blob: 594eea04266e6d05f7256255552a1c4c72c664f3 [file] [log] [blame]
Jens Axboe75bb4622014-05-28 10:15:41 -06001/*
2 * Block multiqueue core code
3 *
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
6 */
Jens Axboe320ae512013-10-24 09:20:05 +01007#include <linux/kernel.h>
8#include <linux/module.h>
9#include <linux/backing-dev.h>
10#include <linux/bio.h>
11#include <linux/blkdev.h>
12#include <linux/mm.h>
13#include <linux/init.h>
14#include <linux/slab.h>
15#include <linux/workqueue.h>
16#include <linux/smp.h>
17#include <linux/llist.h>
18#include <linux/list_sort.h>
19#include <linux/cpu.h>
20#include <linux/cache.h>
21#include <linux/sched/sysctl.h>
22#include <linux/delay.h>
Jens Axboeaedcd722014-09-17 08:27:03 -060023#include <linux/crash_dump.h>
Jens Axboe320ae512013-10-24 09:20:05 +010024
25#include <trace/events/block.h>
26
27#include <linux/blk-mq.h>
28#include "blk.h"
29#include "blk-mq.h"
30#include "blk-mq-tag.h"
31
32static DEFINE_MUTEX(all_q_mutex);
33static LIST_HEAD(all_q_list);
34
35static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
36
Jens Axboe320ae512013-10-24 09:20:05 +010037/*
38 * Check if any of the ctx's have pending work in this hardware queue
39 */
40static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
41{
42 unsigned int i;
43
Jens Axboe569fd0c2015-04-17 08:28:50 -060044 for (i = 0; i < hctx->ctx_map.size; i++)
Jens Axboe1429d7c2014-05-19 09:23:55 -060045 if (hctx->ctx_map.map[i].word)
Jens Axboe320ae512013-10-24 09:20:05 +010046 return true;
47
48 return false;
49}
50
Jens Axboe1429d7c2014-05-19 09:23:55 -060051static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
52 struct blk_mq_ctx *ctx)
53{
54 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
55}
56
57#define CTX_TO_BIT(hctx, ctx) \
58 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
59
Jens Axboe320ae512013-10-24 09:20:05 +010060/*
61 * Mark this ctx as having pending work in this hardware queue
62 */
63static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
64 struct blk_mq_ctx *ctx)
65{
Jens Axboe1429d7c2014-05-19 09:23:55 -060066 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
67
68 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
69 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
70}
71
72static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
73 struct blk_mq_ctx *ctx)
74{
75 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
76
77 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
Jens Axboe320ae512013-10-24 09:20:05 +010078}
79
Keith Buschbfd343a2015-03-11 23:56:39 -040080static int blk_mq_queue_enter(struct request_queue *q, gfp_t gfp)
Jens Axboe320ae512013-10-24 09:20:05 +010081{
Tejun Heoadd703f2014-07-01 10:34:38 -060082 while (true) {
83 int ret;
Jens Axboe320ae512013-10-24 09:20:05 +010084
Tejun Heoadd703f2014-07-01 10:34:38 -060085 if (percpu_ref_tryget_live(&q->mq_usage_counter))
86 return 0;
Keith Busch3b632cf2014-06-06 10:22:07 -060087
Keith Buschbfd343a2015-03-11 23:56:39 -040088 if (!(gfp & __GFP_WAIT))
89 return -EBUSY;
90
Tejun Heoadd703f2014-07-01 10:34:38 -060091 ret = wait_event_interruptible(q->mq_freeze_wq,
92 !q->mq_freeze_depth || blk_queue_dying(q));
93 if (blk_queue_dying(q))
94 return -ENODEV;
95 if (ret)
96 return ret;
97 }
Jens Axboe320ae512013-10-24 09:20:05 +010098}
99
100static void blk_mq_queue_exit(struct request_queue *q)
101{
Tejun Heoadd703f2014-07-01 10:34:38 -0600102 percpu_ref_put(&q->mq_usage_counter);
103}
104
105static void blk_mq_usage_counter_release(struct percpu_ref *ref)
106{
107 struct request_queue *q =
108 container_of(ref, struct request_queue, mq_usage_counter);
109
110 wake_up_all(&q->mq_freeze_wq);
Jens Axboe320ae512013-10-24 09:20:05 +0100111}
112
Keith Buschb4c6a022014-12-19 17:54:14 -0700113void blk_mq_freeze_queue_start(struct request_queue *q)
Ming Lei43a5e4e2013-12-26 21:31:35 +0800114{
Tejun Heocddd5d12014-08-16 08:02:24 -0400115 bool freeze;
116
Tejun Heo72d6f022014-07-01 10:33:02 -0600117 spin_lock_irq(q->queue_lock);
Tejun Heocddd5d12014-08-16 08:02:24 -0400118 freeze = !q->mq_freeze_depth++;
Tejun Heo72d6f022014-07-01 10:33:02 -0600119 spin_unlock_irq(q->queue_lock);
120
Tejun Heocddd5d12014-08-16 08:02:24 -0400121 if (freeze) {
Tejun Heo9eca8042014-09-24 13:07:33 -0400122 percpu_ref_kill(&q->mq_usage_counter);
Mike Snitzerb94ec292015-03-11 23:56:38 -0400123 blk_mq_run_hw_queues(q, false);
Tejun Heocddd5d12014-08-16 08:02:24 -0400124 }
Tejun Heof3af0202014-11-04 13:52:27 -0500125}
Keith Buschb4c6a022014-12-19 17:54:14 -0700126EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
Tejun Heof3af0202014-11-04 13:52:27 -0500127
128static void blk_mq_freeze_queue_wait(struct request_queue *q)
129{
Tejun Heoadd703f2014-07-01 10:34:38 -0600130 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
Ming Lei43a5e4e2013-12-26 21:31:35 +0800131}
132
Tejun Heof3af0202014-11-04 13:52:27 -0500133/*
134 * Guarantee no request is in use, so we can change any data structure of
135 * the queue afterward.
136 */
137void blk_mq_freeze_queue(struct request_queue *q)
138{
139 blk_mq_freeze_queue_start(q);
140 blk_mq_freeze_queue_wait(q);
141}
Jens Axboec761d962015-01-02 15:05:12 -0700142EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
Tejun Heof3af0202014-11-04 13:52:27 -0500143
Keith Buschb4c6a022014-12-19 17:54:14 -0700144void blk_mq_unfreeze_queue(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +0100145{
Tejun Heocddd5d12014-08-16 08:02:24 -0400146 bool wake;
Jens Axboe320ae512013-10-24 09:20:05 +0100147
148 spin_lock_irq(q->queue_lock);
Tejun Heo780db202014-07-01 10:31:13 -0600149 wake = !--q->mq_freeze_depth;
150 WARN_ON_ONCE(q->mq_freeze_depth < 0);
Jens Axboe320ae512013-10-24 09:20:05 +0100151 spin_unlock_irq(q->queue_lock);
Tejun Heoadd703f2014-07-01 10:34:38 -0600152 if (wake) {
153 percpu_ref_reinit(&q->mq_usage_counter);
Jens Axboe320ae512013-10-24 09:20:05 +0100154 wake_up_all(&q->mq_freeze_wq);
Tejun Heoadd703f2014-07-01 10:34:38 -0600155 }
Jens Axboe320ae512013-10-24 09:20:05 +0100156}
Keith Buschb4c6a022014-12-19 17:54:14 -0700157EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
Jens Axboe320ae512013-10-24 09:20:05 +0100158
Jens Axboeaed3ea92014-12-22 14:04:42 -0700159void blk_mq_wake_waiters(struct request_queue *q)
160{
161 struct blk_mq_hw_ctx *hctx;
162 unsigned int i;
163
164 queue_for_each_hw_ctx(q, hctx, i)
165 if (blk_mq_hw_queue_mapped(hctx))
166 blk_mq_tag_wakeup_all(hctx->tags, true);
Keith Busch3fd59402015-01-08 08:53:56 -0700167
168 /*
169 * If we are called because the queue has now been marked as
170 * dying, we need to ensure that processes currently waiting on
171 * the queue are notified as well.
172 */
173 wake_up_all(&q->mq_freeze_wq);
Jens Axboeaed3ea92014-12-22 14:04:42 -0700174}
175
Jens Axboe320ae512013-10-24 09:20:05 +0100176bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
177{
178 return blk_mq_has_free_tags(hctx->tags);
179}
180EXPORT_SYMBOL(blk_mq_can_queue);
181
Jens Axboe94eddfb2013-11-19 09:25:07 -0700182static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
183 struct request *rq, unsigned int rw_flags)
Jens Axboe320ae512013-10-24 09:20:05 +0100184{
Jens Axboe94eddfb2013-11-19 09:25:07 -0700185 if (blk_queue_io_stat(q))
186 rw_flags |= REQ_IO_STAT;
187
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200188 INIT_LIST_HEAD(&rq->queuelist);
189 /* csd/requeue_work/fifo_time is initialized before use */
190 rq->q = q;
Jens Axboe320ae512013-10-24 09:20:05 +0100191 rq->mq_ctx = ctx;
Jens Axboe0d2602c2014-05-13 15:10:52 -0600192 rq->cmd_flags |= rw_flags;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200193 /* do not touch atomic flags, it needs atomic ops against the timer */
194 rq->cpu = -1;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200195 INIT_HLIST_NODE(&rq->hash);
196 RB_CLEAR_NODE(&rq->rb_node);
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200197 rq->rq_disk = NULL;
198 rq->part = NULL;
Jens Axboe3ee32372014-06-09 09:36:53 -0600199 rq->start_time = jiffies;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200200#ifdef CONFIG_BLK_CGROUP
201 rq->rl = NULL;
Ming Lei0fec08b2014-01-03 10:00:08 -0700202 set_start_time_ns(rq);
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200203 rq->io_start_time_ns = 0;
204#endif
205 rq->nr_phys_segments = 0;
206#if defined(CONFIG_BLK_DEV_INTEGRITY)
207 rq->nr_integrity_segments = 0;
208#endif
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200209 rq->special = NULL;
210 /* tag was already set */
211 rq->errors = 0;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200212
Tony Battersby6f4a16262014-08-22 15:53:39 -0400213 rq->cmd = rq->__cmd;
214
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200215 rq->extra_len = 0;
216 rq->sense_len = 0;
217 rq->resid_len = 0;
218 rq->sense = NULL;
219
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200220 INIT_LIST_HEAD(&rq->timeout_list);
Jens Axboef6be4fb2014-06-06 11:03:48 -0600221 rq->timeout = 0;
222
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200223 rq->end_io = NULL;
224 rq->end_io_data = NULL;
225 rq->next_rq = NULL;
226
Jens Axboe320ae512013-10-24 09:20:05 +0100227 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
228}
229
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200230static struct request *
Ming Leicb96a422014-06-01 00:43:37 +0800231__blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200232{
233 struct request *rq;
234 unsigned int tag;
235
Ming Leicb96a422014-06-01 00:43:37 +0800236 tag = blk_mq_get_tag(data);
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200237 if (tag != BLK_MQ_TAG_FAIL) {
Ming Leicb96a422014-06-01 00:43:37 +0800238 rq = data->hctx->tags->rqs[tag];
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200239
Ming Leicb96a422014-06-01 00:43:37 +0800240 if (blk_mq_tag_busy(data->hctx)) {
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200241 rq->cmd_flags = REQ_MQ_INFLIGHT;
Ming Leicb96a422014-06-01 00:43:37 +0800242 atomic_inc(&data->hctx->nr_active);
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200243 }
244
245 rq->tag = tag;
Ming Leicb96a422014-06-01 00:43:37 +0800246 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200247 return rq;
248 }
249
250 return NULL;
251}
252
Christoph Hellwig4ce01dd2014-05-27 20:59:46 +0200253struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
254 bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100255{
Christoph Hellwigd8525642014-05-27 20:59:50 +0200256 struct blk_mq_ctx *ctx;
257 struct blk_mq_hw_ctx *hctx;
Jens Axboe320ae512013-10-24 09:20:05 +0100258 struct request *rq;
Ming Leicb96a422014-06-01 00:43:37 +0800259 struct blk_mq_alloc_data alloc_data;
Joe Lawrencea492f072014-08-28 08:15:21 -0600260 int ret;
Jens Axboe320ae512013-10-24 09:20:05 +0100261
Keith Buschbfd343a2015-03-11 23:56:39 -0400262 ret = blk_mq_queue_enter(q, gfp);
Joe Lawrencea492f072014-08-28 08:15:21 -0600263 if (ret)
264 return ERR_PTR(ret);
Jens Axboe320ae512013-10-24 09:20:05 +0100265
Christoph Hellwigd8525642014-05-27 20:59:50 +0200266 ctx = blk_mq_get_ctx(q);
267 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Ming Leicb96a422014-06-01 00:43:37 +0800268 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
269 reserved, ctx, hctx);
Christoph Hellwigd8525642014-05-27 20:59:50 +0200270
Ming Leicb96a422014-06-01 00:43:37 +0800271 rq = __blk_mq_alloc_request(&alloc_data, rw);
Christoph Hellwigd8525642014-05-27 20:59:50 +0200272 if (!rq && (gfp & __GFP_WAIT)) {
