blob: 050f7a13021baca7347df79d8d26dad2b89e327e [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>
Catalin Marinasf75782e2015-09-14 18:16:02 +010012#include <linux/kmemleak.h>
Jens Axboe320ae512013-10-24 09:20:05 +010013#include <linux/mm.h>
14#include <linux/init.h>
15#include <linux/slab.h>
16#include <linux/workqueue.h>
17#include <linux/smp.h>
18#include <linux/llist.h>
19#include <linux/list_sort.h>
20#include <linux/cpu.h>
21#include <linux/cache.h>
22#include <linux/sched/sysctl.h>
23#include <linux/delay.h>
Jens Axboeaedcd722014-09-17 08:27:03 -060024#include <linux/crash_dump.h>
Jens Axboe320ae512013-10-24 09:20:05 +010025
26#include <trace/events/block.h>
27
28#include <linux/blk-mq.h>
29#include "blk.h"
30#include "blk-mq.h"
31#include "blk-mq-tag.h"
32
33static DEFINE_MUTEX(all_q_mutex);
34static LIST_HEAD(all_q_list);
35
36static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
37
Jens Axboe320ae512013-10-24 09:20:05 +010038/*
39 * Check if any of the ctx's have pending work in this hardware queue
40 */
41static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
42{
43 unsigned int i;
44
Jens Axboe569fd0c2015-04-17 08:28:50 -060045 for (i = 0; i < hctx->ctx_map.size; i++)
Jens Axboe1429d7c2014-05-19 09:23:55 -060046 if (hctx->ctx_map.map[i].word)
Jens Axboe320ae512013-10-24 09:20:05 +010047 return true;
48
49 return false;
50}
51
Jens Axboe1429d7c2014-05-19 09:23:55 -060052static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
53 struct blk_mq_ctx *ctx)
54{
55 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
56}
57
58#define CTX_TO_BIT(hctx, ctx) \
59 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
60
Jens Axboe320ae512013-10-24 09:20:05 +010061/*
62 * Mark this ctx as having pending work in this hardware queue
63 */
64static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
65 struct blk_mq_ctx *ctx)
66{
Jens Axboe1429d7c2014-05-19 09:23:55 -060067 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
68
69 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
70 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
71}
72
73static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
74 struct blk_mq_ctx *ctx)
75{
76 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
77
78 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
Jens Axboe320ae512013-10-24 09:20:05 +010079}
80
Keith Buschb4c6a022014-12-19 17:54:14 -070081void blk_mq_freeze_queue_start(struct request_queue *q)
Ming Lei43a5e4e2013-12-26 21:31:35 +080082{
Christoph Hellwig4ecd4fe2015-05-07 09:38:13 +020083 int freeze_depth;
Tejun Heocddd5d12014-08-16 08:02:24 -040084
Christoph Hellwig4ecd4fe2015-05-07 09:38:13 +020085 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
86 if (freeze_depth == 1) {
Dan Williams3ef28e82015-10-21 13:20:12 -040087 percpu_ref_kill(&q->q_usage_counter);
Mike Snitzerb94ec292015-03-11 23:56:38 -040088 blk_mq_run_hw_queues(q, false);
Tejun Heocddd5d12014-08-16 08:02:24 -040089 }
Tejun Heof3af0202014-11-04 13:52:27 -050090}
Keith Buschb4c6a022014-12-19 17:54:14 -070091EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
Tejun Heof3af0202014-11-04 13:52:27 -050092
93static void blk_mq_freeze_queue_wait(struct request_queue *q)
94{
Dan Williams3ef28e82015-10-21 13:20:12 -040095 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
Ming Lei43a5e4e2013-12-26 21:31:35 +080096}
97
Tejun Heof3af0202014-11-04 13:52:27 -050098/*
99 * Guarantee no request is in use, so we can change any data structure of
100 * the queue afterward.
101 */
Dan Williams3ef28e82015-10-21 13:20:12 -0400102void blk_freeze_queue(struct request_queue *q)
Tejun Heof3af0202014-11-04 13:52:27 -0500103{
Dan Williams3ef28e82015-10-21 13:20:12 -0400104 /*
105 * In the !blk_mq case we are only calling this to kill the
106 * q_usage_counter, otherwise this increases the freeze depth
107 * and waits for it to return to zero. For this reason there is
108 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
109 * exported to drivers as the only user for unfreeze is blk_mq.
110 */
Tejun Heof3af0202014-11-04 13:52:27 -0500111 blk_mq_freeze_queue_start(q);
112 blk_mq_freeze_queue_wait(q);
113}
Dan Williams3ef28e82015-10-21 13:20:12 -0400114
115void blk_mq_freeze_queue(struct request_queue *q)
116{
117 /*
118 * ...just an alias to keep freeze and unfreeze actions balanced
119 * in the blk_mq_* namespace
120 */
121 blk_freeze_queue(q);
122}
Jens Axboec761d962015-01-02 15:05:12 -0700123EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
Tejun Heof3af0202014-11-04 13:52:27 -0500124
Keith Buschb4c6a022014-12-19 17:54:14 -0700125void blk_mq_unfreeze_queue(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +0100126{
Christoph Hellwig4ecd4fe2015-05-07 09:38:13 +0200127 int freeze_depth;
Jens Axboe320ae512013-10-24 09:20:05 +0100128
Christoph Hellwig4ecd4fe2015-05-07 09:38:13 +0200129 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
130 WARN_ON_ONCE(freeze_depth < 0);
131 if (!freeze_depth) {
Dan Williams3ef28e82015-10-21 13:20:12 -0400132 percpu_ref_reinit(&q->q_usage_counter);
Jens Axboe320ae512013-10-24 09:20:05 +0100133 wake_up_all(&q->mq_freeze_wq);
Tejun Heoadd703f2014-07-01 10:34:38 -0600134 }
Jens Axboe320ae512013-10-24 09:20:05 +0100135}
Keith Buschb4c6a022014-12-19 17:54:14 -0700136EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
Jens Axboe320ae512013-10-24 09:20:05 +0100137
Jens Axboeaed3ea92014-12-22 14:04:42 -0700138void blk_mq_wake_waiters(struct request_queue *q)
139{
140 struct blk_mq_hw_ctx *hctx;
141 unsigned int i;
142
143 queue_for_each_hw_ctx(q, hctx, i)
144 if (blk_mq_hw_queue_mapped(hctx))
145 blk_mq_tag_wakeup_all(hctx->tags, true);
Keith Busch3fd59402015-01-08 08:53:56 -0700146
147 /*
148 * If we are called because the queue has now been marked as
149 * dying, we need to ensure that processes currently waiting on
150 * the queue are notified as well.
