blob: ade8a2d1b0aa8600ad31413b59db37392628bffc [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 {
680 queue_for_each_hw_ctx(q, hctx, i)
681 blk_mq_tag_idle(hctx);
682 }
Jens Axboe320ae512013-10-24 09:20:05 +0100683}
684
685/*
686 * Reverse check our software queue for entries that we could potentially
687 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
688 * too much time checking for merges.
689 */
690static bool blk_mq_attempt_merge(struct request_queue *q,
691 struct blk_mq_ctx *ctx, struct bio *bio)
692{
693 struct request *rq;
694 int checked = 8;
695
696 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
697 int el_ret;
698
699 if (!checked--)
700 break;
701
702 if (!blk_rq_merge_ok(rq, bio))
703 continue;
704
705 el_ret = blk_try_merge(rq, bio);
706 if (el_ret == ELEVATOR_BACK_MERGE) {
707 if (bio_attempt_back_merge(q, rq, bio)) {
708 ctx->rq_merged++;
709 return true;
710 }
711 break;
712 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
713 if (bio_attempt_front_merge(q, rq, bio)) {
714 ctx->rq_merged++;
715 return true;
716 }
717 break;
718 }
719 }
720
721 return false;
722}
723
Jens Axboe320ae512013-10-24 09:20:05 +0100724/*
Jens Axboe1429d7c2014-05-19 09:23:55 -0600725 * Process software queues that have been marked busy, splicing them
726 * to the for-dispatch
727 */
728static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
729{
730 struct blk_mq_ctx *ctx;
731 int i;
732
Jens Axboe569fd0c2015-04-17 08:28:50 -0600733 for (i = 0; i < hctx->ctx_map.size; i++) {
Jens Axboe1429d7c2014-05-19 09:23:55 -0600734 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
735 unsigned int off, bit;
736
737 if (!bm->word)
738 continue;
739
740 bit = 0;
741 off = i * hctx->ctx_map.bits_per_word;
742 do {
743 bit = find_next_bit(&bm->word, bm->depth, bit);
744 if (bit >= bm->depth)
745 break;
746
747 ctx = hctx->ctxs[bit + off];
748 clear_bit(bit, &bm->word);
749 spin_lock(&ctx->lock);
750 list_splice_tail_init(&ctx->rq_list, list);
751 spin_unlock(&ctx->lock);
752
753 bit++;
754 } while (1);
755 }
756}
757
758/*
Jens Axboe320ae512013-10-24 09:20:05 +0100759 * Run this hardware queue, pulling any software queues mapped to it in.
760 * Note that this function currently has various problems around ordering
761 * of IO. In particular, we'd like FIFO behaviour on handling existing
762 * items on the hctx->dispatch list. Ignore that for now.
763 */
764static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
765{
766 struct request_queue *q = hctx->queue;
Jens Axboe320ae512013-10-24 09:20:05 +0100767 struct request *rq;
768 LIST_HEAD(rq_list);
Jens Axboe74c45052014-10-29 11:14:52 -0600769 LIST_HEAD(driver_list);
770 struct list_head *dptr;
Jens Axboe1429d7c2014-05-19 09:23:55 -0600771 int queued;
Jens Axboe320ae512013-10-24 09:20:05 +0100772
Jens Axboefd1270d2014-04-16 09:23:48 -0600773 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
Jens Axboee4043dc2014-04-09 10:18:23 -0600774
Jens Axboe5d12f902014-03-19 15:25:02 -0600775 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
Jens Axboe320ae512013-10-24 09:20:05 +0100776 return;
777
778 hctx->run++;
779
780 /*
781 * Touch any software queue that has pending entries.
782 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600783 flush_busy_ctxs(hctx, &rq_list);
Jens Axboe320ae512013-10-24 09:20:05 +0100784
785 /*
786 * If we have previous entries on our dispatch list, grab them
787 * and stuff them at the front for more fair dispatch.
788 */
789 if (!list_empty_careful(&hctx->dispatch)) {
790 spin_lock(&hctx->lock);
791 if (!list_empty(&hctx->dispatch))
792 list_splice_init(&hctx->dispatch, &rq_list);
793 spin_unlock(&hctx->lock);
794 }
795
796 /*
Jens Axboe74c45052014-10-29 11:14:52 -0600797 * Start off with dptr being NULL, so we start the first request
798 * immediately, even if we have more pending.
799 */
800 dptr = NULL;
801
802 /*
Jens Axboe320ae512013-10-24 09:20:05 +0100803 * Now process all the entries, sending them to the driver.
804 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600805 queued = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100806 while (!list_empty(&rq_list)) {
Jens Axboe74c45052014-10-29 11:14:52 -0600807 struct blk_mq_queue_data bd;
Jens Axboe320ae512013-10-24 09:20:05 +0100808 int ret;
809
810 rq = list_first_entry(&rq_list, struct request, queuelist);
811 list_del_init(&rq->queuelist);
Jens Axboe320ae512013-10-24 09:20:05 +0100812
Jens Axboe74c45052014-10-29 11:14:52 -0600813 bd.rq = rq;
814 bd.list = dptr;
815 bd.last = list_empty(&rq_list);
816
817 ret = q->mq_ops->queue_rq(hctx, &bd);
Jens Axboe320ae512013-10-24 09:20:05 +0100818 switch (ret) {
819 case BLK_MQ_RQ_QUEUE_OK:
820 queued++;
821 continue;
822 case BLK_MQ_RQ_QUEUE_BUSY:
Jens Axboe320ae512013-10-24 09:20:05 +0100823 list_add(&rq->queuelist, &rq_list);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200824 __blk_mq_requeue_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100825 break;
826 default:
827 pr_err("blk-mq: bad return on queue: %d\n", ret);
Jens Axboe320ae512013-10-24 09:20:05 +0100828 case BLK_MQ_RQ_QUEUE_ERROR:
Christoph Hellwig1e93b8c2014-02-11 08:27:13 -0800829 rq->errors = -EIO;
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700830 blk_mq_end_request(rq, rq->errors);
Jens Axboe320ae512013-10-24 09:20:05 +0100831 break;
832 }
833
834 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
835 break;
Jens Axboe74c45052014-10-29 11:14:52 -0600836
837 /*
838 * We've done the first request. If we have more than 1
839 * left in the list, set dptr to defer issue.
840 */
841 if (!dptr && rq_list.next != rq_list.prev)
842 dptr = &driver_list;
Jens Axboe320ae512013-10-24 09:20:05 +0100843 }
844
845 if (!queued)
846 hctx->dispatched[0]++;
847 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
848 hctx->dispatched[ilog2(queued) + 1]++;
849
850 /*
851 * Any items that need requeuing? Stuff them into hctx->dispatch,
852 * that is where we will continue on next queue run.
853 */
854 if (!list_empty(&rq_list)) {
855 spin_lock(&hctx->lock);
856 list_splice(&rq_list, &hctx->dispatch);
857 spin_unlock(&hctx->lock);
858 }
859}
860
Jens Axboe506e9312014-05-07 10:26:44 -0600861/*
862 * It'd be great if the workqueue API had a way to pass
863 * in a mask and had some smarts for more clever placement.
864 * For now we just round-robin here, switching for every
865 * BLK_MQ_CPU_WORK_BATCH queued items.