273 __blk_mq_run_hw_queue(hctx);
274 blk_mq_put_ctx(ctx);
275
276 ctx = blk_mq_get_ctx(q);
277 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Ming Leicb96a422014-06-01 00:43:37 +0800278 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
279 hctx);
280 rq = __blk_mq_alloc_request(&alloc_data, rw);
281 ctx = alloc_data.ctx;
Christoph Hellwigd8525642014-05-27 20:59:50 +0200282 }
283 blk_mq_put_ctx(ctx);
Keith Buschc76541a2014-12-19 17:54:13 -0700284 if (!rq) {
285 blk_mq_queue_exit(q);
Joe Lawrencea492f072014-08-28 08:15:21 -0600286 return ERR_PTR(-EWOULDBLOCK);
Keith Buschc76541a2014-12-19 17:54:13 -0700287 }
Jens Axboe320ae512013-10-24 09:20:05 +0100288 return rq;
289}
Jens Axboe4bb659b2014-05-09 09:36:49 -0600290EXPORT_SYMBOL(blk_mq_alloc_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100291
Jens Axboe320ae512013-10-24 09:20:05 +0100292static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
293 struct blk_mq_ctx *ctx, struct request *rq)
294{
295 const int tag = rq->tag;
296 struct request_queue *q = rq->q;
297
Jens Axboe0d2602c2014-05-13 15:10:52 -0600298 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
299 atomic_dec(&hctx->nr_active);
David Hildenbrand683d0e12014-09-18 11:04:31 +0200300 rq->cmd_flags = 0;
Jens Axboe0d2602c2014-05-13 15:10:52 -0600301
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200302 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
Jens Axboe0d2602c2014-05-13 15:10:52 -0600303 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
Jens Axboe320ae512013-10-24 09:20:05 +0100304 blk_mq_queue_exit(q);
305}
306
Jens Axboe7c7f2f22014-11-17 10:41:57 -0700307void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100308{
309 struct blk_mq_ctx *ctx = rq->mq_ctx;
Jens Axboe7c7f2f22014-11-17 10:41:57 -0700310
311 ctx->rq_completed[rq_is_sync(rq)]++;
312 __blk_mq_free_request(hctx, ctx, rq);
313
314}
315EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
316
317void blk_mq_free_request(struct request *rq)
318{
Jens Axboe320ae512013-10-24 09:20:05 +0100319 struct blk_mq_hw_ctx *hctx;
320 struct request_queue *q = rq->q;
321
Jens Axboe7c7f2f22014-11-17 10:41:57 -0700322 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
323 blk_mq_free_hctx_request(hctx, rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100324}
Jens Axboe1a3b5952014-11-17 10:40:48 -0700325EXPORT_SYMBOL_GPL(blk_mq_free_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100326
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700327inline void __blk_mq_end_request(struct request *rq, int error)
Jens Axboe320ae512013-10-24 09:20:05 +0100328{
Ming Lei0d11e6a2013-12-05 10:50:39 -0700329 blk_account_io_done(rq);
330
Christoph Hellwig91b63632014-04-16 09:44:53 +0200331 if (rq->end_io) {
Jens Axboe320ae512013-10-24 09:20:05 +0100332 rq->end_io(rq, error);
Christoph Hellwig91b63632014-04-16 09:44:53 +0200333 } else {
334 if (unlikely(blk_bidi_rq(rq)))
335 blk_mq_free_request(rq->next_rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100336 blk_mq_free_request(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +0200337 }
Jens Axboe320ae512013-10-24 09:20:05 +0100338}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700339EXPORT_SYMBOL(__blk_mq_end_request);
Christoph Hellwig63151a42014-04-16 09:44:52 +0200340
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700341void blk_mq_end_request(struct request *rq, int error)
Christoph Hellwig63151a42014-04-16 09:44:52 +0200342{
343 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
344 BUG();
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700345 __blk_mq_end_request(rq, error);
Christoph Hellwig63151a42014-04-16 09:44:52 +0200346}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700347EXPORT_SYMBOL(blk_mq_end_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100348
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800349static void __blk_mq_complete_request_remote(void *data)
Jens Axboe320ae512013-10-24 09:20:05 +0100350{
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800351 struct request *rq = data;
Jens Axboe320ae512013-10-24 09:20:05 +0100352
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800353 rq->q->softirq_done_fn(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100354}
355
Jens Axboeed851862014-05-30 21:20:50 -0600356static void blk_mq_ipi_complete_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100357{
358 struct blk_mq_ctx *ctx = rq->mq_ctx;
Christoph Hellwig38535202014-04-25 02:32:53 -0700359 bool shared = false;
Jens Axboe320ae512013-10-24 09:20:05 +0100360 int cpu;
361
Christoph Hellwig38535202014-04-25 02:32:53 -0700362 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800363 rq->q->softirq_done_fn(rq);
364 return;
365 }
Jens Axboe320ae512013-10-24 09:20:05 +0100366
367 cpu = get_cpu();
Christoph Hellwig38535202014-04-25 02:32:53 -0700368 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
369 shared = cpus_share_cache(cpu, ctx->cpu);
370
371 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800372 rq->csd.func = __blk_mq_complete_request_remote;
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800373 rq->csd.info = rq;
374 rq->csd.flags = 0;
Frederic Weisbeckerc46fff22014-02-24 16:40:02 +0100375 smp_call_function_single_async(ctx->cpu, &rq->csd);
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800376 } else {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800377 rq->q->softirq_done_fn(rq);
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800378 }
Jens Axboe320ae512013-10-24 09:20:05 +0100379 put_cpu();
380}
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800381
Jens Axboeed851862014-05-30 21:20:50 -0600382void __blk_mq_complete_request(struct request *rq)
383{
384 struct request_queue *q = rq->q;
385
386 if (!q->softirq_done_fn)
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700387 blk_mq_end_request(rq, rq->errors);
Jens Axboeed851862014-05-30 21:20:50 -0600388 else
389 blk_mq_ipi_complete_request(rq);
390}
391
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800392/**
393 * blk_mq_complete_request - end I/O on a request
394 * @rq: the request being processed
395 *
396 * Description:
397 * Ends all I/O on a request. It does not handle partial completions.
398 * The actual completion happens out-of-order, through a IPI handler.
399 **/
400void blk_mq_complete_request(struct request *rq)
401{
Jens Axboe95f09682014-05-27 17:46:48 -0600402 struct request_queue *q = rq->q;
403
404 if (unlikely(blk_should_fake_timeout(q)))
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800405 return;
Jens Axboeed851862014-05-30 21:20:50 -0600406 if (!blk_mark_rq_complete(rq))
407 __blk_mq_complete_request(rq);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800408}
409EXPORT_SYMBOL(blk_mq_complete_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100410
Keith Busch973c0192015-01-07 18:55:43 -0700411int blk_mq_request_started(struct request *rq)
412{
413 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
414}
415EXPORT_SYMBOL_GPL(blk_mq_request_started);
416
Christoph Hellwige2490072014-09-13 16:40:09 -0700417void blk_mq_start_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100418{
419 struct request_queue *q = rq->q;
420
421 trace_block_rq_issue(q, rq);
422
Christoph Hellwig742ee692014-04-14 10:30:06 +0200423 rq->resid_len = blk_rq_bytes(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +0200424 if (unlikely(blk_bidi_rq(rq)))
425 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
Christoph Hellwig742ee692014-04-14 10:30:06 +0200426
Ming Lei2b8393b2014-06-10 00:16:41 +0800427 blk_add_timer(rq);
Jens Axboe87ee7b12014-04-24 08:51:47 -0600428
429 /*
Jens Axboe538b7532014-09-16 10:37:37 -0600430 * Ensure that ->deadline is visible before set the started
431 * flag and clear the completed flag.
432 */
433 smp_mb__before_atomic();
434
435 /*
Jens Axboe87ee7b12014-04-24 08:51:47 -0600436 * Mark us as started and clear complete. Complete might have been
437 * set if requeue raced with timeout, which then marked it as
438 * complete. So be sure to clear complete again when we start
439 * the request, otherwise we'll ignore the completion event.
440 */
Jens Axboe4b570522014-05-29 11:00:11 -0600441 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
442 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
443 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
444 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800445
446 if (q->dma_drain_size && blk_rq_bytes(rq)) {
447 /*
448 * Make sure space for the drain appears. We know we can do
449 * this because max_hw_segments has been adjusted to be one
450 * fewer than the device can handle.
451 */
452 rq->nr_phys_segments++;
453 }
Jens Axboe320ae512013-10-24 09:20:05 +0100454}
Christoph Hellwige2490072014-09-13 16:40:09 -0700455EXPORT_SYMBOL(blk_mq_start_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100456
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200457static void __blk_mq_requeue_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100458{
459 struct request_queue *q = rq->q;
460
461 trace_block_rq_requeue(q, rq);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800462
Christoph Hellwige2490072014-09-13 16:40:09 -0700463 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
464 if (q->dma_drain_size && blk_rq_bytes(rq))
465 rq->nr_phys_segments--;
466 }
Jens Axboe320ae512013-10-24 09:20:05 +0100467}
468
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200469void blk_mq_requeue_request(struct request *rq)
470{
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200471 __blk_mq_requeue_request(rq);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200472
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200473 BUG_ON(blk_queued_rq(rq));
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600474 blk_mq_add_to_requeue_list(rq, true);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200475}
476EXPORT_SYMBOL(blk_mq_requeue_request);
477
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600478static void blk_mq_requeue_work(struct work_struct *work)
479{
480 struct request_queue *q =
481 container_of(work, struct request_queue, requeue_work);
482 LIST_HEAD(rq_list);
483 struct request *rq, *next;
484 unsigned long flags;
485
486 spin_lock_irqsave(&q->requeue_lock, flags);
487 list_splice_init(&q->requeue_list, &rq_list);
488 spin_unlock_irqrestore(&q->requeue_lock, flags);
489
490 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
491 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
492 continue;
493
494 rq->cmd_flags &= ~REQ_SOFTBARRIER;
495 list_del_init(&rq->queuelist);
496 blk_mq_insert_request(rq, true, false, false);
497 }
498
499 while (!list_empty(&rq_list)) {
500 rq = list_entry(rq_list.next, struct request, queuelist);
501 list_del_init(&rq->queuelist);
502 blk_mq_insert_request(rq, false, false, false);
503 }
504
Jens Axboe8b957412014-09-19 13:10:29 -0600505 /*
506 * Use the start variant of queue running here, so that running
507 * the requeue work will kick stopped queues.
508 */
509 blk_mq_start_hw_queues(q);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600510}
511
512void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
513{
514 struct request_queue *q = rq->q;
515 unsigned long flags;
516
517 /*
518 * We abuse this flag that is otherwise used by the I/O scheduler to
519 * request head insertation from the workqueue.
