blob: 2c49b4151da15d419cc16b4fec553ed6d1939e00 [file] [log] [blame]
Christoph Hellwig3dcf60bc2019-04-30 14:42:43 -04001// SPDX-License-Identifier: GPL-2.0
Jens Axboe75bb4622014-05-28 10:15:41 -06002/*
3 * Block multiqueue core code
4 *
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
7 */
Jens Axboe320ae512013-10-24 09:20:05 +01008#include <linux/kernel.h>
9#include <linux/module.h>
10#include <linux/backing-dev.h>
11#include <linux/bio.h>
12#include <linux/blkdev.h>
Christoph Hellwigfe45e632021-09-20 14:33:27 +020013#include <linux/blk-integrity.h>
Catalin Marinasf75782e2015-09-14 18:16:02 +010014#include <linux/kmemleak.h>
Jens Axboe320ae512013-10-24 09:20:05 +010015#include <linux/mm.h>
16#include <linux/init.h>
17#include <linux/slab.h>
18#include <linux/workqueue.h>
19#include <linux/smp.h>
Christoph Hellwige41d12f2021-09-20 14:33:13 +020020#include <linux/interrupt.h>
Jens Axboe320ae512013-10-24 09:20:05 +010021#include <linux/llist.h>
Jens Axboe320ae512013-10-24 09:20:05 +010022#include <linux/cpu.h>
23#include <linux/cache.h>
24#include <linux/sched/sysctl.h>
Ingo Molnar105ab3d82017-02-01 16:36:40 +010025#include <linux/sched/topology.h>
Ingo Molnar174cd4b2017-02-02 19:15:33 +010026#include <linux/sched/signal.h>
Jens Axboe320ae512013-10-24 09:20:05 +010027#include <linux/delay.h>
Jens Axboeaedcd722014-09-17 08:27:03 -060028#include <linux/crash_dump.h>
Jens Axboe88c7b2b2016-08-25 08:07:30 -060029#include <linux/prefetch.h>
Satya Tangiralaa892c8d2020-05-14 00:37:18 +000030#include <linux/blk-crypto.h>
Christoph Hellwig82d981d2021-11-23 19:53:12 +010031#include <linux/part_stat.h>
Jens Axboe320ae512013-10-24 09:20:05 +010032
33#include <trace/events/block.h>
34
35#include <linux/blk-mq.h>
Max Gurtovoy54d4e6a2019-09-16 18:44:29 +030036#include <linux/t10-pi.h>
Jens Axboe320ae512013-10-24 09:20:05 +010037#include "blk.h"
38#include "blk-mq.h"
Omar Sandoval9c1051a2017-05-04 08:17:21 -060039#include "blk-mq-debugfs.h"
Jens Axboe320ae512013-10-24 09:20:05 +010040#include "blk-mq-tag.h"
Bart Van Assche986d4132018-09-26 14:01:10 -070041#include "blk-pm.h"
Jens Axboecf43e6b2016-11-07 21:32:37 -070042#include "blk-stat.h"
Jens Axboebd166ef2017-01-17 06:03:22 -070043#include "blk-mq-sched.h"
Josef Bacikc1c80382018-07-03 11:14:59 -040044#include "blk-rq-qos.h"
Jan Kara82b74ca2022-06-23 09:48:32 +020045#include "blk-ioprio.h"
Jens Axboe320ae512013-10-24 09:20:05 +010046
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +010047static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
Christoph Hellwigc3077b52020-06-11 08:44:41 +020048
Omar Sandoval34dbad52017-03-21 08:56:08 -070049static void blk_mq_poll_stats_start(struct request_queue *q);
50static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
51
Stephen Bates720b8cc2017-04-07 06:24:03 -060052static int blk_mq_poll_stats_bkt(const struct request *rq)
53{
Hou Tao3d244302019-05-21 15:59:03 +080054 int ddir, sectors, bucket;
Stephen Bates720b8cc2017-04-07 06:24:03 -060055
Jens Axboe99c749a2017-04-21 07:55:42 -060056 ddir = rq_data_dir(rq);
Hou Tao3d244302019-05-21 15:59:03 +080057 sectors = blk_rq_stats_sectors(rq);
Stephen Bates720b8cc2017-04-07 06:24:03 -060058
Hou Tao3d244302019-05-21 15:59:03 +080059 bucket = ddir + 2 * ilog2(sectors);
Stephen Bates720b8cc2017-04-07 06:24:03 -060060
61 if (bucket < 0)
62 return -1;
63 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
64 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
65
66 return bucket;
67}
68
Christoph Hellwig3e087732021-10-12 13:12:24 +020069#define BLK_QC_T_SHIFT 16
70#define BLK_QC_T_INTERNAL (1U << 31)
71
Christoph Hellwigf70299f2021-10-12 13:12:15 +020072static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q,
73 blk_qc_t qc)
74{
Ming Lei4e5cc992022-03-08 15:32:19 +080075 return xa_load(&q->hctx_table,
76 (qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT);
Christoph Hellwigf70299f2021-10-12 13:12:15 +020077}
78
Christoph Hellwigc6699d62021-10-12 13:12:16 +020079static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx,
80 blk_qc_t qc)
81{
Christoph Hellwigefbabbe2021-10-12 13:12:17 +020082 unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1);
83
84 if (qc & BLK_QC_T_INTERNAL)
85 return blk_mq_tag_to_rq(hctx->sched_tags, tag);
86 return blk_mq_tag_to_rq(hctx->tags, tag);
Christoph Hellwigc6699d62021-10-12 13:12:16 +020087}
88
Christoph Hellwig3e087732021-10-12 13:12:24 +020089static inline blk_qc_t blk_rq_to_qc(struct request *rq)
90{
91 return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) |
92 (rq->tag != -1 ?
93 rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL));
94}
95
Jens Axboe320ae512013-10-24 09:20:05 +010096/*
Yufen Yu85fae292019-03-24 17:57:08 +080097 * Check if any of the ctx, dispatch list or elevator
98 * have pending work in this hardware queue.
Jens Axboe320ae512013-10-24 09:20:05 +010099 */
Jens Axboe79f720a2017-11-10 09:13:21 -0700100static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
Jens Axboe320ae512013-10-24 09:20:05 +0100101{
Jens Axboe79f720a2017-11-10 09:13:21 -0700102 return !list_empty_careful(&hctx->dispatch) ||
103 sbitmap_any_bit_set(&hctx->ctx_map) ||
Jens Axboebd166ef2017-01-17 06:03:22 -0700104 blk_mq_sched_has_work(hctx);
Jens Axboe320ae512013-10-24 09:20:05 +0100105}
106
107/*
108 * Mark this ctx as having pending work in this hardware queue
109 */
110static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
111 struct blk_mq_ctx *ctx)
112{
Jens Axboef31967f2018-10-29 13:13:29 -0600113 const int bit = ctx->index_hw[hctx->type];
114
115 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
116 sbitmap_set_bit(&hctx->ctx_map, bit);
Jens Axboe1429d7c2014-05-19 09:23:55 -0600117}
118
119static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
120 struct blk_mq_ctx *ctx)
121{
Jens Axboef31967f2018-10-29 13:13:29 -0600122 const int bit = ctx->index_hw[hctx->type];
123
124 sbitmap_clear_bit(&hctx->ctx_map, bit);
Jens Axboe320ae512013-10-24 09:20:05 +0100125}
126
Jens Axboef299b7c2017-08-08 17:51:45 -0600127struct mq_inflight {
Christoph Hellwig8446fe92020-11-24 09:36:54 +0100128 struct block_device *part;
Pavel Begunkova2e80f62019-09-30 21:55:34 +0300129 unsigned int inflight[2];
Jens Axboef299b7c2017-08-08 17:51:45 -0600130};
131
John Garry2dd65322022-07-06 20:03:53 +0800132static bool blk_mq_check_inflight(struct request *rq, void *priv)
Jens Axboef299b7c2017-08-08 17:51:45 -0600133{
134 struct mq_inflight *mi = priv;
135
Haisu Wangb81c14c2022-05-30 14:40:59 +0800136 if (rq->part && blk_do_io_stat(rq) &&
137 (!mi->part->bd_partno || rq->part == mi->part) &&
Jeffle Xub0d97552020-12-02 19:11:45 +0800138 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
Pavel Begunkovbb4e6b12019-09-30 21:55:33 +0300139 mi->inflight[rq_data_dir(rq)]++;
Jens Axboe7baa8572018-11-08 10:24:07 -0700140
141 return true;
Jens Axboef299b7c2017-08-08 17:51:45 -0600142}
143
Christoph Hellwig8446fe92020-11-24 09:36:54 +0100144unsigned int blk_mq_in_flight(struct request_queue *q,
145 struct block_device *part)
Jens Axboef299b7c2017-08-08 17:51:45 -0600146{
Pavel Begunkova2e80f62019-09-30 21:55:34 +0300147 struct mq_inflight mi = { .part = part };
Jens Axboef299b7c2017-08-08 17:51:45 -0600148
Jens Axboef299b7c2017-08-08 17:51:45 -0600149 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
Mikulas Patockae016b782018-12-06 11:41:21 -0500150
Pavel Begunkova2e80f62019-09-30 21:55:34 +0300151 return mi.inflight[0] + mi.inflight[1];
Omar Sandovalbf0ddab2018-04-26 00:21:59 -0700152}
153
Christoph Hellwig8446fe92020-11-24 09:36:54 +0100154void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
155 unsigned int inflight[2])
Omar Sandovalbf0ddab2018-04-26 00:21:59 -0700156{
Pavel Begunkova2e80f62019-09-30 21:55:34 +0300157 struct mq_inflight mi = { .part = part };
Omar Sandovalbf0ddab2018-04-26 00:21:59 -0700158
Pavel Begunkovbb4e6b12019-09-30 21:55:33 +0300159 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
Pavel Begunkova2e80f62019-09-30 21:55:34 +0300160 inflight[0] = mi.inflight[0];
161 inflight[1] = mi.inflight[1];
Omar Sandovalbf0ddab2018-04-26 00:21:59 -0700162}
163
Ming Lei1671d522017-03-27 20:06:57 +0800164void blk_freeze_queue_start(struct request_queue *q)
Ming Lei43a5e4e2013-12-26 21:31:35 +0800165{
Bob Liu7996a8b2019-05-21 11:25:55 +0800166 mutex_lock(&q->mq_freeze_lock);
167 if (++q->mq_freeze_depth == 1) {
Dan Williams3ef28e82015-10-21 13:20:12 -0400168 percpu_ref_kill(&q->q_usage_counter);
Bob Liu7996a8b2019-05-21 11:25:55 +0800169 mutex_unlock(&q->mq_freeze_lock);
Jens Axboe344e9ff2018-11-15 12:22:51 -0700170 if (queue_is_mq(q))
Ming Lei055f6e12017-11-09 10:49:53 -0800171 blk_mq_run_hw_queues(q, false);
Bob Liu7996a8b2019-05-21 11:25:55 +0800172 } else {
173 mutex_unlock(&q->mq_freeze_lock);
Tejun Heocddd5d12014-08-16 08:02:24 -0400174 }
Tejun Heof3af0202014-11-04 13:52:27 -0500175}
Ming Lei1671d522017-03-27 20:06:57 +0800176EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
Tejun Heof3af0202014-11-04 13:52:27 -0500177
Keith Busch6bae363e2017-03-01 14:22:10 -0500178void blk_mq_freeze_queue_wait(struct request_queue *q)
Tejun Heof3af0202014-11-04 13:52:27 -0500179{
Dan Williams3ef28e82015-10-21 13:20:12 -0400180 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
Ming Lei43a5e4e2013-12-26 21:31:35 +0800181}
Keith Busch6bae363e2017-03-01 14:22:10 -0500182EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
Ming Lei43a5e4e2013-12-26 21:31:35 +0800183
Keith Buschf91328c2017-03-01 14:22:11 -0500184int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
185 unsigned long timeout)
186{
187 return wait_event_timeout(q->mq_freeze_wq,
188 percpu_ref_is_zero(&q->q_usage_counter),
189 timeout);
190}
191EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
Jens Axboe320ae512013-10-24 09:20:05 +0100192
Tejun Heof3af0202014-11-04 13:52:27 -0500193/*
194 * Guarantee no request is in use, so we can change any data structure of
195 * the queue afterward.
196 */
Dan Williams3ef28e82015-10-21 13:20:12 -0400197void blk_freeze_queue(struct request_queue *q)
Tejun Heof3af0202014-11-04 13:52:27 -0500198{
Dan Williams3ef28e82015-10-21 13:20:12 -0400199 /*
200 * In the !blk_mq case we are only calling this to kill the
201 * q_usage_counter, otherwise this increases the freeze depth
202 * and waits for it to return to zero. For this reason there is
203 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
204 * exported to drivers as the only user for unfreeze is blk_mq.
205 */
Ming Lei1671d522017-03-27 20:06:57 +0800206 blk_freeze_queue_start(q);
Tejun Heof3af0202014-11-04 13:52:27 -0500207 blk_mq_freeze_queue_wait(q);
208}
Dan Williams3ef28e82015-10-21 13:20:12 -0400209
210void blk_mq_freeze_queue(struct request_queue *q)
211{
212 /*
213 * ...just an alias to keep freeze and unfreeze actions balanced
214 * in the blk_mq_* namespace
215 */
216 blk_freeze_queue(q);
217}
Jens Axboec761d962015-01-02 15:05:12 -0700218EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
Tejun Heof3af0202014-11-04 13:52:27 -0500219
Christoph Hellwigaec89dc2021-09-29 09:12:41 +0200220void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
Jens Axboe320ae512013-10-24 09:20:05 +0100221{
Bob Liu7996a8b2019-05-21 11:25:55 +0800222 mutex_lock(&q->mq_freeze_lock);
Christoph Hellwigaec89dc2021-09-29 09:12:41 +0200223 if (force_atomic)
224 q->q_usage_counter.data->force_atomic = true;
Bob Liu7996a8b2019-05-21 11:25:55 +0800225 q->mq_freeze_depth--;
226 WARN_ON_ONCE(q->mq_freeze_depth < 0);
227 if (!q->mq_freeze_depth) {
Bart Van Asschebdd63162018-09-26 14:01:08 -0700228 percpu_ref_resurrect(&q->q_usage_counter);
Jens Axboe320ae512013-10-24 09:20:05 +0100229 wake_up_all(&q->mq_freeze_wq);
Tejun Heoadd703f2014-07-01 10:34:38 -0600230 }
Bob Liu7996a8b2019-05-21 11:25:55 +0800231 mutex_unlock(&q->mq_freeze_lock);
Jens Axboe320ae512013-10-24 09:20:05 +0100232}
Christoph Hellwigaec89dc2021-09-29 09:12:41 +0200233
234void blk_mq_unfreeze_queue(struct request_queue *q)
235{
236 __blk_mq_unfreeze_queue(q, false);
237}
Keith Buschb4c6a022014-12-19 17:54:14 -0700238EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
Jens Axboe320ae512013-10-24 09:20:05 +0100239
Bart Van Assche852ec802017-06-21 10:55:47 -0700240/*
241 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
242 * mpt3sas driver such that this function can be removed.
243 */
244void blk_mq_quiesce_queue_nowait(struct request_queue *q)
245{
Ming Leie70feb82021-10-14 16:17:10 +0800246 unsigned long flags;
247
248 spin_lock_irqsave(&q->queue_lock, flags);
249 if (!q->quiesce_depth++)
250 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
251 spin_unlock_irqrestore(&q->queue_lock, flags);
Bart Van Assche852ec802017-06-21 10:55:47 -0700252}
253EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
254
Bart Van Assche6a83e742016-11-02 10:09:51 -0600255/**
Ming Lei9ef4d022021-11-09 15:11:41 +0800256 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
Christoph Hellwig483239c2022-11-01 16:00:48 +0100257 * @set: tag_set to wait on
Ming Lei9ef4d022021-11-09 15:11:41 +0800258 *
259 * Note: it is driver's responsibility for making sure that quiesce has
Christoph Hellwig483239c2022-11-01 16:00:48 +0100260 * been started on or more of the request_queues of the tag_set. This
261 * function only waits for the quiesce on those request_queues that had
262 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
Ming Lei9ef4d022021-11-09 15:11:41 +0800263 */
Christoph Hellwig483239c2022-11-01 16:00:48 +0100264void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
Ming Lei9ef4d022021-11-09 15:11:41 +0800265{
Christoph Hellwig483239c2022-11-01 16:00:48 +0100266 if (set->flags & BLK_MQ_F_BLOCKING)
267 synchronize_srcu(set->srcu);
Ming Lei704b9142021-12-03 21:15:32 +0800268 else
Ming Lei9ef4d022021-11-09 15:11:41 +0800269 synchronize_rcu();
270}
271EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
272
273/**
Ming Lei69e07c42017-06-06 23:22:07 +0800274 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
Bart Van Assche6a83e742016-11-02 10:09:51 -0600275 * @q: request queue.
276 *
277 * Note: this function does not prevent that the struct request end_io()
Ming Lei69e07c42017-06-06 23:22:07 +0800278 * callback function is invoked. Once this function is returned, we make
279 * sure no dispatch can happen until the queue is unquiesced via
280 * blk_mq_unquiesce_queue().
Bart Van Assche6a83e742016-11-02 10:09:51 -0600281 */
282void blk_mq_quiesce_queue(struct request_queue *q)
283{
Ming Lei1d9e9bc2017-06-06 23:22:08 +0800284 blk_mq_quiesce_queue_nowait(q);
Christoph Hellwig85373802022-11-01 16:00:46 +0100285 /* nothing to wait for non-mq queues */
286 if (queue_is_mq(q))
Christoph Hellwig483239c2022-11-01 16:00:48 +0100287 blk_mq_wait_quiesce_done(q->tag_set);
Bart Van Assche6a83e742016-11-02 10:09:51 -0600288}
289EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
290
Ming Leie4e73912017-06-06 23:22:03 +0800291/*
292 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
293 * @q: request queue.
294 *
295 * This function recovers queue into the state before quiescing
296 * which is done by blk_mq_quiesce_queue.
297 */
298void blk_mq_unquiesce_queue(struct request_queue *q)
299{
Ming Leie70feb82021-10-14 16:17:10 +0800300 unsigned long flags;
301 bool run_queue = false;
302
303 spin_lock_irqsave(&q->queue_lock, flags);
304 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
305 ;
306 } else if (!--q->quiesce_depth) {
307 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
308 run_queue = true;
309 }
310 spin_unlock_irqrestore(&q->queue_lock, flags);
Ming Leif4560ff2017-06-18 14:24:27 -0600311
Ming Lei1d9e9bc2017-06-06 23:22:08 +0800312 /* dispatch requests which are inserted during quiescing */
Ming Leie70feb82021-10-14 16:17:10 +0800313 if (run_queue)
314 blk_mq_run_hw_queues(q, true);
Ming Leie4e73912017-06-06 23:22:03 +0800315}
316EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
317
Chao Leng414dd482022-11-01 16:00:49 +0100318void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
319{
320 struct request_queue *q;
321
322 mutex_lock(&set->tag_list_lock);
323 list_for_each_entry(q, &set->tag_list, tag_set_list) {
324 if (!blk_queue_skip_tagset_quiesce(q))
325 blk_mq_quiesce_queue_nowait(q);
326 }
327 blk_mq_wait_quiesce_done(set);
328 mutex_unlock(&set->tag_list_lock);
329}
330EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
331
332void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
333{
334 struct request_queue *q;
335
336 mutex_lock(&set->tag_list_lock);
337 list_for_each_entry(q, &set->tag_list, tag_set_list) {
338 if (!blk_queue_skip_tagset_quiesce(q))
339 blk_mq_unquiesce_queue(q);
340 }
341 mutex_unlock(&set->tag_list_lock);
342}
343EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
344
Jens Axboeaed3ea92014-12-22 14:04:42 -0700345void blk_mq_wake_waiters(struct request_queue *q)
346{
347 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +0800348 unsigned long i;
Jens Axboeaed3ea92014-12-22 14:04:42 -0700349
350 queue_for_each_hw_ctx(q, hctx, i)
351 if (blk_mq_hw_queue_mapped(hctx))
352 blk_mq_tag_wakeup_all(hctx->tags, true);
353}
354
Christoph Hellwig52fdbbc2021-11-17 07:13:59 +0100355void blk_rq_init(struct request_queue *q, struct request *rq)
356{
357 memset(rq, 0, sizeof(*rq));
358
359 INIT_LIST_HEAD(&rq->queuelist);
360 rq->q = q;
361 rq->__sector = (sector_t) -1;
362 INIT_HLIST_NODE(&rq->hash);
363 RB_CLEAR_NODE(&rq->rb_node);
364 rq->tag = BLK_MQ_NO_TAG;
365 rq->internal_tag = BLK_MQ_NO_TAG;
366 rq->start_time_ns = ktime_get_ns();
367 rq->part = NULL;
368 blk_crypto_rq_set_defaults(rq);
369}
370EXPORT_SYMBOL(blk_rq_init);
371
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200372static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
Jens Axboefe6134f2021-10-19 09:32:58 -0600373 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
Jens Axboe320ae512013-10-24 09:20:05 +0100374{
Pavel Begunkov605f7842021-10-18 21:37:28 +0100375 struct blk_mq_ctx *ctx = data->ctx;
376 struct blk_mq_hw_ctx *hctx = data->hctx;
377 struct request_queue *q = data->q;
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200378 struct request *rq = tags->static_rqs[tag];
Bart Van Asschec3a148d2017-06-20 11:15:43 -0700379
Jens Axboec7b84d42021-10-19 09:33:00 -0600380 rq->q = q;
381 rq->mq_ctx = ctx;
382 rq->mq_hctx = hctx;
383 rq->cmd_flags = data->cmd_flags;
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200384
Pavel Begunkov12845902021-10-18 21:37:29 +0100385 if (data->flags & BLK_MQ_REQ_PM)
Jens Axboe56f8da62021-10-19 09:32:57 -0600386 data->rq_flags |= RQF_PM;
Pavel Begunkov12845902021-10-18 21:37:29 +0100387 if (blk_queue_io_stat(q))
Jens Axboe56f8da62021-10-19 09:32:57 -0600388 data->rq_flags |= RQF_IO_STAT;
389 rq->rq_flags = data->rq_flags;
Pavel Begunkov12845902021-10-18 21:37:29 +0100390
Jens Axboec7b84d42021-10-19 09:33:00 -0600391 if (!(data->rq_flags & RQF_ELV)) {
Jens Axboe320ae512013-10-24 09:20:05 +0100392 rq->tag = tag;
393 rq->internal_tag = BLK_MQ_NO_TAG;
Jens Axboec7b84d42021-10-19 09:33:00 -0600394 } else {
395 rq->tag = BLK_MQ_NO_TAG;
396 rq->internal_tag = tag;
Jens Axboe320ae512013-10-24 09:20:05 +0100397 }
Jens Axboec7b84d42021-10-19 09:33:00 -0600398 rq->timeout = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100399
Pavel Begunkov4f266f22021-10-18 21:37:27 +0100400 if (blk_mq_need_time_stamp(rq))
401 rq->start_time_ns = ktime_get_ns();
402 else
403 rq->start_time_ns = 0;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200404 rq->part = NULL;
Tejun Heo6f816b42019-08-28 15:05:57 -0700405#ifdef CONFIG_BLK_RQ_ALLOC_TIME
406 rq->alloc_time_ns = alloc_time_ns;
407#endif
Omar Sandoval544ccc8d2018-05-09 02:08:50 -0700408 rq->io_start_time_ns = 0;
Hou Tao3d244302019-05-21 15:59:03 +0800409 rq->stats_sectors = 0;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200410 rq->nr_phys_segments = 0;
411#if defined(CONFIG_BLK_DEV_INTEGRITY)
412 rq->nr_integrity_segments = 0;
413#endif
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200414 rq->end_io = NULL;
415 rq->end_io_data = NULL;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200416
Pavel Begunkov4f266f22021-10-18 21:37:27 +0100417 blk_crypto_rq_set_defaults(rq);
418 INIT_LIST_HEAD(&rq->queuelist);
419 /* tag was already set */
420 WRITE_ONCE(rq->deadline, 0);
Jens Axboe0a467d02021-10-14 14:39:59 -0600421 req_ref_set(rq, 1);
Christoph Hellwig7ea4d8a2020-05-29 15:53:10 +0200422
Pavel Begunkov4f266f22021-10-18 21:37:27 +0100423 if (rq->rq_flags & RQF_ELV) {
Christoph Hellwig7ea4d8a2020-05-29 15:53:10 +0200424 struct elevator_queue *e = data->q->elevator;
425
Pavel Begunkov4f266f22021-10-18 21:37:27 +0100426 INIT_HLIST_NODE(&rq->hash);
427 RB_CLEAR_NODE(&rq->rb_node);
428
429 if (!op_is_flush(data->cmd_flags) &&
430 e->type->ops.prepare_request) {
Christoph Hellwig7ea4d8a2020-05-29 15:53:10 +0200431 e->type->ops.prepare_request(rq);
432 rq->rq_flags |= RQF_ELVPRIV;
433 }
434 }
435
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200436 return rq;
Jens Axboe320ae512013-10-24 09:20:05 +0100437}
438
Jens Axboe349302d2021-10-09 13:10:39 -0600439static inline struct request *
440__blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
441 u64 alloc_time_ns)
442{
443 unsigned int tag, tag_offset;
Jens Axboefe6134f2021-10-19 09:32:58 -0600444 struct blk_mq_tags *tags;
Jens Axboe349302d2021-10-09 13:10:39 -0600445 struct request *rq;
Jens Axboefe6134f2021-10-19 09:32:58 -0600446 unsigned long tag_mask;
Jens Axboe349302d2021-10-09 13:10:39 -0600447 int i, nr = 0;
448
Jens Axboefe6134f2021-10-19 09:32:58 -0600449 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
450 if (unlikely(!tag_mask))
Jens Axboe349302d2021-10-09 13:10:39 -0600451 return NULL;
452
Jens Axboefe6134f2021-10-19 09:32:58 -0600453 tags = blk_mq_tags_from_data(data);
454 for (i = 0; tag_mask; i++) {
455 if (!(tag_mask & (1UL << i)))
Jens Axboe349302d2021-10-09 13:10:39 -0600456 continue;
457 tag = tag_offset + i;
Jens Axboea22c00b2021-11-01 06:56:09 -0600458 prefetch(tags->static_rqs[tag]);
Jens Axboefe6134f2021-10-19 09:32:58 -0600459 tag_mask &= ~(1UL << i);
460 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
Jens Axboe013a7f92021-10-13 07:58:52 -0600461 rq_list_add(data->cached_rq, rq);
Jens Axboec5fc7b92021-11-03 05:49:07 -0600462 nr++;
Jens Axboe349302d2021-10-09 13:10:39 -0600463 }
Jens Axboec5fc7b92021-11-03 05:49:07 -0600464 /* caller already holds a reference, add for remainder */
465 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
Jens Axboe349302d2021-10-09 13:10:39 -0600466 data->nr_tags -= nr;
467
Jens Axboe013a7f92021-10-13 07:58:52 -0600468 return rq_list_pop(data->cached_rq);
Jens Axboe349302d2021-10-09 13:10:39 -0600469}
470
Christoph Hellwigb90cfae2021-10-12 12:40:44 +0200471static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200472{
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200473 struct request_queue *q = data->q;
Tejun Heo6f816b42019-08-28 15:05:57 -0700474 u64 alloc_time_ns = 0;
Jens Axboe47c122e2021-10-06 06:34:11 -0600475 struct request *rq;
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200476 unsigned int tag;
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200477
Tejun Heo6f816b42019-08-28 15:05:57 -0700478 /* alloc_time includes depth and tag waits */
479 if (blk_queue_rq_alloc_time(q))
480 alloc_time_ns = ktime_get_ns();
481
Jens Axboef9afca42018-10-29 13:11:38 -0600482 if (data->cmd_flags & REQ_NOWAIT)
Goldwyn Rodrigues03a07c92017-06-20 07:05:46 -0500483 data->flags |= BLK_MQ_REQ_NOWAIT;
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200484
Jens Axboe781dd832021-11-02 08:34:09 -0600485 if (q->elevator) {
486 struct elevator_queue *e = q->elevator;
487
488 data->rq_flags |= RQF_ELV;
489
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200490 /*
Lin Feng8d663f32021-04-15 11:39:20 +0800491 * Flush/passthrough requests are special and go directly to the
Jens Axboe17a51192018-05-09 13:28:50 -0600492 * dispatch list. Don't include reserved tags in the
493 * limiting, as it isn't useful.
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200494 */
Jens Axboef9afca42018-10-29 13:11:38 -0600495 if (!op_is_flush(data->cmd_flags) &&
Lin Feng8d663f32021-04-15 11:39:20 +0800496 !blk_op_is_passthrough(data->cmd_flags) &&
Jens Axboef9afca42018-10-29 13:11:38 -0600497 e->type->ops.limit_depth &&
Jens Axboe17a51192018-05-09 13:28:50 -0600498 !(data->flags & BLK_MQ_REQ_RESERVED))
Jens Axboef9afca42018-10-29 13:11:38 -0600499 e->type->ops.limit_depth(data->cmd_flags, data);
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200500 }
501
Ming Leibf0beec2020-05-29 15:53:15 +0200502retry:
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200503 data->ctx = blk_mq_get_ctx(q);
504 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
Jens Axboe781dd832021-11-02 08:34:09 -0600505 if (!(data->rq_flags & RQF_ELV))
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200506 blk_mq_tag_busy(data->hctx);
507
John Garry99e48cd2022-07-06 20:03:50 +0800508 if (data->flags & BLK_MQ_REQ_RESERVED)
509 data->rq_flags |= RQF_RESV;
510
Ming Leibf0beec2020-05-29 15:53:15 +0200511 /*
Jens Axboe349302d2021-10-09 13:10:39 -0600512 * Try batched alloc if we want more than 1 tag.
513 */
514 if (data->nr_tags > 1) {
515 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
516 if (rq)
517 return rq;
518 data->nr_tags = 1;
519 }
520
521 /*
Ming Leibf0beec2020-05-29 15:53:15 +0200522 * Waiting allocations only fail because of an inactive hctx. In that
523 * case just retry the hctx assignment and tag allocation as CPU hotplug
524 * should have migrated us to an online CPU by now.
525 */
Christoph Hellwige4cdf1a2017-06-16 18:15:27 +0200526 tag = blk_mq_get_tag(data);
Ming Leibf0beec2020-05-29 15:53:15 +0200527 if (tag == BLK_MQ_NO_TAG) {
528 if (data->flags & BLK_MQ_REQ_NOWAIT)
529 return NULL;
Ming Leibf0beec2020-05-29 15:53:15 +0200530 /*
Jens Axboe349302d2021-10-09 13:10:39 -0600531 * Give up the CPU and sleep for a random short time to
532 * ensure that thread using a realtime scheduling class
533 * are migrated off the CPU, and thus off the hctx that
534 * is going away.
