axboe0a467d02021-10-14 14:39:59 -06003649 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
Ming Lei364b6182021-05-11 23:22:36 +08003650
3651 for (i = 0; i < queue_depth; i++)
3652 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3653
3654 /*
3655 * Wait until all pending iteration is done.
3656 *
3657 * Request reference is cleared and it is guaranteed to be observed
3658 * after the ->lock is released.
3659 */
3660 spin_lock_irqsave(&tags->lock, flags);
3661 spin_unlock_irqrestore(&tags->lock, flags);
3662}
3663
Ming Leic3b4afc2015-06-04 22:25:04 +08003664/* hctx->ctxs will be freed in queue's release handler */
Ming Lei08e98fc2014-09-25 23:23:38 +08003665static void blk_mq_exit_hctx(struct request_queue *q,
3666 struct blk_mq_tag_set *set,
3667 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3668{
Ming Lei364b6182021-05-11 23:22:36 +08003669 struct request *flush_rq = hctx->fq->flush_rq;
3670
Ming Lei8ab0b7d2018-01-09 21:28:29 +08003671 if (blk_mq_hw_queue_mapped(hctx))
3672 blk_mq_tag_idle(hctx);
Ming Lei08e98fc2014-09-25 23:23:38 +08003673
Ming Lei6cfeadb2022-06-16 09:44:01 +08003674 if (blk_queue_init_done(q))
3675 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3676 set->queue_depth, flush_rq);
Ming Leif70ced02014-09-25 23:23:47 +08003677 if (set->ops->exit_request)
Ming Lei364b6182021-05-11 23:22:36 +08003678 set->ops->exit_request(set, flush_rq, hctx_idx);
Ming Leif70ced02014-09-25 23:23:47 +08003679
Ming Lei08e98fc2014-09-25 23:23:38 +08003680 if (set->ops->exit_hctx)
3681 set->ops->exit_hctx(hctx, hctx_idx);
3682
Thomas Gleixner9467f852016-09-22 08:05:17 -06003683 blk_mq_remove_cpuhp(hctx);
Ming Lei2f8f1332019-04-30 09:52:27 +08003684
Ming Lei4e5cc992022-03-08 15:32:19 +08003685 xa_erase(&q->hctx_table, hctx_idx);
3686
Ming Lei2f8f1332019-04-30 09:52:27 +08003687 spin_lock(&q->unused_hctx_lock);
3688 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3689 spin_unlock(&q->unused_hctx_lock);
Ming Lei08e98fc2014-09-25 23:23:38 +08003690}
3691
Ming Lei624dbe42014-05-27 23:35:13 +08003692static void blk_mq_exit_hw_queues(struct request_queue *q,
3693 struct blk_mq_tag_set *set, int nr_queue)
3694{
3695 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08003696 unsigned long i;
Ming Lei624dbe42014-05-27 23:35:13 +08003697
3698 queue_for_each_hw_ctx(q, hctx, i) {
3699 if (i == nr_queue)
3700 break;
Ming Lei08e98fc2014-09-25 23:23:38 +08003701 blk_mq_exit_hctx(q, set, hctx, i);
Ming Lei624dbe42014-05-27 23:35:13 +08003702 }
Ming Lei624dbe42014-05-27 23:35:13 +08003703}
3704
Ming Lei08e98fc2014-09-25 23:23:38 +08003705static int blk_mq_init_hctx(struct request_queue *q,
3706 struct blk_mq_tag_set *set,
3707 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3708{
Ming Lei7c6c5b72019-04-30 09:52:26 +08003709 hctx->queue_num = hctx_idx;
Ming Lei08e98fc2014-09-25 23:23:38 +08003710
Ming Leibf0beec2020-05-29 15:53:15 +02003711 if (!(hctx->flags & BLK_MQ_F_STACKING))
3712 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3713 &hctx->cpuhp_online);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003714 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3715
3716 hctx->tags = set->tags[hctx_idx];
3717
3718 if (set->ops->init_hctx &&
3719 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3720 goto unregister_cpu_notifier;
3721
3722 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3723 hctx->numa_node))
3724 goto exit_hctx;
Ming Lei4e5cc992022-03-08 15:32:19 +08003725
3726 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3727 goto exit_flush_rq;
3728
Ming Lei7c6c5b72019-04-30 09:52:26 +08003729 return 0;
3730
Ming Lei4e5cc992022-03-08 15:32:19 +08003731 exit_flush_rq:
3732 if (set->ops->exit_request)
3733 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003734 exit_hctx:
3735 if (set->ops->exit_hctx)
3736 set->ops->exit_hctx(hctx, hctx_idx);
3737 unregister_cpu_notifier:
3738 blk_mq_remove_cpuhp(hctx);
3739 return -1;
3740}
3741
3742static struct blk_mq_hw_ctx *
3743blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3744 int node)
3745{
3746 struct blk_mq_hw_ctx *hctx;
3747 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3748
Ming Lei704b9142021-12-03 21:15:32 +08003749 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003750 if (!hctx)
3751 goto fail_alloc_hctx;
3752
3753 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3754 goto free_hctx;
3755
3756 atomic_set(&hctx->nr_active, 0);
Ming Lei08e98fc2014-09-25 23:23:38 +08003757 if (node == NUMA_NO_NODE)
Ming Lei7c6c5b72019-04-30 09:52:26 +08003758 node = set->numa_node;
3759 hctx->numa_node = node;
Ming Lei08e98fc2014-09-25 23:23:38 +08003760
Jens Axboe9f993732017-04-10 09:54:54 -06003761 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
Ming Lei08e98fc2014-09-25 23:23:38 +08003762 spin_lock_init(&hctx->lock);
3763 INIT_LIST_HEAD(&hctx->dispatch);
3764 hctx->queue = q;
Ming Lei51db1c32020-08-19 23:20:19 +08003765 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
Ming Lei08e98fc2014-09-25 23:23:38 +08003766
Ming Lei2f8f1332019-04-30 09:52:27 +08003767 INIT_LIST_HEAD(&hctx->hctx_list);
3768
Ming Lei08e98fc2014-09-25 23:23:38 +08003769 /*
3770 * Allocate space for all possible cpus to avoid allocation at
3771 * runtime
3772 */
Johannes Thumshirnd904bfa2017-11-15 17:32:33 -08003773 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
Ming Lei7c6c5b72019-04-30 09:52:26 +08003774 gfp, node);
Ming Lei08e98fc2014-09-25 23:23:38 +08003775 if (!hctx->ctxs)
Ming Lei7c6c5b72019-04-30 09:52:26 +08003776 goto free_cpumask;
Ming Lei08e98fc2014-09-25 23:23:38 +08003777
Jianchao Wang5b202852018-10-12 18:07:26 +08003778 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
Ming Leic548e622021-01-22 10:33:08 +08003779 gfp, node, false, false))
Ming Lei08e98fc2014-09-25 23:23:38 +08003780 goto free_ctxs;
Ming Lei08e98fc2014-09-25 23:23:38 +08003781 hctx->nr_ctx = 0;
3782
Ming Lei5815839b2018-06-25 19:31:47 +08003783 spin_lock_init(&hctx->dispatch_wait_lock);
Jens Axboeeb619fd2017-11-09 08:32:43 -07003784 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3785 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3786
Guoqing Jiang754a1572020-03-09 22:41:37 +01003787 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
Ming Leif70ced02014-09-25 23:23:47 +08003788 if (!hctx->fq)
Ming Lei7c6c5b72019-04-30 09:52:26 +08003789 goto free_bitmap;
Ming Leif70ced02014-09-25 23:23:47 +08003790
Ming Lei7c6c5b72019-04-30 09:52:26 +08003791 blk_mq_hctx_kobj_init(hctx);
Bart Van Assche6a83e742016-11-02 10:09:51 -06003792
Ming Lei7c6c5b72019-04-30 09:52:26 +08003793 return hctx;
Ming Lei08e98fc2014-09-25 23:23:38 +08003794
3795 free_bitmap:
Omar Sandoval88459642016-09-17 08:38:44 -06003796 sbitmap_free(&hctx->ctx_map);
Ming Lei08e98fc2014-09-25 23:23:38 +08003797 free_ctxs:
3798 kfree(hctx->ctxs);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003799 free_cpumask:
3800 free_cpumask_var(hctx->cpumask);
3801 free_hctx:
3802 kfree(hctx);
3803 fail_alloc_hctx:
3804 return NULL;
Ming Lei08e98fc2014-09-25 23:23:38 +08003805}
3806
Jens Axboe320ae512013-10-24 09:20:05 +01003807static void blk_mq_init_cpu_queues(struct request_queue *q,
3808 unsigned int nr_hw_queues)
3809{
Jens Axboeb3c661b2018-10-30 10:36:06 -06003810 struct blk_mq_tag_set *set = q->tag_set;
3811 unsigned int i, j;
Jens Axboe320ae512013-10-24 09:20:05 +01003812
3813 for_each_possible_cpu(i) {
3814 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3815 struct blk_mq_hw_ctx *hctx;
Ming Leic16d6b52018-12-17 08:44:05 -07003816 int k;
Jens Axboe320ae512013-10-24 09:20:05 +01003817
Jens Axboe320ae512013-10-24 09:20:05 +01003818 __ctx->cpu = i;