520 */
521 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
522
523 spin_lock_irqsave(&q->requeue_lock, flags);
524 if (at_head) {
525 rq->cmd_flags |= REQ_SOFTBARRIER;
526 list_add(&rq->queuelist, &q->requeue_list);
527 } else {
528 list_add_tail(&rq->queuelist, &q->requeue_list);
529 }
530 spin_unlock_irqrestore(&q->requeue_lock, flags);
531}
532EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
533
Keith Buschc68ed592015-01-07 18:55:44 -0700534void blk_mq_cancel_requeue_work(struct request_queue *q)
535{
536 cancel_work_sync(&q->requeue_work);
537}
538EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
539
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600540void blk_mq_kick_requeue_list(struct request_queue *q)
541{
542 kblockd_schedule_work(&q->requeue_work);
543}
544EXPORT_SYMBOL(blk_mq_kick_requeue_list);
545
Jens Axboe1885b242015-01-07 18:55:45 -0700546void blk_mq_abort_requeue_list(struct request_queue *q)
547{
548 unsigned long flags;
549 LIST_HEAD(rq_list);
550
551 spin_lock_irqsave(&q->requeue_lock, flags);
552 list_splice_init(&q->requeue_list, &rq_list);
553 spin_unlock_irqrestore(&q->requeue_lock, flags);
554
555 while (!list_empty(&rq_list)) {
556 struct request *rq;
557
558 rq = list_first_entry(&rq_list, struct request, queuelist);
559 list_del_init(&rq->queuelist);
560 rq->errors = -EIO;
561 blk_mq_end_request(rq, rq->errors);
562 }
563}
564EXPORT_SYMBOL(blk_mq_abort_requeue_list);
565
Ming Lei7c94e1c2014-09-25 23:23:43 +0800566static inline bool is_flush_request(struct request *rq,
567 struct blk_flush_queue *fq, unsigned int tag)
Christoph Hellwig24d2f902014-04-15 14:14:00 -0600568{
Jens Axboe0e62f512014-06-04 10:23:49 -0600569 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
Ming Lei7c94e1c2014-09-25 23:23:43 +0800570 fq->flush_rq->tag == tag);
Jens Axboe0e62f512014-06-04 10:23:49 -0600571}
Shaohua Li22302372014-05-30 08:06:42 -0600572
Jens Axboe0e62f512014-06-04 10:23:49 -0600573struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
574{
575 struct request *rq = tags->rqs[tag];
Ming Leie97c2932014-09-25 23:23:46 +0800576 /* mq_ctx of flush rq is always cloned from the corresponding req */
577 struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
Shaohua Li22302372014-05-30 08:06:42 -0600578
Ming Lei7c94e1c2014-09-25 23:23:43 +0800579 if (!is_flush_request(rq, fq, tag))
Jens Axboe0e62f512014-06-04 10:23:49 -0600580 return rq;
581
Ming Lei7c94e1c2014-09-25 23:23:43 +0800582 return fq->flush_rq;
Christoph Hellwig24d2f902014-04-15 14:14:00 -0600583}
584EXPORT_SYMBOL(blk_mq_tag_to_rq);
585
Jens Axboe320ae512013-10-24 09:20:05 +0100586struct blk_mq_timeout_data {
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700587 unsigned long next;
588 unsigned int next_set;
Jens Axboe320ae512013-10-24 09:20:05 +0100589};
590
Christoph Hellwig90415832014-09-22 10:21:48 -0600591void blk_mq_rq_timed_out(struct request *req, bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100592{
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700593 struct blk_mq_ops *ops = req->q->mq_ops;
594 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
Jens Axboe87ee7b12014-04-24 08:51:47 -0600595
596 /*
597 * We know that complete is set at this point. If STARTED isn't set
598 * anymore, then the request isn't active and the "timeout" should
599 * just be ignored. This can happen due to the bitflag ordering.
600 * Timeout first checks if STARTED is set, and if it is, assumes
601 * the request is active. But if we race with completion, then
602 * we both flags will get cleared. So check here again, and ignore
603 * a timeout event with a request that isn't active.
604 */
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700605 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
606 return;
Jens Axboe87ee7b12014-04-24 08:51:47 -0600607
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700608 if (ops->timeout)
Christoph Hellwig0152fb62014-09-13 16:40:13 -0700609 ret = ops->timeout(req, reserved);
Jens Axboe87ee7b12014-04-24 08:51:47 -0600610
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700611 switch (ret) {
612 case BLK_EH_HANDLED:
613 __blk_mq_complete_request(req);
614 break;
615 case BLK_EH_RESET_TIMER:
616 blk_add_timer(req);
617 blk_clear_rq_complete(req);
618 break;
619 case BLK_EH_NOT_HANDLED:
620 break;
621 default:
622 printk(KERN_ERR "block: bad eh return: %d\n", ret);
623 break;
624 }
Jens Axboe87ee7b12014-04-24 08:51:47 -0600625}
Keith Busch5b3f25f2015-01-07 18:55:46 -0700626
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700627static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
628 struct request *rq, void *priv, bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100629{
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700630 struct blk_mq_timeout_data *data = priv;
631
Keith Buscheb130db2015-01-08 08:59:53 -0700632 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
633 /*
634 * If a request wasn't started before the queue was
635 * marked dying, kill it here or it'll go unnoticed.
636 */
637 if (unlikely(blk_queue_dying(rq->q))) {
638 rq->errors = -EIO;
639 blk_mq_complete_request(rq);
640 }
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700641 return;
Keith Buscheb130db2015-01-08 08:59:53 -0700642 }
Keith Busch5b3f25f2015-01-07 18:55:46 -0700643 if (rq->cmd_flags & REQ_NO_TIMEOUT)
644 return;
Jens Axboe320ae512013-10-24 09:20:05 +0100645
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700646 if (time_after_eq(jiffies, rq->deadline)) {
647 if (!blk_mark_rq_complete(rq))
Christoph Hellwig0152fb62014-09-13 16:40:13 -0700648 blk_mq_rq_timed_out(rq, reserved);
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700649 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
650 data->next = rq->deadline;
651 data->next_set = 1;
652 }
Jens Axboe320ae512013-10-24 09:20:05 +0100653}
654
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700655static void blk_mq_rq_timer(unsigned long priv)
Jens Axboe320ae512013-10-24 09:20:05 +0100656{
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700657 struct request_queue *q = (struct request_queue *)priv;
658 struct blk_mq_timeout_data data = {
659 .next = 0,
660 .next_set = 0,
661 };
Jens Axboe320ae512013-10-24 09:20:05 +0100662 struct blk_mq_hw_ctx *hctx;
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700663 int i;
Jens Axboe320ae512013-10-24 09:20:05 +0100664
Jens Axboe484b4062014-05-21 14:01:15 -0600665 queue_for_each_hw_ctx(q, hctx, i) {
666 /*
667 * If not software queues are currently mapped to this
668 * hardware queue, there's nothing to check
669 */
Ming Lei19c66e52014-12-03 19:38:04 +0800670 if (!blk_mq_hw_queue_mapped(hctx))
Jens Axboe484b4062014-05-21 14:01:15 -0600671 continue;
672
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700673 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
Jens Axboe484b4062014-05-21 14:01:15 -0600674 }
Jens Axboe320ae512013-10-24 09:20:05 +0100675
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700676 if (data.next_set) {
677 data.next = blk_rq_timeout(round_jiffies_up(data.next));
678 mod_timer(&q->timeout, data.next);
Jens Axboe0d2602c2014-05-13 15:10:52 -0600679 } else {
Ming Leif054b562015-04-21 10:00:19 +0800680 queue_for_each_hw_ctx(q, hctx, i) {
681 /* the hctx may be unmapped, so check it here */
682 if (blk_mq_hw_queue_mapped(hctx))
683 blk_mq_tag_idle(hctx);
684 }
Jens Axboe0d2602c2014-05-13 15:10:52 -0600685 }
Jens Axboe320ae512013-10-24 09:20:05 +0100686}
687
688/*
689 * Reverse check our software queue for entries that we could potentially
690 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
691 * too much time checking for merges.
692 */
693static bool blk_mq_attempt_merge(struct request_queue *q,
694 struct blk_mq_ctx *ctx, struct bio *bio)
695{
696 struct request *rq;
697 int checked = 8;
698
699 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
700 int el_ret;
701
702 if (!checked--)
703 break;
704
705 if (!blk_rq_merge_ok(rq, bio))
706 continue;
707
708 el_ret = blk_try_merge(rq, bio);
709 if (el_ret == ELEVATOR_BACK_MERGE) {
710 if (bio_attempt_back_merge(q, rq, bio)) {
711 ctx->rq_merged++;
712 return true;
713 }
714 break;
715 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
716 if (bio_attempt_front_merge(q, rq, bio)) {
717 ctx->rq_merged++;
718 return true;
719 }
720 break;
721 }
722 }
723
724 return false;
725}
726
Jens Axboe320ae512013-10-24 09:20:05 +0100727/*
Jens Axboe1429d7c2014-05-19 09:23:55 -0600728 * Process software queues that have been marked busy, splicing them
729 * to the for-dispatch
730 */
731static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
732{
733 struct blk_mq_ctx *ctx;
734 int i;
735
Jens Axboe569fd0c2015-04-17 08:28:50 -0600736 for (i = 0; i < hctx->ctx_map.size; i++) {
Jens Axboe1429d7c2014-05-19 09:23:55 -0600737 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
738 unsigned int off, bit;
739
740 if (!bm->word)
741 continue;
742
743 bit = 0;
744 off = i * hctx->ctx_map.bits_per_word;
745 do {
746 bit = find_next_bit(&bm->word, bm->depth, bit);
747 if (bit >= bm->depth)
748 break;
749
750 ctx = hctx->ctxs[bit + off];
751 clear_bit(bit, &bm->word);
752 spin_lock(&ctx->lock);
753 list_splice_tail_init(&ctx->rq_list, list);
754 spin_unlock(&ctx->lock);
755
756 bit++;
757 } while (1);
758 }
759}
760
761/*
Jens Axboe320ae512013-10-24 09:20:05 +0100762 * Run this hardware queue, pulling any software queues mapped to it in.
763 * Note that this function currently has various problems around ordering
764 * of IO. In particular, we'd like FIFO behaviour on handling existing
765 * items on the hctx->dispatch list. Ignore that for now.
766 */
767static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
768{
769 struct request_queue *q = hctx->queue;
Jens Axboe320ae512013-10-24 09:20:05 +0100770 struct request *rq;
771 LIST_HEAD(rq_list);
Jens Axboe74c45052014-10-29 11:14:52 -0600772 LIST_HEAD(driver_list);
773 struct list_head *dptr;
Jens Axboe1429d7c2014-05-19 09:23:55 -0600774 int queued;
Jens Axboe320ae512013-10-24 09:20:05 +0100775
Jens Axboefd1270d2014-04-16 09:23:48 -0600776 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
Jens Axboee4043dc2014-04-09 10:18:23 -0600777
Jens Axboe5d12f902014-03-19 15:25:02 -0600778 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
Jens Axboe320ae512013-10-24 09:20:05 +0100779 return;
780
781 hctx->run++;
782
783 /*
784 * Touch any software queue that has pending entries.
785 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600786 flush_busy_ctxs(hctx, &rq_list);
Jens Axboe320ae512013-10-24 09:20:05 +0100787
788 /*
789 * If we have previous entries on our dispatch list, grab them
790 * and stuff them at the front for more fair dispatch.
791 */
792 if (!list_empty_careful(&hctx->dispatch)) {
793 spin_lock(&hctx->lock);
794 if (!list_empty(&hctx->dispatch))
795 list_splice_init(&hctx->dispatch, &rq_list);
796 spin_unlock(&hctx->lock);
797 }
798
799 /*
Jens Axboe74c45052014-10-29 11:14:52 -0600800 * Start off with dptr being NULL, so we start the first request
801 * immediately, even if we have more pending.
802 */
803 dptr = NULL;
804
805 /*
Jens Axboe320ae512013-10-24 09:20:05 +0100806 * Now process all the entries, sending them to the driver.