151 */
152 wake_up_all(&q->mq_freeze_wq);
Jens Axboeaed3ea92014-12-22 14:04:42 -0700153}
154
Jens Axboe320ae512013-10-24 09:20:05 +0100155bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
156{
157 return blk_mq_has_free_tags(hctx->tags);
158}
159EXPORT_SYMBOL(blk_mq_can_queue);
160
Jens Axboe94eddfb2013-11-19 09:25:07 -0700161static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
162 struct request *rq, unsigned int rw_flags)
Jens Axboe320ae512013-10-24 09:20:05 +0100163{
Jens Axboe94eddfb2013-11-19 09:25:07 -0700164 if (blk_queue_io_stat(q))
165 rw_flags |= REQ_IO_STAT;
166
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200167 INIT_LIST_HEAD(&rq->queuelist);
168 /* csd/requeue_work/fifo_time is initialized before use */
169 rq->q = q;
Jens Axboe320ae512013-10-24 09:20:05 +0100170 rq->mq_ctx = ctx;
Jens Axboe0d2602c2014-05-13 15:10:52 -0600171 rq->cmd_flags |= rw_flags;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200172 /* do not touch atomic flags, it needs atomic ops against the timer */
173 rq->cpu = -1;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200174 INIT_HLIST_NODE(&rq->hash);
175 RB_CLEAR_NODE(&rq->rb_node);
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200176 rq->rq_disk = NULL;
177 rq->part = NULL;
Jens Axboe3ee32372014-06-09 09:36:53 -0600178 rq->start_time = jiffies;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200179#ifdef CONFIG_BLK_CGROUP
180 rq->rl = NULL;
Ming Lei0fec08b2014-01-03 10:00:08 -0700181 set_start_time_ns(rq);
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200182 rq->io_start_time_ns = 0;
183#endif
184 rq->nr_phys_segments = 0;
185#if defined(CONFIG_BLK_DEV_INTEGRITY)
186 rq->nr_integrity_segments = 0;
187#endif
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200188 rq->special = NULL;
189 /* tag was already set */
190 rq->errors = 0;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200191
Tony Battersby6f4a16262014-08-22 15:53:39 -0400192 rq->cmd = rq->__cmd;
193
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200194 rq->extra_len = 0;
195 rq->sense_len = 0;
196 rq->resid_len = 0;
197 rq->sense = NULL;
198
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200199 INIT_LIST_HEAD(&rq->timeout_list);
Jens Axboef6be4fb2014-06-06 11:03:48 -0600200 rq->timeout = 0;
201
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200202 rq->end_io = NULL;
203 rq->end_io_data = NULL;
204 rq->next_rq = NULL;
205
Jens Axboe320ae512013-10-24 09:20:05 +0100206 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
207}
208
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200209static struct request *
Ming Leicb96a422014-06-01 00:43:37 +0800210__blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200211{
212 struct request *rq;
213 unsigned int tag;
214
Ming Leicb96a422014-06-01 00:43:37 +0800215 tag = blk_mq_get_tag(data);
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200216 if (tag != BLK_MQ_TAG_FAIL) {
Ming Leicb96a422014-06-01 00:43:37 +0800217 rq = data->hctx->tags->rqs[tag];
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200218
Ming Leicb96a422014-06-01 00:43:37 +0800219 if (blk_mq_tag_busy(data->hctx)) {
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200220 rq->cmd_flags = REQ_MQ_INFLIGHT;
Ming Leicb96a422014-06-01 00:43:37 +0800221 atomic_inc(&data->hctx->nr_active);
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200222 }
223
224 rq->tag = tag;
Ming Leicb96a422014-06-01 00:43:37 +0800225 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200226 return rq;
227 }
228
229 return NULL;
230}
231
Christoph Hellwig6f3b0e82015-11-26 09:13:05 +0100232struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
233 unsigned int flags)
Jens Axboe320ae512013-10-24 09:20:05 +0100234{
Christoph Hellwigd8525642014-05-27 20:59:50 +0200235 struct blk_mq_ctx *ctx;
236 struct blk_mq_hw_ctx *hctx;
Jens Axboe320ae512013-10-24 09:20:05 +0100237 struct request *rq;
Ming Leicb96a422014-06-01 00:43:37 +0800238 struct blk_mq_alloc_data alloc_data;
Joe Lawrencea492f072014-08-28 08:15:21 -0600239 int ret;
Jens Axboe320ae512013-10-24 09:20:05 +0100240
Christoph Hellwig6f3b0e82015-11-26 09:13:05 +0100241 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
Joe Lawrencea492f072014-08-28 08:15:21 -0600242 if (ret)
243 return ERR_PTR(ret);
Jens Axboe320ae512013-10-24 09:20:05 +0100244
Christoph Hellwigd8525642014-05-27 20:59:50 +0200245 ctx = blk_mq_get_ctx(q);
246 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Christoph Hellwig6f3b0e82015-11-26 09:13:05 +0100247 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
Christoph Hellwigd8525642014-05-27 20:59:50 +0200248
Ming Leicb96a422014-06-01 00:43:37 +0800249 rq = __blk_mq_alloc_request(&alloc_data, rw);
Christoph Hellwig6f3b0e82015-11-26 09:13:05 +0100250 if (!rq && !(flags & BLK_MQ_REQ_NOWAIT)) {
Christoph Hellwigd8525642014-05-27 20:59:50 +0200251 __blk_mq_run_hw_queue(hctx);
252 blk_mq_put_ctx(ctx);
253
254 ctx = blk_mq_get_ctx(q);
255 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Christoph Hellwig6f3b0e82015-11-26 09:13:05 +0100256 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
Ming Leicb96a422014-06-01 00:43:37 +0800257 rq = __blk_mq_alloc_request(&alloc_data, rw);
258 ctx = alloc_data.ctx;
Christoph Hellwigd8525642014-05-27 20:59:50 +0200259 }
260 blk_mq_put_ctx(ctx);
Keith Buschc76541a2014-12-19 17:54:13 -0700261 if (!rq) {
Dan Williams3ef28e82015-10-21 13:20:12 -0400262 blk_queue_exit(q);
Joe Lawrencea492f072014-08-28 08:15:21 -0600263 return ERR_PTR(-EWOULDBLOCK);
Keith Buschc76541a2014-12-19 17:54:13 -0700264 }
Jens Axboe320ae512013-10-24 09:20:05 +0100265 return rq;
266}
Jens Axboe4bb659b2014-05-09 09:36:49 -0600267EXPORT_SYMBOL(blk_mq_alloc_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100268
Jens Axboe320ae512013-10-24 09:20:05 +0100269static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
270 struct blk_mq_ctx *ctx, struct request *rq)
271{
272 const int tag = rq->tag;
273 struct request_queue *q = rq->q;
274
Jens Axboe0d2602c2014-05-13 15:10:52 -0600275 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
276 atomic_dec(&hctx->nr_active);
David Hildenbrand683d0e12014-09-18 11:04:31 +0200277 rq->cmd_flags = 0;
Jens Axboe0d2602c2014-05-13 15:10:52 -0600278
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200279 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
Jens Axboe0d2602c2014-05-13 15:10:52 -0600280 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
Dan Williams3ef28e82015-10-21 13:20:12 -0400281 blk_queue_exit(q);
Jens Axboe320ae512013-10-24 09:20:05 +0100282}
283
Jens Axboe7c7f2f22014-11-17 10:41:57 -0700284void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100285{
286 struct blk_mq_ctx *ctx = rq->mq_ctx;
Jens Axboe7c7f2f22014-11-17 10:41:57 -0700287
288 ctx->rq_completed[rq_is_sync(rq)]++;
289 __blk_mq_free_request(hctx, ctx, rq);
290
291}
292EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
293
294void blk_mq_free_request(struct request *rq)