3819 spin_lock_init(&__ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07003820 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3821 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3822
Jens Axboe320ae512013-10-24 09:20:05 +01003823 __ctx->queue = q;
3824
Jens Axboe320ae512013-10-24 09:20:05 +01003825 /*
3826 * Set local node, IFF we have more than one hw queue. If
3827 * not, we remain on the home node of the device
3828 */
Jens Axboeb3c661b2018-10-30 10:36:06 -06003829 for (j = 0; j < set->nr_maps; j++) {
3830 hctx = blk_mq_map_queue_type(q, j, i);
3831 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
Xianting Tian576e85c2020-10-19 16:20:47 +08003832 hctx->numa_node = cpu_to_node(i);
Jens Axboeb3c661b2018-10-30 10:36:06 -06003833 }
Jens Axboe320ae512013-10-24 09:20:05 +01003834 }
3835}
3836
John Garry63064be2021-10-05 18:23:35 +08003837struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3838 unsigned int hctx_idx,
3839 unsigned int depth)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003840{
John Garry63064be2021-10-05 18:23:35 +08003841 struct blk_mq_tags *tags;
3842 int ret;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003843
John Garrye155b0c2021-10-05 18:23:37 +08003844 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
John Garry63064be2021-10-05 18:23:35 +08003845 if (!tags)
3846 return NULL;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003847
John Garry63064be2021-10-05 18:23:35 +08003848 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3849 if (ret) {
John Garrye155b0c2021-10-05 18:23:37 +08003850 blk_mq_free_rq_map(tags);
John Garry63064be2021-10-05 18:23:35 +08003851 return NULL;
3852 }
Jens Axboecc71a6f2017-01-11 14:29:56 -07003853
John Garry63064be2021-10-05 18:23:35 +08003854 return tags;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003855}
3856
John Garry63064be2021-10-05 18:23:35 +08003857static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3858 int hctx_idx)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003859{
John Garry079a2e32021-10-05 18:23:39 +08003860 if (blk_mq_is_shared_tags(set->flags)) {
3861 set->tags[hctx_idx] = set->shared_tags;
John Garry1c0706a2020-08-19 23:20:22 +08003862
John Garrye155b0c2021-10-05 18:23:37 +08003863 return true;
Jens Axboebd166ef2017-01-17 06:03:22 -07003864 }
John Garrye155b0c2021-10-05 18:23:37 +08003865
John Garry63064be2021-10-05 18:23:35 +08003866 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3867 set->queue_depth);
3868
3869 return set->tags[hctx_idx];
Jens Axboecc71a6f2017-01-11 14:29:56 -07003870}
3871
John Garry645db342021-10-05 18:23:36 +08003872void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3873 struct blk_mq_tags *tags,
3874 unsigned int hctx_idx)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003875{
John Garry645db342021-10-05 18:23:36 +08003876 if (tags) {
3877 blk_mq_free_rqs(set, tags, hctx_idx);
John Garrye155b0c2021-10-05 18:23:37 +08003878 blk_mq_free_rq_map(tags);
Jens Axboecc71a6f2017-01-11 14:29:56 -07003879 }
3880}
3881
John Garrye155b0c2021-10-05 18:23:37 +08003882static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3883 unsigned int hctx_idx)
3884{
John Garry079a2e32021-10-05 18:23:39 +08003885 if (!blk_mq_is_shared_tags(set->flags))
John Garrye155b0c2021-10-05 18:23:37 +08003886 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3887
3888 set->tags[hctx_idx] = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003889}
3890
Christoph Hellwig4b855ad2017-06-26 12:20:57 +02003891static void blk_mq_map_swqueue(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01003892{
Ming Lei4f481202022-03-08 15:32:18 +08003893 unsigned int j, hctx_idx;
3894 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01003895 struct blk_mq_hw_ctx *hctx;
3896 struct blk_mq_ctx *ctx;
Ming Lei2a34c082015-04-21 10:00:20 +08003897 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01003898
3899 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboee4043dc2014-04-09 10:18:23 -06003900 cpumask_clear(hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01003901 hctx->nr_ctx = 0;
huhaid416c922018-05-18 08:32:30 -06003902 hctx->dispatch_from = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003903 }
3904
3905 /*
Christoph Hellwig4b855ad2017-06-26 12:20:57 +02003906 * Map software to hardware queues.
Ming Lei4412efe2018-04-25 04:01:44 +08003907 *
3908 * If the cpu isn't present, the cpu is mapped to first hctx.
Jens Axboe320ae512013-10-24 09:20:05 +01003909 */
Christoph Hellwig20e4d8132018-01-12 10:53:06 +08003910 for_each_possible_cpu(i) {
Ming Lei4412efe2018-04-25 04:01:44 +08003911
Thomas Gleixner897bb0c2016-03-19 11:30:33 +01003912 ctx = per_cpu_ptr(q->queue_ctx, i);
Jens Axboeb3c661b2018-10-30 10:36:06 -06003913 for (j = 0; j < set->nr_maps; j++) {
Jianchao Wangbb94aea2019-01-24 18:25:33 +08003914 if (!set->map[j].nr_queues) {
3915 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3916 HCTX_TYPE_DEFAULT, i);
Ming Leie5edd5f2018-12-18 01:28:56 +08003917 continue;
Jianchao Wangbb94aea2019-01-24 18:25:33 +08003918 }
Ming Leifd689872020-05-07 21:04:08 +08003919 hctx_idx = set->map[j].mq_map[i];
3920 /* unmapped hw queue can be remapped after CPU topo changed */
3921 if (!set->tags[hctx_idx] &&
John Garry63064be2021-10-05 18:23:35 +08003922 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
Ming Leifd689872020-05-07 21:04:08 +08003923 /*
3924 * If tags initialization fail for some hctx,
3925 * that hctx won't be brought online. In this
3926 * case, remap the current ctx to hctx[0] which
3927 * is guaranteed to always have tags allocated
3928 */
3929 set->map[j].mq_map[i] = 0;
3930 }
Ming Leie5edd5f2018-12-18 01:28:56 +08003931
Jens Axboeb3c661b2018-10-30 10:36:06 -06003932 hctx = blk_mq_map_queue_type(q, j, i);
Jianchao Wang8ccdf4a2019-01-24 18:25:32 +08003933 ctx->hctxs[j] = hctx;
Jens Axboeb3c661b2018-10-30 10:36:06 -06003934 /*
3935 * If the CPU is already set in the mask, then we've
3936 * mapped this one already. This can happen if
3937 * devices share queues across queue maps.
3938 */
3939 if (cpumask_test_cpu(i, hctx->cpumask))
3940 continue;
3941
3942 cpumask_set_cpu(i, hctx->cpumask);
3943 hctx->type = j;
3944 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3945 hctx->ctxs[hctx->nr_ctx++] = ctx;
3946
3947 /*
3948 * If the nr_ctx type overflows, we have exceeded the
3949 * amount of sw queues we can support.
3950 */
3951 BUG_ON(!hctx->nr_ctx);
3952 }
Jianchao Wangbb94aea2019-01-24 18:25:33 +08003953
3954 for (; j < HCTX_MAX_TYPES; j++)
3955 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3956 HCTX_TYPE_DEFAULT, i);
Jens Axboe320ae512013-10-24 09:20:05 +01003957 }
Jens Axboe506e9312014-05-07 10:26:44 -06003958
3959 queue_for_each_hw_ctx(q, hctx, i) {
Ming Leia46c2702024-03-22 10:12:44 +08003960 int cpu;
3961
Ming Lei4412efe2018-04-25 04:01:44 +08003962 /*
3963 * If no software queues are mapped to this hardware queue,
3964 * disable it and free the request entries.
3965 */
3966 if (!hctx->nr_ctx) {
3967 /* Never unmap queue 0. We need it as a
3968 * fallback in case of a new remap fails
3969 * allocation
3970 */
John Garrye155b0c2021-10-05 18:23:37 +08003971 if (i)
3972 __blk_mq_free_map_and_rqs(set, i);
Ming Lei4412efe2018-04-25 04:01:44 +08003973
3974 hctx->tags = NULL;
3975 continue;
3976 }
Jens Axboe484b4062014-05-21 14:01:15 -06003977
Ming Lei2a34c082015-04-21 10:00:20 +08003978 hctx->tags = set->tags[i];
3979 WARN_ON(!hctx->tags);
3980
Jens Axboe484b4062014-05-21 14:01:15 -06003981 /*
Chong Yuan889fa312015-04-15 11:39:29 -06003982 * Set the map size to the number of mapped software queues.