Ming Leibf0beec2020-05-29 15:53:15 +0200535 */
536 msleep(3);
537 goto retry;
538 }
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200539
Jens Axboefe6134f2021-10-19 09:32:58 -0600540 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
541 alloc_time_ns);
Christoph Hellwigd2c0d382017-06-16 18:15:19 +0200542}
543
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600544static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
545 struct blk_plug *plug,
546 blk_opf_t opf,
547 blk_mq_req_flags_t flags)
Jens Axboe320ae512013-10-24 09:20:05 +0100548{
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200549 struct blk_mq_alloc_data data = {
550 .q = q,
551 .flags = flags,
Bart Van Assche16458cf2022-07-14 11:06:32 -0700552 .cmd_flags = opf,
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600553 .nr_tags = plug->nr_ios,
554 .cached_rq = &plug->cached_rq,
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200555 };
Jens Axboebd166ef2017-01-17 06:03:22 -0700556 struct request *rq;
Jens Axboe320ae512013-10-24 09:20:05 +0100557
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600558 if (blk_queue_enter(q, flags))
559 return NULL;
560
561 plug->nr_ios = 1;
Jens Axboe320ae512013-10-24 09:20:05 +0100562
Christoph Hellwigb90cfae2021-10-12 12:40:44 +0200563 rq = __blk_mq_alloc_requests(&data);
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600564 if (unlikely(!rq))
565 blk_queue_exit(q);
566 return rq;
567}
568
569static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
570 blk_opf_t opf,
571 blk_mq_req_flags_t flags)
572{
573 struct blk_plug *plug = current->plug;
574 struct request *rq;
575
576 if (!plug)
577 return NULL;
Jinlong Chen40467282022-11-02 10:52:30 +0800578
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600579 if (rq_list_empty(plug->cached_rq)) {
580 if (plug->nr_ios == 1)
581 return NULL;
582 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
Jinlong Chen40467282022-11-02 10:52:30 +0800583 if (!rq)
584 return NULL;
585 } else {
586 rq = rq_list_peek(&plug->cached_rq);
587 if (!rq || rq->q != q)
588 return NULL;
589
590 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
591 return NULL;
592 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
593 return NULL;
594
595 plug->cached_rq = rq_list_next(rq);
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600596 }
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600597
Jens Axboe4b6a5d92022-09-21 08:22:09 -0600598 rq->cmd_flags = opf;
599 INIT_LIST_HEAD(&rq->queuelist);
600 return rq;
601}
602
603struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
604 blk_mq_req_flags_t flags)
605{
606 struct request *rq;
607
608 rq = blk_mq_alloc_cached_request(q, opf, flags);
609 if (!rq) {
610 struct blk_mq_alloc_data data = {
611 .q = q,
612 .flags = flags,
613 .cmd_flags = opf,
614 .nr_tags = 1,
615 };
616 int ret;
617
618 ret = blk_queue_enter(q, flags);
619 if (ret)
620 return ERR_PTR(ret);
621
622 rq = __blk_mq_alloc_requests(&data);
623 if (!rq)
624 goto out_queue_exit;
625 }
Christoph Hellwig0c4de0f2016-07-19 11:31:50 +0200626 rq->__data_len = 0;
627 rq->__sector = (sector_t) -1;
628 rq->bio = rq->biotail = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +0100629 return rq;
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200630out_queue_exit:
631 blk_queue_exit(q);
632 return ERR_PTR(-EWOULDBLOCK);
Jens Axboe320ae512013-10-24 09:20:05 +0100633}
Jens Axboe4bb659b2014-05-09 09:36:49 -0600634EXPORT_SYMBOL(blk_mq_alloc_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100635
Bart Van Asschecd6ce142017-06-20 11:15:39 -0700636struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
Bart Van Assche16458cf2022-07-14 11:06:32 -0700637 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
Ming Lin1f5bd332016-06-13 16:45:21 +0200638{
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200639 struct blk_mq_alloc_data data = {
640 .q = q,
641 .flags = flags,
Bart Van Assche16458cf2022-07-14 11:06:32 -0700642 .cmd_flags = opf,
Jens Axboe47c122e2021-10-06 06:34:11 -0600643 .nr_tags = 1,
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200644 };
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200645 u64 alloc_time_ns = 0;
John Garrye3c5a782022-10-26 18:35:13 +0800646 struct request *rq;
Omar Sandoval6d2809d2017-02-27 10:28:27 -0800647 unsigned int cpu;
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200648 unsigned int tag;
Ming Lin1f5bd332016-06-13 16:45:21 +0200649 int ret;
650
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200651 /* alloc_time includes depth and tag waits */
652 if (blk_queue_rq_alloc_time(q))
653 alloc_time_ns = ktime_get_ns();
654
Ming Lin1f5bd332016-06-13 16:45:21 +0200655 /*
656 * If the tag allocator sleeps we could get an allocation for a
657 * different hardware context. No need to complicate the low level
658 * allocator for this for the rare use case of a command tied to
659 * a specific queue.
660 */
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200661 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
Ming Lin1f5bd332016-06-13 16:45:21 +0200662 return ERR_PTR(-EINVAL);
663
664 if (hctx_idx >= q->nr_hw_queues)
665 return ERR_PTR(-EIO);
666
Bart Van Assche3a0a5292017-11-09 10:49:58 -0800667 ret = blk_queue_enter(q, flags);
Ming Lin1f5bd332016-06-13 16:45:21 +0200668 if (ret)
669 return ERR_PTR(ret);
670
Christoph Hellwigc8712c62016-09-23 10:25:48 -0600671 /*
672 * Check if the hardware context is actually mapped to anything.
673 * If not tell the caller that it should skip this queue.
674 */
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200675 ret = -EXDEV;
Ming Lei4e5cc992022-03-08 15:32:19 +0800676 data.hctx = xa_load(&q->hctx_table, hctx_idx);
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200677 if (!blk_mq_hw_queue_mapped(data.hctx))
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200678 goto out_queue_exit;
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200679 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
Bart Van Assche14dc7a12022-06-15 14:00:04 -0700680 if (cpu >= nr_cpu_ids)
681 goto out_queue_exit;
Christoph Hellwige6e7abf2020-05-29 15:53:09 +0200682 data.ctx = __blk_mq_get_ctx(q, cpu);
Ming Lin1f5bd332016-06-13 16:45:21 +0200683
Christoph Hellwig42fdc5e2020-06-29 17:08:34 +0200684 if (!q->elevator)
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200685 blk_mq_tag_busy(data.hctx);
Jens Axboe781dd832021-11-02 08:34:09 -0600686 else
687 data.rq_flags |= RQF_ELV;
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200688
John Garry99e48cd2022-07-06 20:03:50 +0800689 if (flags & BLK_MQ_REQ_RESERVED)
690 data.rq_flags |= RQF_RESV;
691
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200692 ret = -EWOULDBLOCK;
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200693 tag = blk_mq_get_tag(&data);
694 if (tag == BLK_MQ_NO_TAG)
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200695 goto out_queue_exit;
John Garrye3c5a782022-10-26 18:35:13 +0800696 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
Jens Axboefe6134f2021-10-19 09:32:58 -0600697 alloc_time_ns);
John Garrye3c5a782022-10-26 18:35:13 +0800698 rq->__data_len = 0;
699 rq->__sector = (sector_t) -1;
700 rq->bio = rq->biotail = NULL;
701 return rq;
Christoph Hellwig600c3b02020-05-29 15:53:13 +0200702
Christoph Hellwiga5ea581102020-05-16 20:27:58 +0200703out_queue_exit:
704 blk_queue_exit(q);
705 return ERR_PTR(ret);
Ming Lin1f5bd332016-06-13 16:45:21 +0200706}
707EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
708
Keith Busch12f5b932018-05-29 15:52:28 +0200709static void __blk_mq_free_request(struct request *rq)
710{
711 struct request_queue *q = rq->q;
712 struct blk_mq_ctx *ctx = rq->mq_ctx;
Jens Axboeea4f9952018-10-29 15:06:13 -0600713 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Keith Busch12f5b932018-05-29 15:52:28 +0200714 const int sched_tag = rq->internal_tag;
715
Satya Tangiralaa892c8d2020-05-14 00:37:18 +0000716 blk_crypto_free_request(rq);
Bart Van Assche986d4132018-09-26 14:01:10 -0700717 blk_pm_mark_last_busy(rq);
Jens Axboeea4f9952018-10-29 15:06:13 -0600718 rq->mq_hctx = NULL;
Christoph Hellwig766473682020-05-29 15:53:12 +0200719 if (rq->tag != BLK_MQ_NO_TAG)
John Garrycae740a2020-02-26 20:10:15 +0800720 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
Christoph Hellwig766473682020-05-29 15:53:12 +0200721 if (sched_tag != BLK_MQ_NO_TAG)
John Garrycae740a2020-02-26 20:10:15 +0800722 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
Keith Busch12f5b932018-05-29 15:52:28 +0200723 blk_mq_sched_restart(hctx);
724 blk_queue_exit(q);
725}
726
Christoph Hellwig6af54052017-06-16 18:15:22 +0200727void blk_mq_free_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100728{
Jens Axboe320ae512013-10-24 09:20:05 +0100729 struct request_queue *q = rq->q;
Jens Axboeea4f9952018-10-29 15:06:13 -0600730 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Jens Axboe320ae512013-10-24 09:20:05 +0100731
Christoph Hellwig222ee582021-11-26 12:58:11 +0100732 if ((rq->rq_flags & RQF_ELVPRIV) &&
733 q->elevator->type->ops.finish_request)
734 q->elevator->type->ops.finish_request(rq);
Christoph Hellwig6af54052017-06-16 18:15:22 +0200735
Christoph Hellwige8064022016-10-20 15:12:13 +0200736 if (rq->rq_flags & RQF_MQ_INFLIGHT)
John Garrybccf5e22020-08-19 23:20:26 +0800737 __blk_mq_dec_active_requests(hctx);
Jens Axboe87760e52016-11-09 12:38:14 -0700738
Jens Axboe7beb2f82017-09-30 02:08:24 -0600739 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
Christoph Hellwigd152c682021-08-16 15:46:24 +0200740 laptop_io_completion(q->disk->bdi);
Jens Axboe7beb2f82017-09-30 02:08:24 -0600741
Josef Bacika7905042018-07-03 09:32:35 -0600742 rq_qos_done(q, rq);
Jens Axboe0d2602c2014-05-13 15:10:52 -0600743
Keith Busch12f5b932018-05-29 15:52:28 +0200744 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Jens Axboe0a467d02021-10-14 14:39:59 -0600745 if (req_ref_put_and_test(rq))
Keith Busch12f5b932018-05-29 15:52:28 +0200746 __blk_mq_free_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100747}
Jens Axboe1a3b5952014-11-17 10:40:48 -0700748EXPORT_SYMBOL_GPL(blk_mq_free_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100749
Jens Axboe47c122e2021-10-06 06:34:11 -0600750void blk_mq_free_plug_rqs(struct blk_plug *plug)
Jens Axboe320ae512013-10-24 09:20:05 +0100751{
Jens Axboe013a7f92021-10-13 07:58:52 -0600752 struct request *rq;
Jens Axboefe1f4522018-11-28 10:50:07 -0700753
Jens Axboec5fc7b92021-11-03 05:49:07 -0600754 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
Jens Axboe47c122e2021-10-06 06:34:11 -0600755 blk_mq_free_request(rq);
Jens Axboe47c122e2021-10-06 06:34:11 -0600756}
Omar Sandoval522a7772018-05-09 02:08:53 -0700757
Christoph Hellwig22350ad2021-11-17 07:14:02 +0100758void blk_dump_rq_flags(struct request *rq, char *msg)
759{
760 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
Christoph Hellwigf3fa33a2021-11-26 13:18:00 +0100761 rq->q->disk ? rq->q->disk->disk_name : "?",
Bart Van Assche16458cf2022-07-14 11:06:32 -0700762 (__force unsigned long long) rq->cmd_flags);
Christoph Hellwig22350ad2021-11-17 07:14:02 +0100763
764 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
765 (unsigned long long)blk_rq_pos(rq),
766 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
767 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
768 rq->bio, rq->biotail, blk_rq_bytes(rq));
769}
770EXPORT_SYMBOL(blk_dump_rq_flags);
771
Jens Axboe9be3e062021-10-14 09:17:01 -0600772static void req_bio_endio(struct request *rq, struct bio *bio,
773 unsigned int nbytes, blk_status_t error)
774{
Pavel Begunkov478eb722021-10-19 22:24:12 +0100775 if (unlikely(error)) {
Jens Axboe9be3e062021-10-14 09:17:01 -0600776 bio->bi_status = error;
Pavel Begunkov478eb722021-10-19 22:24:12 +0100777 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
Jens Axboe9be3e062021-10-14 09:17:01 -0600778 /*
779 * Partial zone append completions cannot be supported as the
780 * BIO fragments may end up not being written sequentially.
781 */
Pavel Begunkov297db732021-10-22 16:01:44 +0100782 if (bio->bi_iter.bi_size != nbytes)
Jens Axboe9be3e062021-10-14 09:17:01 -0600783 bio->bi_status = BLK_STS_IOERR;
784 else
785 bio->bi_iter.bi_sector = rq->__sector;
786 }
787
Pavel Begunkov478eb722021-10-19 22:24:12 +0100788 bio_advance(bio, nbytes);
789
790 if (unlikely(rq->rq_flags & RQF_QUIET))
791 bio_set_flag(bio, BIO_QUIET);
Jens Axboe9be3e062021-10-14 09:17:01 -0600792 /* don't actually finish bio if it's part of flush sequence */
793 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
794 bio_endio(bio);
795}
796
797static void blk_account_io_completion(struct request *req, unsigned int bytes)
798{
799 if (req->part && blk_do_io_stat(req)) {
800 const int sgrp = op_stat_group(req_op(req));
801
802 part_stat_lock();
803 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
804 part_stat_unlock();
805 }
806}
807
Christoph Hellwig0d7a29a2021-11-17 07:14:03 +0100808static void blk_print_req_error(struct request *req, blk_status_t status)
809{
810 printk_ratelimited(KERN_ERR
811 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
812 "phys_seg %u prio class %u\n",
813 blk_status_to_str(status),
Christoph Hellwigf3fa33a2021-11-26 13:18:00 +0100814 req->q->disk ? req->q->disk->disk_name : "?",
Bart Van Assche16458cf2022-07-14 11:06:32 -0700815 blk_rq_pos(req), (__force u32)req_op(req),
816 blk_op_str(req_op(req)),
817 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
Christoph Hellwig0d7a29a2021-11-17 07:14:03 +0100818 req->nr_phys_segments,
819 IOPRIO_PRIO_CLASS(req->ioprio));
820}
821
Jens Axboe5581a5d2021-12-01 15:01:51 -0700822/*
823 * Fully end IO on a request. Does not support partial completions, or
824 * errors.
825 */
826static void blk_complete_request(struct request *req)
827{
828 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
829 int total_bytes = blk_rq_bytes(req);
830 struct bio *bio = req->bio;
831
832 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
833
834 if (!bio)
835 return;
836
837#ifdef CONFIG_BLK_DEV_INTEGRITY
838 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
839 req->q->integrity.profile->complete_fn(req, total_bytes);
840#endif
841
842 blk_account_io_completion(req, total_bytes);
843
844 do {
845 struct bio *next = bio->bi_next;
846
847 /* Completion has already been traced */
848 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
Pankaj Raghava12821d2022-02-11 10:34:25 +0100849
850 if (req_op(req) == REQ_OP_ZONE_APPEND)
851 bio->bi_iter.bi_sector = req->__sector;
852
Jens Axboe5581a5d2021-12-01 15:01:51 -0700853 if (!is_flush)
854 bio_endio(bio);
855 bio = next;
856 } while (bio);
857
858 /*
859 * Reset counters so that the request stacking driver
860 * can find how many bytes remain in the request
861 * later.
862 */
Jens Axboeab3e1d32022-09-21 08:24:16 -0600863 if (!req->end_io) {
864 req->bio = NULL;
865 req->__data_len = 0;
866 }
Jens Axboe5581a5d2021-12-01 15:01:51 -0700867}
868
Jens Axboe9be3e062021-10-14 09:17:01 -0600869/**
870 * blk_update_request - Complete multiple bytes without completing the request
871 * @req: the request being processed
872 * @error: block status code
873 * @nr_bytes: number of bytes to complete for @req
874 *
875 * Description:
876 * Ends I/O on a number of bytes attached to @req, but doesn't complete
877 * the request structure even if @req doesn't have leftover.
878 * If @req has leftover, sets it up for the next range of segments.
879 *
880 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
881 * %false return from this function.
882 *
883 * Note:
884 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
885 * except in the consistency check at the end of this function.
886 *
887 * Return:
888 * %false - this request doesn't have any more data
889 * %true - this request has more data
890 **/
891bool blk_update_request(struct request *req, blk_status_t error,
892 unsigned int nr_bytes)
893{
894 int total_bytes;
895
Christoph Hellwig8a7d2672021-10-18 10:45:18 +0200896 trace_block_rq_complete(req, error, nr_bytes);
Jens Axboe9be3e062021-10-14 09:17:01 -0600897
898 if (!req->bio)
899 return false;
900
901#ifdef CONFIG_BLK_DEV_INTEGRITY
902 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
903 error == BLK_STS_OK)
904 req->q->integrity.profile->complete_fn(req, nr_bytes);
905#endif
906
907 if (unlikely(error && !blk_rq_is_passthrough(req) &&
Christoph Hellwig3d973a72022-03-23 17:38:15 +0100908 !(req->rq_flags & RQF_QUIET)) &&
909 !test_bit(GD_DEAD, &req->q->disk->state)) {
Jens Axboe9be3e062021-10-14 09:17:01 -0600910 blk_print_req_error(req, error);
Yang Shid5869fd2022-02-10 14:52:22 -0800911 trace_block_rq_error(req, error, nr_bytes);
912 }
Jens Axboe9be3e062021-10-14 09:17:01 -0600913
914 blk_account_io_completion(req, nr_bytes);
915
916 total_bytes = 0;
917 while (req->bio) {
918 struct bio *bio = req->bio;
919 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
920
921 if (bio_bytes == bio->bi_iter.bi_size)
922 req->bio = bio->bi_next;
923
924 /* Completion has already been traced */
925 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
926 req_bio_endio(req, bio, bio_bytes, error);
927
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 void __blk_account_io_done(struct request *req, u64 now)
979{
980 const int sgrp = op_stat_group(req_op(req));
981
982 part_stat_lock();
983 update_io_ticks(req->part, jiffies, true);
984 part_stat_inc(req->part, ios[sgrp]);
985 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
986 part_stat_unlock();
987}
988
989static inline void blk_account_io_done(struct request *req, u64 now)
990{
991 /*
992 * Account IO completion. flush_rq isn't accounted as a
993 * normal IO on queueing nor completion. Accounting the
994 * containing request is enough.
995 */
996 if (blk_do_io_stat(req) && req->part &&
997 !(req->rq_flags & RQF_FLUSH_SEQ))
998 __blk_account_io_done(req, now);
999}
1000
1001static void __blk_account_io_start(struct request *rq)
1002{
Christoph Hellwig41fa7222022-03-08 06:51:47 +01001003 /*
1004 * All non-passthrough requests are created from a bio with one
1005 * exception: when a flush command that is part of a flush sequence
1006 * generated by the state machine in blk-flush.c is cloned onto the
1007 * lower device by dm-multipath we can get here without a bio.
1008 */
1009 if (rq->bio)
Christoph Hellwig450b7872021-11-17 07:14:01 +01001010 rq->part = rq->bio->bi_bdev;
Christoph Hellwig41fa7222022-03-08 06:51:47 +01001011 else
Christoph Hellwigf3fa33a2021-11-26 13:18:00 +01001012 rq->part = rq->q->disk->part0;
Christoph Hellwig450b7872021-11-17 07:14:01 +01001013
1014 part_stat_lock();
1015 update_io_ticks(rq->part, jiffies, false);
1016 part_stat_unlock();
1017}
1018
1019static inline void blk_account_io_start(struct request *req)
1020{
1021 if (blk_do_io_stat(req))
1022 __blk_account_io_start(req);
1023}
1024
Jens Axboef794f332021-10-08 05:50:46 -06001025static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1026{
Omar Sandoval4bc63392018-05-09 02:08:52 -07001027 if (rq->rq_flags & RQF_STATS) {
1028 blk_mq_poll_stats_start(rq->q);
Omar Sandoval522a7772018-05-09 02:08:53 -07001029 blk_stat_add(rq, now);
Omar Sandoval4bc63392018-05-09 02:08:52 -07001030 }
1031
Baolin Wang87890092020-07-04 15:28:21 +08001032 blk_mq_sched_completed_request(rq, now);
Omar Sandoval522a7772018-05-09 02:08:53 -07001033 blk_account_io_done(rq, now);
Jens Axboef794f332021-10-08 05:50:46 -06001034}
1035
Ming Lei0d11e6a2013-12-05 10:50:39 -07001036inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
Christoph Hellwig91b63632014-04-16 09:44:53 +02001037{
Jens Axboef794f332021-10-08 05:50:46 -06001038 if (blk_mq_need_time_stamp(rq))
1039 __blk_mq_end_request_acct(rq, ktime_get_ns());
Jens Axboe320ae512013-10-24 09:20:05 +01001040
1041 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
Ming Lei3b87c6e2021-11-02 23:36:19 +08001066 /*
1067 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1068 * update hctx->nr_active in batch
1069 */
1070 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1071 __blk_mq_sub_active_requests(hctx, nr_tags);
1072
Jens Axboef794f332021-10-08 05:50:46 -06001073 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1074 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1075}
1076
1077void blk_mq_end_request_batch(struct io_comp_batch *iob)
1078{
1079 int tags[TAG_COMP_BATCH], nr_tags = 0;
Jens Axboe02f7eab2021-10-28 12:08:34 -06001080 struct blk_mq_hw_ctx *cur_hctx = NULL;
Jens Axboef794f332021-10-08 05:50:46 -06001081 struct request *rq;
1082 u64 now = 0;
1083
1084 if (iob->need_ts)
1085 now = ktime_get_ns();
1086
1087 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1088 prefetch(rq->bio);
1089 prefetch(rq->rq_next);
1090
Jens Axboe5581a5d2021-12-01 15:01:51 -07001091 blk_complete_request(rq);
Jens Axboef794f332021-10-08 05:50:46 -06001092 if (iob->need_ts)
1093 __blk_mq_end_request_acct(rq, now);
1094
Jens Axboe98b26a02021-11-26 09:53:23 -07001095 rq_qos_done(rq->q, rq);
1096
Jens Axboeab3e1d32022-09-21 08:24:16 -06001097 /*
1098 * If end_io handler returns NONE, then it still has
1099 * ownership of the request.
1100 */
1101 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1102 continue;
1103
Jens Axboef794f332021-10-08 05:50:46 -06001104 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Jens Axboe0a467d02021-10-14 14:39:59 -06001105 if (!req_ref_put_and_test(rq))
Jens Axboef794f332021-10-08 05:50:46 -06001106 continue;
1107
1108 blk_crypto_free_request(rq);
1109 blk_pm_mark_last_busy(rq);
Jens Axboef794f332021-10-08 05:50:46 -06001110
Jens Axboe02f7eab2021-10-28 12:08:34 -06001111 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1112 if (cur_hctx)
1113 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
Jens Axboef794f332021-10-08 05:50:46 -06001114 nr_tags = 0;
Jens Axboe02f7eab2021-10-28 12:08:34 -06001115 cur_hctx = rq->mq_hctx;
Jens Axboef794f332021-10-08 05:50:46 -06001116 }
1117 tags[nr_tags++] = rq->tag;
Jens Axboef794f332021-10-08 05:50:46 -06001118 }
1119
1120 if (nr_tags)
Jens Axboe02f7eab2021-10-28 12:08:34 -06001121 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
Jens Axboef794f332021-10-08 05:50:46 -06001122}
1123EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1124
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001125static void blk_complete_reqs(struct llist_head *list)
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001126{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001127 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1128 struct request *rq, *next;
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001129
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001130 llist_for_each_entry_safe(rq, next, entry, ipi_list)
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001131 rq->q->mq_ops->complete(rq);
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001132}
1133
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001134static __latent_entropy void blk_done_softirq(struct softirq_action *h)
Christoph Hellwig115243f2020-06-11 08:44:42 +02001135{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001136 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001137}
1138
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001139static int blk_softirq_cpu_dead(unsigned int cpu)
1140{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001141 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
Christoph Hellwigc3077b52020-06-11 08:44:41 +02001142 return 0;
1143}
1144
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08001145static void __blk_mq_complete_request_remote(void *data)
Jens Axboe320ae512013-10-24 09:20:05 +01001146{
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001147 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
Jens Axboe320ae512013-10-24 09:20:05 +01001148}
1149
Christoph Hellwig963395262020-06-11 08:44:49 +02001150static inline bool blk_mq_complete_need_ipi(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001151{
Christoph Hellwig963395262020-06-11 08:44:49 +02001152 int cpu = raw_smp_processor_id();
Jens Axboe320ae512013-10-24 09:20:05 +01001153
Christoph Hellwig963395262020-06-11 08:44:49 +02001154 if (!IS_ENABLED(CONFIG_SMP) ||
1155 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1156 return false;
Sebastian Andrzej Siewior71425182020-12-04 20:13:54 +01001157 /*
1158 * With force threaded interrupts enabled, raising softirq from an SMP
1159 * function call will always result in waking the ksoftirqd thread.
1160 * This is probably worse than completing the request on a different
1161 * cache domain.
1162 */
Tanner Love91cc4702021-06-02 14:03:38 -04001163 if (force_irqthreads())
Sebastian Andrzej Siewior71425182020-12-04 20:13:54 +01001164 return false;
Christoph Hellwig963395262020-06-11 08:44:49 +02001165
1166 /* same CPU or cache domain? Complete locally */
1167 if (cpu == rq->mq_ctx->cpu ||
1168 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1169 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1170 return false;
1171
1172 /* don't try to IPI to an offline CPU */
1173 return cpu_online(rq->mq_ctx->cpu);
1174}
1175
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001176static void blk_mq_complete_send_ipi(struct request *rq)
1177{
1178 struct llist_head *list;
1179 unsigned int cpu;
1180
1181 cpu = rq->mq_ctx->cpu;
1182 list = &per_cpu(blk_cpu_done, cpu);
1183 if (llist_add(&rq->ipi_list, list)) {
1184 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1185 smp_call_function_single_async(cpu, &rq->csd);
1186 }
1187}
1188
1189static void blk_mq_raise_softirq(struct request *rq)
1190{
1191 struct llist_head *list;
1192
1193 preempt_disable();
1194 list = this_cpu_ptr(&blk_cpu_done);
1195 if (llist_add(&rq->ipi_list, list))
1196 raise_softirq(BLOCK_SOFTIRQ);
1197 preempt_enable();
1198}
1199
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001200bool blk_mq_complete_request_remote(struct request *rq)
1201{
Keith Buschaf78ff72018-11-26 09:54:30 -07001202 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
Ming Lei36e76532018-09-28 16:42:20 +08001203
Jens Axboe4ab32bf2018-11-18 16:15:35 -07001204 /*
Liu Songf1684202022-09-21 11:32:03 +08001205 * For request which hctx has only one ctx mapping,
1206 * or a polled request, always complete locally,
1207 * it's pointless to redirect the completion.
Jens Axboe4ab32bf2018-11-18 16:15:35 -07001208 */
Liu Songf1684202022-09-21 11:32:03 +08001209 if (rq->mq_hctx->nr_ctx == 1 ||
1210 rq->cmd_flags & REQ_POLLED)
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001211 return false;
Jens Axboe320ae512013-10-24 09:20:05 +01001212
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001213 if (blk_mq_complete_need_ipi(rq)) {
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001214 blk_mq_complete_send_ipi(rq);
1215 return true;
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -08001216 }
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001217
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01001218 if (rq->q->nr_hw_queues == 1) {
1219 blk_mq_raise_softirq(rq);
1220 return true;
1221 }
1222 return false;
Jens Axboe320ae512013-10-24 09:20:05 +01001223}
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001224EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1225
Jens Axboe320ae512013-10-24 09:20:05 +01001226/**
Christoph Hellwig15f73f52020-06-11 08:44:47 +02001227 * blk_mq_complete_request - end I/O on a request
1228 * @rq: the request being processed
Jens Axboe320ae512013-10-24 09:20:05 +01001229 *
Christoph Hellwig15f73f52020-06-11 08:44:47 +02001230 * Description:
1231 * Complete a request by scheduling the ->complete_rq operation.
1232 **/
1233void blk_mq_complete_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001234{
Christoph Hellwig40d09b52020-06-11 08:44:50 +02001235 if (!blk_mq_complete_request_remote(rq))
Christoph Hellwig963395262020-06-11 08:44:49 +02001236 rq->q->mq_ops->complete(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01001237}
Christoph Hellwig15f73f52020-06-11 08:44:47 +02001238EXPORT_SYMBOL(blk_mq_complete_request);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08001239
1240/**
André Almeida105663f2020-01-06 15:08:18 -03001241 * blk_mq_start_request - Start processing a request
1242 * @rq: Pointer to request to be started
1243 *
1244 * Function used by device drivers to notify the block layer that a request
1245 * is going to be processed now, so blk layer can do proper initializations
1246 * such as starting the timeout timer.
1247 */
Christoph Hellwige2490072014-09-13 16:40:09 -07001248void blk_mq_start_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001249{
1250 struct request_queue *q = rq->q;
1251
Christoph Hellwiga54895f2020-12-03 17:21:39 +01001252 trace_block_rq_issue(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01001253
Jens Axboecf43e6b2016-11-07 21:32:37 -07001254 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
Tejun Heo4cddeac2022-04-27 09:49:12 -10001255 rq->io_start_time_ns = ktime_get_ns();
Hou Tao3d244302019-05-21 15:59:03 +08001256 rq->stats_sectors = blk_rq_sectors(rq);
Jens Axboecf43e6b2016-11-07 21:32:37 -07001257 rq->rq_flags |= RQF_STATS;
Josef Bacika7905042018-07-03 09:32:35 -06001258 rq_qos_issue(q, rq);
Jens Axboecf43e6b2016-11-07 21:32:37 -07001259 }
1260
Tejun Heo1d9bd512018-01-09 08:29:48 -08001261 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
Jens Axboe538b7532014-09-16 10:37:37 -06001262
Tejun Heo1d9bd512018-01-09 08:29:48 -08001263 blk_add_timer(rq);
Keith Busch12f5b932018-05-29 15:52:28 +02001264 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -08001265
Max Gurtovoy54d4e6a2019-09-16 18:44:29 +03001266#ifdef CONFIG_BLK_DEV_INTEGRITY
1267 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1268 q->integrity.profile->prepare_fn(rq);
1269#endif
Christoph Hellwig3e087732021-10-12 13:12:24 +02001270 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1271 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
Jens Axboe320ae512013-10-24 09:20:05 +01001272}
Christoph Hellwige2490072014-09-13 16:40:09 -07001273EXPORT_SYMBOL(blk_mq_start_request);
Jens Axboe320ae512013-10-24 09:20:05 +01001274
Ming Leia327c342022-05-12 22:00:10 +08001275/*
1276 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1277 * queues. This is important for md arrays to benefit from merging
1278 * requests.