866 */
867static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
868{
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100869 if (hctx->queue->nr_hw_queues == 1)
870 return WORK_CPU_UNBOUND;
Jens Axboe506e9312014-05-07 10:26:44 -0600871
872 if (--hctx->next_cpu_batch <= 0) {
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100873 int cpu = hctx->next_cpu, next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600874
875 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
876 if (next_cpu >= nr_cpu_ids)
877 next_cpu = cpumask_first(hctx->cpumask);
878
879 hctx->next_cpu = next_cpu;
880 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100881
882 return cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600883 }
884
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100885 return hctx->next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600886}
887
Jens Axboe320ae512013-10-24 09:20:05 +0100888void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
889{
Ming Lei19c66e52014-12-03 19:38:04 +0800890 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
891 !blk_mq_hw_queue_mapped(hctx)))
Jens Axboe320ae512013-10-24 09:20:05 +0100892 return;
893
Paolo Bonzini398205b2014-11-07 23:03:59 +0100894 if (!async) {
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100895 int cpu = get_cpu();
896 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
Paolo Bonzini398205b2014-11-07 23:03:59 +0100897 __blk_mq_run_hw_queue(hctx);
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100898 put_cpu();
Paolo Bonzini398205b2014-11-07 23:03:59 +0100899 return;
900 }
Jens Axboee4043dc2014-04-09 10:18:23 -0600901
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100902 put_cpu();
Jens Axboee4043dc2014-04-09 10:18:23 -0600903 }
Paolo Bonzini398205b2014-11-07 23:03:59 +0100904
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100905 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
906 &hctx->run_work, 0);
Jens Axboe320ae512013-10-24 09:20:05 +0100907}
908
Mike Snitzerb94ec292015-03-11 23:56:38 -0400909void blk_mq_run_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +0100910{
911 struct blk_mq_hw_ctx *hctx;
912 int i;
913
914 queue_for_each_hw_ctx(q, hctx, i) {
915 if ((!blk_mq_hctx_has_pending(hctx) &&
916 list_empty_careful(&hctx->dispatch)) ||
Jens Axboe5d12f902014-03-19 15:25:02 -0600917 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
Jens Axboe320ae512013-10-24 09:20:05 +0100918 continue;
919
Mike Snitzerb94ec292015-03-11 23:56:38 -0400920 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +0100921 }
922}
Mike Snitzerb94ec292015-03-11 23:56:38 -0400923EXPORT_SYMBOL(blk_mq_run_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +0100924
925void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
926{
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600927 cancel_delayed_work(&hctx->run_work);
928 cancel_delayed_work(&hctx->delay_work);
Jens Axboe320ae512013-10-24 09:20:05 +0100929 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
930}
931EXPORT_SYMBOL(blk_mq_stop_hw_queue);
932
Christoph Hellwig280d45f2013-10-25 14:45:58 +0100933void blk_mq_stop_hw_queues(struct request_queue *q)
934{
935 struct blk_mq_hw_ctx *hctx;
936 int i;
937
938 queue_for_each_hw_ctx(q, hctx, i)
939 blk_mq_stop_hw_queue(hctx);
940}
941EXPORT_SYMBOL(blk_mq_stop_hw_queues);
942
Jens Axboe320ae512013-10-24 09:20:05 +0100943void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
944{
945 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboee4043dc2014-04-09 10:18:23 -0600946
Jens Axboe0ffbce82014-06-25 08:22:34 -0600947 blk_mq_run_hw_queue(hctx, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100948}
949EXPORT_SYMBOL(blk_mq_start_hw_queue);
950
Christoph Hellwig2f268552014-04-16 09:44:56 +0200951void blk_mq_start_hw_queues(struct request_queue *q)
952{
953 struct blk_mq_hw_ctx *hctx;
954 int i;
955
956 queue_for_each_hw_ctx(q, hctx, i)
957 blk_mq_start_hw_queue(hctx);
958}
959EXPORT_SYMBOL(blk_mq_start_hw_queues);
960
Christoph Hellwig1b4a3252014-04-16 09:44:54 +0200961void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +0100962{
963 struct blk_mq_hw_ctx *hctx;
964 int i;
965
966 queue_for_each_hw_ctx(q, hctx, i) {
967 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
968 continue;
969
970 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Christoph Hellwig1b4a3252014-04-16 09:44:54 +0200971 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +0100972 }
973}
974EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
975
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600976static void blk_mq_run_work_fn(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +0100977{
978 struct blk_mq_hw_ctx *hctx;
979
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600980 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
Jens Axboee4043dc2014-04-09 10:18:23 -0600981
Jens Axboe320ae512013-10-24 09:20:05 +0100982 __blk_mq_run_hw_queue(hctx);
983}
984
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600985static void blk_mq_delay_work_fn(struct work_struct *work)
986{
987 struct blk_mq_hw_ctx *hctx;
988
989 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
990
991 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
992 __blk_mq_run_hw_queue(hctx);
993}
994
995void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
996{
Ming Lei19c66e52014-12-03 19:38:04 +0800997 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
998 return;
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600999
Christoph Hellwigb657d7e2014-11-24 09:27:23 +01001000 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1001 &hctx->delay_work, msecs_to_jiffies(msecs));
Christoph Hellwig70f4db62014-04-16 10:48:08 -06001002}
1003EXPORT_SYMBOL(blk_mq_delay_queue);
1004
Jens Axboe320ae512013-10-24 09:20:05 +01001005static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig72a0a362014-02-07 10:22:36 -08001006 struct request *rq, bool at_head)
Jens Axboe320ae512013-10-24 09:20:05 +01001007{
1008 struct blk_mq_ctx *ctx = rq->mq_ctx;
1009
Jens Axboe01b983c2013-11-19 18:59:10 -07001010 trace_block_rq_insert(hctx->queue, rq);
1011
Christoph Hellwig72a0a362014-02-07 10:22:36 -08001012 if (at_head)
1013 list_add(&rq->queuelist, &ctx->rq_list);
1014 else
1015 list_add_tail(&rq->queuelist, &ctx->rq_list);
Jens Axboe4bb659b2014-05-09 09:36:49 -06001016
Jens Axboe320ae512013-10-24 09:20:05 +01001017 blk_mq_hctx_mark_pending(hctx, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001018}
1019
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001020void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1021 bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01001022{
1023 struct request_queue *q = rq->q;
1024 struct blk_mq_hw_ctx *hctx;
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001025 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001026
1027 current_ctx = blk_mq_get_ctx(q);
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001028 if (!cpu_online(ctx->cpu))
1029 rq->mq_ctx = ctx = current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001030