295{
Jens Axboe320ae512013-10-24 09:20:05 +0100296 struct blk_mq_hw_ctx *hctx;
297 struct request_queue *q = rq->q;
298
Jens Axboe7c7f2f22014-11-17 10:41:57 -0700299 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
300 blk_mq_free_hctx_request(hctx, rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100301}
Jens Axboe1a3b5952014-11-17 10:40:48 -0700302EXPORT_SYMBOL_GPL(blk_mq_free_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100303
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700304inline void __blk_mq_end_request(struct request *rq, int error)
Jens Axboe320ae512013-10-24 09:20:05 +0100305{
Ming Lei0d11e6a2013-12-05 10:50:39 -0700306 blk_account_io_done(rq);
307
Christoph Hellwig91b63632014-04-16 09:44:53 +0200308 if (rq->end_io) {
Jens Axboe320ae512013-10-24 09:20:05 +0100309 rq->end_io(rq, error);
Christoph Hellwig91b63632014-04-16 09:44:53 +0200310 } else {
311 if (unlikely(blk_bidi_rq(rq)))
312 blk_mq_free_request(rq->next_rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100313 blk_mq_free_request(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +0200314 }
Jens Axboe320ae512013-10-24 09:20:05 +0100315}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700316EXPORT_SYMBOL(__blk_mq_end_request);
Christoph Hellwig63151a42014-04-16 09:44:52 +0200317
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700318void blk_mq_end_request(struct request *rq, int error)
Christoph Hellwig63151a42014-04-16 09:44:52 +0200319{
320 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
321 BUG();
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700322 __blk_mq_end_request(rq, error);
Christoph Hellwig63151a42014-04-16 09:44:52 +0200323}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700324EXPORT_SYMBOL(blk_mq_end_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100325
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800326static void __blk_mq_complete_request_remote(void *data)
Jens Axboe320ae512013-10-24 09:20:05 +0100327{
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800328 struct request *rq = data;
Jens Axboe320ae512013-10-24 09:20:05 +0100329
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800330 rq->q->softirq_done_fn(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100331}
332
Jens Axboeed851862014-05-30 21:20:50 -0600333static void blk_mq_ipi_complete_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100334{
335 struct blk_mq_ctx *ctx = rq->mq_ctx;
Christoph Hellwig38535202014-04-25 02:32:53 -0700336 bool shared = false;
Jens Axboe320ae512013-10-24 09:20:05 +0100337 int cpu;
338
Christoph Hellwig38535202014-04-25 02:32:53 -0700339 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800340 rq->q->softirq_done_fn(rq);
341 return;
342 }
Jens Axboe320ae512013-10-24 09:20:05 +0100343
344 cpu = get_cpu();
Christoph Hellwig38535202014-04-25 02:32:53 -0700345 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
346 shared = cpus_share_cache(cpu, ctx->cpu);
347
348 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800349 rq->csd.func = __blk_mq_complete_request_remote;
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800350 rq->csd.info = rq;
351 rq->csd.flags = 0;
Frederic Weisbeckerc46fff22014-02-24 16:40:02 +0100352 smp_call_function_single_async(ctx->cpu, &rq->csd);
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800353 } else {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800354 rq->q->softirq_done_fn(rq);
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800355 }
Jens Axboe320ae512013-10-24 09:20:05 +0100356 put_cpu();
357}
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800358
Jens Axboe1fa8cc52015-11-05 14:32:55 -0700359static void __blk_mq_complete_request(struct request *rq)
Jens Axboeed851862014-05-30 21:20:50 -0600360{
361 struct request_queue *q = rq->q;
362
363 if (!q->softirq_done_fn)
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700364 blk_mq_end_request(rq, rq->errors);
Jens Axboeed851862014-05-30 21:20:50 -0600365 else
366 blk_mq_ipi_complete_request(rq);
367}
368
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800369/**
370 * blk_mq_complete_request - end I/O on a request
371 * @rq: the request being processed
372 *
373 * Description:
374 * Ends all I/O on a request. It does not handle partial completions.
375 * The actual completion happens out-of-order, through a IPI handler.
376 **/
Christoph Hellwigf4829a92015-09-27 21:01:50 +0200377void blk_mq_complete_request(struct request *rq, int error)
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800378{
Jens Axboe95f09682014-05-27 17:46:48 -0600379 struct request_queue *q = rq->q;
380
381 if (unlikely(blk_should_fake_timeout(q)))
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800382 return;
Christoph Hellwigf4829a92015-09-27 21:01:50 +0200383 if (!blk_mark_rq_complete(rq)) {
384 rq->errors = error;
Jens Axboeed851862014-05-30 21:20:50 -0600385 __blk_mq_complete_request(rq);
Christoph Hellwigf4829a92015-09-27 21:01:50 +0200386 }
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800387}
388EXPORT_SYMBOL(blk_mq_complete_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100389
Keith Busch973c0192015-01-07 18:55:43 -0700390int blk_mq_request_started(struct request *rq)
391{
392 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
393}
394EXPORT_SYMBOL_GPL(blk_mq_request_started);
395
Christoph Hellwige2490072014-09-13 16:40:09 -0700396void blk_mq_start_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100397{
398 struct request_queue *q = rq->q;
399
400 trace_block_rq_issue(q, rq);
401
Christoph Hellwig742ee692014-04-14 10:30:06 +0200402 rq->resid_len = blk_rq_bytes(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +0200403 if (unlikely(blk_bidi_rq(rq)))
404 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
Christoph Hellwig742ee692014-04-14 10:30:06 +0200405
Ming Lei2b8393b2014-06-10 00:16:41 +0800406 blk_add_timer(rq);
Jens Axboe87ee7b12014-04-24 08:51:47 -0600407
408 /*
Jens Axboe538b7532014-09-16 10:37:37 -0600409 * Ensure that ->deadline is visible before set the started
410 * flag and clear the completed flag.
411 */
412 smp_mb__before_atomic();
413
414 /*
Jens Axboe87ee7b12014-04-24 08:51:47 -0600415 * Mark us as started and clear complete. Complete might have been
416 * set if requeue raced with timeout, which then marked it as
417 * complete. So be sure to clear complete again when we start
418 * the request, otherwise we'll ignore the completion event.
419 */
Jens Axboe4b570522014-05-29 11:00:11 -0600420 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
421 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
422 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
423 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800424
425 if (q->dma_drain_size && blk_rq_bytes(rq)) {
426 /*
427 * Make sure space for the drain appears. We know we can do
428 * this because max_hw_segments has been adjusted to be one
429 * fewer than the device can handle.