1279 */
1280static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1281{
1282 if (plug->multiple_queues)
1283 return BLK_MAX_REQUEST_COUNT * 2;
1284 return BLK_MAX_REQUEST_COUNT;
1285}
1286
1287static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1288{
1289 struct request *last = rq_list_peek(&plug->mq_list);
1290
1291 if (!plug->rq_count) {
1292 trace_block_plug(rq->q);
1293 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1294 (!blk_queue_nomerges(rq->q) &&
1295 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1296 blk_mq_flush_plug_list(plug, false);
Al Viro878eb6e2022-11-01 00:54:13 +00001297 last = NULL;
Ming Leia327c342022-05-12 22:00:10 +08001298 trace_block_plug(rq->q);
1299 }
1300
1301 if (!plug->multiple_queues && last && last->q != rq->q)
1302 plug->multiple_queues = true;
1303 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1304 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 Hellwigae948fd2022-05-24 14:15:28 +02001324 WARN_ON(irqs_disabled());
1325 WARN_ON(!blk_rq_is_passthrough(rq));
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001326
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001327 blk_account_io_start(rq);
Pankaj Raghav110fdb42022-09-29 16:41:41 +02001328
1329 /*
1330 * As plugging can be enabled for passthrough requests on a zoned
1331 * device, directly accessing the plug instead of using blk_mq_plug()
1332 * should not have any consequences.
1333 */
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001334 if (current->plug)
1335 blk_add_rq_to_plug(current->plug, rq);
1336 else
1337 blk_mq_sched_insert_request(rq, at_head, true, false);
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;
1361 if (WARN_ON_ONCE(!rq->bio))
1362 return false;
1363 return true;
1364}
Kanchan Joshic6e99ea2022-08-23 21:44:42 +05301365EXPORT_SYMBOL_GPL(blk_rq_is_poll);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001366
1367static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1368{
1369 do {
1370 bio_poll(rq->bio, NULL, 0);
1371 cond_resched();
1372 } while (!completion_done(wait));
1373}
1374
1375/**
1376 * blk_execute_rq - insert a request into queue for execution
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001377 * @rq: request to insert
1378 * @at_head: insert request at head or tail of queue
1379 *
1380 * Description:
1381 * Insert a fully prepared request at the back of the I/O scheduler queue
1382 * for execution and wait for completion.
1383 * Return: The blk_status_t result provided to blk_mq_end_request().
1384 */
Christoph Hellwigb84ba302021-11-26 13:18:01 +01001385blk_status_t blk_execute_rq(struct request *rq, bool at_head)
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001386{
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001387 struct blk_rq_wait wait = {
1388 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1389 };
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001390
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001391 WARN_ON(irqs_disabled());
1392 WARN_ON(!blk_rq_is_passthrough(rq));
1393
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001394 rq->end_io_data = &wait;
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001395 rq->end_io = blk_end_sync_rq;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001396
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001397 blk_account_io_start(rq);
1398 blk_mq_sched_insert_request(rq, at_head, true, false);
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001399
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001400 if (blk_rq_is_poll(rq)) {
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001401 blk_rq_poll_completion(rq, &wait.done);
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001402 } else {
1403 /*
1404 * Prevent hang_check timer from firing at us during very long
1405 * I/O
1406 */
1407 unsigned long hang_check = sysctl_hung_task_timeout_secs;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001408
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001409 if (hang_check)
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001410 while (!wait_for_completion_io_timeout(&wait.done,
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001411 hang_check * (HZ/2)))
1412 ;
1413 else
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001414 wait_for_completion_io(&wait.done);
Christoph Hellwigae948fd2022-05-24 14:15:28 +02001415 }
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001416
Christoph Hellwig32ac5a92022-05-24 14:15:29 +02001417 return wait.ret;
Christoph Hellwig4054cff2021-11-17 07:13:56 +01001418}
1419EXPORT_SYMBOL(blk_execute_rq);
1420
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001421static void __blk_mq_requeue_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001422{
1423 struct request_queue *q = rq->q;
1424
Ming Lei923218f2017-11-02 23:24:38 +08001425 blk_mq_put_driver_tag(rq);
1426
Christoph Hellwiga54895f2020-12-03 17:21:39 +01001427 trace_block_rq_requeue(rq);
Josef Bacika7905042018-07-03 09:32:35 -06001428 rq_qos_requeue(q, rq);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -08001429
Keith Busch12f5b932018-05-29 15:52:28 +02001430 if (blk_mq_request_started(rq)) {
1431 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Christoph Hellwigda661262018-06-14 13:58:45 +02001432 rq->rq_flags &= ~RQF_TIMED_OUT;
Christoph Hellwige2490072014-09-13 16:40:09 -07001433 }
Jens Axboe320ae512013-10-24 09:20:05 +01001434}
1435
Bart Van Assche2b053ac2016-10-28 17:21:41 -07001436void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001437{
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001438 __blk_mq_requeue_request(rq);
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001439
Ming Lei105976f2018-02-23 23:36:56 +08001440 /* this request will be re-inserted to io scheduler queue */
1441 blk_mq_sched_requeue_request(rq);
1442
Bart Van Assche2b053ac2016-10-28 17:21:41 -07001443 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
Christoph Hellwiged0791b2014-04-16 09:44:57 +02001444}
1445EXPORT_SYMBOL(blk_mq_requeue_request);
1446
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001447static void blk_mq_requeue_work(struct work_struct *work)
1448{
1449 struct request_queue *q =
Mike Snitzer28494502016-09-14 13:28:30 -04001450 container_of(work, struct request_queue, requeue_work.work);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001451 LIST_HEAD(rq_list);
1452 struct request *rq, *next;
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001453
Jens Axboe18e97812017-07-27 08:03:57 -06001454 spin_lock_irq(&q->requeue_lock);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001455 list_splice_init(&q->requeue_list, &rq_list);
Jens Axboe18e97812017-07-27 08:03:57 -06001456 spin_unlock_irq(&q->requeue_lock);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001457
1458 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
Jianchao Wangaef18972019-02-12 09:56:25 +08001459 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001460 continue;
1461
Christoph Hellwige8064022016-10-20 15:12:13 +02001462 rq->rq_flags &= ~RQF_SOFTBARRIER;
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001463 list_del_init(&rq->queuelist);
Jianchao Wangaef18972019-02-12 09:56:25 +08001464 /*
1465 * If RQF_DONTPREP, rq has contained some driver specific
1466 * data, so insert it to hctx dispatch list to avoid any
1467 * merge.
1468 */
1469 if (rq->rq_flags & RQF_DONTPREP)
Ming Lei01e99ae2020-02-25 09:04:32 +08001470 blk_mq_request_bypass_insert(rq, false, false);
Jianchao Wangaef18972019-02-12 09:56:25 +08001471 else
1472 blk_mq_sched_insert_request(rq, true, false, false);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001473 }
1474
1475 while (!list_empty(&rq_list)) {
1476 rq = list_entry(rq_list.next, struct request, queuelist);
1477 list_del_init(&rq->queuelist);
Mike Snitzer9e97d292018-01-17 11:25:58 -05001478 blk_mq_sched_insert_request(rq, false, false, false);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001479 }
1480
Bart Van Assche52d7f1b2016-10-28 17:20:32 -07001481 blk_mq_run_hw_queues(q, false);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001482}
1483
Bart Van Assche2b053ac2016-10-28 17:21:41 -07001484void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1485 bool kick_requeue_list)
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001486{
1487 struct request_queue *q = rq->q;
1488 unsigned long flags;
1489
1490 /*
1491 * We abuse this flag that is otherwise used by the I/O scheduler to
Jens Axboeff821d22017-11-10 22:05:12 -07001492 * request head insertion from the workqueue.
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001493 */
Christoph Hellwige8064022016-10-20 15:12:13 +02001494 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001495
1496 spin_lock_irqsave(&q->requeue_lock, flags);
1497 if (at_head) {
Christoph Hellwige8064022016-10-20 15:12:13 +02001498 rq->rq_flags |= RQF_SOFTBARRIER;
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001499 list_add(&rq->queuelist, &q->requeue_list);
1500 } else {
1501 list_add_tail(&rq->queuelist, &q->requeue_list);
1502 }
1503 spin_unlock_irqrestore(&q->requeue_lock, flags);
Bart Van Assche2b053ac2016-10-28 17:21:41 -07001504
1505 if (kick_requeue_list)
1506 blk_mq_kick_requeue_list(q);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001507}
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001508
1509void blk_mq_kick_requeue_list(struct request_queue *q)
1510{
Bart Van Asscheae943d22018-01-19 08:58:55 -08001511 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001512}
1513EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1514
Mike Snitzer28494502016-09-14 13:28:30 -04001515void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1516 unsigned long msecs)
1517{
Bart Van Assched4acf362017-08-09 11:28:06 -07001518 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1519 msecs_to_jiffies(msecs));
Mike Snitzer28494502016-09-14 13:28:30 -04001520}
1521EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1522
John Garry2dd65322022-07-06 20:03:53 +08001523static bool blk_mq_rq_inflight(struct request *rq, void *priv)
Jens Axboeae879912018-11-08 09:03:51 -07001524{
1525 /*
John Garry8ab30a32021-12-06 20:49:48 +08001526 * If we find a request that isn't idle we know the queue is busy
1527 * as it's checked in the iter.
1528 * Return false to stop the iteration.
Jens Axboeae879912018-11-08 09:03:51 -07001529 */
John Garry8ab30a32021-12-06 20:49:48 +08001530 if (blk_mq_request_started(rq)) {
Jens Axboeae879912018-11-08 09:03:51 -07001531 bool *busy = priv;
1532
1533 *busy = true;
1534 return false;
1535 }
1536
1537 return true;
1538}
1539
Jens Axboe3c94d832018-12-17 21:11:17 -07001540bool blk_mq_queue_inflight(struct request_queue *q)
Jens Axboeae879912018-11-08 09:03:51 -07001541{
1542 bool busy = false;
1543
Jens Axboe3c94d832018-12-17 21:11:17 -07001544 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
Jens Axboeae879912018-11-08 09:03:51 -07001545 return busy;
1546}
Jens Axboe3c94d832018-12-17 21:11:17 -07001547EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
Jens Axboeae879912018-11-08 09:03:51 -07001548
John Garry9bdb4832022-07-06 20:03:51 +08001549static void blk_mq_rq_timed_out(struct request *req)
Jens Axboe320ae512013-10-24 09:20:05 +01001550{
Christoph Hellwigda661262018-06-14 13:58:45 +02001551 req->rq_flags |= RQF_TIMED_OUT;
Christoph Hellwigd1210d52018-05-29 15:52:39 +02001552 if (req->q->mq_ops->timeout) {
1553 enum blk_eh_timer_return ret;
Jens Axboe87ee7b12014-04-24 08:51:47 -06001554
John Garry9bdb4832022-07-06 20:03:51 +08001555 ret = req->q->mq_ops->timeout(req);
Christoph Hellwigd1210d52018-05-29 15:52:39 +02001556 if (ret == BLK_EH_DONE)
1557 return;
1558 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
Christoph Hellwig46f92d42014-09-13 16:40:12 -07001559 }
Christoph Hellwigd1210d52018-05-29 15:52:39 +02001560
1561 blk_add_timer(req);
Jens Axboe87ee7b12014-04-24 08:51:47 -06001562}
Keith Busch5b3f25f2015-01-07 18:55:46 -07001563
David Jeffery82c22942022-10-26 13:19:57 +08001564struct blk_expired_data {
1565 bool has_timedout_rq;
1566 unsigned long next;
1567 unsigned long timeout_start;
1568};
1569
1570static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
Keith Busch12f5b932018-05-29 15:52:28 +02001571{
1572 unsigned long deadline;
1573
1574 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1575 return false;
Christoph Hellwigda661262018-06-14 13:58:45 +02001576 if (rq->rq_flags & RQF_TIMED_OUT)
1577 return false;
Keith Busch12f5b932018-05-29 15:52:28 +02001578
Christoph Hellwig079076b2018-11-14 17:02:05 +01001579 deadline = READ_ONCE(rq->deadline);
David Jeffery82c22942022-10-26 13:19:57 +08001580 if (time_after_eq(expired->timeout_start, deadline))
Keith Busch12f5b932018-05-29 15:52:28 +02001581 return true;
1582
David Jeffery82c22942022-10-26 13:19:57 +08001583 if (expired->next == 0)
1584 expired->next = deadline;
1585 else if (time_after(expired->next, deadline))
1586 expired->next = deadline;
Keith Busch12f5b932018-05-29 15:52:28 +02001587 return false;
1588}
1589
Ming Lei2e315dc2021-05-11 23:22:34 +08001590void blk_mq_put_rq_ref(struct request *rq)
1591{
Jens Axboede671d62022-09-21 15:19:54 -06001592 if (is_flush_rq(rq)) {
1593 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1594 blk_mq_free_request(rq);
1595 } else if (req_ref_put_and_test(rq)) {
Ming Lei2e315dc2021-05-11 23:22:34 +08001596 __blk_mq_free_request(rq);
Jens Axboede671d62022-09-21 15:19:54 -06001597 }
Ming Lei2e315dc2021-05-11 23:22:34 +08001598}
1599
John Garry2dd65322022-07-06 20:03:53 +08001600static bool blk_mq_check_expired(struct request *rq, void *priv)
Jens Axboe320ae512013-10-24 09:20:05 +01001601{
David Jeffery82c22942022-10-26 13:19:57 +08001602 struct blk_expired_data *expired = priv;
Christoph Hellwig81481eb2014-09-13 16:40:11 -07001603
Keith Busch12f5b932018-05-29 15:52:28 +02001604 /*
Ming Leic797b402021-08-11 23:52:02 +08001605 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1606 * be reallocated underneath the timeout handler's processing, then
1607 * the expire check is reliable. If the request is not expired, then
1608 * it was completed and reallocated as a new request after returning
1609 * from blk_mq_check_expired().
Keith Busch12f5b932018-05-29 15:52:28 +02001610 */
David Jeffery82c22942022-10-26 13:19:57 +08001611 if (blk_mq_req_expired(rq, expired)) {
1612 expired->has_timedout_rq = true;
1613 return false;
1614 }
1615 return true;
1616}
1617
1618static bool blk_mq_handle_expired(struct request *rq, void *priv)
1619{
1620 struct blk_expired_data *expired = priv;
1621
1622 if (blk_mq_req_expired(rq, expired))
John Garry9bdb4832022-07-06 20:03:51 +08001623 blk_mq_rq_timed_out(rq);
Jens Axboe7baa8572018-11-08 10:24:07 -07001624 return true;
Tejun Heo1d9bd512018-01-09 08:29:48 -08001625}
1626
Christoph Hellwig287922e2015-10-30 20:57:30 +08001627static void blk_mq_timeout_work(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +01001628{
Christoph Hellwig287922e2015-10-30 20:57:30 +08001629 struct request_queue *q =
1630 container_of(work, struct request_queue, timeout_work);
David Jeffery82c22942022-10-26 13:19:57 +08001631 struct blk_expired_data expired = {
1632 .timeout_start = jiffies,
1633 };
Tejun Heo1d9bd512018-01-09 08:29:48 -08001634 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08001635 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01001636
Gabriel Krisman Bertazi71f79fb2016-08-01 08:23:39 -06001637 /* A deadlock might occur if a request is stuck requiring a
1638 * timeout at the same time a queue freeze is waiting
1639 * completion, since the timeout code would not be able to
1640 * acquire the queue reference here.
1641 *
1642 * That's why we don't use blk_queue_enter here; instead, we use
1643 * percpu_ref_tryget directly, because we need to be able to
1644 * obtain a reference even in the short window between the queue
1645 * starting to freeze, by dropping the first reference in
Ming Lei1671d522017-03-27 20:06:57 +08001646 * blk_freeze_queue_start, and the moment the last request is
Gabriel Krisman Bertazi71f79fb2016-08-01 08:23:39 -06001647 * consumed, marked by the instant q_usage_counter reaches
1648 * zero.
1649 */
1650 if (!percpu_ref_tryget(&q->q_usage_counter))
Christoph Hellwig287922e2015-10-30 20:57:30 +08001651 return;
1652
David Jeffery82c22942022-10-26 13:19:57 +08001653 /* check if there is any timed-out request */
1654 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1655 if (expired.has_timedout_rq) {
1656 /*
1657 * Before walking tags, we must ensure any submit started
1658 * before the current time has finished. Since the submit
1659 * uses srcu or rcu, wait for a synchronization point to
1660 * ensure all running submits have finished
1661 */
Christoph Hellwig483239c2022-11-01 16:00:48 +01001662 blk_mq_wait_quiesce_done(q->tag_set);
Jens Axboe320ae512013-10-24 09:20:05 +01001663
David Jeffery82c22942022-10-26 13:19:57 +08001664 expired.next = 0;
1665 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1666 }
1667
1668 if (expired.next != 0) {
1669 mod_timer(&q->timeout, expired.next);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001670 } else {
Bart Van Asschefcd36c32018-01-10 08:33:33 -08001671 /*
1672 * Request timeouts are handled as a forward rolling timer. If
1673 * we end up here it means that no requests are pending and
1674 * also that no request has been pending for a while. Mark
1675 * each hctx as idle.
1676 */
Ming Leif054b562015-04-21 10:00:19 +08001677 queue_for_each_hw_ctx(q, hctx, i) {
1678 /* the hctx may be unmapped, so check it here */
1679 if (blk_mq_hw_queue_mapped(hctx))
1680 blk_mq_tag_idle(hctx);
1681 }
Jens Axboe0d2602c2014-05-13 15:10:52 -06001682 }
Christoph Hellwig287922e2015-10-30 20:57:30 +08001683 blk_queue_exit(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001684}
1685
Omar Sandoval88459642016-09-17 08:38:44 -06001686struct flush_busy_ctx_data {
1687 struct blk_mq_hw_ctx *hctx;
1688 struct list_head *list;
1689};
1690
1691static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1692{
1693 struct flush_busy_ctx_data *flush_data = data;
1694 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1695 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
Ming Leic16d6b52018-12-17 08:44:05 -07001696 enum hctx_type type = hctx->type;
Omar Sandoval88459642016-09-17 08:38:44 -06001697
Omar Sandoval88459642016-09-17 08:38:44 -06001698 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07001699 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
Omar Sandovale9a99a62018-02-27 16:56:42 -08001700 sbitmap_clear_bit(sb, bitnr);
Omar Sandoval88459642016-09-17 08:38:44 -06001701 spin_unlock(&ctx->lock);
1702 return true;
1703}
1704
Jens Axboe320ae512013-10-24 09:20:05 +01001705/*
Jens Axboe1429d7c2014-05-19 09:23:55 -06001706 * Process software queues that have been marked busy, splicing them
1707 * to the for-dispatch
1708 */
Jens Axboe2c3ad662016-12-14 14:34:47 -07001709void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
Jens Axboe1429d7c2014-05-19 09:23:55 -06001710{
Omar Sandoval88459642016-09-17 08:38:44 -06001711 struct flush_busy_ctx_data data = {
1712 .hctx = hctx,
1713 .list = list,
1714 };
Jens Axboe1429d7c2014-05-19 09:23:55 -06001715
Omar Sandoval88459642016-09-17 08:38:44 -06001716 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
Jens Axboe1429d7c2014-05-19 09:23:55 -06001717}
Jens Axboe2c3ad662016-12-14 14:34:47 -07001718EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
Jens Axboe1429d7c2014-05-19 09:23:55 -06001719
Ming Leib3476892017-10-14 17:22:30 +08001720struct dispatch_rq_data {
1721 struct blk_mq_hw_ctx *hctx;
1722 struct request *rq;
1723};
1724
1725static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1726 void *data)
1727{
1728 struct dispatch_rq_data *dispatch_data = data;
1729 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1730 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
Ming Leic16d6b52018-12-17 08:44:05 -07001731 enum hctx_type type = hctx->type;
Ming Leib3476892017-10-14 17:22:30 +08001732
1733 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07001734 if (!list_empty(&ctx->rq_lists[type])) {
1735 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
Ming Leib3476892017-10-14 17:22:30 +08001736 list_del_init(&dispatch_data->rq->queuelist);
Ming Leic16d6b52018-12-17 08:44:05 -07001737 if (list_empty(&ctx->rq_lists[type]))
Ming Leib3476892017-10-14 17:22:30 +08001738 sbitmap_clear_bit(sb, bitnr);
1739 }
1740 spin_unlock(&ctx->lock);
1741
1742 return !dispatch_data->rq;
1743}
1744
1745struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1746 struct blk_mq_ctx *start)
1747{
Jens Axboef31967f2018-10-29 13:13:29 -06001748 unsigned off = start ? start->index_hw[hctx->type] : 0;
Ming Leib3476892017-10-14 17:22:30 +08001749 struct dispatch_rq_data data = {
1750 .hctx = hctx,
1751 .rq = NULL,
1752 };
1753
1754 __sbitmap_for_each_set(&hctx->ctx_map, off,
1755 dispatch_rq_from_ctx, &data);
1756
1757 return data.rq;
1758}
1759
Jens Axboea808a9d2021-10-13 08:28:14 -06001760static bool __blk_mq_alloc_driver_tag(struct request *rq)
Jens Axboe703fd1c2016-09-16 13:59:14 -06001761{
John Garryae0f1a72021-10-05 18:23:38 +08001762 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
Ming Lei570e9b72020-06-30 22:03:55 +08001763 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
Ming Lei570e9b72020-06-30 22:03:55 +08001764 int tag;
1765
Ming Lei568f2702020-07-06 22:41:11 +08001766 blk_mq_tag_busy(rq->mq_hctx);
1767
Ming Lei570e9b72020-06-30 22:03:55 +08001768 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
John Garryae0f1a72021-10-05 18:23:38 +08001769 bt = &rq->mq_hctx->tags->breserved_tags;
Ming Lei570e9b72020-06-30 22:03:55 +08001770 tag_offset = 0;
Ming Lei28500852020-09-11 18:41:14 +08001771 } else {
1772 if (!hctx_may_queue(rq->mq_hctx, bt))
1773 return false;
Ming Lei570e9b72020-06-30 22:03:55 +08001774 }
1775
Ming Lei570e9b72020-06-30 22:03:55 +08001776 tag = __sbitmap_queue_get(bt);
1777 if (tag == BLK_MQ_NO_TAG)
1778 return false;
1779
1780 rq->tag = tag + tag_offset;
Ming Lei570e9b72020-06-30 22:03:55 +08001781 return true;
1782}
1783
Jens Axboea808a9d2021-10-13 08:28:14 -06001784bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
Ming Lei570e9b72020-06-30 22:03:55 +08001785{
Jens Axboea808a9d2021-10-13 08:28:14 -06001786 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
Ming Lei568f2702020-07-06 22:41:11 +08001787 return false;
1788
Ming Lei51db1c32020-08-19 23:20:19 +08001789 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
Ming Lei568f2702020-07-06 22:41:11 +08001790 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1791 rq->rq_flags |= RQF_MQ_INFLIGHT;
John Garrybccf5e22020-08-19 23:20:26 +08001792 __blk_mq_inc_active_requests(hctx);
Ming Lei568f2702020-07-06 22:41:11 +08001793 }
1794 hctx->tags->rqs[rq->tag] = rq;
1795 return true;
Ming Lei570e9b72020-06-30 22:03:55 +08001796}
1797
Jens Axboeeb619fd2017-11-09 08:32:43 -07001798static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1799 int flags, void *key)
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001800{
1801 struct blk_mq_hw_ctx *hctx;
1802
1803 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1804
Ming Lei5815839b2018-06-25 19:31:47 +08001805 spin_lock(&hctx->dispatch_wait_lock);
Jens Axboee8618572019-03-25 12:34:10 -06001806 if (!list_empty(&wait->entry)) {
1807 struct sbitmap_queue *sbq;
1808
1809 list_del_init(&wait->entry);
John Garryae0f1a72021-10-05 18:23:38 +08001810 sbq = &hctx->tags->bitmap_tags;
Jens Axboee8618572019-03-25 12:34:10 -06001811 atomic_dec(&sbq->ws_active);
1812 }
Ming Lei5815839b2018-06-25 19:31:47 +08001813 spin_unlock(&hctx->dispatch_wait_lock);
1814
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001815 blk_mq_run_hw_queue(hctx, true);
1816 return 1;
1817}
1818
Jens Axboef906a6a2017-11-09 16:10:13 -07001819/*
1820 * Mark us waiting for a tag. For shared tags, this involves hooking us into
Bart Van Asscheee3e4de2018-01-09 10:09:15 -08001821 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1822 * restart. For both cases, take care to check the condition again after
Jens Axboef906a6a2017-11-09 16:10:13 -07001823 * marking us as waiting.
1824 */
Ming Lei2278d692018-06-25 19:31:46 +08001825static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
Jens Axboef906a6a2017-11-09 16:10:13 -07001826 struct request *rq)
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001827{
John Garryae0f1a72021-10-05 18:23:38 +08001828 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
Ming Lei5815839b2018-06-25 19:31:47 +08001829 struct wait_queue_head *wq;
Jens Axboef906a6a2017-11-09 16:10:13 -07001830 wait_queue_entry_t *wait;
1831 bool ret;
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001832
Ming Lei51db1c32020-08-19 23:20:19 +08001833 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
Yufen Yu684b7322019-03-15 11:05:10 +08001834 blk_mq_sched_mark_restart_hctx(hctx);
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001835
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001836 /*
1837 * It's possible that a tag was freed in the window between the
1838 * allocation failure and adding the hardware queue to the wait
1839 * queue.
1840 *
1841 * Don't clear RESTART here, someone else could have set it.
1842 * At most this will cost an extra queue run.
1843 */
Ming Lei8ab6bb9e2018-06-25 19:31:45 +08001844 return blk_mq_get_driver_tag(rq);
Jens Axboeeb619fd2017-11-09 08:32:43 -07001845 }
1846
Ming Lei2278d692018-06-25 19:31:46 +08001847 wait = &hctx->dispatch_wait;
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001848 if (!list_empty_careful(&wait->entry))
1849 return false;
1850
Jens Axboee8618572019-03-25 12:34:10 -06001851 wq = &bt_wait_ptr(sbq, hctx)->wait;
Ming Lei5815839b2018-06-25 19:31:47 +08001852
1853 spin_lock_irq(&wq->lock);
1854 spin_lock(&hctx->dispatch_wait_lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001855 if (!list_empty(&wait->entry)) {
Ming Lei5815839b2018-06-25 19:31:47 +08001856 spin_unlock(&hctx->dispatch_wait_lock);
1857 spin_unlock_irq(&wq->lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001858 return false;
1859 }
1860
Jens Axboee8618572019-03-25 12:34:10 -06001861 atomic_inc(&sbq->ws_active);
Ming Lei5815839b2018-06-25 19:31:47 +08001862 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1863 __add_wait_queue(wq, wait);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001864
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001865 /*
Jens Axboeeb619fd2017-11-09 08:32:43 -07001866 * It's possible that a tag was freed in the window between the
1867 * allocation failure and adding the hardware queue to the wait
1868 * queue.
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001869 */
Ming Lei8ab6bb9e2018-06-25 19:31:45 +08001870 ret = blk_mq_get_driver_tag(rq);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001871 if (!ret) {
Ming Lei5815839b2018-06-25 19:31:47 +08001872 spin_unlock(&hctx->dispatch_wait_lock);
1873 spin_unlock_irq(&wq->lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001874 return false;
Jens Axboef906a6a2017-11-09 16:10:13 -07001875 }
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001876
1877 /*
1878 * We got a tag, remove ourselves from the wait queue to ensure
1879 * someone else gets the wakeup.
1880 */
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001881 list_del_init(&wait->entry);
Jens Axboee8618572019-03-25 12:34:10 -06001882 atomic_dec(&sbq->ws_active);
Ming Lei5815839b2018-06-25 19:31:47 +08001883 spin_unlock(&hctx->dispatch_wait_lock);
1884 spin_unlock_irq(&wq->lock);
Bart Van Asschec27d53f2018-01-10 13:41:21 -08001885
1886 return true;
Omar Sandovalda55f2c2017-02-22 10:58:29 -08001887}
1888
Ming Lei6e7687172018-07-03 09:03:16 -06001889#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1890#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1891/*
1892 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1893 * - EWMA is one simple way to compute running average value
1894 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1895 * - take 4 as factor for avoiding to get too small(0) result, and this
1896 * factor doesn't matter because EWMA decreases exponentially
1897 */
1898static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1899{
1900 unsigned int ewma;
1901
Ming Lei6e7687172018-07-03 09:03:16 -06001902 ewma = hctx->dispatch_busy;
1903
1904 if (!ewma && !busy)
1905 return;
1906
1907 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1908 if (busy)
1909 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1910 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1911
1912 hctx->dispatch_busy = ewma;
1913}
1914
Ming Lei86ff7c22018-01-30 22:04:57 -05001915#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1916
Johannes Thumshirnc92a4102020-03-25 00:24:44 +09001917static void blk_mq_handle_dev_resource(struct request *rq,
1918 struct list_head *list)
1919{
1920 struct request *next =
1921 list_first_entry_or_null(list, struct request, queuelist);
1922
1923 /*
1924 * If an I/O scheduler has been configured and we got a driver tag for
1925 * the next request already, free it.
1926 */
1927 if (next)
1928 blk_mq_put_driver_tag(next);
1929
1930 list_add(&rq->queuelist, list);
1931 __blk_mq_requeue_request(rq);
1932}
1933
Keith Busch0512a752020-05-12 17:55:47 +09001934static void blk_mq_handle_zone_resource(struct request *rq,
1935 struct list_head *zone_list)
1936{
1937 /*
1938 * If we end up here it is because we cannot dispatch a request to a
1939 * specific zone due to LLD level zone-write locking or other zone
1940 * related resource not being available. In this case, set the request
1941 * aside in zone_list for retrying it later.
1942 */
1943 list_add(&rq->queuelist, zone_list);
1944 __blk_mq_requeue_request(rq);
1945}
1946
Ming Lei75383522020-06-30 18:24:58 +08001947enum prep_dispatch {
1948 PREP_DISPATCH_OK,
1949 PREP_DISPATCH_NO_TAG,
1950 PREP_DISPATCH_NO_BUDGET,
1951};
1952
1953static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1954 bool need_budget)
1955{
1956 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Ming Lei2a5a24a2021-01-22 10:33:12 +08001957 int budget_token = -1;
Ming Lei75383522020-06-30 18:24:58 +08001958
Ming Lei2a5a24a2021-01-22 10:33:12 +08001959 if (need_budget) {
1960 budget_token = blk_mq_get_dispatch_budget(rq->q);
1961 if (budget_token < 0) {
1962 blk_mq_put_driver_tag(rq);
1963 return PREP_DISPATCH_NO_BUDGET;
1964 }
1965 blk_mq_set_rq_budget_token(rq, budget_token);
Ming Lei75383522020-06-30 18:24:58 +08001966 }
1967
1968 if (!blk_mq_get_driver_tag(rq)) {
1969 /*
1970 * The initial allocation attempt failed, so we need to
1971 * rerun the hardware queue when a tag is freed. The
1972 * waitqueue takes care of that. If the queue is run
1973 * before we add this entry back on the dispatch list,
1974 * we'll re-run it below.