807 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600808 queued = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100809 while (!list_empty(&rq_list)) {
Jens Axboe74c45052014-10-29 11:14:52 -0600810 struct blk_mq_queue_data bd;
Jens Axboe320ae512013-10-24 09:20:05 +0100811 int ret;
812
813 rq = list_first_entry(&rq_list, struct request, queuelist);
814 list_del_init(&rq->queuelist);
Jens Axboe320ae512013-10-24 09:20:05 +0100815
Jens Axboe74c45052014-10-29 11:14:52 -0600816 bd.rq = rq;
817 bd.list = dptr;
818 bd.last = list_empty(&rq_list);
819
820 ret = q->mq_ops->queue_rq(hctx, &bd);
Jens Axboe320ae512013-10-24 09:20:05 +0100821 switch (ret) {
822 case BLK_MQ_RQ_QUEUE_OK:
823 queued++;
824 continue;
825 case BLK_MQ_RQ_QUEUE_BUSY:
Jens Axboe320ae512013-10-24 09:20:05 +0100826 list_add(&rq->queuelist, &rq_list);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200827 __blk_mq_requeue_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100828 break;
829 default:
830 pr_err("blk-mq: bad return on queue: %d\n", ret);
Jens Axboe320ae512013-10-24 09:20:05 +0100831 case BLK_MQ_RQ_QUEUE_ERROR:
Christoph Hellwig1e93b8c2014-02-11 08:27:13 -0800832 rq->errors = -EIO;
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700833 blk_mq_end_request(rq, rq->errors);
Jens Axboe320ae512013-10-24 09:20:05 +0100834 break;
835 }
836
837 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
838 break;
Jens Axboe74c45052014-10-29 11:14:52 -0600839
840 /*
841 * We've done the first request. If we have more than 1
842 * left in the list, set dptr to defer issue.
843 */
844 if (!dptr && rq_list.next != rq_list.prev)
845 dptr = &driver_list;
Jens Axboe320ae512013-10-24 09:20:05 +0100846 }
847
848 if (!queued)
849 hctx->dispatched[0]++;
850 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
851 hctx->dispatched[ilog2(queued) + 1]++;
852
853 /*
854 * Any items that need requeuing? Stuff them into hctx->dispatch,
855 * that is where we will continue on next queue run.
856 */
857 if (!list_empty(&rq_list)) {
858 spin_lock(&hctx->lock);
859 list_splice(&rq_list, &hctx->dispatch);
860 spin_unlock(&hctx->lock);
Shaohua Li9ba52e52015-05-04 14:32:48 -0600861 /*
862 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
863 * it's possible the queue is stopped and restarted again
864 * before this. Queue restart will dispatch requests. And since
865 * requests in rq_list aren't added into hctx->dispatch yet,
866 * the requests in rq_list might get lost.
867 *
868 * blk_mq_run_hw_queue() already checks the STOPPED bit
869 **/
870 blk_mq_run_hw_queue(hctx, true);
Jens Axboe320ae512013-10-24 09:20:05 +0100871 }
872}
873
Jens Axboe506e9312014-05-07 10:26:44 -0600874/*
875 * It'd be great if the workqueue API had a way to pass
876 * in a mask and had some smarts for more clever placement.
877 * For now we just round-robin here, switching for every
878 * BLK_MQ_CPU_WORK_BATCH queued items.
879 */
880static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
881{
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100882 if (hctx->queue->nr_hw_queues == 1)
883 return WORK_CPU_UNBOUND;
Jens Axboe506e9312014-05-07 10:26:44 -0600884
885 if (--hctx->next_cpu_batch <= 0) {
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100886 int cpu = hctx->next_cpu, next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600887
888 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
889 if (next_cpu >= nr_cpu_ids)
890 next_cpu = cpumask_first(hctx->cpumask);
891
892 hctx->next_cpu = next_cpu;
893 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100894
895 return cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600896 }
897
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100898 return hctx->next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600899}
900
Jens Axboe320ae512013-10-24 09:20:05 +0100901void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
902{
Ming Lei19c66e52014-12-03 19:38:04 +0800903 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
904 !blk_mq_hw_queue_mapped(hctx)))
Jens Axboe320ae512013-10-24 09:20:05 +0100905 return;
906
Paolo Bonzini398205b2014-11-07 23:03:59 +0100907 if (!async) {
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100908 int cpu = get_cpu();
909 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
Paolo Bonzini398205b2014-11-07 23:03:59 +0100910 __blk_mq_run_hw_queue(hctx);
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100911 put_cpu();
Paolo Bonzini398205b2014-11-07 23:03:59 +0100912 return;
913 }
Jens Axboee4043dc2014-04-09 10:18:23 -0600914
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100915 put_cpu();
Jens Axboee4043dc2014-04-09 10:18:23 -0600916 }
Paolo Bonzini398205b2014-11-07 23:03:59 +0100917
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100918 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
919 &hctx->run_work, 0);
Jens Axboe320ae512013-10-24 09:20:05 +0100920}
921
Mike Snitzerb94ec292015-03-11 23:56:38 -0400922void blk_mq_run_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +0100923{
924 struct blk_mq_hw_ctx *hctx;
925 int i;
926
927 queue_for_each_hw_ctx(q, hctx, i) {
928 if ((!blk_mq_hctx_has_pending(hctx) &&
929 list_empty_careful(&hctx->dispatch)) ||
Jens Axboe5d12f902014-03-19 15:25:02 -0600930 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
Jens Axboe320ae512013-10-24 09:20:05 +0100931 continue;
932
Mike Snitzerb94ec292015-03-11 23:56:38 -0400933 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +0100934 }
935}
Mike Snitzerb94ec292015-03-11 23:56:38 -0400936EXPORT_SYMBOL(blk_mq_run_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +0100937
938void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
939{
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600940 cancel_delayed_work(&hctx->run_work);
941 cancel_delayed_work(&hctx->delay_work);
Jens Axboe320ae512013-10-24 09:20:05 +0100942 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
943}
944EXPORT_SYMBOL(blk_mq_stop_hw_queue);
945
Christoph Hellwig280d45f2013-10-25 14:45:58 +0100946void blk_mq_stop_hw_queues(struct request_queue *q)
947{
948 struct blk_mq_hw_ctx *hctx;
949 int i;
950
951 queue_for_each_hw_ctx(q, hctx, i)
952 blk_mq_stop_hw_queue(hctx);
953}
954EXPORT_SYMBOL(blk_mq_stop_hw_queues);
955
Jens Axboe320ae512013-10-24 09:20:05 +0100956void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
957{
958 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboee4043dc2014-04-09 10:18:23 -0600959
Jens Axboe0ffbce82014-06-25 08:22:34 -0600960 blk_mq_run_hw_queue(hctx, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100961}
962EXPORT_SYMBOL(blk_mq_start_hw_queue);
963
Christoph Hellwig2f268552014-04-16 09:44:56 +0200964void blk_mq_start_hw_queues(struct request_queue *q)
965{
966 struct blk_mq_hw_ctx *hctx;
967 int i;
968
969 queue_for_each_hw_ctx(q, hctx, i)
970 blk_mq_start_hw_queue(hctx);
971}
972EXPORT_SYMBOL(blk_mq_start_hw_queues);
973
Christoph Hellwig1b4a3252014-04-16 09:44:54 +0200974void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +0100975{
976 struct blk_mq_hw_ctx *hctx;
977 int i;
978
979 queue_for_each_hw_ctx(q, hctx, i) {
980 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
981 continue;
982
983 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Christoph Hellwig1b4a3252014-04-16 09:44:54 +0200984 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +0100985 }
986}
987EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
988
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600989static void blk_mq_run_work_fn(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +0100990{
991 struct blk_mq_hw_ctx *hctx;
992
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600993 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
Jens Axboee4043dc2014-04-09 10:18:23 -0600994
Jens Axboe320ae512013-10-24 09:20:05 +0100995 __blk_mq_run_hw_queue(hctx);
996}
997
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600998static void blk_mq_delay_work_fn(struct work_struct *work)
999{
1000 struct blk_mq_hw_ctx *hctx;
1001
1002 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1003
1004 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1005 __blk_mq_run_hw_queue(hctx);
1006}
1007
1008void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1009{
Ming Lei19c66e52014-12-03 19:38:04 +08001010 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1011 return;
Christoph Hellwig70f4db62014-04-16 10:48:08 -06001012
Christoph Hellwigb657d7e2014-11-24 09:27:23 +01001013 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1014 &hctx->delay_work, msecs_to_jiffies(msecs));
Christoph Hellwig70f4db62014-04-16 10:48:08 -06001015}
1016EXPORT_SYMBOL(blk_mq_delay_queue);
1017
Jens Axboe320ae512013-10-24 09:20:05 +01001018static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig72a0a362014-02-07 10:22:36 -08001019 struct request *rq, bool at_head)
Jens Axboe320ae512013-10-24 09:20:05 +01001020{
1021 struct blk_mq_ctx *ctx = rq->mq_ctx;
1022
Jens Axboe01b983c2013-11-19 18:59:10 -07001023 trace_block_rq_insert(hctx->queue, rq);
1024
Christoph Hellwig72a0a362014-02-07 10:22:36 -08001025 if (at_head)
1026 list_add(&rq->queuelist, &ctx->rq_list);
1027 else
1028 list_add_tail(&rq->queuelist, &ctx->rq_list);
Jens Axboe4bb659b2014-05-09 09:36:49 -06001029
Jens Axboe320ae512013-10-24 09:20:05 +01001030 blk_mq_hctx_mark_pending(hctx, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001031}
1032
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001033void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1034 bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01001035{
1036 struct request_queue *q = rq->q;
1037 struct blk_mq_hw_ctx *hctx;
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001038 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001039
1040 current_ctx = blk_mq_get_ctx(q);
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001041 if (!cpu_online(ctx->cpu))
1042 rq->mq_ctx = ctx = current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001043
Jens Axboe320ae512013-10-24 09:20:05 +01001044 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1045
Christoph Hellwiga57a1782014-09-16 14:44:07 -07001046 spin_lock(&ctx->lock);
1047 __blk_mq_insert_request(hctx, rq, at_head);
1048 spin_unlock(&ctx->lock);
Jens Axboe320ae512013-10-24 09:20:05 +01001049
Jens Axboe320ae512013-10-24 09:20:05 +01001050 if (run_queue)
1051 blk_mq_run_hw_queue(hctx, async);
Jens Axboee4043dc2014-04-09 10:18:23 -06001052
1053 blk_mq_put_ctx(current_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001054}
1055
1056static void blk_mq_insert_requests(struct request_queue *q,
1057 struct blk_mq_ctx *ctx,
1058 struct list_head *list,
1059 int depth,
1060 bool from_schedule)
1061
1062{
1063 struct blk_mq_hw_ctx *hctx;
1064 struct blk_mq_ctx *current_ctx;
1065
1066 trace_block_unplug(q, depth, !from_schedule);
1067
1068 current_ctx = blk_mq_get_ctx(q);
1069
1070 if (!cpu_online(ctx->cpu))
1071 ctx = current_ctx;
1072 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1073
1074 /*
1075 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1076 * offline now
1077 */
1078 spin_lock(&ctx->lock);
1079 while (!list_empty(list)) {
1080 struct request *rq;
1081
1082 rq = list_first_entry(list, struct request, queuelist);
1083 list_del_init(&rq->queuelist);
1084 rq->mq_ctx = ctx;
Christoph Hellwig72a0a362014-02-07 10:22:36 -08001085 __blk_mq_insert_request(hctx, rq, false);
Jens Axboe320ae512013-10-24 09:20:05 +01001086 }
1087 spin_unlock(&ctx->lock);
1088
Jens Axboe320ae512013-10-24 09:20:05 +01001089 blk_mq_run_hw_queue(hctx, from_schedule);
Jens Axboee4043dc2014-04-09 10:18:23 -06001090 blk_mq_put_ctx(current_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001091}
1092
1093static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1094{
1095 struct request *rqa = container_of(a, struct request, queuelist);
1096 struct request *rqb = container_of(b, struct request, queuelist);
1097
1098 return !(rqa->mq_ctx < rqb->mq_ctx ||
1099 (rqa->mq_ctx == rqb->mq_ctx &&
1100 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1101}
1102
1103void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1104{
1105 struct blk_mq_ctx *this_ctx;
1106 struct request_queue *this_q;
1107 struct request *rq;
1108 LIST_HEAD(list);
1109 LIST_HEAD(ctx_list);
1110 unsigned int depth;
1111
1112 list_splice_init(&plug->mq_list, &list);
1113
1114 list_sort(NULL, &list, plug_ctx_cmp);
1115
1116 this_q = NULL;
1117 this_ctx = NULL;
1118 depth = 0;
1119
1120 while (!list_empty(&list)) {
1121 rq = list_entry_rq(list.next);
1122 list_del_init(&rq->queuelist);
1123 BUG_ON(!rq->q);
1124 if (rq->mq_ctx != this_ctx) {
1125 if (this_ctx) {
1126 blk_mq_insert_requests(this_q, this_ctx,
1127 &ctx_list, depth,
1128 from_schedule);
1129 }
1130
1131 this_ctx = rq->mq_ctx;
1132 this_q = rq->q;
1133 depth = 0;
1134 }
1135
1136 depth++;
1137 list_add_tail(&rq->queuelist, &ctx_list);
1138 }
1139
1140 /*
1141 * If 'this_ctx' is set, we know we have entries to complete
1142 * on 'ctx_list'. Do those.