430 */
431 rq->nr_phys_segments++;
432 }
Jens Axboe320ae512013-10-24 09:20:05 +0100433}
Christoph Hellwige2490072014-09-13 16:40:09 -0700434EXPORT_SYMBOL(blk_mq_start_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100435
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200436static void __blk_mq_requeue_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100437{
438 struct request_queue *q = rq->q;
439
440 trace_block_rq_requeue(q, rq);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800441
Christoph Hellwige2490072014-09-13 16:40:09 -0700442 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
443 if (q->dma_drain_size && blk_rq_bytes(rq))
444 rq->nr_phys_segments--;
445 }
Jens Axboe320ae512013-10-24 09:20:05 +0100446}
447
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200448void blk_mq_requeue_request(struct request *rq)
449{
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200450 __blk_mq_requeue_request(rq);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200451
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200452 BUG_ON(blk_queued_rq(rq));
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600453 blk_mq_add_to_requeue_list(rq, true);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200454}
455EXPORT_SYMBOL(blk_mq_requeue_request);
456
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600457static void blk_mq_requeue_work(struct work_struct *work)
458{
459 struct request_queue *q =
460 container_of(work, struct request_queue, requeue_work);
461 LIST_HEAD(rq_list);
462 struct request *rq, *next;
463 unsigned long flags;
464
465 spin_lock_irqsave(&q->requeue_lock, flags);
466 list_splice_init(&q->requeue_list, &rq_list);
467 spin_unlock_irqrestore(&q->requeue_lock, flags);
468
469 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
470 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
471 continue;
472
473 rq->cmd_flags &= ~REQ_SOFTBARRIER;
474 list_del_init(&rq->queuelist);
475 blk_mq_insert_request(rq, true, false, false);
476 }
477
478 while (!list_empty(&rq_list)) {
479 rq = list_entry(rq_list.next, struct request, queuelist);
480 list_del_init(&rq->queuelist);
481 blk_mq_insert_request(rq, false, false, false);
482 }
483
Jens Axboe8b957412014-09-19 13:10:29 -0600484 /*
485 * Use the start variant of queue running here, so that running
486 * the requeue work will kick stopped queues.
487 */
488 blk_mq_start_hw_queues(q);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600489}
490
491void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
492{
493 struct request_queue *q = rq->q;
494 unsigned long flags;
495
496 /*
497 * We abuse this flag that is otherwise used by the I/O scheduler to
498 * request head insertation from the workqueue.
499 */
500 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
501
502 spin_lock_irqsave(&q->requeue_lock, flags);
503 if (at_head) {
504 rq->cmd_flags |= REQ_SOFTBARRIER;
505 list_add(&rq->queuelist, &q->requeue_list);
506 } else {
507 list_add_tail(&rq->queuelist, &q->requeue_list);
508 }
509 spin_unlock_irqrestore(&q->requeue_lock, flags);
510}
511EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
512
Keith Buschc68ed592015-01-07 18:55:44 -0700513void blk_mq_cancel_requeue_work(struct request_queue *q)
514{
515 cancel_work_sync(&q->requeue_work);
516}
517EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
518
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600519void blk_mq_kick_requeue_list(struct request_queue *q)
520{
521 kblockd_schedule_work(&q->requeue_work);
522}
523EXPORT_SYMBOL(blk_mq_kick_requeue_list);
524
Jens Axboe1885b242015-01-07 18:55:45 -0700525void blk_mq_abort_requeue_list(struct request_queue *q)
526{
527 unsigned long flags;
528 LIST_HEAD(rq_list);
529
530 spin_lock_irqsave(&q->requeue_lock, flags);
531 list_splice_init(&q->requeue_list, &rq_list);
532 spin_unlock_irqrestore(&q->requeue_lock, flags);
533
534 while (!list_empty(&rq_list)) {
535 struct request *rq;
536
537 rq = list_first_entry(&rq_list, struct request, queuelist);
538 list_del_init(&rq->queuelist);
539 rq->errors = -EIO;
540 blk_mq_end_request(rq, rq->errors);
541 }
542}
543EXPORT_SYMBOL(blk_mq_abort_requeue_list);
544
Jens Axboe0e62f512014-06-04 10:23:49 -0600545struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
546{
Hannes Reinecke4ee86ba2016-03-15 12:03:28 -0700547 if (tag < tags->nr_tags)
548 return tags->rqs[tag];
549
550 return NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -0600551}
552EXPORT_SYMBOL(blk_mq_tag_to_rq);
553
Jens Axboe320ae512013-10-24 09:20:05 +0100554struct blk_mq_timeout_data {
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700555 unsigned long next;
556 unsigned int next_set;
Jens Axboe320ae512013-10-24 09:20:05 +0100557};
558
Christoph Hellwig90415832014-09-22 10:21:48 -0600559void blk_mq_rq_timed_out(struct request *req, bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100560{
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700561 struct blk_mq_ops *ops = req->q->mq_ops;
562 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
Jens Axboe87ee7b12014-04-24 08:51:47 -0600563
564 /*
565 * We know that complete is set at this point. If STARTED isn't set
566 * anymore, then the request isn't active and the "timeout" should
567 * just be ignored. This can happen due to the bitflag ordering.
568 * Timeout first checks if STARTED is set, and if it is, assumes
569 * the request is active. But if we race with completion, then
570 * we both flags will get cleared. So check here again, and ignore
571 * a timeout event with a request that isn't active.
572 */
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700573 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
574 return;
Jens Axboe87ee7b12014-04-24 08:51:47 -0600575
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700576 if (ops->timeout)
Christoph Hellwig0152fb62014-09-13 16:40:13 -0700577 ret = ops->timeout(req, reserved);
Jens Axboe87ee7b12014-04-24 08:51:47 -0600578
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700579 switch (ret) {
580 case BLK_EH_HANDLED:
581 __blk_mq_complete_request(req);
582 break;
583 case BLK_EH_RESET_TIMER:
584 blk_add_timer(req);
585 blk_clear_rq_complete(req);
586 break;
587 case BLK_EH_NOT_HANDLED:
588 break;
589 default:
590 printk(KERN_ERR "block: bad eh return: %d\n", ret);
591 break;
592 }
Jens Axboe87ee7b12014-04-24 08:51:47 -0600593}
Keith Busch5b3f25f2015-01-07 18:55:46 -0700594
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700595static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
596 struct request *rq, void *priv, bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100597{
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700598 struct blk_mq_timeout_data *data = priv;
599
Keith Buscheb130db2015-01-08 08:59:53 -0700600 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
601 /*
602 * If a request wasn't started before the queue was
603 * marked dying, kill it here or it'll go unnoticed.
604 */
Keith Buscha59e0f52016-02-11 13:05:38 -0700605 if (unlikely(blk_queue_dying(rq->q))) {
606 rq->errors = -EIO;
607 blk_mq_end_request(rq, rq->errors);
608 }
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700609 return;
Keith Buscheb130db2015-01-08 08:59:53 -0700610 }
Jens Axboe320ae512013-10-24 09:20:05 +0100611
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700612 if (time_after_eq(jiffies, rq->deadline)) {
613 if (!blk_mark_rq_complete(rq))
Christoph Hellwig0152fb62014-09-13 16:40:13 -0700614 blk_mq_rq_timed_out(rq, reserved);
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700615 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
616 data->next = rq->deadline;
617 data->next_set = 1;
618 }
Jens Axboe320ae512013-10-24 09:20:05 +0100619}
620
Christoph Hellwig287922e2015-10-30 20:57:30 +0800621static void blk_mq_timeout_work(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +0100622{
Christoph Hellwig287922e2015-10-30 20:57:30 +0800623 struct request_queue *q =
624 container_of(work, struct request_queue, timeout_work);
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700625 struct blk_mq_timeout_data data = {
626 .next = 0,
627 .next_set = 0,
628 };
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700629 int i;
Jens Axboe320ae512013-10-24 09:20:05 +0100630
Christoph Hellwig287922e2015-10-30 20:57:30 +0800631 if (blk_queue_enter(q, true))
632 return;
633
Christoph Hellwig0bf6cd52015-09-27 21:01:51 +0200634 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
Jens Axboe320ae512013-10-24 09:20:05 +0100635
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700636 if (data.next_set) {
637 data.next = blk_rq_timeout(round_jiffies_up(data.next));
638 mod_timer(&q->timeout, data.next);
Jens Axboe0d2602c2014-05-13 15:10:52 -0600639 } else {
Christoph Hellwig0bf6cd52015-09-27 21:01:51 +0200640 struct blk_mq_hw_ctx *hctx;
641
Ming Leif054b562015-04-21 10:00:19 +0800642 queue_for_each_hw_ctx(q, hctx, i) {
643 /* the hctx may be unmapped, so check it here */
644 if (blk_mq_hw_queue_mapped(hctx))
645 blk_mq_tag_idle(hctx);
646 }
Jens Axboe0d2602c2014-05-13 15:10:52 -0600647 }
Christoph Hellwig287922e2015-10-30 20:57:30 +0800648 blk_queue_exit(q);
Jens Axboe320ae512013-10-24 09:20:05 +0100649}
650
651/*
652 * Reverse check our software queue for entries that we could potentially
653 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
654 * too much time checking for merges.