1975 */
1976 if (!blk_mq_mark_tag_wait(hctx, rq)) {
Ming Lei1fd40b52020-06-30 18:25:00 +08001977 /*
1978 * All budgets not got from this function will be put
1979 * together during handling partial dispatch
1980 */
1981 if (need_budget)
Ming Lei2a5a24a2021-01-22 10:33:12 +08001982 blk_mq_put_dispatch_budget(rq->q, budget_token);
Ming Lei75383522020-06-30 18:24:58 +08001983 return PREP_DISPATCH_NO_TAG;
1984 }
1985 }
1986
1987 return PREP_DISPATCH_OK;
1988}
1989
Ming Lei1fd40b52020-06-30 18:25:00 +08001990/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1991static void blk_mq_release_budgets(struct request_queue *q,
Ming Lei2a5a24a2021-01-22 10:33:12 +08001992 struct list_head *list)
Ming Lei1fd40b52020-06-30 18:25:00 +08001993{
Ming Lei2a5a24a2021-01-22 10:33:12 +08001994 struct request *rq;
Ming Lei1fd40b52020-06-30 18:25:00 +08001995
Ming Lei2a5a24a2021-01-22 10:33:12 +08001996 list_for_each_entry(rq, list, queuelist) {
1997 int budget_token = blk_mq_get_rq_budget_token(rq);
1998
1999 if (budget_token >= 0)
2000 blk_mq_put_dispatch_budget(q, budget_token);
2001 }
Ming Lei1fd40b52020-06-30 18:25:00 +08002002}
2003
Jens Axboe1f57f8d2018-06-28 11:54:01 -06002004/*
2005 * Returns true if we did some work AND can potentially do more.
2006 */
Ming Lei445874e2020-06-30 18:24:57 +08002007bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
Ming Lei1fd40b52020-06-30 18:25:00 +08002008 unsigned int nr_budgets)
Jens Axboef04c3df2016-12-07 08:41:17 -07002009{
Ming Lei75383522020-06-30 18:24:58 +08002010 enum prep_dispatch prep;
Ming Lei445874e2020-06-30 18:24:57 +08002011 struct request_queue *q = hctx->queue;
Jianchao Wang6d6f167c2017-11-02 23:24:32 +08002012 struct request *rq, *nxt;
Christoph Hellwigfc17b652017-06-03 09:38:05 +02002013 int errors, queued;
Ming Lei86ff7c22018-01-30 22:04:57 -05002014 blk_status_t ret = BLK_STS_OK;
Keith Busch0512a752020-05-12 17:55:47 +09002015 LIST_HEAD(zone_list);
Naohiro Aota9586e672021-10-27 01:51:27 +09002016 bool needs_resource = false;
Jens Axboef04c3df2016-12-07 08:41:17 -07002017
Omar Sandoval81380ca2017-04-07 08:56:26 -06002018 if (list_empty(list))
2019 return false;
2020
Jens Axboef04c3df2016-12-07 08:41:17 -07002021 /*
Jens Axboef04c3df2016-12-07 08:41:17 -07002022 * Now process all the entries, sending them to the driver.
2023 */
Jens Axboe93efe982017-03-24 12:04:19 -06002024 errors = queued = 0;
Omar Sandoval81380ca2017-04-07 08:56:26 -06002025 do {
Jens Axboef04c3df2016-12-07 08:41:17 -07002026 struct blk_mq_queue_data bd;
2027
2028 rq = list_first_entry(list, struct request, queuelist);
Ming Lei0bca7992018-04-05 00:35:21 +08002029
Ming Lei445874e2020-06-30 18:24:57 +08002030 WARN_ON_ONCE(hctx != rq->mq_hctx);
Ming Lei1fd40b52020-06-30 18:25:00 +08002031 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
Ming Lei75383522020-06-30 18:24:58 +08002032 if (prep != PREP_DISPATCH_OK)
Ming Lei0bca7992018-04-05 00:35:21 +08002033 break;
Ming Leide148292017-10-14 17:22:29 +08002034
Jens Axboef04c3df2016-12-07 08:41:17 -07002035 list_del_init(&rq->queuelist);
2036
2037 bd.rq = rq;
Jens Axboe113285b2017-03-02 13:26:04 -07002038
2039 /*
2040 * Flag last if we have no more requests, or if we have more
2041 * but can't assign a driver tag to it.
2042 */
2043 if (list_empty(list))
2044 bd.last = true;
2045 else {
Jens Axboe113285b2017-03-02 13:26:04 -07002046 nxt = list_first_entry(list, struct request, queuelist);
Ming Lei8ab6bb9e2018-06-25 19:31:45 +08002047 bd.last = !blk_mq_get_driver_tag(nxt);
Jens Axboe113285b2017-03-02 13:26:04 -07002048 }
Jens Axboef04c3df2016-12-07 08:41:17 -07002049
Ming Lei1fd40b52020-06-30 18:25:00 +08002050 /*
2051 * once the request is queued to lld, no need to cover the
2052 * budget any more
2053 */
2054 if (nr_budgets)
2055 nr_budgets--;
Jens Axboef04c3df2016-12-07 08:41:17 -07002056 ret = q->mq_ops->queue_rq(hctx, &bd);
Ming Lei7bf13722020-07-01 21:58:57 +08002057 switch (ret) {
2058 case BLK_STS_OK:
2059 queued++;
Jens Axboef04c3df2016-12-07 08:41:17 -07002060 break;
Ming Lei7bf13722020-07-01 21:58:57 +08002061 case BLK_STS_RESOURCE:
Naohiro Aota9586e672021-10-27 01:51:27 +09002062 needs_resource = true;
2063 fallthrough;
Ming Lei7bf13722020-07-01 21:58:57 +08002064 case BLK_STS_DEV_RESOURCE:
2065 blk_mq_handle_dev_resource(rq, list);
2066 goto out;
2067 case BLK_STS_ZONE_RESOURCE:
Keith Busch0512a752020-05-12 17:55:47 +09002068 /*
2069 * Move the request to zone_list and keep going through
2070 * the dispatch list to find more requests the drive can
2071 * accept.
2072 */
2073 blk_mq_handle_zone_resource(rq, &zone_list);
Naohiro Aota9586e672021-10-27 01:51:27 +09002074 needs_resource = true;
Ming Lei7bf13722020-07-01 21:58:57 +08002075 break;
2076 default:
Christoph Hellwigfc17b652017-06-03 09:38:05 +02002077 errors++;
Hannes Reineckee21ee5a2020-09-30 10:02:53 +02002078 blk_mq_end_request(rq, ret);
Christoph Hellwigfc17b652017-06-03 09:38:05 +02002079 }
Omar Sandoval81380ca2017-04-07 08:56:26 -06002080 } while (!list_empty(list));
Ming Lei7bf13722020-07-01 21:58:57 +08002081out:
Keith Busch0512a752020-05-12 17:55:47 +09002082 if (!list_empty(&zone_list))
2083 list_splice_tail_init(&zone_list, list);
2084
yangerkun632bfb62020-09-05 19:25:56 +08002085 /* If we didn't flush the entire list, we could have told the driver
2086 * there was more coming, but that turned out to be a lie.
2087 */
Yu Kuai65fac0d2022-07-26 20:22:24 +08002088 if ((!list_empty(list) || errors || needs_resource ||
2089 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
yangerkun632bfb62020-09-05 19:25:56 +08002090 q->mq_ops->commit_rqs(hctx);
Jens Axboef04c3df2016-12-07 08:41:17 -07002091 /*
2092 * Any items that need requeuing? Stuff them into hctx->dispatch,
2093 * that is where we will continue on next queue run.
2094 */
2095 if (!list_empty(list)) {
Ming Lei86ff7c22018-01-30 22:04:57 -05002096 bool needs_restart;
Ming Lei75383522020-06-30 18:24:58 +08002097 /* For non-shared tags, the RESTART check will suffice */
2098 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
Ming Lei51db1c32020-08-19 23:20:19 +08002099 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
Ming Lei86ff7c22018-01-30 22:04:57 -05002100
Ming Lei2a5a24a2021-01-22 10:33:12 +08002101 if (nr_budgets)
2102 blk_mq_release_budgets(q, list);
Jens Axboef04c3df2016-12-07 08:41:17 -07002103
2104 spin_lock(&hctx->lock);
Ming Lei01e99ae2020-02-25 09:04:32 +08002105 list_splice_tail_init(list, &hctx->dispatch);
Jens Axboef04c3df2016-12-07 08:41:17 -07002106 spin_unlock(&hctx->lock);
2107
2108 /*
Ming Leid7d85352020-08-17 18:01:15 +08002109 * Order adding requests to hctx->dispatch and checking
2110 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2111 * in blk_mq_sched_restart(). Avoid restart code path to
2112 * miss the new added requests to hctx->dispatch, meantime
2113 * SCHED_RESTART is observed here.
2114 */
2115 smp_mb();
2116
2117 /*
Bart Van Assche710c7852017-04-07 11:16:51 -07002118 * If SCHED_RESTART was set by the caller of this function and
2119 * it is no longer set that means that it was cleared by another
2120 * thread and hence that a queue rerun is needed.
Jens Axboef04c3df2016-12-07 08:41:17 -07002121 *
Jens Axboeeb619fd2017-11-09 08:32:43 -07002122 * If 'no_tag' is set, that means that we failed getting
2123 * a driver tag with an I/O scheduler attached. If our dispatch
2124 * waitqueue is no longer active, ensure that we run the queue
2125 * AFTER adding our entries back to the list.
Jens Axboebd166ef2017-01-17 06:03:22 -07002126 *
Bart Van Assche710c7852017-04-07 11:16:51 -07002127 * If no I/O scheduler has been configured it is possible that
2128 * the hardware queue got stopped and restarted before requests
2129 * were pushed back onto the dispatch list. Rerun the queue to
2130 * avoid starvation. Notes:
2131 * - blk_mq_run_hw_queue() checks whether or not a queue has
2132 * been stopped before rerunning a queue.
2133 * - Some but not all block drivers stop a queue before
Christoph Hellwigfc17b652017-06-03 09:38:05 +02002134 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
Bart Van Assche710c7852017-04-07 11:16:51 -07002135 * and dm-rq.
Ming Lei86ff7c22018-01-30 22:04:57 -05002136 *
2137 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2138 * bit is set, run queue after a delay to avoid IO stalls
Douglas Andersonab3cee32020-04-20 09:24:51 -07002139 * that could otherwise occur if the queue is idle. We'll do
Naohiro Aota9586e672021-10-27 01:51:27 +09002140 * similar if we couldn't get budget or couldn't lock a zone
2141 * and SCHED_RESTART is set.
Jens Axboebd166ef2017-01-17 06:03:22 -07002142 */
Ming Lei86ff7c22018-01-30 22:04:57 -05002143 needs_restart = blk_mq_sched_needs_restart(hctx);
Naohiro Aota9586e672021-10-27 01:51:27 +09002144 if (prep == PREP_DISPATCH_NO_BUDGET)
2145 needs_resource = true;
Ming Lei86ff7c22018-01-30 22:04:57 -05002146 if (!needs_restart ||
Jens Axboeeb619fd2017-11-09 08:32:43 -07002147 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
Jens Axboebd166ef2017-01-17 06:03:22 -07002148 blk_mq_run_hw_queue(hctx, true);
Miaohe Lin6d5e8d22022-09-05 18:19:50 +08002149 else if (needs_resource)
Ming Lei86ff7c22018-01-30 22:04:57 -05002150 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
Jens Axboe1f57f8d2018-06-28 11:54:01 -06002151
Ming Lei6e7687172018-07-03 09:03:16 -06002152 blk_mq_update_dispatch_busy(hctx, true);
Jens Axboe1f57f8d2018-06-28 11:54:01 -06002153 return false;
Ming Lei6e7687172018-07-03 09:03:16 -06002154 } else
2155 blk_mq_update_dispatch_busy(hctx, false);
Jens Axboef04c3df2016-12-07 08:41:17 -07002156
Jens Axboe93efe982017-03-24 12:04:19 -06002157 return (queued + errors) != 0;
Jens Axboef04c3df2016-12-07 08:41:17 -07002158}
2159
André Almeida105663f2020-01-06 15:08:18 -03002160/**
2161 * __blk_mq_run_hw_queue - Run a hardware queue.
2162 * @hctx: Pointer to the hardware queue to run.
2163 *
2164 * Send pending requests to the hardware.
2165 */
Bart Van Assche6a83e742016-11-02 10:09:51 -06002166static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2167{
Jens Axboeb7a71e62017-08-01 09:28:24 -06002168 /*
Jens Axboeb7a71e62017-08-01 09:28:24 -06002169 * We can't run the queue inline with ints disabled. Ensure that
2170 * we catch bad users of this early.
2171 */
2172 WARN_ON_ONCE(in_interrupt());
2173
Ming Leibcc330f2021-12-03 21:15:33 +08002174 blk_mq_run_dispatch_ops(hctx->queue,
2175 blk_mq_sched_dispatch_requests(hctx));
Bart Van Assche6a83e742016-11-02 10:09:51 -06002176}
2177
Ming Leif82ddf12018-04-08 17:48:10 +08002178static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2179{
2180 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2181
2182 if (cpu >= nr_cpu_ids)
2183 cpu = cpumask_first(hctx->cpumask);
2184 return cpu;
2185}
2186
Jens Axboe506e9312014-05-07 10:26:44 -06002187/*
2188 * It'd be great if the workqueue API had a way to pass
2189 * in a mask and had some smarts for more clever placement.
2190 * For now we just round-robin here, switching for every
2191 * BLK_MQ_CPU_WORK_BATCH queued items.
2192 */
2193static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2194{
Ming Lei7bed4592018-01-18 00:41:51 +08002195 bool tried = false;
Ming Lei476f8c92018-04-08 17:48:09 +08002196 int next_cpu = hctx->next_cpu;
Ming Lei7bed4592018-01-18 00:41:51 +08002197
Christoph Hellwigb657d7e2014-11-24 09:27:23 +01002198 if (hctx->queue->nr_hw_queues == 1)
2199 return WORK_CPU_UNBOUND;
Jens Axboe506e9312014-05-07 10:26:44 -06002200
2201 if (--hctx->next_cpu_batch <= 0) {
Ming Lei7bed4592018-01-18 00:41:51 +08002202select_cpu:
Ming Lei476f8c92018-04-08 17:48:09 +08002203 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
Christoph Hellwig20e4d8132018-01-12 10:53:06 +08002204 cpu_online_mask);
Jens Axboe506e9312014-05-07 10:26:44 -06002205 if (next_cpu >= nr_cpu_ids)
Ming Leif82ddf12018-04-08 17:48:10 +08002206 next_cpu = blk_mq_first_mapped_cpu(hctx);
Jens Axboe506e9312014-05-07 10:26:44 -06002207 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2208 }
2209
Ming Lei7bed4592018-01-18 00:41:51 +08002210 /*
2211 * Do unbound schedule if we can't find a online CPU for this hctx,
2212 * and it should only happen in the path of handling CPU DEAD.
2213 */
Ming Lei476f8c92018-04-08 17:48:09 +08002214 if (!cpu_online(next_cpu)) {
Ming Lei7bed4592018-01-18 00:41:51 +08002215 if (!tried) {
2216 tried = true;
2217 goto select_cpu;
2218 }
2219
2220 /*
2221 * Make sure to re-select CPU next time once after CPUs
2222 * in hctx->cpumask become online again.
2223 */
Ming Lei476f8c92018-04-08 17:48:09 +08002224 hctx->next_cpu = next_cpu;
Ming Lei7bed4592018-01-18 00:41:51 +08002225 hctx->next_cpu_batch = 1;
2226 return WORK_CPU_UNBOUND;
2227 }
Ming Lei476f8c92018-04-08 17:48:09 +08002228
2229 hctx->next_cpu = next_cpu;
2230 return next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -06002231}
2232
André Almeida105663f2020-01-06 15:08:18 -03002233/**
2234 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2235 * @hctx: Pointer to the hardware queue to run.
2236 * @async: If we want to run the queue asynchronously.
Minwoo Imfa94ba82020-12-05 00:20:55 +09002237 * @msecs: Milliseconds of delay to wait before running the queue.
André Almeida105663f2020-01-06 15:08:18 -03002238 *
2239 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2240 * with a delay of @msecs.
2241 */
Bart Van Assche7587a5a2017-04-07 11:16:52 -07002242static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2243 unsigned long msecs)
Jens Axboe320ae512013-10-24 09:20:05 +01002244{
Bart Van Assche5435c022017-06-20 11:15:49 -07002245 if (unlikely(blk_mq_hctx_stopped(hctx)))
Jens Axboe320ae512013-10-24 09:20:05 +01002246 return;
2247
Jens Axboe1b792f22016-09-21 10:12:13 -06002248 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
Sebastian Andrzej Siewior3c8f9da2022-06-22 10:25:54 +02002249 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
Paolo Bonzini398205b2014-11-07 23:03:59 +01002250 __blk_mq_run_hw_queue(hctx);
Paolo Bonzini398205b2014-11-07 23:03:59 +01002251 return;
2252 }
Jens Axboee4043dc2014-04-09 10:18:23 -06002253 }
Paolo Bonzini398205b2014-11-07 23:03:59 +01002254
Bart Van Asscheae943d22018-01-19 08:58:55 -08002255 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2256 msecs_to_jiffies(msecs));
Bart Van Assche7587a5a2017-04-07 11:16:52 -07002257}
2258
André Almeida105663f2020-01-06 15:08:18 -03002259/**
2260 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2261 * @hctx: Pointer to the hardware queue to run.
Minwoo Imfa94ba82020-12-05 00:20:55 +09002262 * @msecs: Milliseconds of delay to wait before running the queue.
André Almeida105663f2020-01-06 15:08:18 -03002263 *
2264 * Run a hardware queue asynchronously with a delay of @msecs.
2265 */
Bart Van Assche7587a5a2017-04-07 11:16:52 -07002266void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2267{
2268 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2269}
2270EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2271
André Almeida105663f2020-01-06 15:08:18 -03002272/**
2273 * blk_mq_run_hw_queue - Start to run a hardware queue.
2274 * @hctx: Pointer to the hardware queue to run.
2275 * @async: If we want to run the queue asynchronously.
2276 *
2277 * Check if the request queue is not in a quiesced state and if there are
2278 * pending requests to be sent. If this is true, run the queue to send requests
2279 * to hardware.
2280 */
John Garry626fb732019-10-30 00:59:30 +08002281void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
Bart Van Assche7587a5a2017-04-07 11:16:52 -07002282{
Ming Lei24f5a902018-01-06 16:27:38 +08002283 bool need_run;
2284
2285 /*
2286 * When queue is quiesced, we may be switching io scheduler, or
2287 * updating nr_hw_queues, or other things, and we can't run queue
2288 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2289 *
2290 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2291 * quiesced.
2292 */
Ming Lei41adf532021-12-06 19:12:13 +08002293 __blk_mq_run_dispatch_ops(hctx->queue, false,
Ming Lei2a904d02021-12-03 21:15:31 +08002294 need_run = !blk_queue_quiesced(hctx->queue) &&
2295 blk_mq_hctx_has_pending(hctx));
Ming Lei24f5a902018-01-06 16:27:38 +08002296
John Garry626fb732019-10-30 00:59:30 +08002297 if (need_run)
Jens Axboe79f720a2017-11-10 09:13:21 -07002298 __blk_mq_delay_run_hw_queue(hctx, async, 0);
Jens Axboe320ae512013-10-24 09:20:05 +01002299}
Omar Sandoval5b727272017-04-14 01:00:00 -07002300EXPORT_SYMBOL(blk_mq_run_hw_queue);
Jens Axboe320ae512013-10-24 09:20:05 +01002301
Jan Karab6e68ee2021-01-11 17:47:17 +01002302/*
Jan Karab6e68ee2021-01-11 17:47:17 +01002303 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2304 * scheduler.
2305 */
2306static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2307{
Ming Lei5d054262022-05-22 20:23:50 +08002308 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
Jan Karab6e68ee2021-01-11 17:47:17 +01002309 /*
2310 * If the IO scheduler does not respect hardware queues when
2311 * dispatching, we just don't bother with multiple HW queues and
2312 * dispatch from hctx for the current CPU since running multiple queues
2313 * just causes lock contention inside the scheduler and pointless cache
2314 * bouncing.
2315 */
Bart Van Assche51ab80f2022-06-15 15:55:49 -07002316 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
Ming Lei5d054262022-05-22 20:23:50 +08002317
Jan Karab6e68ee2021-01-11 17:47:17 +01002318 if (!blk_mq_hctx_stopped(hctx))
2319 return hctx;
2320 return NULL;
2321}
2322
André Almeida105663f2020-01-06 15:08:18 -03002323/**
Mauro Carvalho Chehab24f7bb82020-10-23 18:32:54 +02002324 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
André Almeida105663f2020-01-06 15:08:18 -03002325 * @q: Pointer to the request queue to run.
2326 * @async: If we want to run the queue asynchronously.
2327 */
Mike Snitzerb94ec292015-03-11 23:56:38 -04002328void blk_mq_run_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01002329{
Jan Karab6e68ee2021-01-11 17:47:17 +01002330 struct blk_mq_hw_ctx *hctx, *sq_hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002331 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01002332
Jan Karab6e68ee2021-01-11 17:47:17 +01002333 sq_hctx = NULL;
Ming Lei4d337ceb2022-06-16 09:44:00 +08002334 if (blk_queue_sq_sched(q))
Jan Karab6e68ee2021-01-11 17:47:17 +01002335 sq_hctx = blk_mq_get_sq_hctx(q);
Jens Axboe320ae512013-10-24 09:20:05 +01002336 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboe79f720a2017-11-10 09:13:21 -07002337 if (blk_mq_hctx_stopped(hctx))
Jens Axboe320ae512013-10-24 09:20:05 +01002338 continue;
Jan Karab6e68ee2021-01-11 17:47:17 +01002339 /*
2340 * Dispatch from this hctx either if there's no hctx preferred
2341 * by IO scheduler or if it has requests that bypass the
2342 * scheduler.
2343 */
2344 if (!sq_hctx || sq_hctx == hctx ||
2345 !list_empty_careful(&hctx->dispatch))
2346 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +01002347 }
2348}
Mike Snitzerb94ec292015-03-11 23:56:38 -04002349EXPORT_SYMBOL(blk_mq_run_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +01002350
Bart Van Asschefd001442016-10-28 17:19:37 -07002351/**
Douglas Andersonb9151e72020-04-20 09:24:52 -07002352 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2353 * @q: Pointer to the request queue to run.
Minwoo Imfa94ba82020-12-05 00:20:55 +09002354 * @msecs: Milliseconds of delay to wait before running the queues.
Douglas Andersonb9151e72020-04-20 09:24:52 -07002355 */
2356void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2357{
Jan Karab6e68ee2021-01-11 17:47:17 +01002358 struct blk_mq_hw_ctx *hctx, *sq_hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002359 unsigned long i;
Douglas Andersonb9151e72020-04-20 09:24:52 -07002360
Jan Karab6e68ee2021-01-11 17:47:17 +01002361 sq_hctx = NULL;
Ming Lei4d337ceb2022-06-16 09:44:00 +08002362 if (blk_queue_sq_sched(q))
Jan Karab6e68ee2021-01-11 17:47:17 +01002363 sq_hctx = blk_mq_get_sq_hctx(q);
Douglas Andersonb9151e72020-04-20 09:24:52 -07002364 queue_for_each_hw_ctx(q, hctx, i) {
2365 if (blk_mq_hctx_stopped(hctx))
2366 continue;
Jan Karab6e68ee2021-01-11 17:47:17 +01002367 /*
David Jeffery8f5fea62022-01-31 15:33:37 -05002368 * If there is already a run_work pending, leave the
2369 * pending delay untouched. Otherwise, a hctx can stall
2370 * if another hctx is re-delaying the other's work
2371 * before the work executes.
2372 */
2373 if (delayed_work_pending(&hctx->run_work))
2374 continue;
2375 /*
Jan Karab6e68ee2021-01-11 17:47:17 +01002376 * Dispatch from this hctx either if there's no hctx preferred
2377 * by IO scheduler or if it has requests that bypass the
2378 * scheduler.
2379 */
2380 if (!sq_hctx || sq_hctx == hctx ||
2381 !list_empty_careful(&hctx->dispatch))
2382 blk_mq_delay_run_hw_queue(hctx, msecs);
Douglas Andersonb9151e72020-04-20 09:24:52 -07002383 }
2384}
2385EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2386
Ming Lei39a70c72017-06-06 23:22:09 +08002387/*
2388 * This function is often used for pausing .queue_rq() by driver when
2389 * there isn't enough resource or some conditions aren't satisfied, and
Bart Van Assche4d606212017-08-17 16:23:00 -07002390 * BLK_STS_RESOURCE is usually returned.
Ming Lei39a70c72017-06-06 23:22:09 +08002391 *
2392 * We do not guarantee that dispatch can be drained or blocked
2393 * after blk_mq_stop_hw_queue() returns. Please use
2394 * blk_mq_quiesce_queue() for that requirement.
2395 */
Jens Axboe320ae512013-10-24 09:20:05 +01002396void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2397{
Ming Lei641a9ed2017-06-06 23:22:10 +08002398 cancel_delayed_work(&hctx->run_work);
2399
2400 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboe320ae512013-10-24 09:20:05 +01002401}
2402EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2403
Ming Lei39a70c72017-06-06 23:22:09 +08002404/*
2405 * This function is often used for pausing .queue_rq() by driver when
2406 * there isn't enough resource or some conditions aren't satisfied, and
Bart Van Assche4d606212017-08-17 16:23:00 -07002407 * BLK_STS_RESOURCE is usually returned.
Ming Lei39a70c72017-06-06 23:22:09 +08002408 *
2409 * We do not guarantee that dispatch can be drained or blocked
2410 * after blk_mq_stop_hw_queues() returns. Please use
2411 * blk_mq_quiesce_queue() for that requirement.
2412 */
Jens Axboe2719aa22017-05-03 11:08:14 -06002413void blk_mq_stop_hw_queues(struct request_queue *q)
2414{
Ming Lei641a9ed2017-06-06 23:22:10 +08002415 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002416 unsigned long i;
Ming Lei641a9ed2017-06-06 23:22:10 +08002417
2418 queue_for_each_hw_ctx(q, hctx, i)
2419 blk_mq_stop_hw_queue(hctx);
Christoph Hellwig280d45f2013-10-25 14:45:58 +01002420}
2421EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2422
Jens Axboe320ae512013-10-24 09:20:05 +01002423void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2424{
2425 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboee4043dc2014-04-09 10:18:23 -06002426
Jens Axboe0ffbce82014-06-25 08:22:34 -06002427 blk_mq_run_hw_queue(hctx, false);
Jens Axboe320ae512013-10-24 09:20:05 +01002428}
2429EXPORT_SYMBOL(blk_mq_start_hw_queue);
2430
Christoph Hellwig2f268552014-04-16 09:44:56 +02002431void blk_mq_start_hw_queues(struct request_queue *q)
2432{
2433 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002434 unsigned long i;
Christoph Hellwig2f268552014-04-16 09:44:56 +02002435
2436 queue_for_each_hw_ctx(q, hctx, i)
2437 blk_mq_start_hw_queue(hctx);
2438}
2439EXPORT_SYMBOL(blk_mq_start_hw_queues);
2440
Jens Axboeae911c52016-12-08 13:19:30 -07002441void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2442{
2443 if (!blk_mq_hctx_stopped(hctx))
2444 return;
2445
2446 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2447 blk_mq_run_hw_queue(hctx, async);
2448}
2449EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2450
Christoph Hellwig1b4a3252014-04-16 09:44:54 +02002451void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01002452{
2453 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08002454 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01002455
Jens Axboeae911c52016-12-08 13:19:30 -07002456 queue_for_each_hw_ctx(q, hctx, i)
2457 blk_mq_start_stopped_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +01002458}
2459EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2460
Christoph Hellwig70f4db62014-04-16 10:48:08 -06002461static void blk_mq_run_work_fn(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +01002462{
2463 struct blk_mq_hw_ctx *hctx;
2464
Jens Axboe9f993732017-04-10 09:54:54 -06002465 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
Jens Axboe21c6e932017-04-10 09:54:56 -06002466
2467 /*
Ming Lei15fe8a902018-04-08 17:48:11 +08002468 * If we are stopped, don't run the queue.
Jens Axboe21c6e932017-04-10 09:54:56 -06002469 */
Yufen Yu08410312020-10-08 23:26:30 -04002470 if (blk_mq_hctx_stopped(hctx))
Jianchao Wang0196d6b2018-06-04 17:03:55 +08002471 return;
Jens Axboee4043dc2014-04-09 10:18:23 -06002472
Jens Axboe320ae512013-10-24 09:20:05 +01002473 __blk_mq_run_hw_queue(hctx);
2474}
2475
Ming Leicfd0c552015-10-20 23:13:57 +08002476static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
Ming Leicfd0c552015-10-20 23:13:57 +08002477 struct request *rq,
2478 bool at_head)
Jens Axboe320ae512013-10-24 09:20:05 +01002479{
Jens Axboee57690f2016-08-24 15:34:35 -06002480 struct blk_mq_ctx *ctx = rq->mq_ctx;
Ming Leic16d6b52018-12-17 08:44:05 -07002481 enum hctx_type type = hctx->type;
Jens Axboee57690f2016-08-24 15:34:35 -06002482
Bart Van Assche7b607812017-06-20 11:15:47 -07002483 lockdep_assert_held(&ctx->lock);
2484
Christoph Hellwiga54895f2020-12-03 17:21:39 +01002485 trace_block_rq_insert(rq);
Jens Axboe01b983c2013-11-19 18:59:10 -07002486
Christoph Hellwig72a0a362014-02-07 10:22:36 -08002487 if (at_head)
Ming Leic16d6b52018-12-17 08:44:05 -07002488 list_add(&rq->queuelist, &ctx->rq_lists[type]);
Christoph Hellwig72a0a362014-02-07 10:22:36 -08002489 else
Ming Leic16d6b52018-12-17 08:44:05 -07002490 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
Ming Leicfd0c552015-10-20 23:13:57 +08002491}
Jens Axboe4bb659b2014-05-09 09:36:49 -06002492
Jens Axboe2c3ad662016-12-14 14:34:47 -07002493void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2494 bool at_head)
Ming Leicfd0c552015-10-20 23:13:57 +08002495{
2496 struct blk_mq_ctx *ctx = rq->mq_ctx;
2497
Bart Van Assche7b607812017-06-20 11:15:47 -07002498 lockdep_assert_held(&ctx->lock);
2499
Jens Axboee57690f2016-08-24 15:34:35 -06002500 __blk_mq_insert_req_list(hctx, rq, at_head);
Jens Axboe320ae512013-10-24 09:20:05 +01002501 blk_mq_hctx_mark_pending(hctx, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01002502}
2503
André Almeida105663f2020-01-06 15:08:18 -03002504/**
2505 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2506 * @rq: Pointer to request to be inserted.