1143 */
1144 if (this_ctx) {
1145 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1146 from_schedule);
1147 }
1148}
1149
1150static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1151{
1152 init_request_from_bio(rq, bio);
Jens Axboe4b570522014-05-29 11:00:11 -06001153
Jens Axboe3ee32372014-06-09 09:36:53 -06001154 if (blk_do_io_stat(rq))
Jens Axboe4b570522014-05-29 11:00:11 -06001155 blk_account_io_start(rq, 1);
Jens Axboe320ae512013-10-24 09:20:05 +01001156}
1157
Jens Axboe274a5842014-08-15 12:44:08 -06001158static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1159{
1160 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1161 !blk_queue_nomerges(hctx->queue);
1162}
1163
Jens Axboe07068d52014-05-22 10:40:51 -06001164static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1165 struct blk_mq_ctx *ctx,
1166 struct request *rq, struct bio *bio)
1167{
Jens Axboe274a5842014-08-15 12:44:08 -06001168 if (!hctx_allow_merges(hctx)) {
Jens Axboe07068d52014-05-22 10:40:51 -06001169 blk_mq_bio_to_request(rq, bio);
1170 spin_lock(&ctx->lock);
1171insert_rq:
1172 __blk_mq_insert_request(hctx, rq, false);
1173 spin_unlock(&ctx->lock);
1174 return false;
1175 } else {
Jens Axboe274a5842014-08-15 12:44:08 -06001176 struct request_queue *q = hctx->queue;
1177
Jens Axboe07068d52014-05-22 10:40:51 -06001178 spin_lock(&ctx->lock);
1179 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1180 blk_mq_bio_to_request(rq, bio);
1181 goto insert_rq;
1182 }
1183
1184 spin_unlock(&ctx->lock);
1185 __blk_mq_free_request(hctx, ctx, rq);
1186 return true;
1187 }
1188}
1189
1190struct blk_map_ctx {
1191 struct blk_mq_hw_ctx *hctx;
1192 struct blk_mq_ctx *ctx;
1193};
1194
1195static struct request *blk_mq_map_request(struct request_queue *q,
1196 struct bio *bio,
1197 struct blk_map_ctx *data)
Jens Axboe320ae512013-10-24 09:20:05 +01001198{
1199 struct blk_mq_hw_ctx *hctx;
1200 struct blk_mq_ctx *ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001201 struct request *rq;
Jens Axboe07068d52014-05-22 10:40:51 -06001202 int rw = bio_data_dir(bio);
Ming Leicb96a422014-06-01 00:43:37 +08001203 struct blk_mq_alloc_data alloc_data;
Jens Axboe320ae512013-10-24 09:20:05 +01001204
Keith Buschbfd343a2015-03-11 23:56:39 -04001205 if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
Nicholas Bellinger14ec77f2014-02-07 13:45:39 -07001206 bio_endio(bio, -EIO);
Jens Axboe07068d52014-05-22 10:40:51 -06001207 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001208 }
1209
1210 ctx = blk_mq_get_ctx(q);
1211 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1212
Jens Axboe07068d52014-05-22 10:40:51 -06001213 if (rw_is_sync(bio->bi_rw))
Shaohua Li27fbf4e2014-02-19 20:20:21 +08001214 rw |= REQ_SYNC;
Jens Axboe07068d52014-05-22 10:40:51 -06001215
Jens Axboe320ae512013-10-24 09:20:05 +01001216 trace_block_getrq(q, bio, rw);
Ming Leicb96a422014-06-01 00:43:37 +08001217 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1218 hctx);
1219 rq = __blk_mq_alloc_request(&alloc_data, rw);
Christoph Hellwig5dee8572014-05-27 20:59:47 +02001220 if (unlikely(!rq)) {
Christoph Hellwig793597a2014-05-27 20:59:49 +02001221 __blk_mq_run_hw_queue(hctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001222 blk_mq_put_ctx(ctx);
1223 trace_block_sleeprq(q, bio, rw);
Christoph Hellwig793597a2014-05-27 20:59:49 +02001224
1225 ctx = blk_mq_get_ctx(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001226 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Ming Leicb96a422014-06-01 00:43:37 +08001227 blk_mq_set_alloc_data(&alloc_data, q,
1228 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1229 rq = __blk_mq_alloc_request(&alloc_data, rw);
1230 ctx = alloc_data.ctx;
1231 hctx = alloc_data.hctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001232 }
1233
1234 hctx->queued++;
Jens Axboe07068d52014-05-22 10:40:51 -06001235 data->hctx = hctx;
1236 data->ctx = ctx;
1237 return rq;
1238}
1239
1240/*
1241 * Multiple hardware queue variant. This will not use per-process plugs,
1242 * but will attempt to bypass the hctx queueing if we can go straight to
1243 * hardware for SYNC IO.
1244 */
1245static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1246{
1247 const int is_sync = rw_is_sync(bio->bi_rw);
1248 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1249 struct blk_map_ctx data;
1250 struct request *rq;
1251
1252 blk_queue_bounce(q, &bio);
1253
1254 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1255 bio_endio(bio, -EIO);
1256 return;
1257 }
1258
1259 rq = blk_mq_map_request(q, bio, &data);
1260 if (unlikely(!rq))
1261 return;
1262
1263 if (unlikely(is_flush_fua)) {
1264 blk_mq_bio_to_request(rq, bio);
1265 blk_insert_flush(rq);
1266 goto run_queue;
1267 }
1268
Jens Axboee167dfb2014-10-29 11:18:26 -06001269 /*
1270 * If the driver supports defer issued based on 'last', then
1271 * queue it up like normal since we can potentially save some
1272 * CPU this way.
1273 */
1274 if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
Jens Axboe74c45052014-10-29 11:14:52 -06001275 struct blk_mq_queue_data bd = {
1276 .rq = rq,
1277 .list = NULL,
1278 .last = 1
1279 };
Jens Axboe07068d52014-05-22 10:40:51 -06001280 int ret;
1281
1282 blk_mq_bio_to_request(rq, bio);
Jens Axboe07068d52014-05-22 10:40:51 -06001283
1284 /*
1285 * For OK queue, we are done. For error, kill it. Any other
1286 * error (busy), just add it to our list as we previously
1287 * would have done
1288 */
Jens Axboe74c45052014-10-29 11:14:52 -06001289 ret = q->mq_ops->queue_rq(data.hctx, &bd);
Jens Axboe07068d52014-05-22 10:40:51 -06001290 if (ret == BLK_MQ_RQ_QUEUE_OK)
1291 goto done;
1292 else {
1293 __blk_mq_requeue_request(rq);
1294
1295 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1296 rq->errors = -EIO;
Christoph Hellwigc8a446a2014-09-13 16:40:10 -07001297 blk_mq_end_request(rq, rq->errors);
Jens Axboe07068d52014-05-22 10:40:51 -06001298 goto done;
1299 }
1300 }
1301 }
1302
1303 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1304 /*
1305 * For a SYNC request, send it to the hardware immediately. For
1306 * an ASYNC request, just ensure that we run it later on. The
1307 * latter allows for merging opportunities and more efficient
1308 * dispatching.
1309 */
1310run_queue:
1311 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1312 }
1313done:
1314 blk_mq_put_ctx(data.ctx);
1315}
1316
1317/*
1318 * Single hardware queue variant. This will attempt to use any per-process
1319 * plug for merging and IO deferral.
1320 */
1321static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1322{
1323 const int is_sync = rw_is_sync(bio->bi_rw);
1324 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1325 unsigned int use_plug, request_count = 0;
1326 struct blk_map_ctx data;
1327 struct request *rq;
1328
1329 /*
1330 * If we have multiple hardware queues, just go directly to
1331 * one of those for sync IO.
1332 */
1333 use_plug = !is_flush_fua && !is_sync;
1334
1335 blk_queue_bounce(q, &bio);
1336
1337 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1338 bio_endio(bio, -EIO);
1339 return;
1340 }
1341
1342 if (use_plug && !blk_queue_nomerges(q) &&
1343 blk_attempt_plug_merge(q, bio, &request_count))
1344 return;
1345
1346 rq = blk_mq_map_request(q, bio, &data);
Jens Axboeff87bce2014-06-03 11:59:49 -06001347 if (unlikely(!rq))
1348 return;
Jens Axboe320ae512013-10-24 09:20:05 +01001349
1350 if (unlikely(is_flush_fua)) {
1351 blk_mq_bio_to_request(rq, bio);
Jens Axboe320ae512013-10-24 09:20:05 +01001352 blk_insert_flush(rq);
1353 goto run_queue;
1354 }
1355
1356 /*
1357 * A task plug currently exists. Since this is completely lockless,
1358 * utilize that to temporarily store requests until the task is
1359 * either done or scheduled away.
1360 */
1361 if (use_plug) {
1362 struct blk_plug *plug = current->plug;
1363
1364 if (plug) {
1365 blk_mq_bio_to_request(rq, bio);
Shaohua Li92f399c2013-10-29 12:01:03 -06001366 if (list_empty(&plug->mq_list))
Jens Axboe320ae512013-10-24 09:20:05 +01001367 trace_block_plug(q);
1368 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1369 blk_flush_plug_list(plug, false);
1370 trace_block_plug(q);
1371 }
1372 list_add_tail(&rq->queuelist, &plug->mq_list);
Jens Axboe07068d52014-05-22 10:40:51 -06001373 blk_mq_put_ctx(data.ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001374 return;
1375 }
1376 }
1377
Jens Axboe07068d52014-05-22 10:40:51 -06001378 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1379 /*
1380 * For a SYNC request, send it to the hardware immediately. For
1381 * an ASYNC request, just ensure that we run it later on. The
1382 * latter allows for merging opportunities and more efficient
1383 * dispatching.
1384 */
1385run_queue:
1386 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
Jens Axboe320ae512013-10-24 09:20:05 +01001387 }
1388
Jens Axboe07068d52014-05-22 10:40:51 -06001389 blk_mq_put_ctx(data.ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001390}
1391
1392/*
1393 * Default mapping to a software queue, since we use one per CPU.