655 */
656static bool blk_mq_attempt_merge(struct request_queue *q,
657 struct blk_mq_ctx *ctx, struct bio *bio)
658{
659 struct request *rq;
660 int checked = 8;
661
662 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
663 int el_ret;
664
665 if (!checked--)
666 break;
667
668 if (!blk_rq_merge_ok(rq, bio))
669 continue;
670
671 el_ret = blk_try_merge(rq, bio);
672 if (el_ret == ELEVATOR_BACK_MERGE) {
673 if (bio_attempt_back_merge(q, rq, bio)) {
674 ctx->rq_merged++;
675 return true;
676 }
677 break;
678 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
679 if (bio_attempt_front_merge(q, rq, bio)) {
680 ctx->rq_merged++;
681 return true;
682 }
683 break;
684 }
685 }
686
687 return false;
688}
689
Jens Axboe320ae512013-10-24 09:20:05 +0100690/*
Jens Axboe1429d7c2014-05-19 09:23:55 -0600691 * Process software queues that have been marked busy, splicing them
692 * to the for-dispatch
693 */
694static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
695{
696 struct blk_mq_ctx *ctx;
697 int i;
698
Jens Axboe569fd0c2015-04-17 08:28:50 -0600699 for (i = 0; i < hctx->ctx_map.size; i++) {
Jens Axboe1429d7c2014-05-19 09:23:55 -0600700 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
701 unsigned int off, bit;
702
703 if (!bm->word)
704 continue;
705
706 bit = 0;
707 off = i * hctx->ctx_map.bits_per_word;
708 do {
709 bit = find_next_bit(&bm->word, bm->depth, bit);
710 if (bit >= bm->depth)
711 break;
712
713 ctx = hctx->ctxs[bit + off];
714 clear_bit(bit, &bm->word);
715 spin_lock(&ctx->lock);
716 list_splice_tail_init(&ctx->rq_list, list);
717 spin_unlock(&ctx->lock);
718
719 bit++;
720 } while (1);
721 }
722}
723
724/*
Jens Axboe320ae512013-10-24 09:20:05 +0100725 * Run this hardware queue, pulling any software queues mapped to it in.
726 * Note that this function currently has various problems around ordering
727 * of IO. In particular, we'd like FIFO behaviour on handling existing
728 * items on the hctx->dispatch list. Ignore that for now.
729 */
730static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
731{
732 struct request_queue *q = hctx->queue;
Jens Axboe320ae512013-10-24 09:20:05 +0100733 struct request *rq;
734 LIST_HEAD(rq_list);
Jens Axboe74c45052014-10-29 11:14:52 -0600735 LIST_HEAD(driver_list);
736 struct list_head *dptr;
Jens Axboe1429d7c2014-05-19 09:23:55 -0600737 int queued;
Jens Axboe320ae512013-10-24 09:20:05 +0100738
Jens Axboefd1270d2014-04-16 09:23:48 -0600739 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
Jens Axboee4043dc2014-04-09 10:18:23 -0600740
Jens Axboe5d12f902014-03-19 15:25:02 -0600741 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
Jens Axboe320ae512013-10-24 09:20:05 +0100742 return;
743
744 hctx->run++;
745
746 /*
747 * Touch any software queue that has pending entries.
748 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600749 flush_busy_ctxs(hctx, &rq_list);
Jens Axboe320ae512013-10-24 09:20:05 +0100750
751 /*
752 * If we have previous entries on our dispatch list, grab them
753 * and stuff them at the front for more fair dispatch.
754 */
755 if (!list_empty_careful(&hctx->dispatch)) {
756 spin_lock(&hctx->lock);
757 if (!list_empty(&hctx->dispatch))
758 list_splice_init(&hctx->dispatch, &rq_list);
759 spin_unlock(&hctx->lock);
760 }
761
762 /*
Jens Axboe74c45052014-10-29 11:14:52 -0600763 * Start off with dptr being NULL, so we start the first request
764 * immediately, even if we have more pending.
765 */
766 dptr = NULL;
767
768 /*
Jens Axboe320ae512013-10-24 09:20:05 +0100769 * Now process all the entries, sending them to the driver.
770 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600771 queued = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100772 while (!list_empty(&rq_list)) {
Jens Axboe74c45052014-10-29 11:14:52 -0600773 struct blk_mq_queue_data bd;
Jens Axboe320ae512013-10-24 09:20:05 +0100774 int ret;
775
776 rq = list_first_entry(&rq_list, struct request, queuelist);
777 list_del_init(&rq->queuelist);
Jens Axboe320ae512013-10-24 09:20:05 +0100778
Jens Axboe74c45052014-10-29 11:14:52 -0600779 bd.rq = rq;
780 bd.list = dptr;
781 bd.last = list_empty(&rq_list);
782
783 ret = q->mq_ops->queue_rq(hctx, &bd);
Jens Axboe320ae512013-10-24 09:20:05 +0100784 switch (ret) {
785 case BLK_MQ_RQ_QUEUE_OK:
786 queued++;
787 continue;
788 case BLK_MQ_RQ_QUEUE_BUSY:
Jens Axboe320ae512013-10-24 09:20:05 +0100789 list_add(&rq->queuelist, &rq_list);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200790 __blk_mq_requeue_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100791 break;
792 default:
793 pr_err("blk-mq: bad return on queue: %d\n", ret);
Jens Axboe320ae512013-10-24 09:20:05 +0100794 case BLK_MQ_RQ_QUEUE_ERROR:
Christoph Hellwig1e93b8c2014-02-11 08:27:13 -0800795 rq->errors = -EIO;
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700796 blk_mq_end_request(rq, rq->errors);
Jens Axboe320ae512013-10-24 09:20:05 +0100797 break;
798 }
799
800 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
801 break;
Jens Axboe74c45052014-10-29 11:14:52 -0600802
803 /*
804 * We've done the first request. If we have more than 1
805 * left in the list, set dptr to defer issue.
806 */
807 if (!dptr && rq_list.next != rq_list.prev)
808 dptr = &driver_list;
Jens Axboe320ae512013-10-24 09:20:05 +0100809 }
810
811 if (!queued)
812 hctx->dispatched[0]++;
813 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
814 hctx->dispatched[ilog2(queued) + 1]++;
815
816 /*
817 * Any items that need requeuing? Stuff them into hctx->dispatch,
818 * that is where we will continue on next queue run.