Randy Dunlap26bfeb22020-08-16 16:39:34 -07002507 * @at_head: true if the request should be inserted at the head of the list.
André Almeida105663f2020-01-06 15:08:18 -03002508 * @run_queue: If we should run the hardware queue after inserting the request.
2509 *
Jens Axboe157f3772017-09-11 16:43:57 -06002510 * Should only be used carefully, when the caller knows we want to
2511 * bypass a potential IO scheduler on the target device.
2512 */
Ming Lei01e99ae2020-02-25 09:04:32 +08002513void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2514 bool run_queue)
Jens Axboe157f3772017-09-11 16:43:57 -06002515{
Jens Axboeea4f9952018-10-29 15:06:13 -06002516 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Jens Axboe157f3772017-09-11 16:43:57 -06002517
2518 spin_lock(&hctx->lock);
Ming Lei01e99ae2020-02-25 09:04:32 +08002519 if (at_head)
2520 list_add(&rq->queuelist, &hctx->dispatch);
2521 else
2522 list_add_tail(&rq->queuelist, &hctx->dispatch);
Jens Axboe157f3772017-09-11 16:43:57 -06002523 spin_unlock(&hctx->lock);
2524
Ming Leib0850292017-11-02 23:24:34 +08002525 if (run_queue)
2526 blk_mq_run_hw_queue(hctx, false);
Jens Axboe157f3772017-09-11 16:43:57 -06002527}
2528
Jens Axboebd166ef2017-01-17 06:03:22 -07002529void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2530 struct list_head *list)
Jens Axboe320ae512013-10-24 09:20:05 +01002531
2532{
Ming Lei3f0cedc2018-07-02 17:35:58 +08002533 struct request *rq;
Ming Leic16d6b52018-12-17 08:44:05 -07002534 enum hctx_type type = hctx->type;
Ming Lei3f0cedc2018-07-02 17:35:58 +08002535
Jens Axboe320ae512013-10-24 09:20:05 +01002536 /*
2537 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2538 * offline now
2539 */
Ming Lei3f0cedc2018-07-02 17:35:58 +08002540 list_for_each_entry(rq, list, queuelist) {
Jens Axboee57690f2016-08-24 15:34:35 -06002541 BUG_ON(rq->mq_ctx != ctx);
Christoph Hellwiga54895f2020-12-03 17:21:39 +01002542 trace_block_rq_insert(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01002543 }
Ming Lei3f0cedc2018-07-02 17:35:58 +08002544
2545 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07002546 list_splice_tail_init(list, &ctx->rq_lists[type]);
Ming Leicfd0c552015-10-20 23:13:57 +08002547 blk_mq_hctx_mark_pending(hctx, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01002548 spin_unlock(&ctx->lock);
Jens Axboe320ae512013-10-24 09:20:05 +01002549}
2550
Jens Axboedc5fc3612021-10-19 06:02:30 -06002551static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2552 bool from_schedule)
Jens Axboe320ae512013-10-24 09:20:05 +01002553{
Jens Axboedc5fc3612021-10-19 06:02:30 -06002554 if (hctx->queue->mq_ops->commit_rqs) {
2555 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2556 hctx->queue->mq_ops->commit_rqs(hctx);
2557 }
2558 *queued = 0;
2559}
Jens Axboe320ae512013-10-24 09:20:05 +01002560
Christoph Hellwig14ccb662019-06-06 12:29:01 +02002561static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2562 unsigned int nr_segs)
Jens Axboe320ae512013-10-24 09:20:05 +01002563{
Eric Biggers93f221a2020-09-15 20:53:14 -07002564 int err;
2565
Christoph Hellwigf924cdd2019-06-06 12:29:00 +02002566 if (bio->bi_opf & REQ_RAHEAD)
2567 rq->cmd_flags |= REQ_FAILFAST_MASK;
2568
2569 rq->__sector = bio->bi_iter.bi_sector;
Christoph Hellwig14ccb662019-06-06 12:29:01 +02002570 blk_rq_bio_prep(rq, bio, nr_segs);
Eric Biggers93f221a2020-09-15 20:53:14 -07002571
2572 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2573 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2574 WARN_ON_ONCE(err);
Jens Axboe4b570522014-05-29 11:00:11 -06002575
Konstantin Khlebnikovb5af37a2020-05-27 07:24:16 +02002576 blk_account_io_start(rq);
Jens Axboe320ae512013-10-24 09:20:05 +01002577}
2578
Mike Snitzer0f955492018-01-17 11:25:56 -05002579static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig3e087732021-10-12 13:12:24 +02002580 struct request *rq, bool last)
Shaohua Lif984df12015-05-08 10:51:32 -07002581{
Shaohua Lif984df12015-05-08 10:51:32 -07002582 struct request_queue *q = rq->q;
Shaohua Lif984df12015-05-08 10:51:32 -07002583 struct blk_mq_queue_data bd = {
2584 .rq = rq,
Jens Axboebe94f052018-11-24 10:15:46 -07002585 .last = last,
Shaohua Lif984df12015-05-08 10:51:32 -07002586 };
Jens Axboef06345a2017-06-12 11:22:46 -06002587 blk_status_t ret;
Mike Snitzer0f955492018-01-17 11:25:56 -05002588
Mike Snitzer0f955492018-01-17 11:25:56 -05002589 /*
2590 * For OK queue, we are done. For error, caller may kill it.
2591 * Any other error (busy), just add it to our list as we
2592 * previously would have done.
2593 */
2594 ret = q->mq_ops->queue_rq(hctx, &bd);
2595 switch (ret) {
2596 case BLK_STS_OK:
Ming Lei6ce3dd62018-07-10 09:03:31 +08002597 blk_mq_update_dispatch_busy(hctx, false);
Mike Snitzer0f955492018-01-17 11:25:56 -05002598 break;
2599 case BLK_STS_RESOURCE:
Ming Lei86ff7c22018-01-30 22:04:57 -05002600 case BLK_STS_DEV_RESOURCE:
Ming Lei6ce3dd62018-07-10 09:03:31 +08002601 blk_mq_update_dispatch_busy(hctx, true);
Mike Snitzer0f955492018-01-17 11:25:56 -05002602 __blk_mq_requeue_request(rq);
2603 break;
2604 default:
Ming Lei6ce3dd62018-07-10 09:03:31 +08002605 blk_mq_update_dispatch_busy(hctx, false);
Mike Snitzer0f955492018-01-17 11:25:56 -05002606 break;
2607 }
2608
2609 return ret;
2610}
2611
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002612static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
Mike Snitzer0f955492018-01-17 11:25:56 -05002613 struct request *rq,
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002614 bool bypass_insert, bool last)
Mike Snitzer0f955492018-01-17 11:25:56 -05002615{
2616 struct request_queue *q = rq->q;
Ming Leid964f042017-06-06 23:22:00 +08002617 bool run_queue = true;
Ming Lei2a5a24a2021-01-22 10:33:12 +08002618 int budget_token;
Ming Leid964f042017-06-06 23:22:00 +08002619
Ming Lei23d4ee12018-01-18 12:06:59 +08002620 /*
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002621 * RCU or SRCU read lock is needed before checking quiesced flag.
Ming Lei23d4ee12018-01-18 12:06:59 +08002622 *
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002623 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2624 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2625 * and avoid driver to try to dispatch again.
Ming Lei23d4ee12018-01-18 12:06:59 +08002626 */
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002627 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
Ming Leid964f042017-06-06 23:22:00 +08002628 run_queue = false;
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002629 bypass_insert = false;
2630 goto insert;
Ming Leid964f042017-06-06 23:22:00 +08002631 }
Shaohua Lif984df12015-05-08 10:51:32 -07002632
Jens Axboe2ff06822021-10-15 09:44:38 -06002633 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002634 goto insert;
Bart Van Assche2253efc2016-10-28 17:20:02 -07002635
Ming Lei2a5a24a2021-01-22 10:33:12 +08002636 budget_token = blk_mq_get_dispatch_budget(q);
2637 if (budget_token < 0)
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002638 goto insert;
Jens Axboebd166ef2017-01-17 06:03:22 -07002639
Ming Lei2a5a24a2021-01-22 10:33:12 +08002640 blk_mq_set_rq_budget_token(rq, budget_token);
2641
Ming Lei8ab6bb9e2018-06-25 19:31:45 +08002642 if (!blk_mq_get_driver_tag(rq)) {
Ming Lei2a5a24a2021-01-22 10:33:12 +08002643 blk_mq_put_dispatch_budget(q, budget_token);
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002644 goto insert;
Ming Lei88022d72017-11-05 02:21:12 +08002645 }
Ming Leide148292017-10-14 17:22:29 +08002646
Christoph Hellwig3e087732021-10-12 13:12:24 +02002647 return __blk_mq_issue_directly(hctx, rq, last);
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002648insert:
2649 if (bypass_insert)
2650 return BLK_STS_RESOURCE;
2651
Ming Leidb03f882020-08-18 17:07:28 +08002652 blk_mq_sched_insert_request(rq, false, run_queue, false);
2653
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002654 return BLK_STS_OK;
2655}
2656
André Almeida105663f2020-01-06 15:08:18 -03002657/**
2658 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2659 * @hctx: Pointer of the associated hardware queue.
2660 * @rq: Pointer to request to be sent.
André Almeida105663f2020-01-06 15:08:18 -03002661 *
2662 * If the device has enough resources to accept a new request now, send the
2663 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2664 * we can try send it another time in the future. Requests inserted at this
2665 * queue have higher priority.
2666 */
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002667static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig3e087732021-10-12 13:12:24 +02002668 struct request *rq)
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002669{
Ming Lei2a904d02021-12-03 21:15:31 +08002670 blk_status_t ret =
2671 __blk_mq_try_issue_directly(hctx, rq, false, true);
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002672
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002673 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
Ming Lei01e99ae2020-02-25 09:04:32 +08002674 blk_mq_request_bypass_insert(rq, false, true);
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002675 else if (ret != BLK_STS_OK)
2676 blk_mq_end_request(rq, ret);
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002677}
2678
Christoph Hellwig06c8c692021-11-17 07:13:58 +01002679static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002680{
Ming Lei4cafe862021-12-03 21:15:34 +08002681 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
Christoph Hellwig5eb61262017-03-22 15:01:51 -04002682}
2683
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002684static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2685{
2686 struct blk_mq_hw_ctx *hctx = NULL;
2687 struct request *rq;
2688 int queued = 0;
2689 int errors = 0;
2690
2691 while ((rq = rq_list_pop(&plug->mq_list))) {
2692 bool last = rq_list_empty(plug->mq_list);
2693 blk_status_t ret;
2694
2695 if (hctx != rq->mq_hctx) {
2696 if (hctx)
2697 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2698 hctx = rq->mq_hctx;
2699 }
2700
2701 ret = blk_mq_request_issue_directly(rq, last);
2702 switch (ret) {
2703 case BLK_STS_OK:
2704 queued++;
2705 break;
2706 case BLK_STS_RESOURCE:
2707 case BLK_STS_DEV_RESOURCE:
Yufen Yud3b385962022-08-03 10:33:55 +08002708 blk_mq_request_bypass_insert(rq, false, true);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002709 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2710 return;
2711 default:
2712 blk_mq_end_request(rq, ret);
2713 errors++;
2714 break;
2715 }
2716 }
2717
2718 /*
2719 * If we didn't flush the entire list, we could have told the driver
2720 * there was more coming, but that turned out to be a lie.
2721 */
2722 if (errors)
2723 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2724}
2725
Keith Busch518579a2021-12-20 12:59:19 -08002726static void __blk_mq_flush_plug_list(struct request_queue *q,
2727 struct blk_plug *plug)
2728{
2729 if (blk_queue_quiesced(q))
2730 return;
2731 q->mq_ops->queue_rqs(&plug->mq_list);
2732}
2733
Jens Axboe26fed4a2022-03-11 10:24:17 -07002734static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2735{
2736 struct blk_mq_hw_ctx *this_hctx = NULL;
2737 struct blk_mq_ctx *this_ctx = NULL;
2738 struct request *requeue_list = NULL;
2739 unsigned int depth = 0;
2740 LIST_HEAD(list);
2741
2742 do {
2743 struct request *rq = rq_list_pop(&plug->mq_list);
2744
2745 if (!this_hctx) {
2746 this_hctx = rq->mq_hctx;
2747 this_ctx = rq->mq_ctx;
2748 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2749 rq_list_add(&requeue_list, rq);
2750 continue;
2751 }
2752 list_add_tail(&rq->queuelist, &list);
2753 depth++;
2754 } while (!rq_list_empty(plug->mq_list));
2755
2756 plug->mq_list = requeue_list;
2757 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2758 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2759}
2760
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002761void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2762{
Jens Axboe3c67d442021-12-03 06:48:53 -07002763 struct request *rq;
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002764
2765 if (rq_list_empty(plug->mq_list))
2766 return;
2767 plug->rq_count = 0;
2768
2769 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
Jens Axboe3c67d442021-12-03 06:48:53 -07002770 struct request_queue *q;
2771
2772 rq = rq_list_peek(&plug->mq_list);
2773 q = rq->q;
2774
2775 /*
2776 * Peek first request and see if we have a ->queue_rqs() hook.
2777 * If we do, we can dispatch the whole plug list in one go. We
2778 * already know at this point that all requests belong to the
2779 * same queue, caller must ensure that's the case.
2780 *
2781 * Since we pass off the full list to the driver at this point,
2782 * we do not increment the active request count for the queue.
2783 * Bypass shared tags for now because of that.
2784 */
2785 if (q->mq_ops->queue_rqs &&
2786 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2787 blk_mq_run_dispatch_ops(q,
Keith Busch518579a2021-12-20 12:59:19 -08002788 __blk_mq_flush_plug_list(q, plug));
Jens Axboe3c67d442021-12-03 06:48:53 -07002789 if (rq_list_empty(plug->mq_list))
2790 return;
2791 }
Ming Lei73f37602021-12-06 11:33:50 +08002792
2793 blk_mq_run_dispatch_ops(q,
Ming Lei4cafe862021-12-03 21:15:34 +08002794 blk_mq_plug_issue_direct(plug, false));
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002795 if (rq_list_empty(plug->mq_list))
2796 return;
2797 }
2798
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002799 do {
Jens Axboe26fed4a2022-03-11 10:24:17 -07002800 blk_mq_dispatch_plug_list(plug, from_schedule);
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002801 } while (!rq_list_empty(plug->mq_list));
Christoph Hellwigb84c5b52021-11-17 07:13:57 +01002802}
2803
Ming Lei6ce3dd62018-07-10 09:03:31 +08002804void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2805 struct list_head *list)
2806{
Keith Busch536167d42020-04-07 03:13:48 +09002807 int queued = 0;
yangerkun632bfb62020-09-05 19:25:56 +08002808 int errors = 0;
Keith Busch536167d42020-04-07 03:13:48 +09002809
Ming Lei6ce3dd62018-07-10 09:03:31 +08002810 while (!list_empty(list)) {
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002811 blk_status_t ret;
Ming Lei6ce3dd62018-07-10 09:03:31 +08002812 struct request *rq = list_first_entry(list, struct request,
2813 queuelist);
2814
2815 list_del_init(&rq->queuelist);
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002816 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2817 if (ret != BLK_STS_OK) {
Yu Kuai65fac0d2022-07-26 20:22:24 +08002818 errors++;
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002819 if (ret == BLK_STS_RESOURCE ||
2820 ret == BLK_STS_DEV_RESOURCE) {
Ming Lei01e99ae2020-02-25 09:04:32 +08002821 blk_mq_request_bypass_insert(rq, false,
Jens Axboec616cbe2018-12-06 22:17:44 -07002822 list_empty(list));
Bart Van Asschefd9c40f2019-04-04 10:08:43 -07002823 break;
2824 }
2825 blk_mq_end_request(rq, ret);
Keith Busch536167d42020-04-07 03:13:48 +09002826 } else
2827 queued++;
Ming Lei6ce3dd62018-07-10 09:03:31 +08002828 }
Jens Axboed666ba92018-11-27 17:02:25 -07002829
2830 /*
2831 * If we didn't flush the entire list, we could have told
2832 * the driver there was more coming, but that turned out to
2833 * be a lie.
2834 */
yangerkun632bfb62020-09-05 19:25:56 +08002835 if ((!list_empty(list) || errors) &&
2836 hctx->queue->mq_ops->commit_rqs && queued)
Jens Axboed666ba92018-11-27 17:02:25 -07002837 hctx->queue->mq_ops->commit_rqs(hctx);
Ming Lei6ce3dd62018-07-10 09:03:31 +08002838}
2839
Ming Leib131f202021-11-11 16:51:34 +08002840static bool blk_mq_attempt_bio_merge(struct request_queue *q,
Christoph Hellwig0c5bcc92021-11-23 17:04:41 +01002841 struct bio *bio, unsigned int nr_segs)
Jens Axboe900e0802021-11-03 05:47:09 -06002842{
2843 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
Christoph Hellwig0c5bcc92021-11-23 17:04:41 +01002844 if (blk_attempt_plug_merge(q, bio, nr_segs))
Jens Axboe900e0802021-11-03 05:47:09 -06002845 return true;
2846 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2847 return true;
2848 }
2849 return false;
2850}
2851
Jens Axboe71539712021-11-03 05:52:45 -06002852static struct request *blk_mq_get_new_requests(struct request_queue *q,
2853 struct blk_plug *plug,
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002854 struct bio *bio,
2855 unsigned int nsegs)
Jens Axboe71539712021-11-03 05:52:45 -06002856{
2857 struct blk_mq_alloc_data data = {
2858 .q = q,
2859 .nr_tags = 1,
Ming Lei9d497e22022-01-04 21:42:23 +08002860 .cmd_flags = bio->bi_opf,
Jens Axboe71539712021-11-03 05:52:45 -06002861 };
2862 struct request *rq;
2863
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002864 if (unlikely(bio_queue_enter(bio)))
Ming Leib6371082021-11-12 20:47:15 +08002865 return NULL;
Jens Axboe900e0802021-11-03 05:47:09 -06002866
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002867 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2868 goto queue_exit;
2869
2870 rq_qos_throttle(q, bio);
2871
Jens Axboe71539712021-11-03 05:52:45 -06002872 if (plug) {
2873 data.nr_tags = plug->nr_ios;
2874 plug->nr_ios = 1;
2875 data.cached_rq = &plug->cached_rq;
2876 }
2877
2878 rq = __blk_mq_alloc_requests(&data);
Jens Axboe373b5412021-12-02 12:42:58 -07002879 if (rq)
2880 return rq;
Jens Axboe71539712021-11-03 05:52:45 -06002881 rq_qos_cleanup(q, bio);
2882 if (bio->bi_opf & REQ_NOWAIT)
2883 bio_wouldblock_error(bio);
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002884queue_exit:
Jens Axboe95febeb2021-11-15 14:23:08 -07002885 blk_queue_exit(q);
2886 return NULL;
Jens Axboe71539712021-11-03 05:52:45 -06002887}
2888
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002889static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002890 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002891{
2892 struct request *rq;
2893
2894 if (!plug)
2895 return NULL;
2896 rq = rq_list_peek(&plug->cached_rq);
2897 if (!rq || rq->q != q)
2898 return NULL;
2899
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002900 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2901 *bio = NULL;
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002902 return NULL;
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002903 }
2904
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002905 if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type)
2906 return NULL;
2907 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002908 return NULL;
2909
Jens Axboe26456722022-06-21 10:03:57 -06002910 /*
2911 * If any qos ->throttle() end up blocking, we will have flushed the
2912 * plug and hence killed the cached_rq list as well. Pop this entry
2913 * before we throttle.
2914 */
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002915 plug->cached_rq = rq_list_next(rq);
Jens Axboe26456722022-06-21 10:03:57 -06002916 rq_qos_throttle(q, *bio);
2917
2918 rq->cmd_flags = (*bio)->bi_opf;
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002919 INIT_LIST_HEAD(&rq->queuelist);
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002920 return rq;
2921}
2922
Jan Kara82b74ca2022-06-23 09:48:32 +02002923static void bio_set_ioprio(struct bio *bio)
2924{
Jan Karaa78418e2022-06-23 09:48:34 +02002925 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2926 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2927 bio->bi_ioprio = get_current_ioprio();
Jan Kara82b74ca2022-06-23 09:48:32 +02002928 blkcg_set_ioprio(bio);
2929}
2930
André Almeida105663f2020-01-06 15:08:18 -03002931/**
Christoph Hellwigc62b37d2020-07-01 10:59:43 +02002932 * blk_mq_submit_bio - Create and send a request to block device.
André Almeida105663f2020-01-06 15:08:18 -03002933 * @bio: Bio pointer.
2934 *
2935 * Builds up a request structure from @q and @bio and send to the device. The
2936 * request may not be queued directly to hardware if:
2937 * * This request can be merged with another one
2938 * * We want to place request at plug queue for possible future merging
2939 * * There is an IO scheduler active at this queue
2940 *
2941 * It will not queue the request if there is an error with the bio, or at the
2942 * request creation.
André Almeida105663f2020-01-06 15:08:18 -03002943 */
Christoph Hellwig3e087732021-10-12 13:12:24 +02002944void blk_mq_submit_bio(struct bio *bio)
Jens Axboe07068d52014-05-22 10:40:51 -06002945{
Pavel Begunkoved6cdde2021-10-14 15:03:30 +01002946 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
Christoph Hellwig6deacb32022-07-06 09:03:38 +02002947 struct blk_plug *plug = blk_mq_plug(bio);
Christoph Hellwigef295ec2016-10-28 08:48:16 -06002948 const int is_sync = op_is_sync(bio->bi_opf);
Jens Axboe07068d52014-05-22 10:40:51 -06002949 struct request *rq;
Jens Axboeabd45c12021-10-13 12:43:41 -06002950 unsigned int nr_segs = 1;
Satya Tangiralaa892c8d2020-05-14 00:37:18 +00002951 blk_status_t ret;
Jens Axboe07068d52014-05-22 10:40:51 -06002952
Christoph Hellwig51d798c2022-07-27 12:22:56 -04002953 bio = blk_queue_bounce(bio, q);
Jens Axboe613b1482023-01-04 08:51:19 -07002954 if (bio_may_exceed_limits(bio, &q->limits)) {
Christoph Hellwigc55ddd92022-07-27 12:23:00 -04002955 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
Jens Axboe613b1482023-01-04 08:51:19 -07002956 if (!bio)
2957 return;
2958 }
Wen Xiongf36ea502017-05-10 08:54:11 -05002959
Dmitry Monakhove23947b2017-06-29 11:31:11 -07002960 if (!bio_integrity_prep(bio))
Jens Axboe900e0802021-11-03 05:47:09 -06002961 return;
Jens Axboe87760e52016-11-09 12:38:14 -07002962
Jan Kara9c6227e2022-06-23 09:48:33 +02002963 bio_set_ioprio(bio);
2964
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002965 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002966 if (!rq) {
Shin'ichiro Kawasaki0a5aa8d2022-03-08 17:09:15 +09002967 if (!bio)
2968 return;
2969 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
Christoph Hellwig5b13bc82021-11-24 07:28:56 +01002970 if (unlikely(!rq))
2971 return;
2972 }
Jens Axboe87760e52016-11-09 12:38:14 -07002973
Christoph Hellwige8a676d2020-12-03 17:21:36 +01002974 trace_block_getrq(bio);
Xiaoguang Wangd6f1dda2018-10-23 22:30:50 +08002975
Josef Bacikc1c80382018-07-03 11:14:59 -04002976 rq_qos_track(q, rq, bio);
Jens Axboe07068d52014-05-22 10:40:51 -06002977
Bart Van Assche970d1682019-07-01 08:47:30 -07002978 blk_mq_bio_to_request(rq, bio, nr_segs);
2979
Satya Tangiralaa892c8d2020-05-14 00:37:18 +00002980 ret = blk_crypto_init_request(rq);
2981 if (ret != BLK_STS_OK) {
2982 bio->bi_status = ret;
2983 bio_endio(bio);
2984 blk_mq_free_request(rq);
Christoph Hellwig3e087732021-10-12 13:12:24 +02002985 return;
Satya Tangiralaa892c8d2020-05-14 00:37:18 +00002986 }
2987
Ming Lei2b504bd2021-11-18 23:30:41 +08002988 if (op_is_flush(bio->bi_opf)) {
2989 blk_insert_flush(rq);
Christoph Hellwigd92ca9d82021-10-19 14:25:53 +02002990 return;
Ming Lei2b504bd2021-11-18 23:30:41 +08002991 }
Christoph Hellwigd92ca9d82021-10-19 14:25:53 +02002992
Christoph Hellwig1e9c2302021-11-23 17:04:42 +01002993 if (plug)
Jens Axboece5b0092018-11-27 17:13:56 -07002994 blk_add_rq_to_plug(plug, rq);
Christoph Hellwig1e9c2302021-11-23 17:04:42 +01002995 else if ((rq->rq_flags & RQF_ELV) ||
2996 (rq->mq_hctx->dispatch_busy &&
2997 (q->nr_hw_queues == 1 || !is_sync)))
Ming Leia12de1d2019-09-27 15:24:30 +08002998 blk_mq_sched_insert_request(rq, false, true, true);
Christoph Hellwig1e9c2302021-11-23 17:04:42 +01002999 else
Ming Leibcc330f2021-12-03 21:15:33 +08003000 blk_mq_run_dispatch_ops(rq->q,
Ming Lei2a904d02021-12-03 21:15:31 +08003001 blk_mq_try_issue_directly(rq->mq_hctx, rq));
Jens Axboe320ae512013-10-24 09:20:05 +01003002}
3003
Christoph Hellwig248c7932022-02-15 11:05:36 +01003004#ifdef CONFIG_BLK_MQ_STACKING
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003005/**
Christoph Hellwiga5efda32022-02-15 11:05:37 +01003006 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
Christoph Hellwiga5efda32022-02-15 11:05:37 +01003007 * @rq: the request being queued
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003008 */
Christoph Hellwig28db4712022-02-15 11:05:38 +01003009blk_status_t blk_insert_cloned_request(struct request *rq)
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003010{
Christoph Hellwig28db4712022-02-15 11:05:38 +01003011 struct request_queue *q = rq->q;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003012 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
Christoph Hellwiga5efda32022-02-15 11:05:37 +01003013 blk_status_t ret;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003014
3015 if (blk_rq_sectors(rq) > max_sectors) {
3016 /*
3017 * SCSI device does not have a good way to return if
3018 * Write Same/Zero is actually supported. If a device rejects
3019 * a non-read/write command (discard, write same,etc.) the
3020 * low-level device driver will set the relevant queue limit to
3021 * 0 to prevent blk-lib from issuing more of the offending
3022 * operations. Commands queued prior to the queue limit being
3023 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3024 * errors being propagated to upper layers.
3025 */
3026 if (max_sectors == 0)
3027 return BLK_STS_NOTSUPP;
3028
3029 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3030 __func__, blk_rq_sectors(rq), max_sectors);
3031 return BLK_STS_IOERR;
3032 }
3033
3034 /*
3035 * The queue settings related to segment counting may differ from the
3036 * original queue.
3037 */
3038 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3039 if (rq->nr_phys_segments > queue_max_segments(q)) {
3040 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
3041 __func__, rq->nr_phys_segments, queue_max_segments(q));
3042 return BLK_STS_IOERR;
3043 }
3044
Christoph Hellwig28db4712022-02-15 11:05:38 +01003045 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003046 return BLK_STS_IOERR;
3047
3048 if (blk_crypto_insert_cloned_request(rq))
3049 return BLK_STS_IOERR;
3050
3051 blk_account_io_start(rq);
3052
3053 /*
3054 * Since we have a scheduler attached on the top device,
3055 * bypass a potential scheduler on the bottom device for
3056 * insert.
3057 */
Christoph Hellwig28db4712022-02-15 11:05:38 +01003058 blk_mq_run_dispatch_ops(q,
Ming Lei4cafe862021-12-03 21:15:34 +08003059 ret = blk_mq_request_issue_directly(rq, true));
Yu Kuai592ee112022-01-26 09:21:32 +08003060 if (ret)
3061 blk_account_io_done(rq, ktime_get_ns());
Ming Lei4cafe862021-12-03 21:15:34 +08003062 return ret;
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003063}
3064EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3065
3066/**
3067 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3068 * @rq: the clone request to be cleaned up
3069 *
3070 * Description:
3071 * Free all bios in @rq for a cloned request.
3072 */
3073void blk_rq_unprep_clone(struct request *rq)
3074{
3075 struct bio *bio;
3076
3077 while ((bio = rq->bio) != NULL) {
3078 rq->bio = bio->bi_next;
3079
3080 bio_put(bio);
3081 }
3082}
3083EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3084
3085/**
3086 * blk_rq_prep_clone - Helper function to setup clone request
3087 * @rq: the request to be setup
3088 * @rq_src: original request to be cloned
3089 * @bs: bio_set that bios for clone are allocated from
3090 * @gfp_mask: memory allocation mask for bio
3091 * @bio_ctr: setup function to be called for each clone bio.
3092 * Returns %0 for success, non %0 for failure.
3093 * @data: private data to be passed to @bio_ctr
3094 *
3095 * Description:
3096 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3097 * Also, pages which the original bios are pointing to are not copied
3098 * and the cloned bios just point same pages.
3099 * So cloned bios must be completed before original bios, which means
3100 * the caller must complete @rq before @rq_src.
3101 */
3102int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3103 struct bio_set *bs, gfp_t gfp_mask,
3104 int (*bio_ctr)(struct bio *, struct bio *, void *),
3105 void *data)
3106{
3107 struct bio *bio, *bio_src;
3108
3109 if (!bs)
3110 bs = &fs_bio_set;
3111
3112 __rq_for_each_bio(bio_src, rq_src) {
Christoph Hellwigabfc4262022-02-02 17:01:09 +01003113 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3114 bs);
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003115 if (!bio)
3116 goto free_and_out;
3117
3118 if (bio_ctr && bio_ctr(bio, bio_src, data))
3119 goto free_and_out;
3120
3121 if (rq->bio) {
3122 rq->biotail->bi_next = bio;
3123 rq->biotail = bio;
3124 } else {
3125 rq->bio = rq->biotail = bio;
3126 }
3127 bio = NULL;
3128 }
3129
3130 /* Copy attributes of the original request to the clone request. */
3131 rq->__sector = blk_rq_pos(rq_src);
3132 rq->__data_len = blk_rq_bytes(rq_src);
3133 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3134 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3135 rq->special_vec = rq_src->special_vec;
3136 }
3137 rq->nr_phys_segments = rq_src->nr_phys_segments;
3138 rq->ioprio = rq_src->ioprio;
3139
3140 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3141 goto free_and_out;
3142
3143 return 0;
3144
3145free_and_out:
3146 if (bio)
3147 bio_put(bio);
3148 blk_rq_unprep_clone(rq);
3149
3150 return -ENOMEM;
3151}
3152EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
Christoph Hellwig248c7932022-02-15 11:05:36 +01003153#endif /* CONFIG_BLK_MQ_STACKING */
Christoph Hellwig06c8c692021-11-17 07:13:58 +01003154
Christoph Hellwigf2b8f3c2021-11-17 07:14:00 +01003155/*
3156 * Steal bios from a request and add them to a bio list.