Jens Axboe320ae512013-10-24 09:20:05 +01001031 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1032
Christoph Hellwiga57a1782014-09-16 14:44:07 -07001033 spin_lock(&ctx->lock);
1034 __blk_mq_insert_request(hctx, rq, at_head);
1035 spin_unlock(&ctx->lock);
Jens Axboe320ae512013-10-24 09:20:05 +01001036
Jens Axboe320ae512013-10-24 09:20:05 +01001037 if (run_queue)
1038 blk_mq_run_hw_queue(hctx, async);
Jens Axboee4043dc2014-04-09 10:18:23 -06001039
1040 blk_mq_put_ctx(current_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001041}
1042
1043static void blk_mq_insert_requests(struct request_queue *q,
1044 struct blk_mq_ctx *ctx,
1045 struct list_head *list,
1046 int depth,
1047 bool from_schedule)
1048
1049{
1050 struct blk_mq_hw_ctx *hctx;
1051 struct blk_mq_ctx *current_ctx;
1052
1053 trace_block_unplug(q, depth, !from_schedule);
1054
1055 current_ctx = blk_mq_get_ctx(q);
1056
1057 if (!cpu_online(ctx->cpu))
1058 ctx = current_ctx;
1059 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1060
1061 /*
1062 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1063 * offline now
1064 */
1065 spin_lock(&ctx->lock);
1066 while (!list_empty(list)) {
1067 struct request *rq;
1068
1069 rq = list_first_entry(list, struct request, queuelist);
1070 list_del_init(&rq->queuelist);
1071 rq->mq_ctx = ctx;
Christoph Hellwig72a0a362014-02-07 10:22:36 -08001072 __blk_mq_insert_request(hctx, rq, false);
Jens Axboe320ae512013-10-24 09:20:05 +01001073 }
1074 spin_unlock(&ctx->lock);
1075
Jens Axboe320ae512013-10-24 09:20:05 +01001076 blk_mq_run_hw_queue(hctx, from_schedule);
Jens Axboee4043dc2014-04-09 10:18:23 -06001077 blk_mq_put_ctx(current_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001078}
1079
1080static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1081{
1082 struct request *rqa = container_of(a, struct request, queuelist);
1083 struct request *rqb = container_of(b, struct request, queuelist);
1084
1085 return !(rqa->mq_ctx < rqb->mq_ctx ||
1086 (rqa->mq_ctx == rqb->mq_ctx &&
1087 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1088}
1089
1090void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1091{
1092 struct blk_mq_ctx *this_ctx;
1093 struct request_queue *this_q;
1094 struct request *rq;
1095 LIST_HEAD(list);
1096 LIST_HEAD(ctx_list);
1097 unsigned int depth;
1098
1099 list_splice_init(&plug->mq_list, &list);
1100
1101 list_sort(NULL, &list, plug_ctx_cmp);
1102
1103 this_q = NULL;
1104 this_ctx = NULL;
1105 depth = 0;
1106
1107 while (!list_empty(&list)) {
1108 rq = list_entry_rq(list.next);
1109 list_del_init(&rq->queuelist);
1110 BUG_ON(!rq->q);
1111 if (rq->mq_ctx != this_ctx) {
1112 if (this_ctx) {
1113 blk_mq_insert_requests(this_q, this_ctx,
1114 &ctx_list, depth,
1115 from_schedule);
1116 }
1117
1118 this_ctx = rq->mq_ctx;
1119 this_q = rq->q;
1120 depth = 0;
1121 }
1122
1123 depth++;
1124 list_add_tail(&rq->queuelist, &ctx_list);
1125 }
1126
1127 /*
1128 * If 'this_ctx' is set, we know we have entries to complete
1129 * on 'ctx_list'. Do those.
1130 */
1131 if (this_ctx) {
1132 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1133 from_schedule);
1134 }
1135}
1136
1137static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1138{
1139 init_request_from_bio(rq, bio);
Jens Axboe4b570522014-05-29 11:00:11 -06001140
Jens Axboe3ee32372014-06-09 09:36:53 -06001141 if (blk_do_io_stat(rq))
Jens Axboe4b570522014-05-29 11:00:11 -06001142 blk_account_io_start(rq, 1);
Jens Axboe320ae512013-10-24 09:20:05 +01001143}
1144
Jens Axboe274a5842014-08-15 12:44:08 -06001145static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1146{
1147 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1148 !blk_queue_nomerges(hctx->queue);
1149}
1150
Jens Axboe07068d52014-05-22 10:40:51 -06001151static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1152 struct blk_mq_ctx *ctx,
1153 struct request *rq, struct bio *bio)
1154{
Jens Axboe274a5842014-08-15 12:44:08 -06001155 if (!hctx_allow_merges(hctx)) {
Jens Axboe07068d52014-05-22 10:40:51 -06001156 blk_mq_bio_to_request(rq, bio);
1157 spin_lock(&ctx->lock);
1158insert_rq:
1159 __blk_mq_insert_request(hctx, rq, false);
1160 spin_unlock(&ctx->lock);
1161 return false;
1162 } else {
Jens Axboe274a5842014-08-15 12:44:08 -06001163 struct request_queue *q = hctx->queue;
1164
Jens Axboe07068d52014-05-22 10:40:51 -06001165 spin_lock(&ctx->lock);
1166 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1167 blk_mq_bio_to_request(rq, bio);
1168 goto insert_rq;
1169 }
1170
1171 spin_unlock(&ctx->lock);
1172 __blk_mq_free_request(hctx, ctx, rq);
1173 return true;
1174 }
1175}
1176
1177struct blk_map_ctx {
1178 struct blk_mq_hw_ctx *hctx;
1179 struct blk_mq_ctx *ctx;
1180};
1181
1182static struct request *blk_mq_map_request(struct request_queue *q,
1183 struct bio *bio,
1184 struct blk_map_ctx *data)
Jens Axboe320ae512013-10-24 09:20:05 +01001185{
1186 struct blk_mq_hw_ctx *hctx;
1187 struct blk_mq_ctx *ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001188 struct request *rq;
Jens Axboe07068d52014-05-22 10:40:51 -06001189 int rw = bio_data_dir(bio);
Ming Leicb96a422014-06-01 00:43:37 +08001190 struct blk_mq_alloc_data alloc_data;
Jens Axboe320ae512013-10-24 09:20:05 +01001191
Keith Buschbfd343a2015-03-11 23:56:39 -04001192 if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
Nicholas Bellinger14ec77f2014-02-07 13:45:39 -07001193 bio_endio(bio, -EIO);
Jens Axboe07068d52014-05-22 10:40:51 -06001194 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001195 }
1196
1197 ctx = blk_mq_get_ctx(q);
1198 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1199
Jens Axboe07068d52014-05-22 10:40:51 -06001200 if (rw_is_sync(bio->bi_rw))
Shaohua Li27fbf4e2014-02-19 20:20:21 +08001201 rw |= REQ_SYNC;
Jens Axboe07068d52014-05-22 10:40:51 -06001202
Jens Axboe320ae512013-10-24 09:20:05 +01001203 trace_block_getrq(q, bio, rw);
Ming Leicb96a422014-06-01 00:43:37 +08001204 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1205 hctx);
1206 rq = __blk_mq_alloc_request(&alloc_data, rw);
Christoph Hellwig5dee8572014-05-27 20:59:47 +02001207 if (unlikely(!rq)) {
Christoph Hellwig793597a2014-05-27 20:59:49 +02001208 __blk_mq_run_hw_queue(hctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001209 blk_mq_put_ctx(ctx);
1210 trace_block_sleeprq(q, bio, rw);
Christoph Hellwig793597a2014-05-27 20:59:49 +02001211
1212 ctx = blk_mq_get_ctx(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001213 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Ming Leicb96a422014-06-01 00:43:37 +08001214 blk_mq_set_alloc_data(&alloc_data, q,
1215 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1216 rq = __blk_mq_alloc_request(&alloc_data, rw);
1217 ctx = alloc_data.ctx;
1218 hctx = alloc_data.hctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001219 }
1220
1221 hctx->queued++;
Jens Axboe07068d52014-05-22 10:40:51 -06001222 data->hctx = hctx;
1223 data->ctx = ctx;
1224 return rq;
1225}
1226
1227/*
1228 * Multiple hardware queue variant. This will not use per-process plugs,
1229 * but will attempt to bypass the hctx queueing if we can go straight to
1230 * hardware for SYNC IO.