1394 */
1395struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1396{
1397 return q->queue_hw_ctx[q->mq_map[cpu]];
1398}
1399EXPORT_SYMBOL(blk_mq_map_queue);
1400
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001401static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1402 struct blk_mq_tags *tags, unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01001403{
1404 struct page *page;
1405
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001406 if (tags->rqs && set->ops->exit_request) {
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001407 int i;
1408
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001409 for (i = 0; i < tags->nr_tags; i++) {
1410 if (!tags->rqs[i])
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001411 continue;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001412 set->ops->exit_request(set->driver_data, tags->rqs[i],
1413 hctx_idx, i);
Jens Axboea5164402014-09-10 09:02:03 -06001414 tags->rqs[i] = NULL;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001415 }
1416 }
1417
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001418 while (!list_empty(&tags->page_list)) {
1419 page = list_first_entry(&tags->page_list, struct page, lru);
Dave Hansen67534712014-01-08 20:17:46 -07001420 list_del_init(&page->lru);
Jens Axboe320ae512013-10-24 09:20:05 +01001421 __free_pages(page, page->private);
1422 }
1423
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001424 kfree(tags->rqs);
Jens Axboe320ae512013-10-24 09:20:05 +01001425
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001426 blk_mq_free_tags(tags);
Jens Axboe320ae512013-10-24 09:20:05 +01001427}
1428
1429static size_t order_to_size(unsigned int order)
1430{
Ming Lei4ca08502014-04-19 18:00:18 +08001431 return (size_t)PAGE_SIZE << order;
Jens Axboe320ae512013-10-24 09:20:05 +01001432}
1433
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001434static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1435 unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01001436{
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001437 struct blk_mq_tags *tags;
Jens Axboe320ae512013-10-24 09:20:05 +01001438 unsigned int i, j, entries_per_page, max_order = 4;
1439 size_t rq_size, left;
1440
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001441 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
Shaohua Li24391c02015-01-23 14:18:00 -07001442 set->numa_node,
1443 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001444 if (!tags)
1445 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001446
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001447 INIT_LIST_HEAD(&tags->page_list);
1448
Jens Axboea5164402014-09-10 09:02:03 -06001449 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1450 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1451 set->numa_node);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001452 if (!tags->rqs) {
1453 blk_mq_free_tags(tags);
1454 return NULL;
1455 }
Jens Axboe320ae512013-10-24 09:20:05 +01001456
1457 /*
1458 * rq_size is the size of the request plus driver payload, rounded
1459 * to the cacheline size
1460 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001461 rq_size = round_up(sizeof(struct request) + set->cmd_size,
Jens Axboe320ae512013-10-24 09:20:05 +01001462 cache_line_size());
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001463 left = rq_size * set->queue_depth;
Jens Axboe320ae512013-10-24 09:20:05 +01001464
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001465 for (i = 0; i < set->queue_depth; ) {
Jens Axboe320ae512013-10-24 09:20:05 +01001466 int this_order = max_order;
1467 struct page *page;
1468 int to_do;
1469 void *p;
1470
1471 while (left < order_to_size(this_order - 1) && this_order)
1472 this_order--;
1473
1474 do {
Jens Axboea5164402014-09-10 09:02:03 -06001475 page = alloc_pages_node(set->numa_node,
Linus Torvaldsac211172015-04-09 14:12:22 -07001476 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
Jens Axboea5164402014-09-10 09:02:03 -06001477 this_order);
Jens Axboe320ae512013-10-24 09:20:05 +01001478 if (page)
1479 break;
1480 if (!this_order--)
1481 break;
1482 if (order_to_size(this_order) < rq_size)
1483 break;
1484 } while (1);
1485
1486 if (!page)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001487 goto fail;
Jens Axboe320ae512013-10-24 09:20:05 +01001488
1489 page->private = this_order;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001490 list_add_tail(&page->lru, &tags->page_list);
Jens Axboe320ae512013-10-24 09:20:05 +01001491
1492 p = page_address(page);
1493 entries_per_page = order_to_size(this_order) / rq_size;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001494 to_do = min(entries_per_page, set->queue_depth - i);
Jens Axboe320ae512013-10-24 09:20:05 +01001495 left -= to_do * rq_size;
1496 for (j = 0; j < to_do; j++) {
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001497 tags->rqs[i] = p;
1498 if (set->ops->init_request) {
1499 if (set->ops->init_request(set->driver_data,
1500 tags->rqs[i], hctx_idx, i,
Jens Axboea5164402014-09-10 09:02:03 -06001501 set->numa_node)) {
1502 tags->rqs[i] = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001503 goto fail;
Jens Axboea5164402014-09-10 09:02:03 -06001504 }
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001505 }
1506
Jens Axboe320ae512013-10-24 09:20:05 +01001507 p += rq_size;
1508 i++;
1509 }
1510 }
1511
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001512 return tags;
Jens Axboe320ae512013-10-24 09:20:05 +01001513
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001514fail:
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001515 blk_mq_free_rq_map(set, tags, hctx_idx);
1516 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001517}
1518
Jens Axboe1429d7c2014-05-19 09:23:55 -06001519static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1520{
1521 kfree(bitmap->map);
1522}
1523
1524static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1525{
1526 unsigned int bpw = 8, total, num_maps, i;
1527
1528 bitmap->bits_per_word = bpw;
1529
1530 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1531 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1532 GFP_KERNEL, node);
1533 if (!bitmap->map)
1534 return -ENOMEM;
1535
Jens Axboe1429d7c2014-05-19 09:23:55 -06001536 total = nr_cpu_ids;
1537 for (i = 0; i < num_maps; i++) {
1538 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1539 total -= bitmap->map[i].depth;
1540 }
1541
1542 return 0;
1543}
1544
Jens Axboe484b4062014-05-21 14:01:15 -06001545static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1546{
1547 struct request_queue *q = hctx->queue;
1548 struct blk_mq_ctx *ctx;
1549 LIST_HEAD(tmp);
1550
1551 /*
1552 * Move ctx entries to new CPU, if this one is going away.
1553 */
1554 ctx = __blk_mq_get_ctx(q, cpu);
1555
1556 spin_lock(&ctx->lock);
1557 if (!list_empty(&ctx->rq_list)) {
1558 list_splice_init(&ctx->rq_list, &tmp);
1559 blk_mq_hctx_clear_pending(hctx, ctx);
1560 }
1561 spin_unlock(&ctx->lock);
1562
1563 if (list_empty(&tmp))
1564 return NOTIFY_OK;
1565
1566 ctx = blk_mq_get_ctx(q);
1567 spin_lock(&ctx->lock);
1568
1569 while (!list_empty(&tmp)) {
1570 struct request *rq;
1571
1572 rq = list_first_entry(&tmp, struct request, queuelist);
1573 rq->mq_ctx = ctx;
1574 list_move_tail(&rq->queuelist, &ctx->rq_list);
1575 }
1576
1577 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1578 blk_mq_hctx_mark_pending(hctx, ctx);
1579
1580 spin_unlock(&ctx->lock);
1581
1582 blk_mq_run_hw_queue(hctx, true);
1583 blk_mq_put_ctx(ctx);
1584 return NOTIFY_OK;
1585}
1586
Jens Axboe484b4062014-05-21 14:01:15 -06001587static int blk_mq_hctx_notify(void *data, unsigned long action,
1588 unsigned int cpu)
1589{
1590 struct blk_mq_hw_ctx *hctx = data;
1591
1592 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1593 return blk_mq_hctx_cpu_offline(hctx, cpu);
Ming Lei2a34c082015-04-21 10:00:20 +08001594
1595 /*
1596 * In case of CPU online, tags may be reallocated
1597 * in blk_mq_map_swqueue() after mapping is updated.
1598 */
Jens Axboe484b4062014-05-21 14:01:15 -06001599
1600 return NOTIFY_OK;
1601}
1602
Ming Leic3b4afc2015-06-04 22:25:04 +08001603/* hctx->ctxs will be freed in queue's release handler */
Ming Lei08e98fc2014-09-25 23:23:38 +08001604static void blk_mq_exit_hctx(struct request_queue *q,
1605 struct blk_mq_tag_set *set,
1606 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1607{
Ming Leif70ced02014-09-25 23:23:47 +08001608 unsigned flush_start_tag = set->queue_depth;
1609
Ming Lei08e98fc2014-09-25 23:23:38 +08001610 blk_mq_tag_idle(hctx);
1611
Ming Leif70ced02014-09-25 23:23:47 +08001612 if (set->ops->exit_request)
1613 set->ops->exit_request(set->driver_data,
1614 hctx->fq->flush_rq, hctx_idx,
1615 flush_start_tag + hctx_idx);
1616
Ming Lei08e98fc2014-09-25 23:23:38 +08001617 if (set->ops->exit_hctx)
1618 set->ops->exit_hctx(hctx, hctx_idx);
1619
1620 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
Ming Leif70ced02014-09-25 23:23:47 +08001621 blk_free_flush_queue(hctx->fq);
Ming Lei08e98fc2014-09-25 23:23:38 +08001622 blk_mq_free_bitmap(&hctx->ctx_map);
1623}
1624
Ming Lei624dbe42014-05-27 23:35:13 +08001625static void blk_mq_exit_hw_queues(struct request_queue *q,
1626 struct blk_mq_tag_set *set, int nr_queue)
1627{
1628 struct blk_mq_hw_ctx *hctx;
1629 unsigned int i;
1630
1631 queue_for_each_hw_ctx(q, hctx, i) {
1632 if (i == nr_queue)
1633 break;
Ming Lei08e98fc2014-09-25 23:23:38 +08001634 blk_mq_exit_hctx(q, set, hctx, i);
Ming Lei624dbe42014-05-27 23:35:13 +08001635 }
Ming Lei624dbe42014-05-27 23:35:13 +08001636}
1637
1638static void blk_mq_free_hw_queues(struct request_queue *q,
1639 struct blk_mq_tag_set *set)
1640{
1641 struct blk_mq_hw_ctx *hctx;
1642 unsigned int i;
1643
Ming Leie09aae72015-01-29 20:17:27 +08001644 queue_for_each_hw_ctx(q, hctx, i)
Ming Lei624dbe42014-05-27 23:35:13 +08001645 free_cpumask_var(hctx->cpumask);
Ming Lei624dbe42014-05-27 23:35:13 +08001646}
1647
Ming Lei08e98fc2014-09-25 23:23:38 +08001648static int blk_mq_init_hctx(struct request_queue *q,
1649 struct blk_mq_tag_set *set,
1650 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1651{
1652 int node;
Ming Leif70ced02014-09-25 23:23:47 +08001653 unsigned flush_start_tag = set->queue_depth;
Ming Lei08e98fc2014-09-25 23:23:38 +08001654
1655 node = hctx->numa_node;
1656 if (node == NUMA_NO_NODE)
1657 node = hctx->numa_node = set->numa_node;
1658
1659 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1660 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1661 spin_lock_init(&hctx->lock);
1662 INIT_LIST_HEAD(&hctx->dispatch);
1663 hctx->queue = q;
1664 hctx->queue_num = hctx_idx;
1665 hctx->flags = set->flags;
Ming Lei08e98fc2014-09-25 23:23:38 +08001666
1667 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1668 blk_mq_hctx_notify, hctx);
1669 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1670
1671 hctx->tags = set->tags[hctx_idx];
1672
1673 /*
1674 * Allocate space for all possible cpus to avoid allocation at
1675 * runtime
1676 */
1677 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1678 GFP_KERNEL, node);
1679 if (!hctx->ctxs)
1680 goto unregister_cpu_notifier;
1681
1682 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1683 goto free_ctxs;
1684
1685 hctx->nr_ctx = 0;
1686
1687 if (set->ops->init_hctx &&
1688 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1689 goto free_bitmap;
1690
Ming Leif70ced02014-09-25 23:23:47 +08001691 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1692 if (!hctx->fq)
1693 goto exit_hctx;
1694
1695 if (set->ops->init_request &&
1696 set->ops->init_request(set->driver_data,
1697 hctx->fq->flush_rq, hctx_idx,
1698 flush_start_tag + hctx_idx, node))
1699 goto free_fq;
1700
Ming Lei08e98fc2014-09-25 23:23:38 +08001701 return 0;
1702
Ming Leif70ced02014-09-25 23:23:47 +08001703 free_fq:
1704 kfree(hctx->fq);
1705 exit_hctx:
1706 if (set->ops->exit_hctx)
1707 set->ops->exit_hctx(hctx, hctx_idx);
Ming Lei08e98fc2014-09-25 23:23:38 +08001708 free_bitmap:
1709 blk_mq_free_bitmap(&hctx->ctx_map);
1710 free_ctxs:
1711 kfree(hctx->ctxs);
1712 unregister_cpu_notifier:
1713 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1714
1715 return -1;
1716}
1717
Jens Axboe320ae512013-10-24 09:20:05 +01001718static int blk_mq_init_hw_queues(struct request_queue *q,
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001719 struct blk_mq_tag_set *set)
Jens Axboe320ae512013-10-24 09:20:05 +01001720{
1721 struct blk_mq_hw_ctx *hctx;
Ming Lei624dbe42014-05-27 23:35:13 +08001722 unsigned int i;
Jens Axboe320ae512013-10-24 09:20:05 +01001723
1724 /*
1725 * Initialize hardware queues
1726 */
1727 queue_for_each_hw_ctx(q, hctx, i) {
Ming Lei08e98fc2014-09-25 23:23:38 +08001728 if (blk_mq_init_hctx(q, set, hctx, i))
Jens Axboe320ae512013-10-24 09:20:05 +01001729 break;
1730 }
1731
1732 if (i == q->nr_hw_queues)
1733 return 0;
1734
1735 /*
1736 * Init failed
1737 */
Ming Lei624dbe42014-05-27 23:35:13 +08001738 blk_mq_exit_hw_queues(q, set, i);
Jens Axboe320ae512013-10-24 09:20:05 +01001739
1740 return 1;
1741}
1742
1743static void blk_mq_init_cpu_queues(struct request_queue *q,
1744 unsigned int nr_hw_queues)
1745{
1746 unsigned int i;
1747
1748 for_each_possible_cpu(i) {
1749 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1750 struct blk_mq_hw_ctx *hctx;
1751
1752 memset(__ctx, 0, sizeof(*__ctx));
1753 __ctx->cpu = i;
1754 spin_lock_init(&__ctx->lock);
1755 INIT_LIST_HEAD(&__ctx->rq_list);
1756 __ctx->queue = q;
1757
1758 /* If the cpu isn't online, the cpu is mapped to first hctx */
Jens Axboe320ae512013-10-24 09:20:05 +01001759 if (!cpu_online(i))
1760 continue;
1761
Jens Axboee4043dc2014-04-09 10:18:23 -06001762 hctx = q->mq_ops->map_queue(q, i);
Jens Axboee4043dc2014-04-09 10:18:23 -06001763
Jens Axboe320ae512013-10-24 09:20:05 +01001764 /*
1765 * Set local node, IFF we have more than one hw queue. If
1766 * not, we remain on the home node of the device
1767 */
1768 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1769 hctx->numa_node = cpu_to_node(i);
1770 }
1771}
1772
1773static void blk_mq_map_swqueue(struct request_queue *q)
1774{
1775 unsigned int i;
1776 struct blk_mq_hw_ctx *hctx;
1777 struct blk_mq_ctx *ctx;
Ming Lei2a34c082015-04-21 10:00:20 +08001778 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01001779
1780 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboee4043dc2014-04-09 10:18:23 -06001781 cpumask_clear(hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001782 hctx->nr_ctx = 0;
1783 }
1784
1785 /*
1786 * Map software to hardware queues
1787 */
1788 queue_for_each_ctx(q, ctx, i) {
1789 /* If the cpu isn't online, the cpu is mapped to first hctx */
Jens Axboee4043dc2014-04-09 10:18:23 -06001790 if (!cpu_online(i))
1791 continue;
1792
Jens Axboe320ae512013-10-24 09:20:05 +01001793 hctx = q->mq_ops->map_queue(q, i);
Jens Axboee4043dc2014-04-09 10:18:23 -06001794 cpumask_set_cpu(i, hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001795 ctx->index_hw = hctx->nr_ctx;
1796 hctx->ctxs[hctx->nr_ctx++] = ctx;
1797 }
Jens Axboe506e9312014-05-07 10:26:44 -06001798
1799 queue_for_each_hw_ctx(q, hctx, i) {
Chong Yuan889fa312015-04-15 11:39:29 -06001800 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1801
Jens Axboe484b4062014-05-21 14:01:15 -06001802 /*
Jens Axboea68aafa2014-08-15 13:19:15 -06001803 * If no software queues are mapped to this hardware queue,
1804 * disable it and free the request entries.