3157 * The request must not have been partially completed before.
3158 */
3159void blk_steal_bios(struct bio_list *list, struct request *rq)
3160{
3161 if (rq->bio) {
3162 if (list->tail)
3163 list->tail->bi_next = rq->bio;
3164 else
3165 list->head = rq->bio;
3166 list->tail = rq->biotail;
3167
3168 rq->bio = NULL;
3169 rq->biotail = NULL;
3170 }
3171
3172 rq->__data_len = 0;
3173}
3174EXPORT_SYMBOL_GPL(blk_steal_bios);
3175
Ming Leibd631412021-05-11 23:22:35 +08003176static size_t order_to_size(unsigned int order)
3177{
3178 return (size_t)PAGE_SIZE << order;
3179}
3180
3181/* called before freeing request pool in @tags */
John Garryf32e4ea2021-10-05 18:23:32 +08003182static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3183 struct blk_mq_tags *tags)
Ming Leibd631412021-05-11 23:22:35 +08003184{
Ming Leibd631412021-05-11 23:22:35 +08003185 struct page *page;
3186 unsigned long flags;
3187
Yu Kuai76dd2982022-10-11 22:22:53 +08003188 /*
3189 * There is no need to clear mapping if driver tags is not initialized
3190 * or the mapping belongs to the driver tags.
3191 */
3192 if (!drv_tags || drv_tags == tags)
John Garry4f245d52021-10-05 18:23:33 +08003193 return;
3194
Ming Leibd631412021-05-11 23:22:35 +08003195 list_for_each_entry(page, &tags->page_list, lru) {
3196 unsigned long start = (unsigned long)page_address(page);
3197 unsigned long end = start + order_to_size(page->private);
3198 int i;
3199
John Garryf32e4ea2021-10-05 18:23:32 +08003200 for (i = 0; i < drv_tags->nr_tags; i++) {
Ming Leibd631412021-05-11 23:22:35 +08003201 struct request *rq = drv_tags->rqs[i];
3202 unsigned long rq_addr = (unsigned long)rq;
3203
3204 if (rq_addr >= start && rq_addr < end) {
Jens Axboe0a467d02021-10-14 14:39:59 -06003205 WARN_ON_ONCE(req_ref_read(rq) != 0);
Ming Leibd631412021-05-11 23:22:35 +08003206 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3207 }
3208 }
3209 }
3210
3211 /*
3212 * Wait until all pending iteration is done.
3213 *
3214 * Request reference is cleared and it is guaranteed to be observed
3215 * after the ->lock is released.
3216 */
3217 spin_lock_irqsave(&drv_tags->lock, flags);
3218 spin_unlock_irqrestore(&drv_tags->lock, flags);
3219}
3220
Jens Axboecc71a6f2017-01-11 14:29:56 -07003221void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3222 unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01003223{
John Garryf32e4ea2021-10-05 18:23:32 +08003224 struct blk_mq_tags *drv_tags;
Jens Axboe320ae512013-10-24 09:20:05 +01003225 struct page *page;
3226
Ming Leie02657e2022-03-08 06:51:48 +01003227 if (list_empty(&tags->page_list))
3228 return;
3229
John Garry079a2e32021-10-05 18:23:39 +08003230 if (blk_mq_is_shared_tags(set->flags))
3231 drv_tags = set->shared_tags;
John Garrye155b0c2021-10-05 18:23:37 +08003232 else
3233 drv_tags = set->tags[hctx_idx];
John Garryf32e4ea2021-10-05 18:23:32 +08003234
John Garry65de57b2021-10-05 18:23:26 +08003235 if (tags->static_rqs && set->ops->exit_request) {
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003236 int i;
3237
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003238 for (i = 0; i < tags->nr_tags; i++) {
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003239 struct request *rq = tags->static_rqs[i];
3240
3241 if (!rq)
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003242 continue;
Christoph Hellwigd6296d392017-05-01 10:19:08 -06003243 set->ops->exit_request(set, rq, hctx_idx);
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003244 tags->static_rqs[i] = NULL;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003245 }
3246 }
3247
John Garryf32e4ea2021-10-05 18:23:32 +08003248 blk_mq_clear_rq_mapping(drv_tags, tags);
Ming Leibd631412021-05-11 23:22:35 +08003249
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003250 while (!list_empty(&tags->page_list)) {
3251 page = list_first_entry(&tags->page_list, struct page, lru);
Dave Hansen67534712014-01-08 20:17:46 -07003252 list_del_init(&page->lru);
Catalin Marinasf75782e2015-09-14 18:16:02 +01003253 /*
3254 * Remove kmemleak object previously allocated in
Raul E Rangel273938b2019-05-02 13:48:11 -06003255 * blk_mq_alloc_rqs().
Catalin Marinasf75782e2015-09-14 18:16:02 +01003256 */
3257 kmemleak_free(page_address(page));
Jens Axboe320ae512013-10-24 09:20:05 +01003258 __free_pages(page, page->private);
3259 }
Jens Axboecc71a6f2017-01-11 14:29:56 -07003260}
Jens Axboe320ae512013-10-24 09:20:05 +01003261
John Garrye155b0c2021-10-05 18:23:37 +08003262void blk_mq_free_rq_map(struct blk_mq_tags *tags)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003263{
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003264 kfree(tags->rqs);
Jens Axboecc71a6f2017-01-11 14:29:56 -07003265 tags->rqs = NULL;
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003266 kfree(tags->static_rqs);
3267 tags->static_rqs = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003268
John Garrye155b0c2021-10-05 18:23:37 +08003269 blk_mq_free_tags(tags);
Jens Axboe320ae512013-10-24 09:20:05 +01003270}
3271
Ming Lei4d805132022-03-08 15:32:14 +08003272static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3273 unsigned int hctx_idx)
3274{
3275 int i;
3276
3277 for (i = 0; i < set->nr_maps; i++) {
3278 unsigned int start = set->map[i].queue_offset;
3279 unsigned int end = start + set->map[i].nr_queues;
3280
3281 if (hctx_idx >= start && hctx_idx < end)
3282 break;
3283 }
3284
3285 if (i >= set->nr_maps)
3286 i = HCTX_TYPE_DEFAULT;
3287
3288 return i;
3289}
3290
3291static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3292 unsigned int hctx_idx)
3293{
3294 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3295
3296 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3297}
3298
John Garry63064be2021-10-05 18:23:35 +08003299static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3300 unsigned int hctx_idx,
3301 unsigned int nr_tags,
John Garrye155b0c2021-10-05 18:23:37 +08003302 unsigned int reserved_tags)
Jens Axboe320ae512013-10-24 09:20:05 +01003303{
Ming Lei4d805132022-03-08 15:32:14 +08003304 int node = blk_mq_get_hctx_node(set, hctx_idx);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003305 struct blk_mq_tags *tags;
Jens Axboe320ae512013-10-24 09:20:05 +01003306
Shaohua Li59f082e2017-02-01 09:53:14 -08003307 if (node == NUMA_NO_NODE)
3308 node = set->numa_node;
3309
John Garrye155b0c2021-10-05 18:23:37 +08003310 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3311 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003312 if (!tags)
3313 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003314
Kees Cook590b5b72018-06-12 14:04:20 -07003315 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
Gabriel Krisman Bertazi36e1f3d12016-12-06 13:31:44 -02003316 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
Shaohua Li59f082e2017-02-01 09:53:14 -08003317 node);
Jinlong Chen7edfd682022-11-02 10:52:29 +08003318 if (!tags->rqs)
3319 goto err_free_tags;
Jens Axboe320ae512013-10-24 09:20:05 +01003320
Kees Cook590b5b72018-06-12 14:04:20 -07003321 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3322 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3323 node);
Jinlong Chen7edfd682022-11-02 10:52:29 +08003324 if (!tags->static_rqs)
3325 goto err_free_rqs;
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003326
Jens Axboecc71a6f2017-01-11 14:29:56 -07003327 return tags;
Jinlong Chen7edfd682022-11-02 10:52:29 +08003328
3329err_free_rqs:
3330 kfree(tags->rqs);
3331err_free_tags:
3332 blk_mq_free_tags(tags);
3333 return NULL;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003334}
3335
Tejun Heo1d9bd512018-01-09 08:29:48 -08003336static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3337 unsigned int hctx_idx, int node)
3338{
3339 int ret;
3340
3341 if (set->ops->init_request) {
3342 ret = set->ops->init_request(set, rq, hctx_idx, node);
3343 if (ret)
3344 return ret;
3345 }
3346
Keith Busch12f5b932018-05-29 15:52:28 +02003347 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
Tejun Heo1d9bd512018-01-09 08:29:48 -08003348 return 0;
3349}
3350
John Garry63064be2021-10-05 18:23:35 +08003351static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3352 struct blk_mq_tags *tags,
3353 unsigned int hctx_idx, unsigned int depth)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003354{
3355 unsigned int i, j, entries_per_page, max_order = 4;
Ming Lei4d805132022-03-08 15:32:14 +08003356 int node = blk_mq_get_hctx_node(set, hctx_idx);
Jens Axboecc71a6f2017-01-11 14:29:56 -07003357 size_t rq_size, left;
Shaohua Li59f082e2017-02-01 09:53:14 -08003358
Shaohua Li59f082e2017-02-01 09:53:14 -08003359 if (node == NUMA_NO_NODE)
3360 node = set->numa_node;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003361
3362 INIT_LIST_HEAD(&tags->page_list);
3363
Jens Axboe320ae512013-10-24 09:20:05 +01003364 /*
3365 * rq_size is the size of the request plus driver payload, rounded
3366 * to the cacheline size
3367 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003368 rq_size = round_up(sizeof(struct request) + set->cmd_size,
Jens Axboe320ae512013-10-24 09:20:05 +01003369 cache_line_size());
Jens Axboecc71a6f2017-01-11 14:29:56 -07003370 left = rq_size * depth;
Jens Axboe320ae512013-10-24 09:20:05 +01003371
Jens Axboecc71a6f2017-01-11 14:29:56 -07003372 for (i = 0; i < depth; ) {
Jens Axboe320ae512013-10-24 09:20:05 +01003373 int this_order = max_order;
3374 struct page *page;
3375 int to_do;
3376 void *p;
3377
Bartlomiej Zolnierkiewiczb3a834b2016-05-16 09:54:47 -06003378 while (this_order && left < order_to_size(this_order - 1))
Jens Axboe320ae512013-10-24 09:20:05 +01003379 this_order--;
3380
3381 do {
Shaohua Li59f082e2017-02-01 09:53:14 -08003382 page = alloc_pages_node(node,
Gabriel Krisman Bertazi36e1f3d12016-12-06 13:31:44 -02003383 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
Jens Axboea5164402014-09-10 09:02:03 -06003384 this_order);
Jens Axboe320ae512013-10-24 09:20:05 +01003385 if (page)
3386 break;
3387 if (!this_order--)
3388 break;
3389 if (order_to_size(this_order) < rq_size)
3390 break;
3391 } while (1);
3392
3393 if (!page)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003394 goto fail;
Jens Axboe320ae512013-10-24 09:20:05 +01003395
3396 page->private = this_order;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003397 list_add_tail(&page->lru, &tags->page_list);
Jens Axboe320ae512013-10-24 09:20:05 +01003398
3399 p = page_address(page);
Catalin Marinasf75782e2015-09-14 18:16:02 +01003400 /*
3401 * Allow kmemleak to scan these pages as they contain pointers
3402 * to additional allocations like via ops->init_request().
3403 */
Gabriel Krisman Bertazi36e1f3d12016-12-06 13:31:44 -02003404 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
Jens Axboe320ae512013-10-24 09:20:05 +01003405 entries_per_page = order_to_size(this_order) / rq_size;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003406 to_do = min(entries_per_page, depth - i);
Jens Axboe320ae512013-10-24 09:20:05 +01003407 left -= to_do * rq_size;
3408 for (j = 0; j < to_do; j++) {
Jens Axboe2af8cbe2017-01-13 14:39:30 -07003409 struct request *rq = p;
3410
3411 tags->static_rqs[i] = rq;
Tejun Heo1d9bd512018-01-09 08:29:48 -08003412 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3413 tags->static_rqs[i] = NULL;
3414 goto fail;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06003415 }
3416
Jens Axboe320ae512013-10-24 09:20:05 +01003417 p += rq_size;
3418 i++;
3419 }
3420 }
Jens Axboecc71a6f2017-01-11 14:29:56 -07003421 return 0;
Jens Axboe320ae512013-10-24 09:20:05 +01003422
Christoph Hellwig24d2f902014-04-15 14:14:00 -06003423fail:
Jens Axboecc71a6f2017-01-11 14:29:56 -07003424 blk_mq_free_rqs(set, tags, hctx_idx);
3425 return -ENOMEM;
Jens Axboe320ae512013-10-24 09:20:05 +01003426}
3427
Ming Leibf0beec2020-05-29 15:53:15 +02003428struct rq_iter_data {
3429 struct blk_mq_hw_ctx *hctx;
3430 bool has_rq;
3431};
3432
John Garry2dd65322022-07-06 20:03:53 +08003433static bool blk_mq_has_request(struct request *rq, void *data)
Ming Leibf0beec2020-05-29 15:53:15 +02003434{
3435 struct rq_iter_data *iter_data = data;
3436
3437 if (rq->mq_hctx != iter_data->hctx)
3438 return true;
3439 iter_data->has_rq = true;
3440 return false;
3441}
3442
3443static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3444{
3445 struct blk_mq_tags *tags = hctx->sched_tags ?
3446 hctx->sched_tags : hctx->tags;
3447 struct rq_iter_data data = {
3448 .hctx = hctx,
3449 };
3450
3451 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3452 return data.has_rq;
3453}
3454
3455static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3456 struct blk_mq_hw_ctx *hctx)
3457{
Yury Norov9b51d9d2021-08-14 14:17:05 -07003458 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
Ming Leibf0beec2020-05-29 15:53:15 +02003459 return false;
3460 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3461 return false;
3462 return true;
3463}
3464
3465static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3466{
3467 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3468 struct blk_mq_hw_ctx, cpuhp_online);
3469
3470 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3471 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3472 return 0;
3473
3474 /*
3475 * Prevent new request from being allocated on the current hctx.
3476 *
3477 * The smp_mb__after_atomic() Pairs with the implied barrier in
3478 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3479 * seen once we return from the tag allocator.
3480 */
3481 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3482 smp_mb__after_atomic();
3483
3484 /*
3485 * Try to grab a reference to the queue and wait for any outstanding
3486 * requests. If we could not grab a reference the queue has been
3487 * frozen and there are no requests.
3488 */
3489 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3490 while (blk_mq_hctx_has_requests(hctx))
3491 msleep(5);
3492 percpu_ref_put(&hctx->queue->q_usage_counter);
3493 }
3494
3495 return 0;
3496}
3497
3498static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3499{
3500 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3501 struct blk_mq_hw_ctx, cpuhp_online);
3502
3503 if (cpumask_test_cpu(cpu, hctx->cpumask))
3504 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3505 return 0;
3506}
3507
Jens Axboee57690f2016-08-24 15:34:35 -06003508/*
3509 * 'cpu' is going away. splice any existing rq_list entries from this
3510 * software queue to the hw queue dispatch list, and ensure that it
3511 * gets run.
3512 */
Thomas Gleixner9467f852016-09-22 08:05:17 -06003513static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
Jens Axboe484b4062014-05-21 14:01:15 -06003514{
Thomas Gleixner9467f852016-09-22 08:05:17 -06003515 struct blk_mq_hw_ctx *hctx;
Jens Axboe484b4062014-05-21 14:01:15 -06003516 struct blk_mq_ctx *ctx;
3517 LIST_HEAD(tmp);
Ming Leic16d6b52018-12-17 08:44:05 -07003518 enum hctx_type type;
Jens Axboe484b4062014-05-21 14:01:15 -06003519
Thomas Gleixner9467f852016-09-22 08:05:17 -06003520 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
Ming Leibf0beec2020-05-29 15:53:15 +02003521 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3522 return 0;
3523
Jens Axboee57690f2016-08-24 15:34:35 -06003524 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
Ming Leic16d6b52018-12-17 08:44:05 -07003525 type = hctx->type;
Jens Axboe484b4062014-05-21 14:01:15 -06003526
3527 spin_lock(&ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07003528 if (!list_empty(&ctx->rq_lists[type])) {
3529 list_splice_init(&ctx->rq_lists[type], &tmp);
Jens Axboe484b4062014-05-21 14:01:15 -06003530 blk_mq_hctx_clear_pending(hctx, ctx);
3531 }
3532 spin_unlock(&ctx->lock);
3533
3534 if (list_empty(&tmp))
Thomas Gleixner9467f852016-09-22 08:05:17 -06003535 return 0;
Jens Axboe484b4062014-05-21 14:01:15 -06003536
Jens Axboee57690f2016-08-24 15:34:35 -06003537 spin_lock(&hctx->lock);
3538 list_splice_tail_init(&tmp, &hctx->dispatch);
3539 spin_unlock(&hctx->lock);
Jens Axboe484b4062014-05-21 14:01:15 -06003540
3541 blk_mq_run_hw_queue(hctx, true);
Thomas Gleixner9467f852016-09-22 08:05:17 -06003542 return 0;
Jens Axboe484b4062014-05-21 14:01:15 -06003543}
3544
Thomas Gleixner9467f852016-09-22 08:05:17 -06003545static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
Jens Axboe484b4062014-05-21 14:01:15 -06003546{
Ming Leibf0beec2020-05-29 15:53:15 +02003547 if (!(hctx->flags & BLK_MQ_F_STACKING))
3548 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3549 &hctx->cpuhp_online);
Thomas Gleixner9467f852016-09-22 08:05:17 -06003550 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3551 &hctx->cpuhp_dead);
Jens Axboe484b4062014-05-21 14:01:15 -06003552}
3553
Ming Lei364b6182021-05-11 23:22:36 +08003554/*
3555 * Before freeing hw queue, clearing the flush request reference in
3556 * tags->rqs[] for avoiding potential UAF.
3557 */
3558static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3559 unsigned int queue_depth, struct request *flush_rq)
3560{
3561 int i;
3562 unsigned long flags;
3563
3564 /* The hw queue may not be mapped yet */
3565 if (!tags)
3566 return;
3567
Jens Axboe0a467d02021-10-14 14:39:59 -06003568 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
Ming Lei364b6182021-05-11 23:22:36 +08003569
3570 for (i = 0; i < queue_depth; i++)
3571 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3572
3573 /*
3574 * Wait until all pending iteration is done.
3575 *
3576 * Request reference is cleared and it is guaranteed to be observed
3577 * after the ->lock is released.
3578 */
3579 spin_lock_irqsave(&tags->lock, flags);
3580 spin_unlock_irqrestore(&tags->lock, flags);
3581}
3582
Ming Leic3b4afc2015-06-04 22:25:04 +08003583/* hctx->ctxs will be freed in queue's release handler */
Ming Lei08e98fc2014-09-25 23:23:38 +08003584static void blk_mq_exit_hctx(struct request_queue *q,
3585 struct blk_mq_tag_set *set,
3586 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3587{
Ming Lei364b6182021-05-11 23:22:36 +08003588 struct request *flush_rq = hctx->fq->flush_rq;
3589
Ming Lei8ab0b7d2018-01-09 21:28:29 +08003590 if (blk_mq_hw_queue_mapped(hctx))
3591 blk_mq_tag_idle(hctx);
Ming Lei08e98fc2014-09-25 23:23:38 +08003592
Ming Lei6cfeadb2022-06-16 09:44:01 +08003593 if (blk_queue_init_done(q))
3594 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3595 set->queue_depth, flush_rq);
Ming Leif70ced02014-09-25 23:23:47 +08003596 if (set->ops->exit_request)
Ming Lei364b6182021-05-11 23:22:36 +08003597 set->ops->exit_request(set, flush_rq, hctx_idx);
Ming Leif70ced02014-09-25 23:23:47 +08003598
Ming Lei08e98fc2014-09-25 23:23:38 +08003599 if (set->ops->exit_hctx)
3600 set->ops->exit_hctx(hctx, hctx_idx);
3601
Thomas Gleixner9467f852016-09-22 08:05:17 -06003602 blk_mq_remove_cpuhp(hctx);
Ming Lei2f8f1332019-04-30 09:52:27 +08003603
Ming Lei4e5cc992022-03-08 15:32:19 +08003604 xa_erase(&q->hctx_table, hctx_idx);
3605
Ming Lei2f8f1332019-04-30 09:52:27 +08003606 spin_lock(&q->unused_hctx_lock);
3607 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3608 spin_unlock(&q->unused_hctx_lock);
Ming Lei08e98fc2014-09-25 23:23:38 +08003609}
3610
Ming Lei624dbe42014-05-27 23:35:13 +08003611static void blk_mq_exit_hw_queues(struct request_queue *q,
3612 struct blk_mq_tag_set *set, int nr_queue)
3613{
3614 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08003615 unsigned long i;
Ming Lei624dbe42014-05-27 23:35:13 +08003616
3617 queue_for_each_hw_ctx(q, hctx, i) {
3618 if (i == nr_queue)
3619 break;
Ming Lei08e98fc2014-09-25 23:23:38 +08003620 blk_mq_exit_hctx(q, set, hctx, i);
Ming Lei624dbe42014-05-27 23:35:13 +08003621 }
Ming Lei624dbe42014-05-27 23:35:13 +08003622}
3623
Ming Lei08e98fc2014-09-25 23:23:38 +08003624static int blk_mq_init_hctx(struct request_queue *q,
3625 struct blk_mq_tag_set *set,
3626 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3627{
Ming Lei7c6c5b72019-04-30 09:52:26 +08003628 hctx->queue_num = hctx_idx;
Ming Lei08e98fc2014-09-25 23:23:38 +08003629
Ming Leibf0beec2020-05-29 15:53:15 +02003630 if (!(hctx->flags & BLK_MQ_F_STACKING))
3631 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3632 &hctx->cpuhp_online);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003633 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3634
3635 hctx->tags = set->tags[hctx_idx];
3636
3637 if (set->ops->init_hctx &&
3638 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3639 goto unregister_cpu_notifier;
3640
3641 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3642 hctx->numa_node))
3643 goto exit_hctx;
Ming Lei4e5cc992022-03-08 15:32:19 +08003644
3645 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3646 goto exit_flush_rq;
3647
Ming Lei7c6c5b72019-04-30 09:52:26 +08003648 return 0;
3649
Ming Lei4e5cc992022-03-08 15:32:19 +08003650 exit_flush_rq:
3651 if (set->ops->exit_request)
3652 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003653 exit_hctx:
3654 if (set->ops->exit_hctx)
3655 set->ops->exit_hctx(hctx, hctx_idx);
3656 unregister_cpu_notifier:
3657 blk_mq_remove_cpuhp(hctx);
3658 return -1;
3659}
3660
3661static struct blk_mq_hw_ctx *
3662blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3663 int node)
3664{
3665 struct blk_mq_hw_ctx *hctx;
3666 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3667
Ming Lei704b9142021-12-03 21:15:32 +08003668 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003669 if (!hctx)
3670 goto fail_alloc_hctx;
3671
3672 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3673 goto free_hctx;
3674
3675 atomic_set(&hctx->nr_active, 0);
Ming Lei08e98fc2014-09-25 23:23:38 +08003676 if (node == NUMA_NO_NODE)
Ming Lei7c6c5b72019-04-30 09:52:26 +08003677 node = set->numa_node;
3678 hctx->numa_node = node;
Ming Lei08e98fc2014-09-25 23:23:38 +08003679
Jens Axboe9f993732017-04-10 09:54:54 -06003680 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
Ming Lei08e98fc2014-09-25 23:23:38 +08003681 spin_lock_init(&hctx->lock);
3682 INIT_LIST_HEAD(&hctx->dispatch);
3683 hctx->queue = q;
Ming Lei51db1c32020-08-19 23:20:19 +08003684 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
Ming Lei08e98fc2014-09-25 23:23:38 +08003685
Ming Lei2f8f1332019-04-30 09:52:27 +08003686 INIT_LIST_HEAD(&hctx->hctx_list);
3687
Ming Lei08e98fc2014-09-25 23:23:38 +08003688 /*
3689 * Allocate space for all possible cpus to avoid allocation at
3690 * runtime
3691 */
Johannes Thumshirnd904bfa2017-11-15 17:32:33 -08003692 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
Ming Lei7c6c5b72019-04-30 09:52:26 +08003693 gfp, node);
Ming Lei08e98fc2014-09-25 23:23:38 +08003694 if (!hctx->ctxs)
Ming Lei7c6c5b72019-04-30 09:52:26 +08003695 goto free_cpumask;
Ming Lei08e98fc2014-09-25 23:23:38 +08003696
Jianchao Wang5b202852018-10-12 18:07:26 +08003697 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
Ming Leic548e622021-01-22 10:33:08 +08003698 gfp, node, false, false))
Ming Lei08e98fc2014-09-25 23:23:38 +08003699 goto free_ctxs;
Ming Lei08e98fc2014-09-25 23:23:38 +08003700 hctx->nr_ctx = 0;
3701
Ming Lei5815839b2018-06-25 19:31:47 +08003702 spin_lock_init(&hctx->dispatch_wait_lock);
Jens Axboeeb619fd2017-11-09 08:32:43 -07003703 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3704 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3705
Guoqing Jiang754a1572020-03-09 22:41:37 +01003706 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
Ming Leif70ced02014-09-25 23:23:47 +08003707 if (!hctx->fq)
Ming Lei7c6c5b72019-04-30 09:52:26 +08003708 goto free_bitmap;
Ming Leif70ced02014-09-25 23:23:47 +08003709
Ming Lei7c6c5b72019-04-30 09:52:26 +08003710 blk_mq_hctx_kobj_init(hctx);
Bart Van Assche6a83e742016-11-02 10:09:51 -06003711
Ming Lei7c6c5b72019-04-30 09:52:26 +08003712 return hctx;
Ming Lei08e98fc2014-09-25 23:23:38 +08003713
3714 free_bitmap:
Omar Sandoval88459642016-09-17 08:38:44 -06003715 sbitmap_free(&hctx->ctx_map);
Ming Lei08e98fc2014-09-25 23:23:38 +08003716 free_ctxs:
3717 kfree(hctx->ctxs);
Ming Lei7c6c5b72019-04-30 09:52:26 +08003718 free_cpumask:
3719 free_cpumask_var(hctx->cpumask);
3720 free_hctx:
3721 kfree(hctx);
3722 fail_alloc_hctx:
3723 return NULL;
Ming Lei08e98fc2014-09-25 23:23:38 +08003724}
3725
Jens Axboe320ae512013-10-24 09:20:05 +01003726static void blk_mq_init_cpu_queues(struct request_queue *q,
3727 unsigned int nr_hw_queues)
3728{
Jens Axboeb3c661b2018-10-30 10:36:06 -06003729 struct blk_mq_tag_set *set = q->tag_set;
3730 unsigned int i, j;
Jens Axboe320ae512013-10-24 09:20:05 +01003731
3732 for_each_possible_cpu(i) {
3733 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3734 struct blk_mq_hw_ctx *hctx;
Ming Leic16d6b52018-12-17 08:44:05 -07003735 int k;
Jens Axboe320ae512013-10-24 09:20:05 +01003736
Jens Axboe320ae512013-10-24 09:20:05 +01003737 __ctx->cpu = i;
3738 spin_lock_init(&__ctx->lock);
Ming Leic16d6b52018-12-17 08:44:05 -07003739 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3740 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3741
Jens Axboe320ae512013-10-24 09:20:05 +01003742 __ctx->queue = q;
3743
Jens Axboe320ae512013-10-24 09:20:05 +01003744 /*
3745 * Set local node, IFF we have more than one hw queue. If
3746 * not, we remain on the home node of the device
3747 */
Jens Axboeb3c661b2018-10-30 10:36:06 -06003748 for (j = 0; j < set->nr_maps; j++) {
3749 hctx = blk_mq_map_queue_type(q, j, i);
3750 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
Xianting Tian576e85c2020-10-19 16:20:47 +08003751 hctx->numa_node = cpu_to_node(i);
Jens Axboeb3c661b2018-10-30 10:36:06 -06003752 }
Jens Axboe320ae512013-10-24 09:20:05 +01003753 }
3754}
3755
John Garry63064be2021-10-05 18:23:35 +08003756struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3757 unsigned int hctx_idx,
3758 unsigned int depth)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003759{
John Garry63064be2021-10-05 18:23:35 +08003760 struct blk_mq_tags *tags;
3761 int ret;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003762
John Garrye155b0c2021-10-05 18:23:37 +08003763 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
John Garry63064be2021-10-05 18:23:35 +08003764 if (!tags)
3765 return NULL;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003766
John Garry63064be2021-10-05 18:23:35 +08003767 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3768 if (ret) {
John Garrye155b0c2021-10-05 18:23:37 +08003769 blk_mq_free_rq_map(tags);
John Garry63064be2021-10-05 18:23:35 +08003770 return NULL;
3771 }
Jens Axboecc71a6f2017-01-11 14:29:56 -07003772
John Garry63064be2021-10-05 18:23:35 +08003773 return tags;
Jens Axboecc71a6f2017-01-11 14:29:56 -07003774}
3775
John Garry63064be2021-10-05 18:23:35 +08003776static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3777 int hctx_idx)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003778{
John Garry079a2e32021-10-05 18:23:39 +08003779 if (blk_mq_is_shared_tags(set->flags)) {
3780 set->tags[hctx_idx] = set->shared_tags;
John Garry1c0706a2020-08-19 23:20:22 +08003781
John Garrye155b0c2021-10-05 18:23:37 +08003782 return true;
Jens Axboebd166ef2017-01-17 06:03:22 -07003783 }
John Garrye155b0c2021-10-05 18:23:37 +08003784
John Garry63064be2021-10-05 18:23:35 +08003785 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3786 set->queue_depth);
3787
3788 return set->tags[hctx_idx];
Jens Axboecc71a6f2017-01-11 14:29:56 -07003789}
3790
John Garry645db342021-10-05 18:23:36 +08003791void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3792 struct blk_mq_tags *tags,
3793 unsigned int hctx_idx)
Jens Axboecc71a6f2017-01-11 14:29:56 -07003794{
John Garry645db342021-10-05 18:23:36 +08003795 if (tags) {
3796 blk_mq_free_rqs(set, tags, hctx_idx);
John Garrye155b0c2021-10-05 18:23:37 +08003797 blk_mq_free_rq_map(tags);
Jens Axboecc71a6f2017-01-11 14:29:56 -07003798 }
3799}
3800
John Garrye155b0c2021-10-05 18:23:37 +08003801static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3802 unsigned int hctx_idx)
3803{
John Garry079a2e32021-10-05 18:23:39 +08003804 if (!blk_mq_is_shared_tags(set->flags))
John Garrye155b0c2021-10-05 18:23:37 +08003805 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3806
3807 set->tags[hctx_idx] = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003808}
3809
Christoph Hellwig4b855ad2017-06-26 12:20:57 +02003810static void blk_mq_map_swqueue(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01003811{
Ming Lei4f481202022-03-08 15:32:18 +08003812 unsigned int j, hctx_idx;
3813 unsigned long i;
Jens Axboe320ae512013-10-24 09:20:05 +01003814 struct blk_mq_hw_ctx *hctx;
3815 struct blk_mq_ctx *ctx;
Ming Lei2a34c082015-04-21 10:00:20 +08003816 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01003817
3818 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboee4043dc2014-04-09 10:18:23 -06003819 cpumask_clear(hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01003820 hctx->nr_ctx = 0;
huhaid416c922018-05-18 08:32:30 -06003821 hctx->dispatch_from = NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01003822 }
3823
3824 /*
Christoph Hellwig4b855ad2017-06-26 12:20:57 +02003825 * Map software to hardware queues.