3983 * This is more accurate and more efficient than looping
3984 * over all possibly mapped software queues.
3985 */
Omar Sandoval88459642016-09-17 08:38:44 -06003986 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
Chong Yuan889fa312015-04-15 11:39:29 -06003987
3988 /*
Ming Leia46c2702024-03-22 10:12:44 +08003989 * Rule out isolated CPUs from hctx->cpumask to avoid
3990 * running block kworker on isolated CPUs
3991 */
3992 for_each_cpu(cpu, hctx->cpumask) {
3993 if (cpu_is_isolated(cpu))
3994 cpumask_clear_cpu(cpu, hctx->cpumask);
3995 }
3996
3997 /*
Jens Axboe484b4062014-05-21 14:01:15 -06003998 * Initialize batch roundrobin counts
3999 */
Ming Leif82ddf12018-04-08 17:48:10 +08004000 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
Jens Axboe506e9312014-05-07 10:26:44 -06004001 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
4002 }
Jens Axboe320ae512013-10-24 09:20:05 +01004003}
4004
Jens Axboe8e8320c2017-06-20 17:56:13 -06004005/*
4006 * Caller needs to ensure that we're either frozen/quiesced, or that
4007 * the queue isn't live yet.
4008 */
Jeff Moyer2404e602015-11-03 10:40:06 -05004009static void queue_set_hctx_shared(struct request_queue *q, bool shared)
Jens Axboe0d2602c2014-05-13 15:10:52 -06004010{
4011 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08004012 unsigned long i;
Jens Axboe0d2602c2014-05-13 15:10:52 -06004013
Jeff Moyer2404e602015-11-03 10:40:06 -05004014 queue_for_each_hw_ctx(q, hctx, i) {
Yu Kuai454bb672021-07-31 14:21:30 +08004015 if (shared) {
Ming Lei51db1c32020-08-19 23:20:19 +08004016 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
Yu Kuai454bb672021-07-31 14:21:30 +08004017 } else {
4018 blk_mq_tag_idle(hctx);
Ming Lei51db1c32020-08-19 23:20:19 +08004019 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
Yu Kuai454bb672021-07-31 14:21:30 +08004020 }
Jeff Moyer2404e602015-11-03 10:40:06 -05004021 }
4022}
4023
Hannes Reinecke655ac302020-08-19 23:20:20 +08004024static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
4025 bool shared)
Jeff Moyer2404e602015-11-03 10:40:06 -05004026{
4027 struct request_queue *q;
Jens Axboe0d2602c2014-05-13 15:10:52 -06004028
Bart Van Assche705cda92017-04-07 11:16:49 -07004029 lockdep_assert_held(&set->tag_list_lock);
4030
Jens Axboe0d2602c2014-05-13 15:10:52 -06004031 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4032 blk_mq_freeze_queue(q);
Jeff Moyer2404e602015-11-03 10:40:06 -05004033 queue_set_hctx_shared(q, shared);
Jens Axboe0d2602c2014-05-13 15:10:52 -06004034 blk_mq_unfreeze_queue(q);
4035 }
4036}
4037
4038static void blk_mq_del_queue_tag_set(struct request_queue *q)
4039{
4040 struct blk_mq_tag_set *set = q->tag_set;
4041
Jens Axboe0d2602c2014-05-13 15:10:52 -06004042 mutex_lock(&set->tag_list_lock);
Daniel Wagner08c875c2020-07-28 15:29:51 +02004043 list_del(&q->tag_set_list);
Jeff Moyer2404e602015-11-03 10:40:06 -05004044 if (list_is_singular(&set->tag_list)) {
4045 /* just transitioned to unshared */
Ming Lei51db1c32020-08-19 23:20:19 +08004046 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
Jeff Moyer2404e602015-11-03 10:40:06 -05004047 /* update existing queue */
Hannes Reinecke655ac302020-08-19 23:20:20 +08004048 blk_mq_update_tag_set_shared(set, false);
Jeff Moyer2404e602015-11-03 10:40:06 -05004049 }
Jens Axboe0d2602c2014-05-13 15:10:52 -06004050 mutex_unlock(&set->tag_list_lock);
Roman Pena347c7a2018-06-10 22:38:24 +02004051 INIT_LIST_HEAD(&q->tag_set_list);
Jens Axboe0d2602c2014-05-13 15:10:52 -06004052}
4053
4054static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
4055 struct request_queue *q)
4056{
Jens Axboe0d2602c2014-05-13 15:10:52 -06004057 mutex_lock(&set->tag_list_lock);
Jeff Moyer2404e602015-11-03 10:40:06 -05004058
Jens Axboeff821d22017-11-10 22:05:12 -07004059 /*
4060 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4061 */
4062 if (!list_empty(&set->tag_list) &&
Ming Lei51db1c32020-08-19 23:20:19 +08004063 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
4064 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
Jeff Moyer2404e602015-11-03 10:40:06 -05004065 /* update existing queue */
Hannes Reinecke655ac302020-08-19 23:20:20 +08004066 blk_mq_update_tag_set_shared(set, true);
Jeff Moyer2404e602015-11-03 10:40:06 -05004067 }
Ming Lei51db1c32020-08-19 23:20:19 +08004068 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
Jeff Moyer2404e602015-11-03 10:40:06 -05004069 queue_set_hctx_shared(q, true);
Daniel Wagner08c875c2020-07-28 15:29:51 +02004070 list_add_tail(&q->tag_set_list, &set->tag_list);
Jeff Moyer2404e602015-11-03 10:40:06 -05004071
Jens Axboe0d2602c2014-05-13 15:10:52 -06004072 mutex_unlock(&set->tag_list_lock);
4073}
4074
Ming Lei1db49092018-11-20 09:44:35 +08004075/* All allocations will be freed in release handler of q->mq_kobj */
4076static int blk_mq_alloc_ctxs(struct request_queue *q)
4077{
4078 struct blk_mq_ctxs *ctxs;
4079 int cpu;
4080
4081 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4082 if (!ctxs)
4083 return -ENOMEM;
4084
4085 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4086 if (!ctxs->queue_ctx)
4087 goto fail;
4088
4089 for_each_possible_cpu(cpu) {
4090 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4091 ctx->ctxs = ctxs;
4092 }
4093
4094 q->mq_kobj = &ctxs->kobj;
4095 q->queue_ctx = ctxs->queue_ctx;
4096
4097 return 0;
4098 fail:
4099 kfree(ctxs);
4100 return -ENOMEM;
4101}
4102
Ming Leie09aae72015-01-29 20:17:27 +08004103/*
4104 * It is the actual release handler for mq, but we do it from
4105 * request queue's release handler for avoiding use-after-free
4106 * and headache because q->mq_kobj shouldn't have been introduced,
4107 * but we can't group ctx/kctx kobj without it.
4108 */
4109void blk_mq_release(struct request_queue *q)
4110{
Ming Lei2f8f1332019-04-30 09:52:27 +08004111 struct blk_mq_hw_ctx *hctx, *next;
Ming Lei4f481202022-03-08 15:32:18 +08004112 unsigned long i;
Ming Leie09aae72015-01-29 20:17:27 +08004113
Ming Lei2f8f1332019-04-30 09:52:27 +08004114 queue_for_each_hw_ctx(q, hctx, i)
4115 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4116
4117 /* all hctx are in .unused_hctx_list now */
4118 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4119 list_del_init(&hctx->hctx_list);
Ming Lei6c8b2322017-02-22 18:14:01 +08004120 kobject_put(&hctx->kobj);
Ming Leic3b4afc2015-06-04 22:25:04 +08004121 }
Ming Leie09aae72015-01-29 20:17:27 +08004122
Ming Lei4e5cc992022-03-08 15:32:19 +08004123 xa_destroy(&q->hctx_table);
Ming Leie09aae72015-01-29 20:17:27 +08004124
Ming Lei7ea5fe32017-02-22 18:14:00 +08004125 /*
4126 * release .mq_kobj and sw queue's kobject now because
4127 * both share lifetime with request queue.