819 */
820 if (!list_empty(&rq_list)) {
821 spin_lock(&hctx->lock);
822 list_splice(&rq_list, &hctx->dispatch);
823 spin_unlock(&hctx->lock);
Shaohua Li9ba52e52015-05-04 14:32:48 -0600824 /*
825 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
826 * it's possible the queue is stopped and restarted again
827 * before this. Queue restart will dispatch requests. And since
828 * requests in rq_list aren't added into hctx->dispatch yet,
829 * the requests in rq_list might get lost.
830 *
831 * blk_mq_run_hw_queue() already checks the STOPPED bit
832 **/
833 blk_mq_run_hw_queue(hctx, true);
Jens Axboe320ae512013-10-24 09:20:05 +0100834 }
835}
836
Jens Axboe506e9312014-05-07 10:26:44 -0600837/*
838 * It'd be great if the workqueue API had a way to pass
839 * in a mask and had some smarts for more clever placement.
840 * For now we just round-robin here, switching for every
841 * BLK_MQ_CPU_WORK_BATCH queued items.
842 */
843static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
844{
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100845 if (hctx->queue->nr_hw_queues == 1)
846 return WORK_CPU_UNBOUND;
Jens Axboe506e9312014-05-07 10:26:44 -0600847
848 if (--hctx->next_cpu_batch <= 0) {
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100849 int cpu = hctx->next_cpu, next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600850
851 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
852 if (next_cpu >= nr_cpu_ids)
853 next_cpu = cpumask_first(hctx->cpumask);
854
855 hctx->next_cpu = next_cpu;
856 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100857
858 return cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600859 }
860
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100861 return hctx->next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600862}
863
Jens Axboe320ae512013-10-24 09:20:05 +0100864void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
865{
Ming Lei19c66e52014-12-03 19:38:04 +0800866 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
867 !blk_mq_hw_queue_mapped(hctx)))
Jens Axboe320ae512013-10-24 09:20:05 +0100868 return;
869
Paolo Bonzini398205b2014-11-07 23:03:59 +0100870 if (!async) {
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100871 int cpu = get_cpu();
872 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
Paolo Bonzini398205b2014-11-07 23:03:59 +0100873 __blk_mq_run_hw_queue(hctx);
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100874 put_cpu();
Paolo Bonzini398205b2014-11-07 23:03:59 +0100875 return;
876 }
Jens Axboee4043dc2014-04-09 10:18:23 -0600877
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100878 put_cpu();
Jens Axboee4043dc2014-04-09 10:18:23 -0600879 }
Paolo Bonzini398205b2014-11-07 23:03:59 +0100880
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100881 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
882 &hctx->run_work, 0);
Jens Axboe320ae512013-10-24 09:20:05 +0100883}
884
Mike Snitzerb94ec292015-03-11 23:56:38 -0400885void blk_mq_run_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +0100886{
887 struct blk_mq_hw_ctx *hctx;
888 int i;
889
890 queue_for_each_hw_ctx(q, hctx, i) {
891 if ((!blk_mq_hctx_has_pending(hctx) &&
892 list_empty_careful(&hctx->dispatch)) ||
Jens Axboe5d12f902014-03-19 15:25:02 -0600893 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
Jens Axboe320ae512013-10-24 09:20:05 +0100894 continue;
895
Mike Snitzerb94ec292015-03-11 23:56:38 -0400896 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +0100897 }
898}
Mike Snitzerb94ec292015-03-11 23:56:38 -0400899EXPORT_SYMBOL(blk_mq_run_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +0100900
901void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
902{
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600903 cancel_delayed_work(&hctx->run_work);
904 cancel_delayed_work(&hctx->delay_work);
Jens Axboe320ae512013-10-24 09:20:05 +0100905 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
906}
907EXPORT_SYMBOL(blk_mq_stop_hw_queue);
908
Christoph Hellwig280d45f2013-10-25 14:45:58 +0100909void blk_mq_stop_hw_queues(struct request_queue *q)
910{
911 struct blk_mq_hw_ctx *hctx;
912 int i;
913
914 queue_for_each_hw_ctx(q, hctx, i)
915 blk_mq_stop_hw_queue(hctx);
916}
917EXPORT_SYMBOL(blk_mq_stop_hw_queues);
918
Jens Axboe320ae512013-10-24 09:20:05 +0100919void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
920{
921 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboee4043dc2014-04-09 10:18:23 -0600922
Jens Axboe0ffbce82014-06-25 08:22:34 -0600923 blk_mq_run_hw_queue(hctx, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100924}
925EXPORT_SYMBOL(blk_mq_start_hw_queue);
926
Christoph Hellwig2f268552014-04-16 09:44:56 +0200927void blk_mq_start_hw_queues(struct request_queue *q)
928{
929 struct blk_mq_hw_ctx *hctx;
930 int i;
931
932 queue_for_each_hw_ctx(q, hctx, i)
933 blk_mq_start_hw_queue(hctx);
934}
935EXPORT_SYMBOL(blk_mq_start_hw_queues);
936
Christoph Hellwig1b4a3252014-04-16 09:44:54 +0200937void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +0100938{
939 struct blk_mq_hw_ctx *hctx;
940 int i;
941
942 queue_for_each_hw_ctx(q, hctx, i) {
943 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
944 continue;
945
946 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Christoph Hellwig1b4a3252014-04-16 09:44:54 +0200947 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +0100948 }
949}
950EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
951
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600952static void blk_mq_run_work_fn(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +0100953{
954 struct blk_mq_hw_ctx *hctx;
955
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600956 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
Jens Axboee4043dc2014-04-09 10:18:23 -0600957
Jens Axboe320ae512013-10-24 09:20:05 +0100958 __blk_mq_run_hw_queue(hctx);
959}
960
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600961static void blk_mq_delay_work_fn(struct work_struct *work)
962{
963 struct blk_mq_hw_ctx *hctx;
964
965 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
966
967 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
968 __blk_mq_run_hw_queue(hctx);
969}
970
971void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
972{
Ming Lei19c66e52014-12-03 19:38:04 +0800973 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
974 return;
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600975
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100976 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
977 &hctx->delay_work, msecs_to_jiffies(msecs));
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600978}
979EXPORT_SYMBOL(blk_mq_delay_queue);
980