Ming Lei4412efe2018-04-25 04:01:44 +08003826 *
3827 * If the cpu isn't present, the cpu is mapped to first hctx.
Jens Axboe320ae512013-10-24 09:20:05 +01003828 */
Christoph Hellwig20e4d8132018-01-12 10:53:06 +08003829 for_each_possible_cpu(i) {
Ming Lei4412efe2018-04-25 04:01:44 +08003830
Thomas Gleixner897bb0c2016-03-19 11:30:33 +01003831 ctx = per_cpu_ptr(q->queue_ctx, i);
Jens Axboeb3c661b2018-10-30 10:36:06 -06003832 for (j = 0; j < set->nr_maps; j++) {
Jianchao Wangbb94aea2019-01-24 18:25:33 +08003833 if (!set->map[j].nr_queues) {
3834 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3835 HCTX_TYPE_DEFAULT, i);
Ming Leie5edd5f2018-12-18 01:28:56 +08003836 continue;
Jianchao Wangbb94aea2019-01-24 18:25:33 +08003837 }
Ming Leifd689872020-05-07 21:04:08 +08003838 hctx_idx = set->map[j].mq_map[i];
3839 /* unmapped hw queue can be remapped after CPU topo changed */
3840 if (!set->tags[hctx_idx] &&
John Garry63064be2021-10-05 18:23:35 +08003841 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
Ming Leifd689872020-05-07 21:04:08 +08003842 /*
3843 * If tags initialization fail for some hctx,
3844 * that hctx won't be brought online. In this
3845 * case, remap the current ctx to hctx[0] which
3846 * is guaranteed to always have tags allocated
3847 */
3848 set->map[j].mq_map[i] = 0;
3849 }
Ming Leie5edd5f2018-12-18 01:28:56 +08003850
Jens Axboeb3c661b2018-10-30 10:36:06 -06003851 hctx = blk_mq_map_queue_type(q, j, i);
Jianchao Wang8ccdf4a2019-01-24 18:25:32 +08003852 ctx->hctxs[j] = hctx;
Jens Axboeb3c661b2018-10-30 10:36:06 -06003853 /*
3854 * If the CPU is already set in the mask, then we've
3855 * mapped this one already. This can happen if
3856 * devices share queues across queue maps.
3857 */
3858 if (cpumask_test_cpu(i, hctx->cpumask))
3859 continue;
3860
3861 cpumask_set_cpu(i, hctx->cpumask);
3862 hctx->type = j;
3863 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3864 hctx->ctxs[hctx->nr_ctx++] = ctx;
3865
3866 /*
3867 * If the nr_ctx type overflows, we have exceeded the
3868 * amount of sw queues we can support.
3869 */
3870 BUG_ON(!hctx->nr_ctx);
3871 }
Jianchao Wangbb94aea2019-01-24 18:25:33 +08003872
3873 for (; j < HCTX_MAX_TYPES; j++)
3874 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3875 HCTX_TYPE_DEFAULT, i);
Jens Axboe320ae512013-10-24 09:20:05 +01003876 }
Jens Axboe506e9312014-05-07 10:26:44 -06003877
3878 queue_for_each_hw_ctx(q, hctx, i) {
Ming Lei4412efe2018-04-25 04:01:44 +08003879 /*
3880 * If no software queues are mapped to this hardware queue,
3881 * disable it and free the request entries.
3882 */
3883 if (!hctx->nr_ctx) {
3884 /* Never unmap queue 0. We need it as a
3885 * fallback in case of a new remap fails
3886 * allocation
3887 */
John Garrye155b0c2021-10-05 18:23:37 +08003888 if (i)
3889 __blk_mq_free_map_and_rqs(set, i);
Ming Lei4412efe2018-04-25 04:01:44 +08003890
3891 hctx->tags = NULL;
3892 continue;
3893 }
Jens Axboe484b4062014-05-21 14:01:15 -06003894
Ming Lei2a34c082015-04-21 10:00:20 +08003895 hctx->tags = set->tags[i];
3896 WARN_ON(!hctx->tags);
3897
Jens Axboe484b4062014-05-21 14:01:15 -06003898 /*
Chong Yuan889fa312015-04-15 11:39:29 -06003899 * Set the map size to the number of mapped software queues.
3900 * This is more accurate and more efficient than looping
3901 * over all possibly mapped software queues.
3902 */
Omar Sandoval88459642016-09-17 08:38:44 -06003903 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
Chong Yuan889fa312015-04-15 11:39:29 -06003904
3905 /*
Jens Axboe484b4062014-05-21 14:01:15 -06003906 * Initialize batch roundrobin counts
3907 */
Ming Leif82ddf12018-04-08 17:48:10 +08003908 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
Jens Axboe506e9312014-05-07 10:26:44 -06003909 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3910 }
Jens Axboe320ae512013-10-24 09:20:05 +01003911}
3912
Jens Axboe8e8320c2017-06-20 17:56:13 -06003913/*
3914 * Caller needs to ensure that we're either frozen/quiesced, or that
3915 * the queue isn't live yet.
3916 */
Jeff Moyer2404e602015-11-03 10:40:06 -05003917static void queue_set_hctx_shared(struct request_queue *q, bool shared)
Jens Axboe0d2602c2014-05-13 15:10:52 -06003918{
3919 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08003920 unsigned long i;
Jens Axboe0d2602c2014-05-13 15:10:52 -06003921
Jeff Moyer2404e602015-11-03 10:40:06 -05003922 queue_for_each_hw_ctx(q, hctx, i) {
Yu Kuai454bb672021-07-31 14:21:30 +08003923 if (shared) {
Ming Lei51db1c32020-08-19 23:20:19 +08003924 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
Yu Kuai454bb672021-07-31 14:21:30 +08003925 } else {
3926 blk_mq_tag_idle(hctx);
Ming Lei51db1c32020-08-19 23:20:19 +08003927 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
Yu Kuai454bb672021-07-31 14:21:30 +08003928 }
Jeff Moyer2404e602015-11-03 10:40:06 -05003929 }
3930}
3931
Hannes Reinecke655ac302020-08-19 23:20:20 +08003932static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3933 bool shared)
Jeff Moyer2404e602015-11-03 10:40:06 -05003934{
3935 struct request_queue *q;
Jens Axboe0d2602c2014-05-13 15:10:52 -06003936
Bart Van Assche705cda92017-04-07 11:16:49 -07003937 lockdep_assert_held(&set->tag_list_lock);
3938
Jens Axboe0d2602c2014-05-13 15:10:52 -06003939 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3940 blk_mq_freeze_queue(q);
Jeff Moyer2404e602015-11-03 10:40:06 -05003941 queue_set_hctx_shared(q, shared);
Jens Axboe0d2602c2014-05-13 15:10:52 -06003942 blk_mq_unfreeze_queue(q);
3943 }
3944}
3945
3946static void blk_mq_del_queue_tag_set(struct request_queue *q)
3947{
3948 struct blk_mq_tag_set *set = q->tag_set;
3949
Jens Axboe0d2602c2014-05-13 15:10:52 -06003950 mutex_lock(&set->tag_list_lock);
Daniel Wagner08c875c2020-07-28 15:29:51 +02003951 list_del(&q->tag_set_list);
Jeff Moyer2404e602015-11-03 10:40:06 -05003952 if (list_is_singular(&set->tag_list)) {
3953 /* just transitioned to unshared */
Ming Lei51db1c32020-08-19 23:20:19 +08003954 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
Jeff Moyer2404e602015-11-03 10:40:06 -05003955 /* update existing queue */
Hannes Reinecke655ac302020-08-19 23:20:20 +08003956 blk_mq_update_tag_set_shared(set, false);
Jeff Moyer2404e602015-11-03 10:40:06 -05003957 }
Jens Axboe0d2602c2014-05-13 15:10:52 -06003958 mutex_unlock(&set->tag_list_lock);
Roman Pena347c7a2018-06-10 22:38:24 +02003959 INIT_LIST_HEAD(&q->tag_set_list);
Jens Axboe0d2602c2014-05-13 15:10:52 -06003960}
3961
3962static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3963 struct request_queue *q)
3964{
Jens Axboe0d2602c2014-05-13 15:10:52 -06003965 mutex_lock(&set->tag_list_lock);
Jeff Moyer2404e602015-11-03 10:40:06 -05003966
Jens Axboeff821d22017-11-10 22:05:12 -07003967 /*
3968 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3969 */
3970 if (!list_empty(&set->tag_list) &&
Ming Lei51db1c32020-08-19 23:20:19 +08003971 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3972 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
Jeff Moyer2404e602015-11-03 10:40:06 -05003973 /* update existing queue */
Hannes Reinecke655ac302020-08-19 23:20:20 +08003974 blk_mq_update_tag_set_shared(set, true);
Jeff Moyer2404e602015-11-03 10:40:06 -05003975 }
Ming Lei51db1c32020-08-19 23:20:19 +08003976 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
Jeff Moyer2404e602015-11-03 10:40:06 -05003977 queue_set_hctx_shared(q, true);
Daniel Wagner08c875c2020-07-28 15:29:51 +02003978 list_add_tail(&q->tag_set_list, &set->tag_list);
Jeff Moyer2404e602015-11-03 10:40:06 -05003979
Jens Axboe0d2602c2014-05-13 15:10:52 -06003980 mutex_unlock(&set->tag_list_lock);
3981}
3982
Ming Lei1db49092018-11-20 09:44:35 +08003983/* All allocations will be freed in release handler of q->mq_kobj */
3984static int blk_mq_alloc_ctxs(struct request_queue *q)
3985{
3986 struct blk_mq_ctxs *ctxs;
3987 int cpu;
3988
3989 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3990 if (!ctxs)
3991 return -ENOMEM;
3992
3993 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3994 if (!ctxs->queue_ctx)
3995 goto fail;
3996
3997 for_each_possible_cpu(cpu) {
3998 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3999 ctx->ctxs = ctxs;
4000 }
4001
4002 q->mq_kobj = &ctxs->kobj;
4003 q->queue_ctx = ctxs->queue_ctx;
4004
4005 return 0;
4006 fail:
4007 kfree(ctxs);
4008 return -ENOMEM;
4009}
4010
Ming Leie09aae72015-01-29 20:17:27 +08004011/*
4012 * It is the actual release handler for mq, but we do it from
4013 * request queue's release handler for avoiding use-after-free
4014 * and headache because q->mq_kobj shouldn't have been introduced,
4015 * but we can't group ctx/kctx kobj without it.
4016 */
4017void blk_mq_release(struct request_queue *q)
4018{
Ming Lei2f8f1332019-04-30 09:52:27 +08004019 struct blk_mq_hw_ctx *hctx, *next;
Ming Lei4f481202022-03-08 15:32:18 +08004020 unsigned long i;
Ming Leie09aae72015-01-29 20:17:27 +08004021
Ming Lei2f8f1332019-04-30 09:52:27 +08004022 queue_for_each_hw_ctx(q, hctx, i)
4023 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4024
4025 /* all hctx are in .unused_hctx_list now */
4026 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4027 list_del_init(&hctx->hctx_list);
Ming Lei6c8b2322017-02-22 18:14:01 +08004028 kobject_put(&hctx->kobj);
Ming Leic3b4afc2015-06-04 22:25:04 +08004029 }
Ming Leie09aae72015-01-29 20:17:27 +08004030
Ming Lei4e5cc992022-03-08 15:32:19 +08004031 xa_destroy(&q->hctx_table);
Ming Leie09aae72015-01-29 20:17:27 +08004032
Ming Lei7ea5fe32017-02-22 18:14:00 +08004033 /*
4034 * release .mq_kobj and sw queue's kobject now because
4035 * both share lifetime with request queue.
4036 */
4037 blk_mq_sysfs_deinit(q);
Ming Leie09aae72015-01-29 20:17:27 +08004038}
4039
Christoph Hellwig5ec780a2021-06-24 10:10:12 +02004040static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
Christoph Hellwig2f227bb2020-03-27 09:30:08 +01004041 void *queuedata)
Jens Axboe320ae512013-10-24 09:20:05 +01004042{
Christoph Hellwig26a97502021-06-02 09:53:17 +03004043 struct request_queue *q;
4044 int ret;
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004045
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004046 q = blk_alloc_queue(set->numa_node);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004047 if (!q)
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004048 return ERR_PTR(-ENOMEM);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004049 q->queuedata = queuedata;
4050 ret = blk_mq_init_allocated_queue(set, q);
4051 if (ret) {
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004052 blk_put_queue(q);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004053 return ERR_PTR(ret);
4054 }
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004055 return q;
4056}
Christoph Hellwig2f227bb2020-03-27 09:30:08 +01004057
4058struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4059{
4060 return blk_mq_init_queue_data(set, NULL);
4061}
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004062EXPORT_SYMBOL(blk_mq_init_queue);
4063
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004064/**
4065 * blk_mq_destroy_queue - shutdown a request queue
4066 * @q: request queue to shutdown
4067 *
4068 * This shuts down a request queue allocated by blk_mq_init_queue() and drops
4069 * the initial reference. All future requests will failed with -ENODEV.
4070 *
4071 * Context: can sleep
4072 */
4073void blk_mq_destroy_queue(struct request_queue *q)
4074{
4075 WARN_ON_ONCE(!queue_is_mq(q));
4076 WARN_ON_ONCE(blk_queue_registered(q));
4077
4078 might_sleep();
4079
4080 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4081 blk_queue_start_drain(q);
Jinlong Chen56c1ee92022-10-30 16:32:12 +08004082 blk_mq_freeze_queue_wait(q);
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004083
4084 blk_sync_queue(q);
4085 blk_mq_cancel_work_sync(q);
4086 blk_mq_exit_queue(q);
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004087}
4088EXPORT_SYMBOL(blk_mq_destroy_queue);
4089
Christoph Hellwig4dcc48742021-08-16 15:19:05 +02004090struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4091 struct lock_class_key *lkclass)
Jens Axboe9316a9e2018-10-15 08:40:37 -06004092{
4093 struct request_queue *q;
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004094 struct gendisk *disk;
Jens Axboe9316a9e2018-10-15 08:40:37 -06004095
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004096 q = blk_mq_init_queue_data(set, queuedata);
4097 if (IS_ERR(q))
4098 return ERR_CAST(q);
Jens Axboe9316a9e2018-10-15 08:40:37 -06004099
Christoph Hellwig4a1fa412021-08-16 15:19:08 +02004100 disk = __alloc_disk_node(q, set->numa_node, lkclass);
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004101 if (!disk) {
Christoph Hellwig0a3e5cc2022-07-20 15:05:40 +02004102 blk_mq_destroy_queue(q);
Christoph Hellwig2b3f0562022-10-18 15:57:17 +02004103 blk_put_queue(q);
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004104 return ERR_PTR(-ENOMEM);
Jens Axboe9316a9e2018-10-15 08:40:37 -06004105 }
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004106 set_bit(GD_OWNS_QUEUE, &disk->state);
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004107 return disk;
Jens Axboe9316a9e2018-10-15 08:40:37 -06004108}
Christoph Hellwigb461dfc2021-06-02 09:53:18 +03004109EXPORT_SYMBOL(__blk_mq_alloc_disk);
Jens Axboe9316a9e2018-10-15 08:40:37 -06004110
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004111struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4112 struct lock_class_key *lkclass)
4113{
Christoph Hellwig22c17e22022-11-22 08:27:53 +01004114 struct gendisk *disk;
4115
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004116 if (!blk_get_queue(q))
4117 return NULL;
Christoph Hellwig22c17e22022-11-22 08:27:53 +01004118 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4119 if (!disk)
4120 blk_put_queue(q);
4121 return disk;
Christoph Hellwig6f8191f2022-06-19 08:05:51 +02004122}
4123EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4124
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004125static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4126 struct blk_mq_tag_set *set, struct request_queue *q,
4127 int hctx_idx, int node)
4128{
Ming Lei2f8f1332019-04-30 09:52:27 +08004129 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004130
Ming Lei2f8f1332019-04-30 09:52:27 +08004131 /* reuse dead hctx first */
4132 spin_lock(&q->unused_hctx_lock);
4133 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4134 if (tmp->numa_node == node) {
4135 hctx = tmp;
4136 break;
4137 }
4138 }
4139 if (hctx)
4140 list_del_init(&hctx->hctx_list);
4141 spin_unlock(&q->unused_hctx_lock);
4142
4143 if (!hctx)
4144 hctx = blk_mq_alloc_hctx(q, set, node);
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004145 if (!hctx)
Ming Lei7c6c5b72019-04-30 09:52:26 +08004146 goto fail;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004147
Ming Lei7c6c5b72019-04-30 09:52:26 +08004148 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4149 goto free_hctx;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004150
4151 return hctx;
Ming Lei7c6c5b72019-04-30 09:52:26 +08004152
4153 free_hctx:
4154 kobject_put(&hctx->kobj);
4155 fail:
4156 return NULL;
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004157}
4158
Keith Busch868f2f02015-12-17 17:08:14 -07004159static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4160 struct request_queue *q)
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004161{
Ming Lei4e5cc992022-03-08 15:32:19 +08004162 struct blk_mq_hw_ctx *hctx;
4163 unsigned long i, j;
Bart Van Asscheac0d6b92019-10-25 09:50:09 -07004164
Ming Leifb350e02018-01-06 16:27:40 +08004165 /* protect against switching io scheduler */
4166 mutex_lock(&q->sysfs_lock);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004167 for (i = 0; i < set->nr_hw_queues; i++) {
Ming Lei306f13ee2022-03-08 15:32:15 +08004168 int old_node;
Ming Lei4d805132022-03-08 15:32:14 +08004169 int node = blk_mq_get_hctx_node(set, i);
Ming Lei4e5cc992022-03-08 15:32:19 +08004170 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
Keith Busch868f2f02015-12-17 17:08:14 -07004171
Ming Lei306f13ee2022-03-08 15:32:15 +08004172 if (old_hctx) {
4173 old_node = old_hctx->numa_node;
4174 blk_mq_exit_hctx(q, set, old_hctx, i);
4175 }
Jens Axboe320ae512013-10-24 09:20:05 +01004176
Ming Lei4e5cc992022-03-08 15:32:19 +08004177 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
Ming Lei306f13ee2022-03-08 15:32:15 +08004178 if (!old_hctx)
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004179 break;
Ming Lei306f13ee2022-03-08 15:32:15 +08004180 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4181 node, old_node);
Ming Lei4e5cc992022-03-08 15:32:19 +08004182 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4183 WARN_ON_ONCE(!hctx);
Keith Busch868f2f02015-12-17 17:08:14 -07004184 }
Jens Axboe320ae512013-10-24 09:20:05 +01004185 }
Jianchao Wange01ad462018-10-12 18:07:28 +08004186 /*
4187 * Increasing nr_hw_queues fails. Free the newly allocated
4188 * hctxs and keep the previous q->nr_hw_queues.
4189 */
4190 if (i != set->nr_hw_queues) {
4191 j = q->nr_hw_queues;
Jianchao Wange01ad462018-10-12 18:07:28 +08004192 } else {
4193 j = i;
Jianchao Wange01ad462018-10-12 18:07:28 +08004194 q->nr_hw_queues = set->nr_hw_queues;
4195 }
Jianchao Wang34d11ff2018-10-12 18:07:27 +08004196
Ming Lei4e5cc992022-03-08 15:32:19 +08004197 xa_for_each_start(&q->hctx_table, j, hctx, j)
4198 blk_mq_exit_hctx(q, set, hctx, j);
Ming Leifb350e02018-01-06 16:27:40 +08004199 mutex_unlock(&q->sysfs_lock);
Keith Busch868f2f02015-12-17 17:08:14 -07004200}
4201
Ming Lei42ee3062022-03-08 15:32:16 +08004202static void blk_mq_update_poll_flag(struct request_queue *q)
4203{
4204 struct blk_mq_tag_set *set = q->tag_set;
4205
4206 if (set->nr_maps > HCTX_TYPE_POLL &&
4207 set->map[HCTX_TYPE_POLL].nr_queues)
4208 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4209 else
4210 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4211}
4212
Christoph Hellwig26a97502021-06-02 09:53:17 +03004213int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4214 struct request_queue *q)
Keith Busch868f2f02015-12-17 17:08:14 -07004215{
Ming Lei66841672016-02-12 15:27:00 +08004216 /* mark the queue as mq asap */
4217 q->mq_ops = set->ops;
4218
Omar Sandoval34dbad52017-03-21 08:56:08 -07004219 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
Stephen Bates720b8cc2017-04-07 06:24:03 -06004220 blk_mq_poll_stats_bkt,
4221 BLK_MQ_POLL_STATS_BKTS, q);
Omar Sandoval34dbad52017-03-21 08:56:08 -07004222 if (!q->poll_cb)
4223 goto err_exit;
4224
Ming Lei1db49092018-11-20 09:44:35 +08004225 if (blk_mq_alloc_ctxs(q))
Jes Sorensen41de54c2019-04-19 16:35:44 -04004226 goto err_poll;
Keith Busch868f2f02015-12-17 17:08:14 -07004227
Ming Lei737f98c2017-02-22 18:13:59 +08004228 /* init q->mq_kobj and sw queues' kobjects */
4229 blk_mq_sysfs_init(q);
4230
Ming Lei2f8f1332019-04-30 09:52:27 +08004231 INIT_LIST_HEAD(&q->unused_hctx_list);
4232 spin_lock_init(&q->unused_hctx_lock);
4233
Ming Lei4e5cc992022-03-08 15:32:19 +08004234 xa_init(&q->hctx_table);
4235
Keith Busch868f2f02015-12-17 17:08:14 -07004236 blk_mq_realloc_hw_ctxs(set, q);
4237 if (!q->nr_hw_queues)
4238 goto err_hctxs;
Jens Axboe320ae512013-10-24 09:20:05 +01004239
Christoph Hellwig287922e2015-10-30 20:57:30 +08004240 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
Ming Leie56f6982015-07-16 19:53:22 +08004241 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
Jens Axboe320ae512013-10-24 09:20:05 +01004242
Jens Axboea8908932018-10-16 14:23:06 -06004243 q->tag_set = set;
Jens Axboe320ae512013-10-24 09:20:05 +01004244
Jens Axboe94eddfb2013-11-19 09:25:07 -07004245 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
Ming Lei42ee3062022-03-08 15:32:16 +08004246 blk_mq_update_poll_flag(q);
Jens Axboe320ae512013-10-24 09:20:05 +01004247
Mike Snitzer28494502016-09-14 13:28:30 -04004248 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06004249 INIT_LIST_HEAD(&q->requeue_list);
4250 spin_lock_init(&q->requeue_lock);
4251
Jens Axboeeba71762014-05-20 15:17:27 -06004252 q->nr_requests = set->queue_depth;
4253
Jens Axboe64f1c212016-11-14 13:03:03 -07004254 /*
4255 * Default to classic polling
4256 */
Yufen Yu29ece8b2019-03-18 22:44:41 +08004257 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
Jens Axboe64f1c212016-11-14 13:03:03 -07004258
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004259 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
Jens Axboe0d2602c2014-05-13 15:10:52 -06004260 blk_mq_add_queue_tag_set(set, q);
Christoph Hellwig4b855ad2017-06-26 12:20:57 +02004261 blk_mq_map_swqueue(q);
Christoph Hellwig26a97502021-06-02 09:53:17 +03004262 return 0;
Christoph Hellwig18741982014-02-10 09:29:00 -07004263
Jens Axboe320ae512013-10-24 09:20:05 +01004264err_hctxs:
Chen Jun943f45b2022-10-31 03:12:42 +00004265 blk_mq_release(q);
Jes Sorensen41de54c2019-04-19 16:35:44 -04004266err_poll:
4267 blk_stat_free_callback(q->poll_cb);
4268 q->poll_cb = NULL;
Ming Linc7de5722016-05-25 23:23:27 -07004269err_exit:
4270 q->mq_ops = NULL;
Christoph Hellwig26a97502021-06-02 09:53:17 +03004271 return -ENOMEM;
Jens Axboe320ae512013-10-24 09:20:05 +01004272}
Mike Snitzerb62c21b2015-03-12 23:56:02 -04004273EXPORT_SYMBOL(blk_mq_init_allocated_queue);
Jens Axboe320ae512013-10-24 09:20:05 +01004274
Ming Leic7e2d942019-04-30 09:52:25 +08004275/* tags can _not_ be used after returning from blk_mq_exit_queue */
4276void blk_mq_exit_queue(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01004277{
Bart Van Assche630ef622021-05-13 10:15:29 -07004278 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01004279
Bart Van Assche630ef622021-05-13 10:15:29 -07004280 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
Ming Lei624dbe42014-05-27 23:35:13 +08004281 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
Bart Van Assche630ef622021-05-13 10:15:29 -07004282 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4283 blk_mq_del_queue_tag_set(q);
Jens Axboe320ae512013-10-24 09:20:05 +01004284}
Jens Axboe320ae512013-10-24 09:20:05 +01004285
Jens Axboea5164402014-09-10 09:02:03 -06004286static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4287{
4288 int i;
4289
John Garry079a2e32021-10-05 18:23:39 +08004290 if (blk_mq_is_shared_tags(set->flags)) {
4291 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
John Garrye155b0c2021-10-05 18:23:37 +08004292 BLK_MQ_NO_HCTX_IDX,
4293 set->queue_depth);
John Garry079a2e32021-10-05 18:23:39 +08004294 if (!set->shared_tags)
John Garrye155b0c2021-10-05 18:23:37 +08004295 return -ENOMEM;
4296 }
4297
Xianting Tian8229cca2020-09-26 10:39:47 +08004298 for (i = 0; i < set->nr_hw_queues; i++) {
John Garry63064be2021-10-05 18:23:35 +08004299 if (!__blk_mq_alloc_map_and_rqs(set, i))
Jens Axboea5164402014-09-10 09:02:03 -06004300 goto out_unwind;
Xianting Tian8229cca2020-09-26 10:39:47 +08004301 cond_resched();
4302 }
Jens Axboea5164402014-09-10 09:02:03 -06004303
4304 return 0;
4305
4306out_unwind:
4307 while (--i >= 0)
John Garrye155b0c2021-10-05 18:23:37 +08004308 __blk_mq_free_map_and_rqs(set, i);
4309
John Garry079a2e32021-10-05 18:23:39 +08004310 if (blk_mq_is_shared_tags(set->flags)) {
4311 blk_mq_free_map_and_rqs(set, set->shared_tags,
John Garrye155b0c2021-10-05 18:23:37 +08004312 BLK_MQ_NO_HCTX_IDX);
John Garry645db342021-10-05 18:23:36 +08004313 }
Jens Axboea5164402014-09-10 09:02:03 -06004314
Jens Axboea5164402014-09-10 09:02:03 -06004315 return -ENOMEM;
4316}
4317
4318/*
4319 * Allocate the request maps associated with this tag_set. Note that this
4320 * may reduce the depth asked for, if memory is tight. set->queue_depth
4321 * will be updated to reflect the allocated depth.
4322 */
John Garry63064be2021-10-05 18:23:35 +08004323static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
Jens Axboea5164402014-09-10 09:02:03 -06004324{
4325 unsigned int depth;
4326 int err;
4327
4328 depth = set->queue_depth;
4329 do {
4330 err = __blk_mq_alloc_rq_maps(set);
4331 if (!err)
4332 break;
4333
4334 set->queue_depth >>= 1;
4335 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4336 err = -ENOMEM;
4337 break;
4338 }
4339 } while (set->queue_depth);
4340
4341 if (!set->queue_depth || err) {
4342 pr_err("blk-mq: failed to allocate request map\n");
4343 return -ENOMEM;
4344 }
4345
4346 if (depth != set->queue_depth)
4347 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4348 depth, set->queue_depth);
4349
4350 return 0;
4351}
4352
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004353static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
Omar Sandovalebe8bdd2017-04-07 08:53:11 -06004354{
Bart Van Assche6e66b492020-03-09 21:26:17 -07004355 /*
4356 * blk_mq_map_queues() and multiple .map_queues() implementations
4357 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4358 * number of hardware queues.
4359 */
4360 if (set->nr_maps == 1)
4361 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4362
Ming Lei59388702018-12-07 11:03:53 +08004363 if (set->ops->map_queues && !is_kdump_kernel()) {
Jens Axboeb3c661b2018-10-30 10:36:06 -06004364 int i;
4365
Ming Lei7d4901a2018-01-06 16:27:39 +08004366 /*
4367 * transport .map_queues is usually done in the following
4368 * way:
4369 *
4370 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4371 * mask = get_cpu_mask(queue)
4372 * for_each_cpu(cpu, mask)
Jens Axboeb3c661b2018-10-30 10:36:06 -06004373 * set->map[x].mq_map[cpu] = queue;
Ming Lei7d4901a2018-01-06 16:27:39 +08004374 * }
4375 *
4376 * When we need to remap, the table has to be cleared for
4377 * killing stale mapping since one CPU may not be mapped
4378 * to any hw queue.