Jens Axboe484b4062014-05-21 14:01:15 -06001805 */
1806 if (!hctx->nr_ctx) {
Jens Axboe484b4062014-05-21 14:01:15 -06001807 if (set->tags[i]) {
1808 blk_mq_free_rq_map(set, set->tags[i], i);
1809 set->tags[i] = NULL;
Jens Axboe484b4062014-05-21 14:01:15 -06001810 }
Ming Lei2a34c082015-04-21 10:00:20 +08001811 hctx->tags = NULL;
Jens Axboe484b4062014-05-21 14:01:15 -06001812 continue;
1813 }
1814
Ming Lei2a34c082015-04-21 10:00:20 +08001815 /* unmapped hw queue can be remapped after CPU topo changed */
1816 if (!set->tags[i])
1817 set->tags[i] = blk_mq_init_rq_map(set, i);
1818 hctx->tags = set->tags[i];
1819 WARN_ON(!hctx->tags);
1820
Jens Axboe484b4062014-05-21 14:01:15 -06001821 /*
Chong Yuan889fa312015-04-15 11:39:29 -06001822 * Set the map size to the number of mapped software queues.
1823 * This is more accurate and more efficient than looping
1824 * over all possibly mapped software queues.
1825 */
Jens Axboe569fd0c2015-04-17 08:28:50 -06001826 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
Chong Yuan889fa312015-04-15 11:39:29 -06001827
1828 /*
Jens Axboe484b4062014-05-21 14:01:15 -06001829 * Initialize batch roundrobin counts
1830 */
Jens Axboe506e9312014-05-07 10:26:44 -06001831 hctx->next_cpu = cpumask_first(hctx->cpumask);
1832 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1833 }
Jens Axboe320ae512013-10-24 09:20:05 +01001834}
1835
Jens Axboe0d2602c2014-05-13 15:10:52 -06001836static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1837{
1838 struct blk_mq_hw_ctx *hctx;
1839 struct request_queue *q;
1840 bool shared;
1841 int i;
1842
1843 if (set->tag_list.next == set->tag_list.prev)
1844 shared = false;
1845 else
1846 shared = true;
1847
1848 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1849 blk_mq_freeze_queue(q);
1850
1851 queue_for_each_hw_ctx(q, hctx, i) {
1852 if (shared)
1853 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1854 else
1855 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1856 }
1857 blk_mq_unfreeze_queue(q);
1858 }
1859}
1860
1861static void blk_mq_del_queue_tag_set(struct request_queue *q)
1862{
1863 struct blk_mq_tag_set *set = q->tag_set;
1864
Jens Axboe0d2602c2014-05-13 15:10:52 -06001865 mutex_lock(&set->tag_list_lock);
1866 list_del_init(&q->tag_set_list);
1867 blk_mq_update_tag_set_depth(set);
1868 mutex_unlock(&set->tag_list_lock);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001869}
1870
1871static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1872 struct request_queue *q)
1873{
1874 q->tag_set = set;
1875
1876 mutex_lock(&set->tag_list_lock);
1877 list_add_tail(&q->tag_set_list, &set->tag_list);
1878 blk_mq_update_tag_set_depth(set);
1879 mutex_unlock(&set->tag_list_lock);
1880}
1881
Ming Leie09aae72015-01-29 20:17:27 +08001882/*
1883 * It is the actual release handler for mq, but we do it from
1884 * request queue's release handler for avoiding use-after-free
1885 * and headache because q->mq_kobj shouldn't have been introduced,
1886 * but we can't group ctx/kctx kobj without it.
1887 */
1888void blk_mq_release(struct request_queue *q)
1889{
1890 struct blk_mq_hw_ctx *hctx;
1891 unsigned int i;
1892
1893 /* hctx kobj stays in hctx */
Ming Leic3b4afc2015-06-04 22:25:04 +08001894 queue_for_each_hw_ctx(q, hctx, i) {
1895 if (!hctx)
1896 continue;
1897 kfree(hctx->ctxs);
Ming Leie09aae72015-01-29 20:17:27 +08001898 kfree(hctx);
Ming Leic3b4afc2015-06-04 22:25:04 +08001899 }
Ming Leie09aae72015-01-29 20:17:27 +08001900
1901 kfree(q->queue_hw_ctx);
1902
1903 /* ctx kobj stays in queue_ctx */
1904 free_percpu(q->queue_ctx);
1905}
1906
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001907struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
Jens Axboe320ae512013-10-24 09:20:05 +01001908{
Mike Snitzerb62c21b2015-03-12 23:56:02 -04001909 struct request_queue *uninit_q, *q;
1910
1911 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1912 if (!uninit_q)
1913 return ERR_PTR(-ENOMEM);
1914
1915 q = blk_mq_init_allocated_queue(set, uninit_q);
1916 if (IS_ERR(q))
1917 blk_cleanup_queue(uninit_q);
1918
1919 return q;
1920}
1921EXPORT_SYMBOL(blk_mq_init_queue);
1922
1923struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1924 struct request_queue *q)
1925{
Jens Axboe320ae512013-10-24 09:20:05 +01001926 struct blk_mq_hw_ctx **hctxs;
Ming Leie6cdb092014-06-03 11:24:06 +08001927 struct blk_mq_ctx __percpu *ctx;
Jens Axboef14bbe72014-05-27 12:06:53 -06001928 unsigned int *map;
Jens Axboe320ae512013-10-24 09:20:05 +01001929 int i;
1930
Jens Axboe320ae512013-10-24 09:20:05 +01001931 ctx = alloc_percpu(struct blk_mq_ctx);
1932 if (!ctx)
1933 return ERR_PTR(-ENOMEM);
1934
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001935 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1936 set->numa_node);
Jens Axboe320ae512013-10-24 09:20:05 +01001937
1938 if (!hctxs)
1939 goto err_percpu;
1940
Jens Axboef14bbe72014-05-27 12:06:53 -06001941 map = blk_mq_make_queue_map(set);
1942 if (!map)
1943 goto err_map;
1944
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001945 for (i = 0; i < set->nr_hw_queues; i++) {
Jens Axboef14bbe72014-05-27 12:06:53 -06001946 int node = blk_mq_hw_queue_to_node(map, i);
1947
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02001948 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1949 GFP_KERNEL, node);
Jens Axboe320ae512013-10-24 09:20:05 +01001950 if (!hctxs[i])
1951 goto err_hctxs;
1952
Jens Axboea86073e2014-10-13 15:41:54 -06001953 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1954 node))
Jens Axboee4043dc2014-04-09 10:18:23 -06001955 goto err_hctxs;
1956
Jens Axboe0d2602c2014-05-13 15:10:52 -06001957 atomic_set(&hctxs[i]->nr_active, 0);
Jens Axboef14bbe72014-05-27 12:06:53 -06001958 hctxs[i]->numa_node = node;
Jens Axboe320ae512013-10-24 09:20:05 +01001959 hctxs[i]->queue_num = i;
1960 }
1961
Tejun Heo17497ac2014-09-24 13:31:50 -04001962 /*
1963 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1964 * See blk_register_queue() for details.
1965 */
Tejun Heoa34375e2014-09-08 09:51:30 +09001966 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
Tejun Heo17497ac2014-09-24 13:31:50 -04001967 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
Mike Snitzerb62c21b2015-03-12 23:56:02 -04001968 goto err_hctxs;
Ming Lei3d2936f2014-05-27 23:35:14 +08001969
Jens Axboe320ae512013-10-24 09:20:05 +01001970 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
Wei Fangc76cbbc2015-03-30 09:07:00 -06001971 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30000);
Jens Axboe320ae512013-10-24 09:20:05 +01001972
1973 q->nr_queues = nr_cpu_ids;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001974 q->nr_hw_queues = set->nr_hw_queues;
Jens Axboef14bbe72014-05-27 12:06:53 -06001975 q->mq_map = map;
Jens Axboe320ae512013-10-24 09:20:05 +01001976
1977 q->queue_ctx = ctx;
1978 q->queue_hw_ctx = hctxs;
1979
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001980 q->mq_ops = set->ops;
Jens Axboe94eddfb2013-11-19 09:25:07 -07001981 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
Jens Axboe320ae512013-10-24 09:20:05 +01001982
Jens Axboe05f1dd52014-05-29 09:53:32 -06001983 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1984 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1985
Christoph Hellwig1be036e2014-02-07 10:22:39 -08001986 q->sg_reserved_size = INT_MAX;
1987
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001988 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1989 INIT_LIST_HEAD(&q->requeue_list);
1990 spin_lock_init(&q->requeue_lock);
1991
Jens Axboe07068d52014-05-22 10:40:51 -06001992 if (q->nr_hw_queues > 1)
1993 blk_queue_make_request(q, blk_mq_make_request);
1994 else
1995 blk_queue_make_request(q, blk_sq_make_request);
1996
Jens Axboeeba71762014-05-20 15:17:27 -06001997 /*
1998 * Do this after blk_queue_make_request() overrides it...