4128 */
4129 blk_mq_sysfs_deinit(q);
Ming Leie09aae72015-01-29 20:17:27 +08004130}
4131
Christoph Hellwig8023e142024-06-17 08:04:48 +02004132static bool blk_mq_can_poll(struct blk_mq_tag_set *set)
4133{
4134 return set->nr_maps > HCTX_TYPE_POLL &&
4135 set->map[HCTX_TYPE_POLL].nr_queues;
4136}
4137
Christoph Hellwig9ac4dd82024-02-13 08:34:19 +01004138struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set,
4139 struct queue_limits *lim, void *queuedata)
Jens Axboe320ae512013-10-24 09:20:05 +01004140{
Christoph Hellwig9ac4dd82024-02-13 08:34:19 +01004141 struct queue_limits default_lim = { };
Christoph Hellwig26a97502021-06-02 09:53:17 +03004142 struct request_queue *q;
4143 int ret;
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004144
Christoph Hellwigcdb24972024-06-17 08:04:43 +02004145 if (!lim)
4146 lim = &default_lim;
Christoph Hellwigf76af422024-06-17 08:04:46 +02004147 lim->features |= BLK_FEAT_IO_STAT | BLK_FEAT_NOWAIT;
Christoph Hellwig8023e142024-06-17 08:04:48 +02004148 if (blk_mq_can_poll(set))
4149 lim->features |= BLK_FEAT_POLL;
Christoph Hellwigcdb24972024-06-17 08:04:43 +02004150
4151 q = blk_alloc_queue(lim, set->numa_node);
Christoph Hellwigad751ba2024-02-13 08:34:18 +01004152 if (IS_ERR(q))
4153 return q;
Christoph Hellwig26a97502021-06-02 09:53:17 +03004154 q->queuedata = queuedata;
4155 ret = blk_mq_init_allocated_queue(set, q);
4156 if (ret) {
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004157 blk_put_queue(q);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004158 return ERR_PTR(ret);
4159 }
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004160 return q;
4161}
Christoph Hellwig9ac4dd82024-02-13 08:34:19 +01004162EXPORT_SYMBOL(blk_mq_alloc_queue);
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004163
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004164/**
4165 * blk_mq_destroy_queue - shutdown a request queue
4166 * @q: request queue to shutdown
4167 *
Christoph Hellwig9ac4dd82024-02-13 08:34:19 +01004168 * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future
Bart Van Assche81ea42b2023-01-30 13:12:33 -08004169 * requests will be failed with -ENODEV. The caller is responsible for dropping
Christoph Hellwig9ac4dd82024-02-13 08:34:19 +01004170 * the reference from blk_mq_alloc_queue() by calling blk_put_queue().
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004171 *
4172 * Context: can sleep
4173 */
4174void blk_mq_destroy_queue(struct request_queue *q)
4175{
4176 WARN_ON_ONCE(!queue_is_mq(q));
4177 WARN_ON_ONCE(blk_queue_registered(q));
4178
4179 might_sleep();
4180
4181 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4182 blk_queue_start_drain(q);
Jinlong Chen56c1ee92022-10-30 16:32:12 +08004183 blk_mq_freeze_queue_wait(q);
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004184
4185 blk_sync_queue(q);
4186 blk_mq_cancel_work_sync(q);
4187 blk_mq_exit_queue(q);
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004188}
4189EXPORT_SYMBOL(blk_mq_destroy_queue);
4190
Christoph Hellwig27e32cd2024-02-13 08:34:20 +01004191struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set,
4192 struct queue_limits *lim, void *queuedata,
Christoph Hellwig4dcc48742021-08-16 15:19:05 +02004193 struct lock_class_key *lkclass)
Jens Axboe9316a9e2018-10-15 08:40:37 -06004194{
4195 struct request_queue *q;
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004196 struct gendisk *disk;
Jens Axboe9316a9e2018-10-15 08:40:37 -06004197
Christoph Hellwig27e32cd2024-02-13 08:34:20 +01004198 q = blk_mq_alloc_queue(set, lim, queuedata);
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004199 if (IS_ERR(q))
4200 return ERR_CAST(q);
Jens Axboe9316a9e2018-10-15 08:40:37 -06004201
Christoph Hellwig4a1fa412021-08-16 15:19:08 +02004202 disk = __alloc_disk_node(q, set->numa_node, lkclass);
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004203 if (!disk) {
Christoph Hellwig0a3e5cc2022-07-20 15:05:40 +02004204 blk_mq_destroy_queue(q);
Christoph Hellwig2b3f0562022-10-18 15:57:17 +02004205 blk_put_queue(q);
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004206 return ERR_PTR(-ENOMEM);
Jens Axboe9316a9e2018-10-15 08:40:37 -06004207 }
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004208 set_bit(GD_OWNS_QUEUE, &disk->state);
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004209 return disk;
Jens Axboe9316a9e2018-10-15 08:40:37 -06004210}
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004211EXPORT_SYMBOL(__blk_mq_alloc_disk);
Jens Axboe9316a9e2018-10-15 08:40:37 -06004212
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004213struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4214 struct lock_class_key *lkclass)
4215{
Christoph Hellwig22c17e22022-11-22 08:27:53 +01004216 struct gendisk *disk;
4217
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004218 if (!blk_get_queue(q))
4219 return NULL;
Christoph Hellwig22c17e22022-11-22 08:27:53 +01004220 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4221 if (!disk)
4222 blk_put_queue(q);
4223 return disk;
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004224}
4225EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4226
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004227static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4228 struct blk_mq_tag_set *set, struct request_queue *q,
4229 int hctx_idx, int node)
4230{
Ming Lei2f8f1332019-04-30 09:52:27 +08004231 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004232
Ming Lei2f8f1332019-04-30 09:52:27 +08004233 /* reuse dead hctx first */
4234 spin_lock(&q->unused_hctx_lock);
4235 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4236 if (tmp->numa_node == node) {
4237 hctx = tmp;
4238 break;
4239 }
4240 }
4241 if (hctx)
4242 list_del_init(&hctx->hctx_list);
4243 spin_unlock(&q->unused_hctx_lock);
4244
4245 if (!hctx)
4246 hctx = blk_mq_alloc_hctx(q, set, node);
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004247 if (!hctx)
Ming Lei7c6c5b72019-04-30 09:52:26 +08004248 goto fail;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004249
Ming Lei7c6c5b72019-04-30 09:52:26 +08004250 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4251 goto free_hctx;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004252
4253 return hctx;
Ming Lei7c6c5b72019-04-30 09:52:26 +08004254
4255 free_hctx:
4256 kobject_put(&hctx->kobj);
4257 fail:
4258 return NULL;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004259}
4260
Keith Busch868f2f02015-12-17 17:08:14 -07004261static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4262 struct request_queue *q)
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004263{
Ming Lei4e5cc992022-03-08 15:32:19 +08004264 struct blk_mq_hw_ctx *hctx;
4265 unsigned long i, j;
Bart Van Asscheac0d6b92019-10-25 09:50:09 -07004266
Ming Leifb350e02018-01-06 16:27:40 +08004267 /* protect against switching io scheduler */
4268 mutex_lock(&q->sysfs_lock);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004269 for (i = 0; i < set->nr_hw_queues; i++) {
Ming Lei306f13ee2022-03-08 15:32:15 +08004270 int old_node;
Ming Lei4d805132022-03-08 15:32:14 +08004271 int node = blk_mq_get_hctx_node(set, i);
Ming Lei4e5cc992022-03-08 15:32:19 +08004272 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
Keith Busch868f2f02015-12-17 17:08:14 -07004273
Ming Lei306f13ee2022-03-08 15:32:15 +08004274 if (old_hctx) {
4275 old_node = old_hctx->numa_node;
4276 blk_mq_exit_hctx(q, set, old_hctx, i);
4277 }
Jens Axboe320ae512013-10-24 09:20:05 +01004278
Ming Lei4e5cc992022-03-08 15:32:19 +08004279 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
Ming Lei306f13ee2022-03-08 15:32:15 +08004280 if (!old_hctx)
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004281 break;
Ming Lei306f13ee2022-03-08 15:32:15 +08004282 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4283 node, old_node);
Ming Lei4e5cc992022-03-08 15:32:19 +08004284 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4285 WARN_ON_ONCE(!hctx);
Keith Busch868f2f02015-12-17 17:08:14 -07004286 }
Jens Axboe320ae512013-10-24 09:20:05 +01004287 }
Jianchao Wange01ad462018-10-12 18:07:28 +08004288 /*
4289 * Increasing nr_hw_queues fails. Free the newly allocated
4290 * hctxs and keep the previous q->nr_hw_queues.