Ming Leicfd0c552015-10-20 23:13:57 +0800981static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
982 struct blk_mq_ctx *ctx,
983 struct request *rq,
984 bool at_head)
Jens Axboe320ae512013-10-24 09:20:05 +0100985{
Jens Axboe01b983c2013-11-19 18:59:10 -0700986 trace_block_rq_insert(hctx->queue, rq);
987
Christoph Hellwig72a0a362014-02-07 10:22:36 -0800988 if (at_head)
989 list_add(&rq->queuelist, &ctx->rq_list);
990 else
991 list_add_tail(&rq->queuelist, &ctx->rq_list);
Ming Leicfd0c552015-10-20 23:13:57 +0800992}
Jens Axboe4bb659b2014-05-09 09:36:49 -0600993
Ming Leicfd0c552015-10-20 23:13:57 +0800994static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
995 struct request *rq, bool at_head)
996{
997 struct blk_mq_ctx *ctx = rq->mq_ctx;
998
999 __blk_mq_insert_req_list(hctx, ctx, rq, at_head);
Jens Axboe320ae512013-10-24 09:20:05 +01001000 blk_mq_hctx_mark_pending(hctx, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001001}
1002
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001003void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1004 bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01001005{
1006 struct request_queue *q = rq->q;
1007 struct blk_mq_hw_ctx *hctx;
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001008 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001009
1010 current_ctx = blk_mq_get_ctx(q);
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001011 if (!cpu_online(ctx->cpu))
1012 rq->mq_ctx = ctx = current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001013
Jens Axboe320ae512013-10-24 09:20:05 +01001014 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1015
Christoph Hellwiga57a1782014-09-16 14:44:07 -07001016 spin_lock(&ctx->lock);
1017 __blk_mq_insert_request(hctx, rq, at_head);
1018 spin_unlock(&ctx->lock);
Jens Axboe320ae512013-10-24 09:20:05 +01001019
Jens Axboe320ae512013-10-24 09:20:05 +01001020 if (run_queue)
1021 blk_mq_run_hw_queue(hctx, async);
Jens Axboee4043dc2014-04-09 10:18:23 -06001022
1023 blk_mq_put_ctx(current_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001024}
1025
1026static void blk_mq_insert_requests(struct request_queue *q,
1027 struct blk_mq_ctx *ctx,
1028 struct list_head *list,
1029 int depth,
1030 bool from_schedule)
1031
1032{
1033 struct blk_mq_hw_ctx *hctx;
1034 struct blk_mq_ctx *current_ctx;
1035
1036 trace_block_unplug(q, depth, !from_schedule);
1037
1038 current_ctx = blk_mq_get_ctx(q);
1039
1040 if (!cpu_online(ctx->cpu))
1041 ctx = current_ctx;
1042 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1043
1044 /*
1045 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1046 * offline now
1047 */
1048 spin_lock(&ctx->lock);
1049 while (!list_empty(list)) {
1050 struct request *rq;
1051
1052 rq = list_first_entry(list, struct request, queuelist);
1053 list_del_init(&rq->queuelist);
1054 rq->mq_ctx = ctx;
Ming Leicfd0c552015-10-20 23:13:57 +08001055 __blk_mq_insert_req_list(hctx, ctx, rq, false);
Jens Axboe320ae512013-10-24 09:20:05 +01001056 }
Ming Leicfd0c552015-10-20 23:13:57 +08001057 blk_mq_hctx_mark_pending(hctx, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001058 spin_unlock(&ctx->lock);
1059
Jens Axboe320ae512013-10-24 09:20:05 +01001060 blk_mq_run_hw_queue(hctx, from_schedule);
Jens Axboee4043dc2014-04-09 10:18:23 -06001061 blk_mq_put_ctx(current_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001062}
1063
1064static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1065{
1066 struct request *rqa = container_of(a, struct request, queuelist);
1067 struct request *rqb = container_of(b, struct request, queuelist);
1068
1069 return !(rqa->mq_ctx < rqb->mq_ctx ||
1070 (rqa->mq_ctx == rqb->mq_ctx &&
1071 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1072}
1073
1074void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1075{
1076 struct blk_mq_ctx *this_ctx;
1077 struct request_queue *this_q;
1078 struct request *rq;
1079 LIST_HEAD(list);
1080 LIST_HEAD(ctx_list);
1081 unsigned int depth;
1082
1083 list_splice_init(&plug->mq_list, &list);
1084
1085 list_sort(NULL, &list, plug_ctx_cmp);
1086
1087 this_q = NULL;
1088 this_ctx = NULL;
1089 depth = 0;
1090
1091 while (!list_empty(&list)) {
1092 rq = list_entry_rq(list.next);
1093 list_del_init(&rq->queuelist);
1094 BUG_ON(!rq->q);
1095 if (rq->mq_ctx != this_ctx) {
1096 if (this_ctx) {
1097 blk_mq_insert_requests(this_q, this_ctx,
1098 &ctx_list, depth,
1099 from_schedule);
1100 }
1101
1102 this_ctx = rq->mq_ctx;
1103 this_q = rq->q;
1104 depth = 0;
1105 }
1106
1107 depth++;
1108 list_add_tail(&rq->queuelist, &ctx_list);
1109 }
1110
1111 /*
1112 * If 'this_ctx' is set, we know we have entries to complete
1113 * on 'ctx_list'. Do those.
1114 */
1115 if (this_ctx) {
1116 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1117 from_schedule);
1118 }
1119}
1120
1121static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1122{
1123 init_request_from_bio(rq, bio);
Jens Axboe4b570522014-05-29 11:00:11 -06001124
Jens Axboe3ee32372014-06-09 09:36:53 -06001125 if (blk_do_io_stat(rq))
Jens Axboe4b570522014-05-29 11:00:11 -06001126 blk_account_io_start(rq, 1);
Jens Axboe320ae512013-10-24 09:20:05 +01001127}
1128
Jens Axboe274a5842014-08-15 12:44:08 -06001129static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1130{
1131 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1132 !blk_queue_nomerges(hctx->queue);
1133}
1134
Jens Axboe07068d52014-05-22 10:40:51 -06001135static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1136 struct blk_mq_ctx *ctx,
1137 struct request *rq, struct bio *bio)
1138{
Ming Leie18378a2015-10-20 23:13:54 +08001139 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
Jens Axboe07068d52014-05-22 10:40:51 -06001140 blk_mq_bio_to_request(rq, bio);
1141 spin_lock(&ctx->lock);
1142insert_rq:
1143 __blk_mq_insert_request(hctx, rq, false);
1144 spin_unlock(&ctx->lock);
1145 return false;
1146 } else {
Jens Axboe274a5842014-08-15 12:44:08 -06001147 struct request_queue *q = hctx->queue;
1148
Jens Axboe07068d52014-05-22 10:40:51 -06001149 spin_lock(&ctx->lock);
1150 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1151 blk_mq_bio_to_request(rq, bio);
1152 goto insert_rq;
1153 }
1154
1155 spin_unlock(&ctx->lock);
1156 __blk_mq_free_request(hctx, ctx, rq);
1157 return true;
1158 }
1159}
1160
1161struct blk_map_ctx {
1162 struct blk_mq_hw_ctx *hctx;
1163 struct blk_mq_ctx *ctx;
1164};
1165
1166static struct request *blk_mq_map_request(struct request_queue *q,
1167 struct bio *bio,
1168 struct blk_map_ctx *data)
Jens Axboe320ae512013-10-24 09:20:05 +01001169{
1170 struct blk_mq_hw_ctx *hctx;
1171 struct blk_mq_ctx *ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001172 struct request *rq;
Jens Axboe07068d52014-05-22 10:40:51 -06001173 int rw = bio_data_dir(bio);