1999 */
2000 q->nr_requests = set->queue_depth;
2001
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002002 if (set->ops->complete)
2003 blk_queue_softirq_done(q, set->ops->complete);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08002004
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002005 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +01002006
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002007 if (blk_mq_init_hw_queues(q, set))
Mike Snitzerb62c21b2015-03-12 23:56:02 -04002008 goto err_hctxs;
Christoph Hellwig18741982014-02-10 09:29:00 -07002009
Jens Axboe320ae512013-10-24 09:20:05 +01002010 mutex_lock(&all_q_mutex);
2011 list_add_tail(&q->all_q_node, &all_q_list);
2012 mutex_unlock(&all_q_mutex);
2013
Jens Axboe0d2602c2014-05-13 15:10:52 -06002014 blk_mq_add_queue_tag_set(set, q);
2015
Jens Axboe484b4062014-05-21 14:01:15 -06002016 blk_mq_map_swqueue(q);
2017
Jens Axboe320ae512013-10-24 09:20:05 +01002018 return q;
Christoph Hellwig18741982014-02-10 09:29:00 -07002019
Jens Axboe320ae512013-10-24 09:20:05 +01002020err_hctxs:
Jens Axboef14bbe72014-05-27 12:06:53 -06002021 kfree(map);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002022 for (i = 0; i < set->nr_hw_queues; i++) {
Jens Axboe320ae512013-10-24 09:20:05 +01002023 if (!hctxs[i])
2024 break;
Jens Axboee4043dc2014-04-09 10:18:23 -06002025 free_cpumask_var(hctxs[i]->cpumask);
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02002026 kfree(hctxs[i]);
Jens Axboe320ae512013-10-24 09:20:05 +01002027 }
Jens Axboef14bbe72014-05-27 12:06:53 -06002028err_map:
Jens Axboe320ae512013-10-24 09:20:05 +01002029 kfree(hctxs);
2030err_percpu:
2031 free_percpu(ctx);
2032 return ERR_PTR(-ENOMEM);
2033}
Mike Snitzerb62c21b2015-03-12 23:56:02 -04002034EXPORT_SYMBOL(blk_mq_init_allocated_queue);
Jens Axboe320ae512013-10-24 09:20:05 +01002035
2036void blk_mq_free_queue(struct request_queue *q)
2037{
Ming Lei624dbe42014-05-27 23:35:13 +08002038 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01002039
Jens Axboe0d2602c2014-05-13 15:10:52 -06002040 blk_mq_del_queue_tag_set(q);
2041
Ming Lei624dbe42014-05-27 23:35:13 +08002042 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2043 blk_mq_free_hw_queues(q, set);
Jens Axboe320ae512013-10-24 09:20:05 +01002044
Tejun Heoadd703f2014-07-01 10:34:38 -06002045 percpu_ref_exit(&q->mq_usage_counter);
Ming Lei3d2936f2014-05-27 23:35:14 +08002046
Jens Axboe320ae512013-10-24 09:20:05 +01002047 kfree(q->mq_map);
2048
Jens Axboe320ae512013-10-24 09:20:05 +01002049 q->mq_map = NULL;
2050
2051 mutex_lock(&all_q_mutex);
2052 list_del_init(&q->all_q_node);
2053 mutex_unlock(&all_q_mutex);
2054}
Jens Axboe320ae512013-10-24 09:20:05 +01002055
2056/* Basically redo blk_mq_init_queue with queue frozen */
Paul Gortmakerf618ef72013-11-14 08:26:02 -07002057static void blk_mq_queue_reinit(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01002058{
Tejun Heof3af0202014-11-04 13:52:27 -05002059 WARN_ON_ONCE(!q->mq_freeze_depth);
Jens Axboe320ae512013-10-24 09:20:05 +01002060
Jens Axboe67aec142014-05-30 08:25:36 -06002061 blk_mq_sysfs_unregister(q);
2062
Jens Axboe320ae512013-10-24 09:20:05 +01002063 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2064
2065 /*
2066 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2067 * we should change hctx numa_node according to new topology (this
2068 * involves free and re-allocate memory, worthy doing?)
2069 */
2070
2071 blk_mq_map_swqueue(q);
2072
Jens Axboe67aec142014-05-30 08:25:36 -06002073 blk_mq_sysfs_register(q);
Jens Axboe320ae512013-10-24 09:20:05 +01002074}
2075
Paul Gortmakerf618ef72013-11-14 08:26:02 -07002076static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2077 unsigned long action, void *hcpu)
Jens Axboe320ae512013-10-24 09:20:05 +01002078{
2079 struct request_queue *q;
2080
2081 /*
Jens Axboe9fccfed2014-05-08 14:50:19 -06002082 * Before new mappings are established, hotadded cpu might already
2083 * start handling requests. This doesn't break anything as we map
2084 * offline CPUs to first hardware queue. We will re-init the queue
2085 * below to get optimal settings.
Jens Axboe320ae512013-10-24 09:20:05 +01002086 */
2087 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2088 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2089 return NOTIFY_OK;
2090
2091 mutex_lock(&all_q_mutex);
Tejun Heof3af0202014-11-04 13:52:27 -05002092
2093 /*
2094 * We need to freeze and reinit all existing queues. Freezing
2095 * involves synchronous wait for an RCU grace period and doing it
2096 * one by one may take a long time. Start freezing all queues in
2097 * one swoop and then wait for the completions so that freezing can
2098 * take place in parallel.
2099 */
2100 list_for_each_entry(q, &all_q_list, all_q_node)
2101 blk_mq_freeze_queue_start(q);
Ming Leif054b562015-04-21 10:00:19 +08002102 list_for_each_entry(q, &all_q_list, all_q_node) {
Tejun Heof3af0202014-11-04 13:52:27 -05002103 blk_mq_freeze_queue_wait(q);
2104
Ming Leif054b562015-04-21 10:00:19 +08002105 /*
2106 * timeout handler can't touch hw queue during the
2107 * reinitialization
2108 */
2109 del_timer_sync(&q->timeout);
2110 }
2111
Jens Axboe320ae512013-10-24 09:20:05 +01002112 list_for_each_entry(q, &all_q_list, all_q_node)
2113 blk_mq_queue_reinit(q);
Tejun Heof3af0202014-11-04 13:52:27 -05002114
2115 list_for_each_entry(q, &all_q_list, all_q_node)
2116 blk_mq_unfreeze_queue(q);
2117
Jens Axboe320ae512013-10-24 09:20:05 +01002118 mutex_unlock(&all_q_mutex);
2119 return NOTIFY_OK;
2120}
2121
Jens Axboea5164402014-09-10 09:02:03 -06002122static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2123{
2124 int i;
2125
2126 for (i = 0; i < set->nr_hw_queues; i++) {
2127 set->tags[i] = blk_mq_init_rq_map(set, i);
2128 if (!set->tags[i])
2129 goto out_unwind;
2130 }
2131
2132 return 0;
2133
2134out_unwind:
2135 while (--i >= 0)
2136 blk_mq_free_rq_map(set, set->tags[i], i);
2137
Jens Axboea5164402014-09-10 09:02:03 -06002138 return -ENOMEM;
2139}
2140
2141/*
2142 * Allocate the request maps associated with this tag_set. Note that this
2143 * may reduce the depth asked for, if memory is tight. set->queue_depth
2144 * will be updated to reflect the allocated depth.
2145 */
2146static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2147{
2148 unsigned int depth;
2149 int err;
2150
2151 depth = set->queue_depth;
2152 do {
2153 err = __blk_mq_alloc_rq_maps(set);
2154 if (!err)
2155 break;
2156
2157 set->queue_depth >>= 1;
2158 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2159 err = -ENOMEM;
2160 break;
2161 }
2162 } while (set->queue_depth);
2163
2164 if (!set->queue_depth || err) {
2165 pr_err("blk-mq: failed to allocate request map\n");
2166 return -ENOMEM;
2167 }
2168
2169 if (depth != set->queue_depth)
2170 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2171 depth, set->queue_depth);
2172
2173 return 0;
2174}
2175
Jens Axboea4391c62014-06-05 15:21:56 -06002176/*
2177 * Alloc a tag set to be associated with one or more request queues.
2178 * May fail with EINVAL for various error conditions. May adjust the
2179 * requested depth down, if if it too large. In that case, the set
2180 * value will be stored in set->queue_depth.
2181 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002182int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2183{
Bart Van Assche205fb5f2014-10-30 14:45:11 +01002184 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2185
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002186 if (!set->nr_hw_queues)
2187 return -EINVAL;
Jens Axboea4391c62014-06-05 15:21:56 -06002188 if (!set->queue_depth)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002189 return -EINVAL;
2190 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2191 return -EINVAL;
2192
Xiaoguang Wangf9018ac2015-03-30 13:19:14 +08002193 if (!set->ops->queue_rq || !set->ops->map_queue)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002194 return -EINVAL;
2195
Jens Axboea4391c62014-06-05 15:21:56 -06002196 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2197 pr_info("blk-mq: reduced tag depth to %u\n",
2198 BLK_MQ_MAX_DEPTH);
2199 set->queue_depth = BLK_MQ_MAX_DEPTH;
2200 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002201
Shaohua Li6637fad2014-11-30 16:00:58 -08002202 /*
2203 * If a crashdump is active, then we are potentially in a very
2204 * memory constrained environment. Limit us to 1 queue and
2205 * 64 tags to prevent using too much memory.
2206 */
2207 if (is_kdump_kernel()) {
2208 set->nr_hw_queues = 1;
2209 set->queue_depth = min(64U, set->queue_depth);
2210 }
2211
Ming Lei48479002014-04-19 18:00:17 +08002212 set->tags = kmalloc_node(set->nr_hw_queues *
2213 sizeof(struct blk_mq_tags *),
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002214 GFP_KERNEL, set->numa_node);
2215 if (!set->tags)
Jens Axboea5164402014-09-10 09:02:03 -06002216 return -ENOMEM;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002217
Jens Axboea5164402014-09-10 09:02:03 -06002218 if (blk_mq_alloc_rq_maps(set))
2219 goto enomem;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002220
Jens Axboe0d2602c2014-05-13 15:10:52 -06002221 mutex_init(&set->tag_list_lock);
2222 INIT_LIST_HEAD(&set->tag_list);
2223
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002224 return 0;
Jens Axboea5164402014-09-10 09:02:03 -06002225enomem:
Robert Elliott5676e7b2014-09-02 11:38:44 -05002226 kfree(set->tags);
2227 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002228 return -ENOMEM;
2229}
2230EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2231
2232void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2233{
2234 int i;
2235
Jens Axboe484b4062014-05-21 14:01:15 -06002236 for (i = 0; i < set->nr_hw_queues; i++) {
2237 if (set->tags[i])
2238 blk_mq_free_rq_map(set, set->tags[i], i);
2239 }
2240
Ming Lei981bd182014-04-24 00:07:34 +08002241 kfree(set->tags);
Robert Elliott5676e7b2014-09-02 11:38:44 -05002242 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002243}
2244EXPORT_SYMBOL(blk_mq_free_tag_set);
2245
Jens Axboee3a2b3f2014-05-20 11:49:02 -06002246int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2247{
2248 struct blk_mq_tag_set *set = q->tag_set;
2249 struct blk_mq_hw_ctx *hctx;
2250 int i, ret;
2251
2252 if (!set || nr > set->queue_depth)
2253 return -EINVAL;
2254
2255 ret = 0;
2256 queue_for_each_hw_ctx(q, hctx, i) {
2257 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2258 if (ret)
2259 break;
2260 }
2261
2262 if (!ret)
2263 q->nr_requests = nr;
2264
2265 return ret;
2266}
2267
Jens Axboe676141e2014-03-20 13:29:18 -06002268void blk_mq_disable_hotplug(void)
2269{
2270 mutex_lock(&all_q_mutex);
2271}
2272
2273void blk_mq_enable_hotplug(void)
2274{
2275 mutex_unlock(&all_q_mutex);
2276}
2277
Jens Axboe320ae512013-10-24 09:20:05 +01002278static int __init blk_mq_init(void)
2279{
Jens Axboe320ae512013-10-24 09:20:05 +01002280 blk_mq_cpu_init();
2281
Tejun Heoadd703f2014-07-01 10:34:38 -06002282 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
Jens Axboe320ae512013-10-24 09:20:05 +01002283
2284 return 0;
2285}
2286subsys_initcall(blk_mq_init);