4291 */
4292 if (i != set->nr_hw_queues) {
4293 j = q->nr_hw_queues;
Jianchao Wange01ad462018-10-12 18:07:28 +08004294 } else {
4295 j = i;
Jianchao Wange01ad462018-10-12 18:07:28 +08004296 q->nr_hw_queues = set->nr_hw_queues;
4297 }
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004298
Ming Lei4e5cc992022-03-08 15:32:19 +08004299 xa_for_each_start(&q->hctx_table, j, hctx, j)
4300 blk_mq_exit_hctx(q, set, hctx, j);
Ming Leifb350e02018-01-06 16:27:40 +08004301 mutex_unlock(&q->sysfs_lock);
Keith Busch868f2f02015-12-17 17:08:14 -07004302}
4303
Christoph Hellwig26a97502021-06-02 09:53:17 +03004304int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4305 struct request_queue *q)
Keith Busch868f2f02015-12-17 17:08:14 -07004306{
Ming Lei66841672016-02-12 15:27:00 +08004307 /* mark the queue as mq asap */
4308 q->mq_ops = set->ops;
4309
Ming Lei1db49092018-11-20 09:44:35 +08004310 if (blk_mq_alloc_ctxs(q))
Keith Busch54bdd672023-03-20 12:49:26 -07004311 goto err_exit;
Keith Busch868f2f02015-12-17 17:08:14 -07004312
Ming Lei737f98c2017-02-22 18:13:59 +08004313 /* init q->mq_kobj and sw queues' kobjects */
4314 blk_mq_sysfs_init(q);
4315
Ming Lei2f8f1332019-04-30 09:52:27 +08004316 INIT_LIST_HEAD(&q->unused_hctx_list);
4317 spin_lock_init(&q->unused_hctx_lock);
4318
Ming Lei4e5cc992022-03-08 15:32:19 +08004319 xa_init(&q->hctx_table);
4320
Keith Busch868f2f02015-12-17 17:08:14 -07004321 blk_mq_realloc_hw_ctxs(set, q);
4322 if (!q->nr_hw_queues)
4323 goto err_hctxs;
Jens Axboe320ae512013-10-24 09:20:05 +01004324
Christoph Hellwig287922e2015-10-30 20:57:30 +08004325 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
Ming Leie56f6982015-07-16 19:53:22 +08004326 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
Jens Axboe320ae512013-10-24 09:20:05 +01004327
Jens Axboea8908932018-10-16 14:23:06 -06004328 q->tag_set = set;
Jens Axboe320ae512013-10-24 09:20:05 +01004329
Jens Axboe94eddfb2013-11-19 09:25:07 -07004330 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
Jens Axboe320ae512013-10-24 09:20:05 +01004331
Mike Snitzer28494502016-09-14 13:28:30 -04004332 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
Christoph Hellwig9a67aa52023-05-19 06:40:50 +02004333 INIT_LIST_HEAD(&q->flush_list);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06004334 INIT_LIST_HEAD(&q->requeue_list);
4335 spin_lock_init(&q->requeue_lock);
4336
Jens Axboeeba71762014-05-20 15:17:27 -06004337 q->nr_requests = set->queue_depth;
4338
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004339 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
Jens Axboe0d2602c2014-05-13 15:10:52 -06004340 blk_mq_add_queue_tag_set(set, q);
Christoph Hellwig4b855ad2017-06-26 12:20:57 +02004341 blk_mq_map_swqueue(q);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004342 return 0;
Christoph Hellwig18741982014-02-10 09:29:00 -07004343
Jens Axboe320ae512013-10-24 09:20:05 +01004344err_hctxs:
Chen Jun943f45b2022-10-31 03:12:42 +00004345 blk_mq_release(q);
Ming Linc7de5722016-05-25 23:23:27 -07004346err_exit:
4347 q->mq_ops = NULL;
Christoph Hellwig26a97502021-06-02 09:53:17 +03004348 return -ENOMEM;
Jens Axboe320ae512013-10-24 09:20:05 +01004349}
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004350EXPORT_SYMBOL(blk_mq_init_allocated_queue);
Jens Axboe320ae512013-10-24 09:20:05 +01004351
Ming Leic7e2d942019-04-30 09:52:25 +08004352/* tags can _not_ be used after returning from blk_mq_exit_queue */
4353void blk_mq_exit_queue(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01004354{
Bart Van Assche630ef622021-05-13 10:15:29 -07004355 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01004356
Bart Van Assche630ef622021-05-13 10:15:29 -07004357 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
Ming Lei624dbe42014-05-27 23:35:13 +08004358 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
Bart Van Assche630ef622021-05-13 10:15:29 -07004359 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4360 blk_mq_del_queue_tag_set(q);
Jens Axboe320ae512013-10-24 09:20:05 +01004361}
Jens Axboe320ae512013-10-24 09:20:05 +01004362
Jens Axboea5164402014-09-10 09:02:03 -06004363static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4364{
4365 int i;
4366
John Garry079a2e32021-10-05 18:23:39 +08004367 if (blk_mq_is_shared_tags(set->flags)) {
4368 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
John Garrye155b0c2021-10-05 18:23:37 +08004369 BLK_MQ_NO_HCTX_IDX,
4370 set->queue_depth);
John Garry079a2e32021-10-05 18:23:39 +08004371 if (!set->shared_tags)
John Garrye155b0c2021-10-05 18:23:37 +08004372 return -ENOMEM;
4373 }
4374
Xianting Tian8229cca2020-09-26 10:39:47 +08004375 for (i = 0; i < set->nr_hw_queues; i++) {
John Garry63064be2021-10-05 18:23:35 +08004376 if (!__blk_mq_alloc_map_and_rqs(set, i))
Jens Axboea5164402014-09-10 09:02:03 -06004377 goto out_unwind;
Xianting Tian8229cca2020-09-26 10:39:47 +08004378 cond_resched();
4379 }
Jens Axboea5164402014-09-10 09:02:03 -06004380
4381 return 0;
4382
4383out_unwind:
4384 while (--i >= 0)
John Garrye155b0c2021-10-05 18:23:37 +08004385 __blk_mq_free_map_and_rqs(set, i);
4386
John Garry079a2e32021-10-05 18:23:39 +08004387 if (blk_mq_is_shared_tags(set->flags)) {
4388 blk_mq_free_map_and_rqs(set, set->shared_tags,
John Garrye155b0c2021-10-05 18:23:37 +08004389 BLK_MQ_NO_HCTX_IDX);
John Garry645db342021-10-05 18:23:36 +08004390 }
Jens Axboea5164402014-09-10 09:02:03 -06004391
Jens Axboea5164402014-09-10 09:02:03 -06004392 return -ENOMEM;
4393}
4394
4395/*
4396 * Allocate the request maps associated with this tag_set. Note that this
4397 * may reduce the depth asked for, if memory is tight. set->queue_depth
4398 * will be updated to reflect the allocated depth.
4399 */
John Garry63064be2021-10-05 18:23:35 +08004400static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
Jens Axboea5164402014-09-10 09:02:03 -06004401{
4402 unsigned int depth;
4403 int err;
4404
4405 depth = set->queue_depth;
4406 do {
4407 err = __blk_mq_alloc_rq_maps(set);
4408 if (!err)
4409 break;
4410
4411 set->queue_depth >>= 1;
4412 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4413 err = -ENOMEM;
4414 break;
4415 }
4416 } while (set->queue_depth);
4417
4418 if (!set->queue_depth || err) {
4419 pr_err("blk-mq: failed to allocate request map\n");
4420 return -ENOMEM;
4421 }
4422
4423 if (depth != set->queue_depth)
4424 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4425 depth, set->queue_depth);
4426
4427 return 0;
4428}
4429
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004430static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
Omar Sandovalebe8bdd2017-04-07 08:53:11 -06004431{
Bart Van Assche6e66b492020-03-09 21:26:17 -07004432 /*
4433 * blk_mq_map_queues() and multiple .map_queues() implementations
4434 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4435 * number of hardware queues.
4436 */
4437 if (set->nr_maps == 1)
4438 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4439
Ming Leiec30b462024-02-28 12:08:57 +08004440 if (set->ops->map_queues) {
Jens Axboeb3c661b2018-10-30 10:36:06 -06004441 int i;
4442
Ming Lei7d4901a2018-01-06 16:27:39 +08004443 /*
4444 * transport .map_queues is usually done in the following
4445 * way:
4446 *
4447 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4448 * mask = get_cpu_mask(queue)
4449 * for_each_cpu(cpu, mask)
Jens Axboeb3c661b2018-10-30 10:36:06 -06004450 * set->map[x].mq_map[cpu] = queue;
Ming Lei7d4901a2018-01-06 16:27:39 +08004451 * }
4452 *
4453 * When we need to remap, the table has to be cleared for
4454 * killing stale mapping since one CPU may not be mapped
4455 * to any hw queue.