1231 */
1232static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1233{
1234 const int is_sync = rw_is_sync(bio->bi_rw);
1235 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1236 struct blk_map_ctx data;
1237 struct request *rq;
1238
1239 blk_queue_bounce(q, &bio);
1240
1241 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1242 bio_endio(bio, -EIO);
1243 return;
1244 }
1245
1246 rq = blk_mq_map_request(q, bio, &data);
1247 if (unlikely(!rq))
1248 return;
1249
1250 if (unlikely(is_flush_fua)) {
1251 blk_mq_bio_to_request(rq, bio);
1252 blk_insert_flush(rq);
1253 goto run_queue;
1254 }
1255
Jens Axboee167dfb2014-10-29 11:18:26 -06001256 /*
1257 * If the driver supports defer issued based on 'last', then
1258 * queue it up like normal since we can potentially save some
1259 * CPU this way.
1260 */
1261 if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
Jens Axboe74c45052014-10-29 11:14:52 -06001262 struct blk_mq_queue_data bd = {
1263 .rq = rq,
1264 .list = NULL,
1265 .last = 1
1266 };
Jens Axboe07068d52014-05-22 10:40:51 -06001267 int ret;
1268
1269 blk_mq_bio_to_request(rq, bio);
Jens Axboe07068d52014-05-22 10:40:51 -06001270
1271 /*
1272 * For OK queue, we are done. For error, kill it. Any other
1273 * error (busy), just add it to our list as we previously
1274 * would have done
1275 */
Jens Axboe74c45052014-10-29 11:14:52 -06001276 ret = q->mq_ops->queue_rq(data.hctx, &bd);
Jens Axboe07068d52014-05-22 10:40:51 -06001277 if (ret == BLK_MQ_RQ_QUEUE_OK)
1278 goto done;
1279 else {
1280 __blk_mq_requeue_request(rq);
1281
1282 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1283 rq->errors = -EIO;
Christoph Hellwigc8a446a2014-09-13 16:40:10 -07001284 blk_mq_end_request(rq, rq->errors);
Jens Axboe07068d52014-05-22 10:40:51 -06001285 goto done;
1286 }
1287 }
1288 }
1289
1290 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1291 /*
1292 * For a SYNC request, send it to the hardware immediately. For
1293 * an ASYNC request, just ensure that we run it later on. The
1294 * latter allows for merging opportunities and more efficient
1295 * dispatching.
1296 */
1297run_queue:
1298 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1299 }
1300done:
1301 blk_mq_put_ctx(data.ctx);
1302}
1303
1304/*
1305 * Single hardware queue variant. This will attempt to use any per-process
1306 * plug for merging and IO deferral.
1307 */
1308static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1309{
1310 const int is_sync = rw_is_sync(bio->bi_rw);
1311 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1312 unsigned int use_plug, request_count = 0;
1313 struct blk_map_ctx data;
1314 struct request *rq;
1315
1316 /*
1317 * If we have multiple hardware queues, just go directly to
1318 * one of those for sync IO.
1319 */
1320 use_plug = !is_flush_fua && !is_sync;
1321
1322 blk_queue_bounce(q, &bio);
1323
1324 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1325 bio_endio(bio, -EIO);
1326 return;
1327 }
1328
1329 if (use_plug && !blk_queue_nomerges(q) &&
1330 blk_attempt_plug_merge(q, bio, &request_count))
1331 return;
1332
1333 rq = blk_mq_map_request(q, bio, &data);
Jens Axboeff87bce2014-06-03 11:59:49 -06001334 if (unlikely(!rq))
1335 return;
Jens Axboe320ae512013-10-24 09:20:05 +01001336
1337 if (unlikely(is_flush_fua)) {
1338 blk_mq_bio_to_request(rq, bio);
Jens Axboe320ae512013-10-24 09:20:05 +01001339 blk_insert_flush(rq);
1340 goto run_queue;
1341 }
1342
1343 /*
1344 * A task plug currently exists. Since this is completely lockless,
1345 * utilize that to temporarily store requests until the task is
1346 * either done or scheduled away.
1347 */
1348 if (use_plug) {
1349 struct blk_plug *plug = current->plug;
1350
1351 if (plug) {
1352 blk_mq_bio_to_request(rq, bio);
Shaohua Li92f399c2013-10-29 12:01:03 -06001353 if (list_empty(&plug->mq_list))
Jens Axboe320ae512013-10-24 09:20:05 +01001354 trace_block_plug(q);
1355 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1356 blk_flush_plug_list(plug, false);
1357 trace_block_plug(q);
1358 }
1359 list_add_tail(&rq->queuelist, &plug->mq_list);
Jens Axboe07068d52014-05-22 10:40:51 -06001360 blk_mq_put_ctx(data.ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001361 return;
1362 }
1363 }
1364
Jens Axboe07068d52014-05-22 10:40:51 -06001365 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1366 /*
1367 * For a SYNC request, send it to the hardware immediately. For
1368 * an ASYNC request, just ensure that we run it later on. The
1369 * latter allows for merging opportunities and more efficient
1370 * dispatching.
1371 */
1372run_queue:
1373 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
Jens Axboe320ae512013-10-24 09:20:05 +01001374 }
1375
Jens Axboe07068d52014-05-22 10:40:51 -06001376 blk_mq_put_ctx(data.ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001377}
1378
1379/*
1380 * Default mapping to a software queue, since we use one per CPU.