4379 */
Jens Axboeb3c661b2018-10-30 10:36:06 -06004380 for (i = 0; i < set->nr_maps; i++)
4381 blk_mq_clear_mq_map(&set->map[i]);
Ming Lei7d4901a2018-01-06 16:27:39 +08004382
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004383 set->ops->map_queues(set);
Jens Axboeb3c661b2018-10-30 10:36:06 -06004384 } else {
4385 BUG_ON(set->nr_maps > 1);
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004386 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
Jens Axboeb3c661b2018-10-30 10:36:06 -06004387 }
Omar Sandovalebe8bdd2017-04-07 08:53:11 -06004388}
4389
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004390static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004391 int new_nr_hw_queues)
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004392{
4393 struct blk_mq_tags **new_tags;
4394
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004395 if (set->nr_hw_queues >= new_nr_hw_queues)
Shin'ichiro Kawasakid4b2e0d2022-11-22 17:49:17 +09004396 goto done;
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004397
4398 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4399 GFP_KERNEL, set->numa_node);
4400 if (!new_tags)
4401 return -ENOMEM;
4402
4403 if (set->tags)
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004404 memcpy(new_tags, set->tags, set->nr_hw_queues *
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004405 sizeof(*set->tags));
4406 kfree(set->tags);
4407 set->tags = new_tags;
Shin'ichiro Kawasakid4b2e0d2022-11-22 17:49:17 +09004408done:
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004409 set->nr_hw_queues = new_nr_hw_queues;
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004410 return 0;
4411}
4412
Jens Axboea4391c62014-06-05 15:21:56 -06004413/*
4414 * Alloc a tag set to be associated with one or more request queues.
4415 * May fail with EINVAL for various error conditions. May adjust the
Minwoo Imc018c842018-06-30 22:12:41 +09004416 * requested depth down, if it's too large. In that case, the set
Jens Axboea4391c62014-06-05 15:21:56 -06004417 * value will be stored in set->queue_depth.
4418 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004419int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4420{
Jens Axboeb3c661b2018-10-30 10:36:06 -06004421 int i, ret;
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004422
Bart Van Assche205fb5f52014-10-30 14:45:11 +01004423 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4424
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004425 if (!set->nr_hw_queues)
4426 return -EINVAL;
Jens Axboea4391c62014-06-05 15:21:56 -06004427 if (!set->queue_depth)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004428 return -EINVAL;
4429 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4430 return -EINVAL;
4431
Christoph Hellwig7d7e0f92016-09-14 16:18:54 +02004432 if (!set->ops->queue_rq)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004433 return -EINVAL;
4434
Ming Leide148292017-10-14 17:22:29 +08004435 if (!set->ops->get_budget ^ !set->ops->put_budget)
4436 return -EINVAL;
4437
Jens Axboea4391c62014-06-05 15:21:56 -06004438 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4439 pr_info("blk-mq: reduced tag depth to %u\n",
4440 BLK_MQ_MAX_DEPTH);
4441 set->queue_depth = BLK_MQ_MAX_DEPTH;
4442 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004443
Jens Axboeb3c661b2018-10-30 10:36:06 -06004444 if (!set->nr_maps)
4445 set->nr_maps = 1;
4446 else if (set->nr_maps > HCTX_MAX_TYPES)
4447 return -EINVAL;
4448
Shaohua Li6637fad2014-11-30 16:00:58 -08004449 /*
4450 * If a crashdump is active, then we are potentially in a very
4451 * memory constrained environment. Limit us to 1 queue and
4452 * 64 tags to prevent using too much memory.
4453 */
4454 if (is_kdump_kernel()) {
4455 set->nr_hw_queues = 1;
Ming Lei59388702018-12-07 11:03:53 +08004456 set->nr_maps = 1;
Shaohua Li6637fad2014-11-30 16:00:58 -08004457 set->queue_depth = min(64U, set->queue_depth);
4458 }
Keith Busch868f2f02015-12-17 17:08:14 -07004459 /*
Jens Axboe392546a2018-10-29 13:25:27 -06004460 * There is no use for more h/w queues than cpus if we just have
4461 * a single map
Keith Busch868f2f02015-12-17 17:08:14 -07004462 */
Jens Axboe392546a2018-10-29 13:25:27 -06004463 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
Keith Busch868f2f02015-12-17 17:08:14 -07004464 set->nr_hw_queues = nr_cpu_ids;
Shaohua Li6637fad2014-11-30 16:00:58 -08004465
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004466 if (set->flags & BLK_MQ_F_BLOCKING) {
4467 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4468 if (!set->srcu)
4469 return -ENOMEM;
4470 ret = init_srcu_struct(set->srcu);
4471 if (ret)
4472 goto out_free_srcu;
4473 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004474
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004475 ret = -ENOMEM;
Christoph Hellwig5ee20292022-11-09 11:08:10 +01004476 set->tags = kcalloc_node(set->nr_hw_queues,
4477 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4478 set->numa_node);
4479 if (!set->tags)
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004480 goto out_cleanup_srcu;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004481
Jens Axboeb3c661b2018-10-30 10:36:06 -06004482 for (i = 0; i < set->nr_maps; i++) {
4483 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
Ming Lei07b35eb2018-12-17 18:42:45 +08004484 sizeof(set->map[i].mq_map[0]),
Jens Axboeb3c661b2018-10-30 10:36:06 -06004485 GFP_KERNEL, set->numa_node);
4486 if (!set->map[i].mq_map)
4487 goto out_free_mq_map;
Ming Lei59388702018-12-07 11:03:53 +08004488 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
Jens Axboeb3c661b2018-10-30 10:36:06 -06004489 }
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004490
Bart Van Asschea4e1d0b2022-08-15 10:00:43 -07004491 blk_mq_update_queue_map(set);
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004492
John Garry63064be2021-10-05 18:23:35 +08004493 ret = blk_mq_alloc_set_map_and_rqs(set);
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004494 if (ret)
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004495 goto out_free_mq_map;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004496
Jens Axboe0d2602c2014-05-13 15:10:52 -06004497 mutex_init(&set->tag_list_lock);
4498 INIT_LIST_HEAD(&set->tag_list);
4499
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004500 return 0;
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004501
4502out_free_mq_map:
Jens Axboeb3c661b2018-10-30 10:36:06 -06004503 for (i = 0; i < set->nr_maps; i++) {
4504 kfree(set->map[i].mq_map);
4505 set->map[i].mq_map = NULL;
4506 }
Robert Elliott5676e7b2014-09-02 11:38:44 -05004507 kfree(set->tags);
4508 set->tags = NULL;
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004509out_cleanup_srcu:
4510 if (set->flags & BLK_MQ_F_BLOCKING)
4511 cleanup_srcu_struct(set->srcu);
4512out_free_srcu:
4513 if (set->flags & BLK_MQ_F_BLOCKING)
4514 kfree(set->srcu);
Christoph Hellwigda695ba2016-09-14 16:18:55 +02004515 return ret;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004516}
4517EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4518
Christoph Hellwigcdb14e02021-06-02 09:53:16 +03004519/* allocate and initialize a tagset for a simple single-queue device */
4520int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4521 const struct blk_mq_ops *ops, unsigned int queue_depth,
4522 unsigned int set_flags)
4523{
4524 memset(set, 0, sizeof(*set));
4525 set->ops = ops;
4526 set->nr_hw_queues = 1;
4527 set->nr_maps = 1;
4528 set->queue_depth = queue_depth;
4529 set->numa_node = NUMA_NO_NODE;
4530 set->flags = set_flags;
4531 return blk_mq_alloc_tag_set(set);
4532}
4533EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4534
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004535void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4536{
Jens Axboeb3c661b2018-10-30 10:36:06 -06004537 int i, j;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004538
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004539 for (i = 0; i < set->nr_hw_queues; i++)
John Garrye155b0c2021-10-05 18:23:37 +08004540 __blk_mq_free_map_and_rqs(set, i);
Jens Axboe484b4062014-05-21 14:01:15 -06004541
John Garry079a2e32021-10-05 18:23:39 +08004542 if (blk_mq_is_shared_tags(set->flags)) {
4543 blk_mq_free_map_and_rqs(set, set->shared_tags,
John Garrye155b0c2021-10-05 18:23:37 +08004544 BLK_MQ_NO_HCTX_IDX);
4545 }
John Garry32bc15a2020-08-19 23:20:24 +08004546
Jens Axboeb3c661b2018-10-30 10:36:06 -06004547 for (j = 0; j < set->nr_maps; j++) {
4548 kfree(set->map[j].mq_map);
4549 set->map[j].mq_map = NULL;
4550 }
Christoph Hellwigbdd17e72016-09-14 16:18:53 +02004551
Ming Lei981bd182014-04-24 00:07:34 +08004552 kfree(set->tags);
Robert Elliott5676e7b2014-09-02 11:38:44 -05004553 set->tags = NULL;
Christoph Hellwig80bd4a72022-11-01 16:00:47 +01004554 if (set->flags & BLK_MQ_F_BLOCKING) {
4555 cleanup_srcu_struct(set->srcu);
4556 kfree(set->srcu);
4557 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06004558}
4559EXPORT_SYMBOL(blk_mq_free_tag_set);
4560
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004561int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4562{
4563 struct blk_mq_tag_set *set = q->tag_set;
4564 struct blk_mq_hw_ctx *hctx;
Ming Lei4f481202022-03-08 15:32:18 +08004565 int ret;
4566 unsigned long i;
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004567
Jens Axboebd166ef2017-01-17 06:03:22 -07004568 if (!set)
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004569 return -EINVAL;
4570
Aleksei Zakharove5fa8142019-02-08 19:14:05 +03004571 if (q->nr_requests == nr)
4572 return 0;
4573
Jens Axboe70f36b62017-01-19 10:59:07 -07004574 blk_mq_freeze_queue(q);
Ming Lei24f5a902018-01-06 16:27:38 +08004575 blk_mq_quiesce_queue(q);
Jens Axboe70f36b62017-01-19 10:59:07 -07004576
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004577 ret = 0;
4578 queue_for_each_hw_ctx(q, hctx, i) {
Keith Busche9137d42016-02-18 14:56:35 -07004579 if (!hctx->tags)
4580 continue;
Jens Axboebd166ef2017-01-17 06:03:22 -07004581 /*
4582 * If we're using an MQ scheduler, just update the scheduler
4583 * queue depth. This is similar to what the old code would do.
4584 */
John Garryf6adcef2021-10-05 18:23:29 +08004585 if (hctx->sched_tags) {
Jens Axboe70f36b62017-01-19 10:59:07 -07004586 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
John Garryf6adcef2021-10-05 18:23:29 +08004587 nr, true);
John Garryf6adcef2021-10-05 18:23:29 +08004588 } else {
4589 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4590 false);
Jens Axboe70f36b62017-01-19 10:59:07 -07004591 }
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004592 if (ret)
4593 break;
Jens Axboe77f1e0a2019-01-18 10:34:16 -07004594 if (q->elevator && q->elevator->type->ops.depth_updated)
4595 q->elevator->type->ops.depth_updated(hctx);
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004596 }
John Garryd97e5942021-05-13 20:00:58 +08004597 if (!ret) {
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004598 q->nr_requests = nr;
John Garry079a2e32021-10-05 18:23:39 +08004599 if (blk_mq_is_shared_tags(set->flags)) {
John Garry8fa04462021-10-05 18:23:28 +08004600 if (q->elevator)
John Garry079a2e32021-10-05 18:23:39 +08004601 blk_mq_tag_update_sched_shared_tags(q);
John Garry8fa04462021-10-05 18:23:28 +08004602 else
John Garry079a2e32021-10-05 18:23:39 +08004603 blk_mq_tag_resize_shared_tags(set, nr);
John Garry8fa04462021-10-05 18:23:28 +08004604 }
John Garryd97e5942021-05-13 20:00:58 +08004605 }
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004606
Ming Lei24f5a902018-01-06 16:27:38 +08004607 blk_mq_unquiesce_queue(q);
Jens Axboe70f36b62017-01-19 10:59:07 -07004608 blk_mq_unfreeze_queue(q);
Jens Axboe70f36b62017-01-19 10:59:07 -07004609
Jens Axboee3a2b3f2014-05-20 11:49:02 -06004610 return ret;
4611}
4612
Jianchao Wangd48ece22018-08-21 15:15:03 +08004613/*
4614 * request_queue and elevator_type pair.
4615 * It is just used by __blk_mq_update_nr_hw_queues to cache
4616 * the elevator_type associated with a request_queue.
4617 */
4618struct blk_mq_qe_pair {
4619 struct list_head node;
4620 struct request_queue *q;
4621 struct elevator_type *type;
4622};
4623
4624/*
4625 * Cache the elevator_type in qe pair list and switch the
4626 * io scheduler to 'none'
4627 */
4628static bool blk_mq_elv_switch_none(struct list_head *head,
4629 struct request_queue *q)
4630{
4631 struct blk_mq_qe_pair *qe;
4632
4633 if (!q->elevator)
4634 return true;
4635
4636 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4637 if (!qe)
4638 return false;
4639
Ming Lei5fd7a842022-06-16 09:43:59 +08004640 /* q->elevator needs protection from ->sysfs_lock */
4641 mutex_lock(&q->sysfs_lock);
4642
Jianchao Wangd48ece22018-08-21 15:15:03 +08004643 INIT_LIST_HEAD(&qe->node);
4644 qe->q = q;
4645 qe->type = q->elevator->type;
Christoph Hellwigdd6f7f12022-10-20 08:48:16 +02004646 /* keep a reference to the elevator module as we'll switch back */
4647 __elevator_get(qe->type);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004648 list_add(&qe->node, head);
Christoph Hellwig64b36072022-10-30 11:07:14 +01004649 elevator_disable(q);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004650 mutex_unlock(&q->sysfs_lock);
4651
4652 return true;
4653}
4654
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004655static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4656 struct request_queue *q)
Jianchao Wangd48ece22018-08-21 15:15:03 +08004657{
4658 struct blk_mq_qe_pair *qe;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004659
4660 list_for_each_entry(qe, head, node)
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004661 if (qe->q == q)
4662 return qe;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004663
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004664 return NULL;
4665}
4666
4667static void blk_mq_elv_switch_back(struct list_head *head,
4668 struct request_queue *q)
4669{
4670 struct blk_mq_qe_pair *qe;
4671 struct elevator_type *t;
4672
4673 qe = blk_lookup_qe_pair(head, q);
4674 if (!qe)
Jianchao Wangd48ece22018-08-21 15:15:03 +08004675 return;
Jakob Koschel4a3b666e2022-03-31 11:12:18 +02004676 t = qe->type;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004677 list_del(&qe->node);
4678 kfree(qe);
4679
4680 mutex_lock(&q->sysfs_lock);
Keith Busch8237c012022-09-27 08:56:52 -07004681 elevator_switch(q, t);
Jinlong Chen8ed40ee2022-10-20 08:48:19 +02004682 /* drop the reference acquired in blk_mq_elv_switch_none */
4683 elevator_put(t);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004684 mutex_unlock(&q->sysfs_lock);
4685}
4686
Keith Busche4dc2b32017-05-30 14:39:11 -04004687static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4688 int nr_hw_queues)
Keith Busch868f2f02015-12-17 17:08:14 -07004689{
4690 struct request_queue *q;
Jianchao Wangd48ece22018-08-21 15:15:03 +08004691 LIST_HEAD(head);
Jianchao Wange01ad462018-10-12 18:07:28 +08004692 int prev_nr_hw_queues;
Keith Busch868f2f02015-12-17 17:08:14 -07004693
Bart Van Assche705cda92017-04-07 11:16:49 -07004694 lockdep_assert_held(&set->tag_list_lock);
4695
Jens Axboe392546a2018-10-29 13:25:27 -06004696 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
Keith Busch868f2f02015-12-17 17:08:14 -07004697 nr_hw_queues = nr_cpu_ids;
Weiping Zhangfe35ec52020-06-17 14:18:37 +08004698 if (nr_hw_queues < 1)
4699 return;
4700 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
Keith Busch868f2f02015-12-17 17:08:14 -07004701 return;
4702
4703 list_for_each_entry(q, &set->tag_list, tag_set_list)
4704 blk_mq_freeze_queue(q);
Jianchao Wangd48ece22018-08-21 15:15:03 +08004705 /*
4706 * Switch IO scheduler to 'none', cleaning up the data associated
4707 * with the previous scheduler. We will switch back once we are done
4708 * updating the new sw to hw queue mappings.
4709 */
4710 list_for_each_entry(q, &set->tag_list, tag_set_list)
4711 if (!blk_mq_elv_switch_none(&head, q))
4712 goto switch_back;
Keith Busch868f2f02015-12-17 17:08:14 -07004713
Jianchao Wang477e19d2018-10-12 18:07:25 +08004714 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4715 blk_mq_debugfs_unregister_hctxs(q);
Christoph Hellwigeaa870f2022-06-28 19:18:49 +02004716 blk_mq_sysfs_unregister_hctxs(q);
Jianchao Wang477e19d2018-10-12 18:07:25 +08004717 }
4718
Weiping Zhanga2584e42020-05-07 21:03:56 +08004719 prev_nr_hw_queues = set->nr_hw_queues;
Christoph Hellwigee9d5522022-11-09 11:08:11 +01004720 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004721 goto reregister;
4722
Jianchao Wange01ad462018-10-12 18:07:28 +08004723fallback:
Weiping Zhangaa880ad2020-05-13 08:44:05 +08004724 blk_mq_update_queue_map(set);
Keith Busch868f2f02015-12-17 17:08:14 -07004725 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4726 blk_mq_realloc_hw_ctxs(set, q);
Ming Lei42ee3062022-03-08 15:32:16 +08004727 blk_mq_update_poll_flag(q);
Jianchao Wange01ad462018-10-12 18:07:28 +08004728 if (q->nr_hw_queues != set->nr_hw_queues) {
Ye Bina846a8e2021-11-08 15:40:19 +08004729 int i = prev_nr_hw_queues;
4730
Jianchao Wange01ad462018-10-12 18:07:28 +08004731 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4732 nr_hw_queues, prev_nr_hw_queues);
Ye Bina846a8e2021-11-08 15:40:19 +08004733 for (; i < set->nr_hw_queues; i++)
4734 __blk_mq_free_map_and_rqs(set, i);
4735
Jianchao Wange01ad462018-10-12 18:07:28 +08004736 set->nr_hw_queues = prev_nr_hw_queues;
Dongli Zhang7d76f852019-02-27 21:35:01 +08004737 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
Jianchao Wange01ad462018-10-12 18:07:28 +08004738 goto fallback;
4739 }
Jianchao Wang477e19d2018-10-12 18:07:25 +08004740 blk_mq_map_swqueue(q);
4741 }
4742
Bart Van Asschef7e76db2019-10-25 09:50:10 -07004743reregister:
Jianchao Wang477e19d2018-10-12 18:07:25 +08004744 list_for_each_entry(q, &set->tag_list, tag_set_list) {
Christoph Hellwigeaa870f2022-06-28 19:18:49 +02004745 blk_mq_sysfs_register_hctxs(q);
Jianchao Wang477e19d2018-10-12 18:07:25 +08004746 blk_mq_debugfs_register_hctxs(q);
Keith Busch868f2f02015-12-17 17:08:14 -07004747 }
4748
Jianchao Wangd48ece22018-08-21 15:15:03 +08004749switch_back:
4750 list_for_each_entry(q, &set->tag_list, tag_set_list)
4751 blk_mq_elv_switch_back(&head, q);
4752
Keith Busch868f2f02015-12-17 17:08:14 -07004753 list_for_each_entry(q, &set->tag_list, tag_set_list)
4754 blk_mq_unfreeze_queue(q);
4755}
Keith Busche4dc2b32017-05-30 14:39:11 -04004756
4757void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4758{
4759 mutex_lock(&set->tag_list_lock);
4760 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4761 mutex_unlock(&set->tag_list_lock);
4762}
Keith Busch868f2f02015-12-17 17:08:14 -07004763EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4764
Omar Sandoval34dbad52017-03-21 08:56:08 -07004765/* Enable polling stats and return whether they were already enabled. */
4766static bool blk_poll_stats_enable(struct request_queue *q)
4767{
Jens Axboe48b5c1f2021-11-13 14:03:26 -07004768 if (q->poll_stat)
Omar Sandoval34dbad52017-03-21 08:56:08 -07004769 return true;
Jens Axboe48b5c1f2021-11-13 14:03:26 -07004770
4771 return blk_stats_alloc_enable(q);
Omar Sandoval34dbad52017-03-21 08:56:08 -07004772}
4773
4774static void blk_mq_poll_stats_start(struct request_queue *q)
4775{
4776 /*
4777 * We don't arm the callback if polling stats are not enabled or the
4778 * callback is already active.
4779 */
Jens Axboe48b5c1f2021-11-13 14:03:26 -07004780 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
Omar Sandoval34dbad52017-03-21 08:56:08 -07004781 return;
4782
4783 blk_stat_activate_msecs(q->poll_cb, 100);
4784}
4785
4786static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4787{
4788 struct request_queue *q = cb->data;
Stephen Bates720b8cc2017-04-07 06:24:03 -06004789 int bucket;
Omar Sandoval34dbad52017-03-21 08:56:08 -07004790
Stephen Bates720b8cc2017-04-07 06:24:03 -06004791 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4792 if (cb->stat[bucket].nr_samples)
4793 q->poll_stat[bucket] = cb->stat[bucket];
4794 }
Omar Sandoval34dbad52017-03-21 08:56:08 -07004795}
4796
Jens Axboe64f1c212016-11-14 13:03:03 -07004797static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
Jens Axboe64f1c212016-11-14 13:03:03 -07004798 struct request *rq)
4799{
Jens Axboe64f1c212016-11-14 13:03:03 -07004800 unsigned long ret = 0;
Stephen Bates720b8cc2017-04-07 06:24:03 -06004801 int bucket;
Jens Axboe64f1c212016-11-14 13:03:03 -07004802
4803 /*
4804 * If stats collection isn't on, don't sleep but turn it on for
4805 * future users
4806 */
Omar Sandoval34dbad52017-03-21 08:56:08 -07004807 if (!blk_poll_stats_enable(q))
Jens Axboe64f1c212016-11-14 13:03:03 -07004808 return 0;
4809
4810 /*
Jens Axboe64f1c212016-11-14 13:03:03 -07004811 * As an optimistic guess, use half of the mean service time
4812 * for this type of request. We can (and should) make this smarter.
4813 * For instance, if the completion latencies are tight, we can
4814 * get closer than just half the mean. This is especially
4815 * important on devices where the completion latencies are longer
Stephen Bates720b8cc2017-04-07 06:24:03 -06004816 * than ~10 usec. We do use the stats for the relevant IO size
4817 * if available which does lead to better estimates.
Jens Axboe64f1c212016-11-14 13:03:03 -07004818 */
Stephen Bates720b8cc2017-04-07 06:24:03 -06004819 bucket = blk_mq_poll_stats_bkt(rq);
4820 if (bucket < 0)
4821 return ret;
4822
4823 if (q->poll_stat[bucket].nr_samples)
4824 ret = (q->poll_stat[bucket].mean + 1) / 2;
Jens Axboe64f1c212016-11-14 13:03:03 -07004825
4826 return ret;
4827}
4828
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004829static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
Jens Axboe06426ad2016-11-14 13:01:59 -07004830{
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004831 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4832 struct request *rq = blk_qc_to_rq(hctx, qc);
Jens Axboe06426ad2016-11-14 13:01:59 -07004833 struct hrtimer_sleeper hs;
4834 enum hrtimer_mode mode;
Jens Axboe64f1c212016-11-14 13:03:03 -07004835 unsigned int nsecs;
Jens Axboe06426ad2016-11-14 13:01:59 -07004836 ktime_t kt;
4837
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004838 /*
4839 * If a request has completed on queue that uses an I/O scheduler, we
4840 * won't get back a request from blk_qc_to_rq.
4841 */
4842 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
Jens Axboe64f1c212016-11-14 13:03:03 -07004843 return false;
4844
4845 /*
Jens Axboe1052b8a2018-11-26 08:21:49 -07004846 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
Jens Axboe64f1c212016-11-14 13:03:03 -07004847 *
Jens Axboe64f1c212016-11-14 13:03:03 -07004848 * 0: use half of prev avg
4849 * >0: use this specific value
4850 */
Jens Axboe1052b8a2018-11-26 08:21:49 -07004851 if (q->poll_nsec > 0)
Jens Axboe64f1c212016-11-14 13:03:03 -07004852 nsecs = q->poll_nsec;
4853 else
John Garrycae740a2020-02-26 20:10:15 +08004854 nsecs = blk_mq_poll_nsecs(q, rq);
Jens Axboe64f1c212016-11-14 13:03:03 -07004855
4856 if (!nsecs)
Jens Axboe06426ad2016-11-14 13:01:59 -07004857 return false;
4858
Jens Axboe76a86f92018-01-10 11:30:56 -07004859 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
Jens Axboe06426ad2016-11-14 13:01:59 -07004860
4861 /*
4862 * This will be replaced with the stats tracking code, using
4863 * 'avg_completion_time / 2' as the pre-sleep target.
4864 */
Thomas Gleixner8b0e1952016-12-25 12:30:41 +01004865 kt = nsecs;
Jens Axboe06426ad2016-11-14 13:01:59 -07004866
4867 mode = HRTIMER_MODE_REL;
Sebastian Andrzej Siewiordbc16252019-07-26 20:30:50 +02004868 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
Jens Axboe06426ad2016-11-14 13:01:59 -07004869 hrtimer_set_expires(&hs.timer, kt);
4870
Jens Axboe06426ad2016-11-14 13:01:59 -07004871 do {
Tejun Heo5a61c362018-01-09 08:29:52 -08004872 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
Jens Axboe06426ad2016-11-14 13:01:59 -07004873 break;
4874 set_current_state(TASK_UNINTERRUPTIBLE);
Thomas Gleixner9dd88132019-07-30 21:16:55 +02004875 hrtimer_sleeper_start_expires(&hs, mode);
Jens Axboe06426ad2016-11-14 13:01:59 -07004876 if (hs.task)
4877 io_schedule();
4878 hrtimer_cancel(&hs.timer);
4879 mode = HRTIMER_MODE_ABS;
4880 } while (hs.task && !signal_pending(current));
4881
4882 __set_current_state(TASK_RUNNING);
4883 destroy_hrtimer_on_stack(&hs.timer);
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004884
4885 /*
4886 * If we sleep, have the caller restart the poll loop to reset the
4887 * state. Like for the other success return cases, the caller is
4888 * responsible for checking if the IO completed. If the IO isn't
4889 * complete, we'll get called again and will go straight to the busy
4890 * poll loop.
4891 */
Jens Axboe06426ad2016-11-14 13:01:59 -07004892 return true;
4893}
4894
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004895static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
Jens Axboe5a72e892021-10-12 09:24:29 -06004896 struct io_comp_batch *iob, unsigned int flags)
Jens Axboebbd7bb72016-11-04 09:34:34 -06004897{
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004898 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4899 long state = get_current_state();
4900 int ret;
Jens Axboe1052b8a2018-11-26 08:21:49 -07004901
Jens Axboeaa61bec2018-11-13 21:32:10 -07004902 do {
Jens Axboe5a72e892021-10-12 09:24:29 -06004903 ret = q->mq_ops->poll(hctx, iob);
Jens Axboebbd7bb72016-11-04 09:34:34 -06004904 if (ret > 0) {
Jens Axboe849a3702018-11-16 08:37:34 -07004905 __set_current_state(TASK_RUNNING);
Jens Axboe85f4d4b2018-11-06 13:30:55 -07004906 return ret;
Jens Axboebbd7bb72016-11-04 09:34:34 -06004907 }
4908
4909 if (signal_pending_state(state, current))
Jens Axboe849a3702018-11-16 08:37:34 -07004910 __set_current_state(TASK_RUNNING);
Peter Zijlstrab03fbd42021-06-11 10:28:12 +02004911 if (task_is_running(current))
Jens Axboe85f4d4b2018-11-06 13:30:55 -07004912 return 1;
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004913
Christoph Hellwigef99b2d2021-10-12 13:12:19 +02004914 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
Jens Axboebbd7bb72016-11-04 09:34:34 -06004915 break;
4916 cpu_relax();
Jens Axboeaa61bec2018-11-13 21:32:10 -07004917 } while (!need_resched());
Jens Axboebbd7bb72016-11-04 09:34:34 -06004918
Nitesh Shetty67b41102018-02-13 21:18:12 +05304919 __set_current_state(TASK_RUNNING);
Jens Axboe85f4d4b2018-11-06 13:30:55 -07004920 return 0;
Jens Axboebbd7bb72016-11-04 09:34:34 -06004921}
Jens Axboebbd7bb72016-11-04 09:34:34 -06004922
Jens Axboe5a72e892021-10-12 09:24:29 -06004923int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4924 unsigned int flags)
Jens Axboe676141e2014-03-20 13:29:18 -06004925{
Christoph Hellwigd729cf92021-10-12 13:12:20 +02004926 if (!(flags & BLK_POLL_NOSLEEP) &&
4927 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004928 if (blk_mq_poll_hybrid(q, cookie))
Jens Axboe320ae512013-10-24 09:20:05 +01004929 return 1;
Christoph Hellwigc6699d62021-10-12 13:12:16 +02004930 }
Jens Axboe5a72e892021-10-12 09:24:29 -06004931 return blk_mq_poll_classic(q, cookie, iob, flags);
Jens Axboe320ae512013-10-24 09:20:05 +01004932}
4933
Jens Axboe9cf2bab2018-10-31 17:01:22 -06004934unsigned int blk_mq_rq_cpu(struct request *rq)
4935{
4936 return rq->mq_ctx->cpu;
4937}
4938EXPORT_SYMBOL(blk_mq_rq_cpu);
4939
Ming Lei2a19b282021-11-16 09:43:43 +08004940void blk_mq_cancel_work_sync(struct request_queue *q)
4941{
Jinlong Chen219cf432022-10-30 17:47:30 +08004942 struct blk_mq_hw_ctx *hctx;
4943 unsigned long i;
Ming Lei2a19b282021-11-16 09:43:43 +08004944
Jinlong Chen219cf432022-10-30 17:47:30 +08004945 cancel_delayed_work_sync(&q->requeue_work);
Ming Lei2a19b282021-11-16 09:43:43 +08004946
Jinlong Chen219cf432022-10-30 17:47:30 +08004947 queue_for_each_hw_ctx(q, hctx, i)
4948 cancel_delayed_work_sync(&hctx->run_work);
Ming Lei2a19b282021-11-16 09:43:43 +08004949}
4950
Jens Axboe320ae512013-10-24 09:20:05 +01004951static int __init blk_mq_init(void)
4952{
Christoph Hellwigc3077b52020-06-11 08:44:41 +02004953 int i;
4954
4955 for_each_possible_cpu(i)
Sebastian Andrzej Siewiorf9ab4912021-01-23 21:10:27 +01004956 init_llist_head(&per_cpu(blk_cpu_done, i));
Christoph Hellwigc3077b52020-06-11 08:44:41 +02004957 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4958
4959 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4960 "block/softirq:dead", NULL,
4961 blk_softirq_cpu_dead);
Jens Axboe320ae512013-10-24 09:20:05 +01004962 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4963 blk_mq_hctx_notify_dead);
Ming Leibf0beec2020-05-29 15:53:15 +02004964 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4965 blk_mq_hctx_notify_online,
4966 blk_mq_hctx_notify_offline);
Jens Axboe320ae512013-10-24 09:20:05 +01004967 return 0;
4968}
4969subsys_initcall(blk_mq_init);