4456 */
Jens Axboeb3c661b2018-10-30 10:36:06 -06004457 for (i = 0; i < set->nr_maps; i++)
4458 blk_mq_clear_mq_map(&set->map[i]);
Ming Lei7d4901a2018-01-06 16:27:39 +08004459
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004460 set->ops->map_queues(set);
Jens Axboeb3c661b2018-10-30 10:36:06 -06004461 } else {
4462 BUG_ON(set->nr_maps > 1);
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004463 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
Jens Axboeb3c661b2018-10-30 10:36:06 -06004464 }
Omar Sandovalebe8bdd2017-04-07 08:53:11 -06004465}
4466
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004467static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004468 int new_nr_hw_queues)
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004469{
4470 struct blk_mq_tags **new_tags;
Chengming Zhoue1dd7bc92023-08-21 17:56:00 +08004471 int i;
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004472
Chengming Zhou6be6d112023-09-08 08:57:02 +08004473 if (set->nr_hw_queues >= new_nr_hw_queues)
Shin'ichiro Kawasakid4b2e0d2022-11-22 17:49:17 +09004474 goto done;
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004475
4476 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4477 GFP_KERNEL, set->numa_node);
4478 if (!new_tags)
4479 return -ENOMEM;
4480
4481 if (set->tags)
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004482 memcpy(new_tags, set->tags, set->nr_hw_queues *
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004483 sizeof(*set->tags));
4484 kfree(set->tags);
4485 set->tags = new_tags;
Chengming Zhou72226572023-08-21 17:56:02 +08004486
4487 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4488 if (!__blk_mq_alloc_map_and_rqs(set, i)) {
4489 while (--i >= set->nr_hw_queues)
4490 __blk_mq_free_map_and_rqs(set, i);
4491 return -ENOMEM;
4492 }
4493 cond_resched();
4494 }
4495
Shin'ichiro Kawasakid4b2e0d2022-11-22 17:49:17 +09004496done:
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004497 set->nr_hw_queues = new_nr_hw_queues;
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004498 return 0;
4499}
4500
Jens Axboea4391c62014-06-05 15:21:56 -06004501/*
4502 * Alloc a tag set to be associated with one or more request queues.
4503 * May fail with EINVAL for various error conditions. May adjust the
Minwoo Imc018c842018-06-30 22:12:41 +09004504 * requested depth down, if it's too large. In that case, the set
Jens Axboea4391c62014-06-05 15:21:56 -06004505 * value will be stored in set->queue_depth.
4506 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004507int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4508{
Jens Axboeb3c661b2018-10-30 10:36:06 -06004509 int i, ret;
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004510
Bart Van Assche205fb5f52014-10-30 14:45:11 +01004511 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4512
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004513 if (!set->nr_hw_queues)
4514 return -EINVAL;
Jens Axboea4391c62014-06-05 15:21:56 -06004515 if (!set->queue_depth)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004516 return -EINVAL;
4517 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4518 return -EINVAL;
4519
Christoph Hellwig7d7e0f92016-09-14 16:18:54 +02004520 if (!set->ops->queue_rq)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004521 return -EINVAL;
4522
Ming Leide148292017-10-14 17:22:29 +08004523 if (!set->ops->get_budget ^ !set->ops->put_budget)
4524 return -EINVAL;
4525
Jens Axboea4391c62014-06-05 15:21:56 -06004526 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4527 pr_info("blk-mq: reduced tag depth to %u\n",
4528 BLK_MQ_MAX_DEPTH);
4529 set->queue_depth = BLK_MQ_MAX_DEPTH;
4530 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004531
Jens Axboeb3c661b2018-10-30 10:36:06 -06004532 if (!set->nr_maps)
4533 set->nr_maps = 1;
4534 else if (set->nr_maps > HCTX_MAX_TYPES)
4535 return -EINVAL;
4536
Shaohua Li6637fad2014-11-30 16:00:58 -08004537 /*
4538 * If a crashdump is active, then we are potentially in a very
Ming Leiec30b462024-02-28 12:08:57 +08004539 * memory constrained environment. Limit us to 64 tags to prevent
4540 * using too much memory.
Shaohua Li6637fad2014-11-30 16:00:58 -08004541 */
Ming Leiec30b462024-02-28 12:08:57 +08004542 if (is_kdump_kernel())
Shaohua Li6637fad2014-11-30 16:00:58 -08004543 set->queue_depth = min(64U, set->queue_depth);
Ming Leiec30b462024-02-28 12:08:57 +08004544
Keith Busch868f2f02015-12-17 17:08:14 -07004545 /*
Jens Axboe392546a2018-10-29 13:25:27 -06004546 * There is no use for more h/w queues than cpus if we just have
4547 * a single map
Keith Busch868f2f02015-12-17 17:08:14 -07004548 */
Jens Axboe392546a2018-10-29 13:25:27 -06004549 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
Keith Busch868f2f02015-12-17 17:08:14 -07004550 set->nr_hw_queues = nr_cpu_ids;
Shaohua Li6637fad2014-11-30 16:00:58 -08004551
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004552 if (set->flags & BLK_MQ_F_BLOCKING) {
4553 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4554 if (!set->srcu)
4555 return -ENOMEM;
4556 ret = init_srcu_struct(set->srcu);
4557 if (ret)
4558 goto out_free_srcu;
4559 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004560
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004561 ret = -ENOMEM;
Christoph Hellwig5ee20292022-11-09 11:08:10 +01004562 set->tags = kcalloc_node(set->nr_hw_queues,
4563 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4564 set->numa_node);
4565 if (!set->tags)
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004566 goto out_cleanup_srcu;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004567
Jens Axboeb3c661b2018-10-30 10:36:06 -06004568 for (i = 0; i < set->nr_maps; i++) {
4569 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
Ming Lei07b35eb2018-12-17 18:42:45 +08004570 sizeof(set->map[i].mq_map[0]),
Jens Axboeb3c661b2018-10-30 10:36:06 -06004571 GFP_KERNEL, set->numa_node);
4572 if (!set->map[i].mq_map)
4573 goto out_free_mq_map;
Ming Leiec30b462024-02-28 12:08:57 +08004574 set->map[i].nr_queues = set->nr_hw_queues;
Jens Axboeb3c661b2018-10-30 10:36:06 -06004575 }
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004576
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004577 blk_mq_update_queue_map(set);
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004578
John Garry63064be2021-10-05 18:23:35 +08004579 ret = blk_mq_alloc_set_map_and_rqs(set);
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004580 if (ret)
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004581 goto out_free_mq_map;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004582
Jens Axboe0d2602c2014-05-13 15:10:52 -06004583 mutex_init(&set->tag_list_lock);
4584 INIT_LIST_HEAD(&set->tag_list);
4585
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004586 return 0;
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004587
4588out_free_mq_map:
Jens Axboeb3c661b2018-10-30 10:36:06 -06004589 for (i = 0; i < set->nr_maps; i++) {
4590 kfree(set->map[i].mq_map);
4591 set->map[i].mq_map = NULL;
4592 }
Robert Elliott5676e7b2014-09-02 11:38:44 -05004593 kfree(set->tags);
4594 set->tags = NULL;
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004595out_cleanup_srcu:
4596 if (set->flags & BLK_MQ_F_BLOCKING)
4597 cleanup_srcu_struct(set->srcu);
4598out_free_srcu:
4599 if (set->flags & BLK_MQ_F_BLOCKING)
4600 kfree(set->srcu);
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004601 return ret;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004602}
4603EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4604
Christoph Hellwigcdb14e02021-06-02 09:53:16 +03004605/* allocate and initialize a tagset for a simple single-queue device */
4606int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4607 const struct blk_mq_ops *ops, unsigned int queue_depth,
4608 unsigned int set_flags)
4609{
4610 memset(set, 0, sizeof(*set));
4611 set->ops = ops;
4612 set->nr_hw_queues = 1;
4613 set->nr_maps = 1;
4614 set->queue_depth = queue_depth;
4615 set->numa_node = NUMA_NO_NODE;
4616 set->flags = set_flags;
4617 return blk_mq_alloc_tag_set(set);
4618}
4619EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4620
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004621void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4622{
Jens Axboeb3c661b2018-10-30 10:36:06 -06004623 int i, j;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004624
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004625 for (i = 0; i < set->nr_hw_queues; i++)
John Garrye155b0c2021-10-05 18:23:37 +08004626 __blk_mq_free_map_and_rqs(set, i);
Jens Axboe484b4062014-05-21 14:01:15 -06004627
John Garry079a2e32021-10-05 18:23:39 +08004628 if (blk_mq_is_shared_tags(set->flags)) {
4629 blk_mq_free_map_and_rqs(set, set->shared_tags,
John Garrye155b0c2021-10-05 18:23:37 +08004630 BLK_MQ_NO_HCTX_IDX);
4631 }
John Garry32bc15a2020-08-19 23:20:24 +08004632
Jens Axboeb3c661b2018-10-30 10:36:06 -06004633 for (j = 0; j < set->nr_maps; j++) {
4634 kfree(set->map[j].mq_map);
4635 set->map[j].mq_map = NULL;
4636 }
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004637
Ming Lei981bd182014-04-24 00:07:34 +08004638 kfree(set->tags);
Robert Elliott5676e7b2014-09-02 11:38:44 -05004639 set->tags = NULL;
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004640 if (set->flags & BLK_MQ_F_BLOCKING) {
4641 cleanup_srcu_struct(set->srcu);
4642 kfree(set->srcu);
4643 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004644}
4645EXPORT_SYMBOL(blk_mq_free_tag_set);
4646
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004647int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4648{
4649 struct blk_mq_tag_set *set = q->tag_set;
4650 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08004651 int ret;
4652 unsigned long i;
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004653
Christoph Hellwigaf281412024-06-17 08:04:38 +02004654 if (WARN_ON_ONCE(!q->mq_freeze_depth))
4655 return -EINVAL;
4656
Jens Axboebd166ef2017-01-17 06:03:22 -07004657 if (!set)
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004658 return -EINVAL;
4659
Aleksei Zakharove5fa8142019-02-08 19:14:05 +03004660 if (q->nr_requests == nr)
4661 return 0;
4662
Ming Lei24f5a902018-01-06 16:27:38 +08004663 blk_mq_quiesce_queue(q);
Jens Axboe70f36b62017-01-19 10:59:07 -07004664
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004665 ret = 0;
4666 queue_for_each_hw_ctx(q, hctx, i) {
Keith Busche9137d42016-02-18 14:56:35 -07004667 if (!hctx->tags)
4668 continue;
Jens Axboebd166ef2017-01-17 06:03:22 -07004669 /*
4670 * If we're using an MQ scheduler, just update the scheduler
4671 * queue depth. This is similar to what the old code would do.