1381 */
1382struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1383{
1384 return q->queue_hw_ctx[q->mq_map[cpu]];
1385}
1386EXPORT_SYMBOL(blk_mq_map_queue);
1387
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001388static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1389 struct blk_mq_tags *tags, unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01001390{
1391 struct page *page;
1392
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001393 if (tags->rqs && set->ops->exit_request) {
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001394 int i;
1395
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001396 for (i = 0; i < tags->nr_tags; i++) {
1397 if (!tags->rqs[i])
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001398 continue;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001399 set->ops->exit_request(set->driver_data, tags->rqs[i],
1400 hctx_idx, i);
Jens Axboea5164402014-09-10 09:02:03 -06001401 tags->rqs[i] = NULL;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001402 }
1403 }
1404
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001405 while (!list_empty(&tags->page_list)) {
1406 page = list_first_entry(&tags->page_list, struct page, lru);
Dave Hansen67534712014-01-08 20:17:46 -07001407 list_del_init(&page->lru);
Jens Axboe320ae512013-10-24 09:20:05 +01001408 __free_pages(page, page->private);
1409 }
1410
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001411 kfree(tags->rqs);
Jens Axboe320ae512013-10-24 09:20:05 +01001412
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001413 blk_mq_free_tags(tags);
Jens Axboe320ae512013-10-24 09:20:05 +01001414}
1415
1416static size_t order_to_size(unsigned int order)
1417{
Ming Lei4ca08502014-04-19 18:00:18 +08001418 return (size_t)PAGE_SIZE << order;
Jens Axboe320ae512013-10-24 09:20:05 +01001419}
1420
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001421static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1422 unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01001423{
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001424 struct blk_mq_tags *tags;
Jens Axboe320ae512013-10-24 09:20:05 +01001425 unsigned int i, j, entries_per_page, max_order = 4;
1426 size_t rq_size, left;
1427
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001428 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
Shaohua Li24391c02015-01-23 14:18:00 -07001429 set->numa_node,
1430 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001431 if (!tags)
1432 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001433
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001434 INIT_LIST_HEAD(&tags->page_list);
1435
Jens Axboea5164402014-09-10 09:02:03 -06001436 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1437 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1438 set->numa_node);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001439 if (!tags->rqs) {
1440 blk_mq_free_tags(tags);
1441 return NULL;
1442 }
Jens Axboe320ae512013-10-24 09:20:05 +01001443
1444 /*
1445 * rq_size is the size of the request plus driver payload, rounded
1446 * to the cacheline size
1447 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001448 rq_size = round_up(sizeof(struct request) + set->cmd_size,
Jens Axboe320ae512013-10-24 09:20:05 +01001449 cache_line_size());
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001450 left = rq_size * set->queue_depth;
Jens Axboe320ae512013-10-24 09:20:05 +01001451
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001452 for (i = 0; i < set->queue_depth; ) {
Jens Axboe320ae512013-10-24 09:20:05 +01001453 int this_order = max_order;
1454 struct page *page;
1455 int to_do;
1456 void *p;
1457
1458 while (left < order_to_size(this_order - 1) && this_order)
1459 this_order--;
1460
1461 do {
Jens Axboea5164402014-09-10 09:02:03 -06001462 page = alloc_pages_node(set->numa_node,
Linus Torvaldsac211172015-04-09 14:12:22 -07001463 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
Jens Axboea5164402014-09-10 09:02:03 -06001464 this_order);
Jens Axboe320ae512013-10-24 09:20:05 +01001465 if (page)
1466 break;
1467 if (!this_order--)
1468 break;
1469 if (order_to_size(this_order) < rq_size)
1470 break;
1471 } while (1);
1472
1473 if (!page)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001474 goto fail;
Jens Axboe320ae512013-10-24 09:20:05 +01001475
1476 page->private = this_order;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001477 list_add_tail(&page->lru, &tags->page_list);
Jens Axboe320ae512013-10-24 09:20:05 +01001478
1479 p = page_address(page);
1480 entries_per_page = order_to_size(this_order) / rq_size;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001481 to_do = min(entries_per_page, set->queue_depth - i);
Jens Axboe320ae512013-10-24 09:20:05 +01001482 left -= to_do * rq_size;
1483 for (j = 0; j < to_do; j++) {
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001484 tags->rqs[i] = p;
1485 if (set->ops->init_request) {
1486 if (set->ops->init_request(set->driver_data,
1487 tags->rqs[i], hctx_idx, i,
Jens Axboea5164402014-09-10 09:02:03 -06001488 set->numa_node)) {
1489 tags->rqs[i] = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001490 goto fail;
Jens Axboea5164402014-09-10 09:02:03 -06001491 }
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001492 }
1493
Jens Axboe320ae512013-10-24 09:20:05 +01001494 p += rq_size;
1495 i++;
1496 }
1497 }
1498
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001499 return tags;
Jens Axboe320ae512013-10-24 09:20:05 +01001500
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001501fail:
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001502 blk_mq_free_rq_map(set, tags, hctx_idx);
1503 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001504}
1505
Jens Axboe1429d7c2014-05-19 09:23:55 -06001506static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1507{
1508 kfree(bitmap->map);
1509}
1510
1511static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1512{
1513 unsigned int bpw = 8, total, num_maps, i;
1514
1515 bitmap->bits_per_word = bpw;
1516
1517 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1518 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1519 GFP_KERNEL, node);
1520 if (!bitmap->map)
1521 return -ENOMEM;
1522
Jens Axboe1429d7c2014-05-19 09:23:55 -06001523 total = nr_cpu_ids;
1524 for (i = 0; i < num_maps; i++) {
1525 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1526 total -= bitmap->map[i].depth;
1527 }
1528
1529 return 0;
1530}
1531
Jens Axboe484b4062014-05-21 14:01:15 -06001532static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1533{
1534 struct request_queue *q = hctx->queue;
1535 struct blk_mq_ctx *ctx;
1536 LIST_HEAD(tmp);
1537
1538 /*
1539 * Move ctx entries to new CPU, if this one is going away.
1540 */
1541 ctx = __blk_mq_get_ctx(q, cpu);
1542
1543 spin_lock(&ctx->lock);
1544 if (!list_empty(&ctx->rq_list)) {
1545 list_splice_init(&ctx->rq_list, &tmp);
1546 blk_mq_hctx_clear_pending(hctx, ctx);
1547 }
1548 spin_unlock(&ctx->lock);
1549
1550 if (list_empty(&tmp))
1551 return NOTIFY_OK;
1552
1553 ctx = blk_mq_get_ctx(q);
1554 spin_lock(&ctx->lock);
1555
1556 while (!list_empty(&tmp)) {
1557 struct request *rq;
1558
1559 rq = list_first_entry(&tmp, struct request, queuelist);
1560 rq->mq_ctx = ctx;
1561 list_move_tail(&rq->queuelist, &ctx->rq_list);
1562 }
1563
1564 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1565 blk_mq_hctx_mark_pending(hctx, ctx);
1566
1567 spin_unlock(&ctx->lock);
1568
1569 blk_mq_run_hw_queue(hctx, true);
1570 blk_mq_put_ctx(ctx);
1571 return NOTIFY_OK;
1572}
1573
1574static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1575{
1576 struct request_queue *q = hctx->queue;
1577 struct blk_mq_tag_set *set = q->tag_set;
1578
1579 if (set->tags[hctx->queue_num])
1580 return NOTIFY_OK;
1581
1582 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1583 if (!set->tags[hctx->queue_num])
1584 return NOTIFY_STOP;
1585
1586 hctx->tags = set->tags[hctx->queue_num];
1587 return NOTIFY_OK;
1588}
1589
1590static int blk_mq_hctx_notify(void *data, unsigned long action,
1591 unsigned int cpu)
1592{
1593 struct blk_mq_hw_ctx *hctx = data;
1594
1595 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1596 return blk_mq_hctx_cpu_offline(hctx, cpu);
1597 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1598 return blk_mq_hctx_cpu_online(hctx, cpu);
1599
1600 return NOTIFY_OK;
1601}
1602
Ming Lei08e98fc2014-09-25 23:23:38 +08001603static void blk_mq_exit_hctx(struct request_queue *q,
1604 struct blk_mq_tag_set *set,
1605 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1606{
Ming Leif70ced02014-09-25 23:23:47 +08001607 unsigned flush_start_tag = set->queue_depth;
1608
Ming Lei08e98fc2014-09-25 23:23:38 +08001609 blk_mq_tag_idle(hctx);
1610
Ming Leif70ced02014-09-25 23:23:47 +08001611 if (set->ops->exit_request)
1612 set->ops->exit_request(set->driver_data,
1613 hctx->fq->flush_rq, hctx_idx,
1614 flush_start_tag + hctx_idx);
1615
Ming Lei08e98fc2014-09-25 23:23:38 +08001616 if (set->ops->exit_hctx)
1617 set->ops->exit_hctx(hctx, hctx_idx);
1618
1619 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
Ming Leif70ced02014-09-25 23:23:47 +08001620 blk_free_flush_queue(hctx->fq);
Ming Lei08e98fc2014-09-25 23:23:38 +08001621 kfree(hctx->ctxs);
1622 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;
1778
1779 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboee4043dc2014-04-09 10:18:23 -06001780 cpumask_clear(hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001781 hctx->nr_ctx = 0;
1782 }
1783
1784 /*
1785 * Map software to hardware queues
1786 */
1787 queue_for_each_ctx(q, ctx, i) {
1788 /* If the cpu isn't online, the cpu is mapped to first hctx */
Jens Axboee4043dc2014-04-09 10:18:23 -06001789 if (!cpu_online(i))
1790 continue;
1791
Jens Axboe320ae512013-10-24 09:20:05 +01001792 hctx = q->mq_ops->map_queue(q, i);
Jens Axboee4043dc2014-04-09 10:18:23 -06001793 cpumask_set_cpu(i, hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001794 ctx->index_hw = hctx->nr_ctx;
1795 hctx->ctxs[hctx->nr_ctx++] = ctx;
1796 }
Jens Axboe506e9312014-05-07 10:26:44 -06001797
1798 queue_for_each_hw_ctx(q, hctx, i) {
Chong Yuan889fa312015-04-15 11:39:29 -06001799 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1800
Jens Axboe484b4062014-05-21 14:01:15 -06001801 /*
Jens Axboea68aafa2014-08-15 13:19:15 -06001802 * If no software queues are mapped to this hardware queue,
1803 * disable it and free the request entries.