4672 */
John Garryf6adcef2021-10-05 18:23:29 +08004673 if (hctx->sched_tags) {
Jens Axboe70f36b62017-01-19 10:59:07 -07004674 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
John Garryf6adcef2021-10-05 18:23:29 +08004675 nr, true);
John Garryf6adcef2021-10-05 18:23:29 +08004676 } else {
4677 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4678 false);
Jens Axboe70f36b62017-01-19 10:59:07 -07004679 }
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004680 if (ret)
4681 break;
Jens Axboe77f1e0a2019-01-18 10:34:16 -07004682 if (q->elevator && q->elevator->type->ops.depth_updated)
4683 q->elevator->type->ops.depth_updated(hctx);
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004684 }
John Garryd97e5942021-05-13 20:00:58 +08004685 if (!ret) {
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004686 q->nr_requests = nr;
John Garry079a2e32021-10-05 18:23:39 +08004687 if (blk_mq_is_shared_tags(set->flags)) {
John Garry8fa04462021-10-05 18:23:28 +08004688 if (q->elevator)
John Garry079a2e32021-10-05 18:23:39 +08004689 blk_mq_tag_update_sched_shared_tags(q);
John Garry8fa04462021-10-05 18:23:28 +08004690 else
John Garry079a2e32021-10-05 18:23:39 +08004691 blk_mq_tag_resize_shared_tags(set, nr);
John Garry8fa04462021-10-05 18:23:28 +08004692 }
John Garryd97e5942021-05-13 20:00:58 +08004693 }
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004694
Ming Lei24f5a902018-01-06 16:27:38 +08004695 blk_mq_unquiesce_queue(q);
Jens Axboe70f36b62017-01-19 10:59:07 -07004696
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004697 return ret;
4698}
4699
Jianchao Wangd48ece22018-08-21 15:15:03 +08004700/*
4701 * request_queue and elevator_type pair.
4702 * It is just used by __blk_mq_update_nr_hw_queues to cache
4703 * the elevator_type associated with a request_queue.
4704 */
4705struct blk_mq_qe_pair {
4706 struct list_head node;
4707 struct request_queue *q;
4708 struct elevator_type *type;
4709};
4710
4711/*
4712 * Cache the elevator_type in qe pair list and switch the
4713 * io scheduler to 'none'
4714 */
4715static bool blk_mq_elv_switch_none(struct list_head *head,
4716 struct request_queue *q)
4717{
4718 struct blk_mq_qe_pair *qe;
4719
Jianchao Wangd48ece22018-08-21 15:15:03 +08004720 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4721 if (!qe)
4722 return false;
4723
Ming Lei5fd7a842022-06-16 09:43:59 +08004724 /* q->elevator needs protection from ->sysfs_lock */
4725 mutex_lock(&q->sysfs_lock);
4726
Ming Lei24516562023-06-16 21:23:54 +08004727 /* the check has to be done with holding sysfs_lock */
4728 if (!q->elevator) {
4729 kfree(qe);
4730 goto unlock;
4731 }
4732
Jianchao Wangd48ece22018-08-21 15:15:03 +08004733 INIT_LIST_HEAD(&qe->node);
4734 qe->q = q;
4735 qe->type = q->elevator->type;
Christoph Hellwigdd6f7f12022-10-20 08:48:16 +02004736 /* keep a reference to the elevator module as we'll switch back */
4737 __elevator_get(qe->type);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004738 list_add(&qe->node, head);
Christoph Hellwig64b36072022-10-30 11:07:14 +01004739 elevator_disable(q);
Ming Lei24516562023-06-16 21:23:54 +08004740unlock:
Jianchao Wangd48ece22018-08-21 15:15:03 +08004741 mutex_unlock(&q->sysfs_lock);
4742
4743 return true;
4744}
4745
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004746static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4747 struct request_queue *q)
Jianchao Wangd48ece22018-08-21 15:15:03 +08004748{
4749 struct blk_mq_qe_pair *qe;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004750
4751 list_for_each_entry(qe, head, node)
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004752 if (qe->q == q)
4753 return qe;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004754
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004755 return NULL;
4756}
4757
4758static void blk_mq_elv_switch_back(struct list_head *head,
4759 struct request_queue *q)
4760{
4761 struct blk_mq_qe_pair *qe;
4762 struct elevator_type *t;
4763
4764 qe = blk_lookup_qe_pair(head, q);
4765 if (!qe)
Jianchao Wangd48ece22018-08-21 15:15:03 +08004766 return;
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004767 t = qe->type;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004768 list_del(&qe->node);
4769 kfree(qe);
4770
4771 mutex_lock(&q->sysfs_lock);
Keith Busch8237c012022-09-27 08:56:52 -07004772 elevator_switch(q, t);
Jinlong Chen8ed40ee2022-10-20 08:48:19 +02004773 /* drop the reference acquired in blk_mq_elv_switch_none */
4774 elevator_put(t);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004775 mutex_unlock(&q->sysfs_lock);
4776}
4777
Keith Busche4dc2b32017-05-30 14:39:11 -04004778static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4779 int nr_hw_queues)
Keith Busch868f2f02015-12-17 17:08:14 -07004780{
4781 struct request_queue *q;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004782 LIST_HEAD(head);
Chengming Zhou6be6d112023-09-08 08:57:02 +08004783 int prev_nr_hw_queues = set->nr_hw_queues;
4784 int i;
Keith Busch868f2f02015-12-17 17:08:14 -07004785
Bart Van Assche705cda92017-04-07 11:16:49 -07004786 lockdep_assert_held(&set->tag_list_lock);
4787
Jens Axboe392546a2018-10-29 13:25:27 -06004788 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
Keith Busch868f2f02015-12-17 17:08:14 -07004789 nr_hw_queues = nr_cpu_ids;
Weiping Zhangfe35ec52020-06-17 14:18:37 +08004790 if (nr_hw_queues < 1)
4791 return;
4792 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
Keith Busch868f2f02015-12-17 17:08:14 -07004793 return;
4794
4795 list_for_each_entry(q, &set->tag_list, tag_set_list)
4796 blk_mq_freeze_queue(q);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004797 /*
4798 * Switch IO scheduler to 'none', cleaning up the data associated
4799 * with the previous scheduler. We will switch back once we are done
4800 * updating the new sw to hw queue mappings.
4801 */
4802 list_for_each_entry(q, &set->tag_list, tag_set_list)
4803 if (!blk_mq_elv_switch_none(&head, q))
4804 goto switch_back;
Keith Busch868f2f02015-12-17 17:08:14 -07004805
Jianchao Wang477e19d2018-10-12 18:07:25 +08004806 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4807 blk_mq_debugfs_unregister_hctxs(q);
Christoph Hellwigeaa870f2022-06-28 19:18:49 +02004808 blk_mq_sysfs_unregister_hctxs(q);
Jianchao Wang477e19d2018-10-12 18:07:25 +08004809 }
4810
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004811 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004812 goto reregister;
4813
Jianchao Wange01ad462018-10-12 18:07:28 +08004814fallback:
Weiping Zhangaa880ad2020-05-13 08:44:05 +08004815 blk_mq_update_queue_map(set);
Keith Busch868f2f02015-12-17 17:08:14 -07004816 list_for_each_entry(q, &set->tag_list, tag_set_list) {
Christoph Hellwig8023e142024-06-17 08:04:48 +02004817 struct queue_limits lim;
4818
Keith Busch868f2f02015-12-17 17:08:14 -07004819 blk_mq_realloc_hw_ctxs(set, q);
Christoph Hellwig8023e142024-06-17 08:04:48 +02004820
Jianchao Wange01ad462018-10-12 18:07:28 +08004821 if (q->nr_hw_queues != set->nr_hw_queues) {
Ye Bina846a8e2021-11-08 15:40:19 +08004822 int i = prev_nr_hw_queues;
4823
Jianchao Wange01ad462018-10-12 18:07:28 +08004824 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4825 nr_hw_queues, prev_nr_hw_queues);
Ye Bina846a8e2021-11-08 15:40:19 +08004826 for (; i < set->nr_hw_queues; i++)
4827 __blk_mq_free_map_and_rqs(set, i);
4828
Jianchao Wange01ad462018-10-12 18:07:28 +08004829 set->nr_hw_queues = prev_nr_hw_queues;
Jianchao Wange01ad462018-10-12 18:07:28 +08004830 goto fallback;
4831 }
Christoph Hellwig8023e142024-06-17 08:04:48 +02004832 lim = queue_limits_start_update(q);
4833 if (blk_mq_can_poll(set))
4834 lim.features |= BLK_FEAT_POLL;
4835 else
4836 lim.features &= ~BLK_FEAT_POLL;
4837 if (queue_limits_commit_update(q, &lim) < 0)
4838 pr_warn("updating the poll flag failed\n");
Jianchao Wang477e19d2018-10-12 18:07:25 +08004839 blk_mq_map_swqueue(q);
4840 }
4841
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004842reregister:
Jianchao Wang477e19d2018-10-12 18:07:25 +08004843 list_for_each_entry(q, &set->tag_list, tag_set_list) {
Christoph Hellwigeaa870f2022-06-28 19:18:49 +02004844 blk_mq_sysfs_register_hctxs(q);
Jianchao Wang477e19d2018-10-12 18:07:25 +08004845 blk_mq_debugfs_register_hctxs(q);
Keith Busch868f2f02015-12-17 17:08:14 -07004846 }
4847
Jianchao Wangd48ece22018-08-21 15:15:03 +08004848switch_back:
4849 list_for_each_entry(q, &set->tag_list, tag_set_list)
4850 blk_mq_elv_switch_back(&head, q);
4851
Keith Busch868f2f02015-12-17 17:08:14 -07004852 list_for_each_entry(q, &set->tag_list, tag_set_list)
4853 blk_mq_unfreeze_queue(q);
Chengming Zhou6be6d112023-09-08 08:57:02 +08004854
4855 /* Free the excess tags when nr_hw_queues shrink. */
4856 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
4857 __blk_mq_free_map_and_rqs(set, i);
Keith Busch868f2f02015-12-17 17:08:14 -07004858}
Keith Busche4dc2b32017-05-30 14:39:11 -04004859
4860void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4861{
4862 mutex_lock(&set->tag_list_lock);
4863 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4864 mutex_unlock(&set->tag_list_lock);
4865}
Keith Busch868f2f02015-12-17 17:08:14 -07004866EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4867
Keith Buschf6c80cf2023-06-12 12:03:42 -07004868static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4869 struct io_comp_batch *iob, unsigned int flags)
Jens Axboebbd7bb72016-11-04 09:34:34 -06004870{
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004871 long state = get_current_state();
4872 int ret;
Jens Axboe1052b8a2018-11-26 08:21:49 -07004873
Jens Axboeaa61bec2018-11-13 21:32:10 -07004874 do {
Jens Axboe5a72e892021-10-12 09:24:29 -06004875 ret = q->mq_ops->poll(hctx, iob);
Jens Axboebbd7bb72016-11-04 09:34:34 -06004876 if (ret > 0) {
Jens Axboe849a3702018-11-16 08:37:34 -07004877 __set_current_state(TASK_RUNNING);
Jens Axboe85f4d4b2018-11-06 13:30:55 -07004878 return ret;
Jens Axboebbd7bb72016-11-04 09:34:34 -06004879 }
4880
4881 if (signal_pending_state(state, current))
Jens Axboe849a3702018-11-16 08:37:34 -07004882 __set_current_state(TASK_RUNNING);
Peter Zijlstrab03fbd42021-06-11 10:28:12 +02004883 if (task_is_running(current))
Jens Axboe85f4d4b2018-11-06 13:30:55 -07004884 return 1;
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004885
Christoph Hellwigef99b2d2021-10-12 13:12:19 +02004886 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
Jens Axboebbd7bb72016-11-04 09:34:34 -06004887 break;
4888 cpu_relax();
Jens Axboeaa61bec2018-11-13 21:32:10 -07004889 } while (!need_resched());
Jens Axboebbd7bb72016-11-04 09:34:34 -06004890
Nitesh Shetty67b41102018-02-13 21:18:12 +05304891 __set_current_state(TASK_RUNNING);
Jens Axboe85f4d4b2018-11-06 13:30:55 -07004892 return 0;
Jens Axboebbd7bb72016-11-04 09:34:34 -06004893}
Jens Axboebbd7bb72016-11-04 09:34:34 -06004894
Keith Buschf6c80cf2023-06-12 12:03:42 -07004895int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4896 struct io_comp_batch *iob, unsigned int flags)
4897{
4898 struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4899
4900 return blk_hctx_poll(q, hctx, iob, flags);
4901}
4902
4903int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4904 unsigned int poll_flags)
4905{
4906 struct request_queue *q = rq->q;
4907 int ret;
4908
4909 if (!blk_rq_is_poll(rq))
4910 return 0;
4911 if (!percpu_ref_tryget(&q->q_usage_counter))
4912 return 0;
4913
4914 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4915 blk_queue_exit(q);
4916
4917 return ret;
4918}
4919EXPORT_SYMBOL_GPL(blk_rq_poll);
4920
Jens Axboe9cf2bab2018-10-31 17:01:22 -06004921unsigned int blk_mq_rq_cpu(struct request *rq)
4922{
4923 return rq->mq_ctx->cpu;
4924}
4925EXPORT_SYMBOL(blk_mq_rq_cpu);
4926
Ming Lei2a19b282021-11-16 09:43:43 +08004927void blk_mq_cancel_work_sync(struct request_queue *q)
4928{
Jinlong Chen219cf432022-10-30 17:47:30 +08004929 struct blk_mq_hw_ctx *hctx;
4930 unsigned long i;
Ming Lei2a19b282021-11-16 09:43:43 +08004931
Jinlong Chen219cf432022-10-30 17:47:30 +08004932 cancel_delayed_work_sync(&q->requeue_work);
Ming Lei2a19b282021-11-16 09:43:43 +08004933
Jinlong Chen219cf432022-10-30 17:47:30 +08004934 queue_for_each_hw_ctx(q, hctx, i)
4935 cancel_delayed_work_sync(&hctx->run_work);
Ming Lei2a19b282021-11-16 09:43:43 +08004936}
4937
Jens Axboe320ae512013-10-24 09:20:05 +01004938static int __init blk_mq_init(void)
4939{
Christoph Hellwigc3077b52020-06-11 08:44:41 +02004940 int i;
4941
4942 for_each_possible_cpu(i)
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01004943 init_llist_head(&per_cpu(blk_cpu_done, i));
Chengming Zhou660e8022023-07-17 12:00:55 +08004944 for_each_possible_cpu(i)
4945 INIT_CSD(&per_cpu(blk_cpu_csd, i),
4946 __blk_mq_complete_request_remote, NULL);
Christoph Hellwigc3077b52020-06-11 08:44:41 +02004947 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4948
4949 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4950 "block/softirq:dead", NULL,
4951 blk_softirq_cpu_dead);
Thomas Gleixner9467f852016-09-22 08:05:17 -06004952 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4953 blk_mq_hctx_notify_dead);
Ming Leibf0beec2020-05-29 15:53:15 +02004954 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4955 blk_mq_hctx_notify_online,
4956 blk_mq_hctx_notify_offline);
Jens Axboe320ae512013-10-24 09:20:05 +01004957 return 0;
4958}
4959subsys_initcall(blk_mq_init);