Jens Axboe484b4062014-05-21 14:01:15 -06001804 */
1805 if (!hctx->nr_ctx) {
1806 struct blk_mq_tag_set *set = q->tag_set;
1807
1808 if (set->tags[i]) {
1809 blk_mq_free_rq_map(set, set->tags[i], i);
1810 set->tags[i] = NULL;
1811 hctx->tags = NULL;
1812 }
1813 continue;
1814 }
1815
1816 /*
Chong Yuan889fa312015-04-15 11:39:29 -06001817 * Set the map size to the number of mapped software queues.
1818 * This is more accurate and more efficient than looping
1819 * over all possibly mapped software queues.
1820 */
Jens Axboe569fd0c2015-04-17 08:28:50 -06001821 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
Chong Yuan889fa312015-04-15 11:39:29 -06001822
1823 /*
Jens Axboe484b4062014-05-21 14:01:15 -06001824 * Initialize batch roundrobin counts
1825 */
Jens Axboe506e9312014-05-07 10:26:44 -06001826 hctx->next_cpu = cpumask_first(hctx->cpumask);
1827 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1828 }
Jens Axboe320ae512013-10-24 09:20:05 +01001829}
1830
Jens Axboe0d2602c2014-05-13 15:10:52 -06001831static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1832{
1833 struct blk_mq_hw_ctx *hctx;
1834 struct request_queue *q;
1835 bool shared;
1836 int i;
1837
1838 if (set->tag_list.next == set->tag_list.prev)
1839 shared = false;
1840 else
1841 shared = true;
1842
1843 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1844 blk_mq_freeze_queue(q);
1845
1846 queue_for_each_hw_ctx(q, hctx, i) {
1847 if (shared)
1848 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1849 else
1850 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1851 }
1852 blk_mq_unfreeze_queue(q);
1853 }
1854}
1855
1856static void blk_mq_del_queue_tag_set(struct request_queue *q)
1857{
1858 struct blk_mq_tag_set *set = q->tag_set;
1859
Jens Axboe0d2602c2014-05-13 15:10:52 -06001860 mutex_lock(&set->tag_list_lock);
1861 list_del_init(&q->tag_set_list);
1862 blk_mq_update_tag_set_depth(set);
1863 mutex_unlock(&set->tag_list_lock);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001864}
1865
1866static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1867 struct request_queue *q)
1868{
1869 q->tag_set = set;
1870
1871 mutex_lock(&set->tag_list_lock);
1872 list_add_tail(&q->tag_set_list, &set->tag_list);
1873 blk_mq_update_tag_set_depth(set);
1874 mutex_unlock(&set->tag_list_lock);
1875}
1876
Ming Leie09aae72015-01-29 20:17:27 +08001877/*
1878 * It is the actual release handler for mq, but we do it from
1879 * request queue's release handler for avoiding use-after-free
1880 * and headache because q->mq_kobj shouldn't have been introduced,
1881 * but we can't group ctx/kctx kobj without it.
1882 */
1883void blk_mq_release(struct request_queue *q)
1884{
1885 struct blk_mq_hw_ctx *hctx;
1886 unsigned int i;
1887
1888 /* hctx kobj stays in hctx */
1889 queue_for_each_hw_ctx(q, hctx, i)
1890 kfree(hctx);
1891
1892 kfree(q->queue_hw_ctx);
1893
1894 /* ctx kobj stays in queue_ctx */
1895 free_percpu(q->queue_ctx);
1896}
1897
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001898struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
Jens Axboe320ae512013-10-24 09:20:05 +01001899{
Mike Snitzerb62c21b2015-03-12 23:56:02 -04001900 struct request_queue *uninit_q, *q;
1901
1902 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1903 if (!uninit_q)
1904 return ERR_PTR(-ENOMEM);
1905
1906 q = blk_mq_init_allocated_queue(set, uninit_q);
1907 if (IS_ERR(q))
1908 blk_cleanup_queue(uninit_q);
1909
1910 return q;
1911}
1912EXPORT_SYMBOL(blk_mq_init_queue);
1913
1914struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1915 struct request_queue *q)
1916{
Jens Axboe320ae512013-10-24 09:20:05 +01001917 struct blk_mq_hw_ctx **hctxs;
Ming Leie6cdb092014-06-03 11:24:06 +08001918 struct blk_mq_ctx __percpu *ctx;
Jens Axboef14bbe72014-05-27 12:06:53 -06001919 unsigned int *map;
Jens Axboe320ae512013-10-24 09:20:05 +01001920 int i;
1921
Jens Axboe320ae512013-10-24 09:20:05 +01001922 ctx = alloc_percpu(struct blk_mq_ctx);
1923 if (!ctx)
1924 return ERR_PTR(-ENOMEM);
1925
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001926 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1927 set->numa_node);
Jens Axboe320ae512013-10-24 09:20:05 +01001928
1929 if (!hctxs)
1930 goto err_percpu;
1931
Jens Axboef14bbe72014-05-27 12:06:53 -06001932 map = blk_mq_make_queue_map(set);
1933 if (!map)
1934 goto err_map;
1935
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001936 for (i = 0; i < set->nr_hw_queues; i++) {
Jens Axboef14bbe72014-05-27 12:06:53 -06001937 int node = blk_mq_hw_queue_to_node(map, i);
1938
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02001939 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1940 GFP_KERNEL, node);
Jens Axboe320ae512013-10-24 09:20:05 +01001941 if (!hctxs[i])
1942 goto err_hctxs;
1943
Jens Axboea86073e2014-10-13 15:41:54 -06001944 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1945 node))
Jens Axboee4043dc2014-04-09 10:18:23 -06001946 goto err_hctxs;
1947
Jens Axboe0d2602c2014-05-13 15:10:52 -06001948 atomic_set(&hctxs[i]->nr_active, 0);
Jens Axboef14bbe72014-05-27 12:06:53 -06001949 hctxs[i]->numa_node = node;
Jens Axboe320ae512013-10-24 09:20:05 +01001950 hctxs[i]->queue_num = i;
1951 }
1952
Tejun Heo17497ac2014-09-24 13:31:50 -04001953 /*
1954 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1955 * See blk_register_queue() for details.
1956 */
Tejun Heoa34375e2014-09-08 09:51:30 +09001957 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
Tejun Heo17497ac2014-09-24 13:31:50 -04001958 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
Mike Snitzerb62c21b2015-03-12 23:56:02 -04001959 goto err_hctxs;
Ming Lei3d2936f2014-05-27 23:35:14 +08001960
Jens Axboe320ae512013-10-24 09:20:05 +01001961 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
Wei Fangc76cbbc2015-03-30 09:07:00 -06001962 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30000);
Jens Axboe320ae512013-10-24 09:20:05 +01001963
1964 q->nr_queues = nr_cpu_ids;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001965 q->nr_hw_queues = set->nr_hw_queues;
Jens Axboef14bbe72014-05-27 12:06:53 -06001966 q->mq_map = map;
Jens Axboe320ae512013-10-24 09:20:05 +01001967
1968 q->queue_ctx = ctx;
1969 q->queue_hw_ctx = hctxs;
1970
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001971 q->mq_ops = set->ops;
Jens Axboe94eddfb2013-11-19 09:25:07 -07001972 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
Jens Axboe320ae512013-10-24 09:20:05 +01001973
Jens Axboe05f1dd52014-05-29 09:53:32 -06001974 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1975 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1976
Christoph Hellwig1be036e2014-02-07 10:22:39 -08001977 q->sg_reserved_size = INT_MAX;
1978
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001979 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1980 INIT_LIST_HEAD(&q->requeue_list);
1981 spin_lock_init(&q->requeue_lock);
1982
Jens Axboe07068d52014-05-22 10:40:51 -06001983 if (q->nr_hw_queues > 1)
1984 blk_queue_make_request(q, blk_mq_make_request);
1985 else
1986 blk_queue_make_request(q, blk_sq_make_request);
1987
Jens Axboeeba71762014-05-20 15:17:27 -06001988 /*
1989 * Do this after blk_queue_make_request() overrides it...
1990 */
1991 q->nr_requests = set->queue_depth;
1992
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001993 if (set->ops->complete)
1994 blk_queue_softirq_done(q, set->ops->complete);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08001995
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001996 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +01001997
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001998 if (blk_mq_init_hw_queues(q, set))
Mike Snitzerb62c21b2015-03-12 23:56:02 -04001999 goto err_hctxs;
Christoph Hellwig18741982014-02-10 09:29:00 -07002000
Jens Axboe320ae512013-10-24 09:20:05 +01002001 mutex_lock(&all_q_mutex);
2002 list_add_tail(&q->all_q_node, &all_q_list);
2003 mutex_unlock(&all_q_mutex);
2004
Jens Axboe0d2602c2014-05-13 15:10:52 -06002005 blk_mq_add_queue_tag_set(set, q);
2006
Jens Axboe484b4062014-05-21 14:01:15 -06002007 blk_mq_map_swqueue(q);
2008
Jens Axboe320ae512013-10-24 09:20:05 +01002009 return q;
Christoph Hellwig18741982014-02-10 09:29:00 -07002010
Jens Axboe320ae512013-10-24 09:20:05 +01002011err_hctxs:
Jens Axboef14bbe72014-05-27 12:06:53 -06002012 kfree(map);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002013 for (i = 0; i < set->nr_hw_queues; i++) {
Jens Axboe320ae512013-10-24 09:20:05 +01002014 if (!hctxs[i])
2015 break;
Jens Axboee4043dc2014-04-09 10:18:23 -06002016 free_cpumask_var(hctxs[i]->cpumask);
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02002017 kfree(hctxs[i]);
Jens Axboe320ae512013-10-24 09:20:05 +01002018 }
Jens Axboef14bbe72014-05-27 12:06:53 -06002019err_map:
Jens Axboe320ae512013-10-24 09:20:05 +01002020 kfree(hctxs);
2021err_percpu:
2022 free_percpu(ctx);
2023 return ERR_PTR(-ENOMEM);
2024}
Mike Snitzerb62c21b2015-03-12 23:56:02 -04002025EXPORT_SYMBOL(blk_mq_init_allocated_queue);
Jens Axboe320ae512013-10-24 09:20:05 +01002026
2027void blk_mq_free_queue(struct request_queue *q)
2028{
Ming Lei624dbe42014-05-27 23:35:13 +08002029 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01002030
Jens Axboe0d2602c2014-05-13 15:10:52 -06002031 blk_mq_del_queue_tag_set(q);
2032
Ming Lei624dbe42014-05-27 23:35:13 +08002033 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2034 blk_mq_free_hw_queues(q, set);
Jens Axboe320ae512013-10-24 09:20:05 +01002035
Tejun Heoadd703f2014-07-01 10:34:38 -06002036 percpu_ref_exit(&q->mq_usage_counter);
Ming Lei3d2936f2014-05-27 23:35:14 +08002037
Jens Axboe320ae512013-10-24 09:20:05 +01002038 kfree(q->mq_map);
2039
Jens Axboe320ae512013-10-24 09:20:05 +01002040 q->mq_map = NULL;
2041
2042 mutex_lock(&all_q_mutex);
2043 list_del_init(&q->all_q_node);
2044 mutex_unlock(&all_q_mutex);
2045}
Jens Axboe320ae512013-10-24 09:20:05 +01002046
2047/* Basically redo blk_mq_init_queue with queue frozen */
Paul Gortmakerf618ef72013-11-14 08:26:02 -07002048static void blk_mq_queue_reinit(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01002049{
Tejun Heof3af0202014-11-04 13:52:27 -05002050 WARN_ON_ONCE(!q->mq_freeze_depth);
Jens Axboe320ae512013-10-24 09:20:05 +01002051
Jens Axboe67aec142014-05-30 08:25:36 -06002052 blk_mq_sysfs_unregister(q);
2053
Jens Axboe320ae512013-10-24 09:20:05 +01002054 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2055
2056 /*
2057 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2058 * we should change hctx numa_node according to new topology (this
2059 * involves free and re-allocate memory, worthy doing?)
2060 */
2061
2062 blk_mq_map_swqueue(q);
2063
Jens Axboe67aec142014-05-30 08:25:36 -06002064 blk_mq_sysfs_register(q);
Jens Axboe320ae512013-10-24 09:20:05 +01002065}
2066
Paul Gortmakerf618ef72013-11-14 08:26:02 -07002067static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2068 unsigned long action, void *hcpu)
Jens Axboe320ae512013-10-24 09:20:05 +01002069{
2070 struct request_queue *q;
2071
2072 /*
Jens Axboe9fccfed2014-05-08 14:50:19 -06002073 * Before new mappings are established, hotadded cpu might already
2074 * start handling requests. This doesn't break anything as we map
2075 * offline CPUs to first hardware queue. We will re-init the queue
2076 * below to get optimal settings.
Jens Axboe320ae512013-10-24 09:20:05 +01002077 */
2078 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2079 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2080 return NOTIFY_OK;
2081
2082 mutex_lock(&all_q_mutex);
Tejun Heof3af0202014-11-04 13:52:27 -05002083
2084 /*
2085 * We need to freeze and reinit all existing queues. Freezing
2086 * involves synchronous wait for an RCU grace period and doing it
2087 * one by one may take a long time. Start freezing all queues in
2088 * one swoop and then wait for the completions so that freezing can
2089 * take place in parallel.
2090 */
2091 list_for_each_entry(q, &all_q_list, all_q_node)
2092 blk_mq_freeze_queue_start(q);
2093 list_for_each_entry(q, &all_q_list, all_q_node)
2094 blk_mq_freeze_queue_wait(q);
2095
Jens Axboe320ae512013-10-24 09:20:05 +01002096 list_for_each_entry(q, &all_q_list, all_q_node)
2097 blk_mq_queue_reinit(q);
Tejun Heof3af0202014-11-04 13:52:27 -05002098
2099 list_for_each_entry(q, &all_q_list, all_q_node)
2100 blk_mq_unfreeze_queue(q);
2101
Jens Axboe320ae512013-10-24 09:20:05 +01002102 mutex_unlock(&all_q_mutex);
2103 return NOTIFY_OK;
2104}
2105
Jens Axboea5164402014-09-10 09:02:03 -06002106static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2107{
2108 int i;
2109
2110 for (i = 0; i < set->nr_hw_queues; i++) {
2111 set->tags[i] = blk_mq_init_rq_map(set, i);
2112 if (!set->tags[i])
2113 goto out_unwind;
2114 }
2115
2116 return 0;
2117
2118out_unwind:
2119 while (--i >= 0)
2120 blk_mq_free_rq_map(set, set->tags[i], i);
2121
Jens Axboea5164402014-09-10 09:02:03 -06002122 return -ENOMEM;
2123}
2124
2125/*
2126 * Allocate the request maps associated with this tag_set. Note that this
2127 * may reduce the depth asked for, if memory is tight. set->queue_depth
2128 * will be updated to reflect the allocated depth.
2129 */
2130static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2131{
2132 unsigned int depth;
2133 int err;
2134
2135 depth = set->queue_depth;
2136 do {
2137 err = __blk_mq_alloc_rq_maps(set);
2138 if (!err)
2139 break;
2140
2141 set->queue_depth >>= 1;
2142 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2143 err = -ENOMEM;
2144 break;
2145 }
2146 } while (set->queue_depth);
2147
2148 if (!set->queue_depth || err) {
2149 pr_err("blk-mq: failed to allocate request map\n");
2150 return -ENOMEM;
2151 }
2152
2153 if (depth != set->queue_depth)
2154 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2155 depth, set->queue_depth);
2156
2157 return 0;
2158}
2159
Jens Axboea4391c62014-06-05 15:21:56 -06002160/*
2161 * Alloc a tag set to be associated with one or more request queues.
2162 * May fail with EINVAL for various error conditions. May adjust the
2163 * requested depth down, if if it too large. In that case, the set
2164 * value will be stored in set->queue_depth.
2165 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002166int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2167{
Bart Van Assche205fb5f2014-10-30 14:45:11 +01002168 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2169
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002170 if (!set->nr_hw_queues)
2171 return -EINVAL;
Jens Axboea4391c62014-06-05 15:21:56 -06002172 if (!set->queue_depth)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002173 return -EINVAL;
2174 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2175 return -EINVAL;
2176
Xiaoguang Wangf9018ac2015-03-30 13:19:14 +08002177 if (!set->ops->queue_rq || !set->ops->map_queue)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002178 return -EINVAL;
2179
Jens Axboea4391c62014-06-05 15:21:56 -06002180 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2181 pr_info("blk-mq: reduced tag depth to %u\n",
2182 BLK_MQ_MAX_DEPTH);
2183 set->queue_depth = BLK_MQ_MAX_DEPTH;
2184 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002185
Shaohua Li6637fad2014-11-30 16:00:58 -08002186 /*
2187 * If a crashdump is active, then we are potentially in a very
2188 * memory constrained environment. Limit us to 1 queue and
2189 * 64 tags to prevent using too much memory.
2190 */
2191 if (is_kdump_kernel()) {
2192 set->nr_hw_queues = 1;
2193 set->queue_depth = min(64U, set->queue_depth);
2194 }
2195
Ming Lei48479002014-04-19 18:00:17 +08002196 set->tags = kmalloc_node(set->nr_hw_queues *
2197 sizeof(struct blk_mq_tags *),
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002198 GFP_KERNEL, set->numa_node);
2199 if (!set->tags)
Jens Axboea5164402014-09-10 09:02:03 -06002200 return -ENOMEM;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002201
Jens Axboea5164402014-09-10 09:02:03 -06002202 if (blk_mq_alloc_rq_maps(set))
2203 goto enomem;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002204
Jens Axboe0d2602c2014-05-13 15:10:52 -06002205 mutex_init(&set->tag_list_lock);
2206 INIT_LIST_HEAD(&set->tag_list);
2207
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002208 return 0;
Jens Axboea5164402014-09-10 09:02:03 -06002209enomem:
Robert Elliott5676e7b2014-09-02 11:38:44 -05002210 kfree(set->tags);
2211 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002212 return -ENOMEM;
2213}
2214EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2215
2216void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2217{
2218 int i;
2219
Jens Axboe484b4062014-05-21 14:01:15 -06002220 for (i = 0; i < set->nr_hw_queues; i++) {
2221 if (set->tags[i])
2222 blk_mq_free_rq_map(set, set->tags[i], i);
2223 }
2224
Ming Lei981bd182014-04-24 00:07:34 +08002225 kfree(set->tags);
Robert Elliott5676e7b2014-09-02 11:38:44 -05002226 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002227}
2228EXPORT_SYMBOL(blk_mq_free_tag_set);
2229
Jens Axboee3a2b3f2014-05-20 11:49:02 -06002230int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2231{
2232 struct blk_mq_tag_set *set = q->tag_set;
2233 struct blk_mq_hw_ctx *hctx;
2234 int i, ret;
2235
2236 if (!set || nr > set->queue_depth)
2237 return -EINVAL;
2238
2239 ret = 0;
2240 queue_for_each_hw_ctx(q, hctx, i) {
2241 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2242 if (ret)
2243 break;
2244 }
2245
2246 if (!ret)
2247 q->nr_requests = nr;
2248
2249 return ret;
2250}
2251
Jens Axboe676141e2014-03-20 13:29:18 -06002252void blk_mq_disable_hotplug(void)
2253{
2254 mutex_lock(&all_q_mutex);
2255}
2256
2257void blk_mq_enable_hotplug(void)
2258{
2259 mutex_unlock(&all_q_mutex);
2260}
2261
Jens Axboe320ae512013-10-24 09:20:05 +01002262static int __init blk_mq_init(void)
2263{
Jens Axboe320ae512013-10-24 09:20:05 +01002264 blk_mq_cpu_init();
2265
Tejun Heoadd703f2014-07-01 10:34:38 -06002266 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
Jens Axboe320ae512013-10-24 09:20:05 +01002267
2268 return 0;
2269}
2270subsys_initcall(blk_mq_init);