Ming Leicb96a422014-06-01 00:43:37 +08001174 struct blk_mq_alloc_data alloc_data;
Jens Axboe320ae512013-10-24 09:20:05 +01001175
Dan Williams3ef28e82015-10-21 13:20:12 -04001176 blk_queue_enter_live(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001177 ctx = blk_mq_get_ctx(q);
1178 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1179
Jens Axboe07068d52014-05-22 10:40:51 -06001180 if (rw_is_sync(bio->bi_rw))
Shaohua Li27fbf4e2014-02-19 20:20:21 +08001181 rw |= REQ_SYNC;
Jens Axboe07068d52014-05-22 10:40:51 -06001182
Jens Axboe320ae512013-10-24 09:20:05 +01001183 trace_block_getrq(q, bio, rw);
Christoph Hellwig6f3b0e82015-11-26 09:13:05 +01001184 blk_mq_set_alloc_data(&alloc_data, q, BLK_MQ_REQ_NOWAIT, ctx, hctx);
Ming Leicb96a422014-06-01 00:43:37 +08001185 rq = __blk_mq_alloc_request(&alloc_data, rw);
Christoph Hellwig5dee8572014-05-27 20:59:47 +02001186 if (unlikely(!rq)) {
Christoph Hellwig793597a2014-05-27 20:59:49 +02001187 __blk_mq_run_hw_queue(hctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001188 blk_mq_put_ctx(ctx);
1189 trace_block_sleeprq(q, bio, rw);
Christoph Hellwig793597a2014-05-27 20:59:49 +02001190
1191 ctx = blk_mq_get_ctx(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001192 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Christoph Hellwig6f3b0e82015-11-26 09:13:05 +01001193 blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
Ming Leicb96a422014-06-01 00:43:37 +08001194 rq = __blk_mq_alloc_request(&alloc_data, rw);
1195 ctx = alloc_data.ctx;
1196 hctx = alloc_data.hctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001197 }
1198
1199 hctx->queued++;
Jens Axboe07068d52014-05-22 10:40:51 -06001200 data->hctx = hctx;
1201 data->ctx = ctx;
1202 return rq;
1203}
1204
Jens Axboe7b371632015-11-05 10:41:40 -07001205static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
Shaohua Lif984df12015-05-08 10:51:32 -07001206{
1207 int ret;
1208 struct request_queue *q = rq->q;
1209 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1210 rq->mq_ctx->cpu);
1211 struct blk_mq_queue_data bd = {
1212 .rq = rq,
1213 .list = NULL,
1214 .last = 1
1215 };
Jens Axboe7b371632015-11-05 10:41:40 -07001216 blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
Shaohua Lif984df12015-05-08 10:51:32 -07001217
1218 /*
1219 * For OK queue, we are done. For error, kill it. Any other
1220 * error (busy), just add it to our list as we previously
1221 * would have done
1222 */
1223 ret = q->mq_ops->queue_rq(hctx, &bd);
Jens Axboe7b371632015-11-05 10:41:40 -07001224 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1225 *cookie = new_cookie;
Shaohua Lif984df12015-05-08 10:51:32 -07001226 return 0;
Shaohua Lif984df12015-05-08 10:51:32 -07001227 }
Jens Axboe7b371632015-11-05 10:41:40 -07001228
1229 __blk_mq_requeue_request(rq);
1230
1231 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1232 *cookie = BLK_QC_T_NONE;
1233 rq->errors = -EIO;
1234 blk_mq_end_request(rq, rq->errors);
1235 return 0;
1236 }
1237
1238 return -1;
Shaohua Lif984df12015-05-08 10:51:32 -07001239}
1240
Jens Axboe07068d52014-05-22 10:40:51 -06001241/*
1242 * Multiple hardware queue variant. This will not use per-process plugs,
1243 * but will attempt to bypass the hctx queueing if we can go straight to
1244 * hardware for SYNC IO.
1245 */
Jens Axboedece1632015-11-05 10:41:16 -07001246static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
Jens Axboe07068d52014-05-22 10:40:51 -06001247{
1248 const int is_sync = rw_is_sync(bio->bi_rw);
1249 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1250 struct blk_map_ctx data;
1251 struct request *rq;
Shaohua Lif984df12015-05-08 10:51:32 -07001252 unsigned int request_count = 0;
1253 struct blk_plug *plug;
Shaohua Li5b3f3412015-05-08 10:51:33 -07001254 struct request *same_queue_rq = NULL;
Jens Axboe7b371632015-11-05 10:41:40 -07001255 blk_qc_t cookie;
Jens Axboe07068d52014-05-22 10:40:51 -06001256
1257 blk_queue_bounce(q, &bio);
1258
1259 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001260 bio_io_error(bio);
Jens Axboedece1632015-11-05 10:41:16 -07001261 return BLK_QC_T_NONE;
Jens Axboe07068d52014-05-22 10:40:51 -06001262 }
1263
Kent Overstreet54efd502015-04-23 22:37:18 -07001264 blk_queue_split(q, &bio, q->bio_split);
1265
Jeff Moyer0809e3a2015-10-20 23:13:51 +08001266 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1267 if (blk_attempt_plug_merge(q, bio, &request_count,
1268 &same_queue_rq))
Jens Axboedece1632015-11-05 10:41:16 -07001269 return BLK_QC_T_NONE;
Jeff Moyer0809e3a2015-10-20 23:13:51 +08001270 } else
1271 request_count = blk_plug_queued_count(q);
Shaohua Lif984df12015-05-08 10:51:32 -07001272
Jens Axboe07068d52014-05-22 10:40:51 -06001273 rq = blk_mq_map_request(q, bio, &data);
1274 if (unlikely(!rq))
Jens Axboedece1632015-11-05 10:41:16 -07001275 return BLK_QC_T_NONE;
Jens Axboe07068d52014-05-22 10:40:51 -06001276
Jens Axboe7b371632015-11-05 10:41:40 -07001277 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
Jens Axboe07068d52014-05-22 10:40:51 -06001278
1279 if (unlikely(is_flush_fua)) {
1280 blk_mq_bio_to_request(rq, bio);
1281 blk_insert_flush(rq);
1282 goto run_queue;
1283 }
1284
Shaohua Lif984df12015-05-08 10:51:32 -07001285 plug = current->plug;
Jens Axboee167dfb2014-10-29 11:18:26 -06001286 /*
1287 * If the driver supports defer issued based on 'last', then
1288 * queue it up like normal since we can potentially save some
1289 * CPU this way.
1290 */
Shaohua Lif984df12015-05-08 10:51:32 -07001291 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1292 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1293 struct request *old_rq = NULL;
Jens Axboe07068d52014-05-22 10:40:51 -06001294
1295 blk_mq_bio_to_request(rq, bio);
Jens Axboe07068d52014-05-22 10:40:51 -06001296
1297 /*
Jens Axboeb094f892015-11-20 20:29:45 -07001298 * We do limited pluging. If the bio can be merged, do that.
Shaohua Lif984df12015-05-08 10:51:32 -07001299 * Otherwise the existing request in the plug list will be
1300 * issued. So the plug list will have one request at most
Jens Axboe07068d52014-05-22 10:40:51 -06001301 */
Shaohua Lif984df12015-05-08 10:51:32 -07001302 if (plug) {
Shaohua Li5b3f3412015-05-08 10:51:33 -07001303 /*
1304 * The plug list might get flushed before this. If that
Jens Axboeb094f892015-11-20 20:29:45 -07001305 * happens, same_queue_rq is invalid and plug list is
1306 * empty
1307 */
Shaohua Li5b3f3412015-05-08 10:51:33 -07001308 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1309 old_rq = same_queue_rq;
Shaohua Lif984df12015-05-08 10:51:32 -07001310 list_del_init(&old_rq->queuelist);
Jens Axboe07068d52014-05-22 10:40:51 -06001311 }
Shaohua Lif984df12015-05-08 10:51:32 -07001312 list_add_tail(&rq->queuelist, &plug->mq_list);
1313 } else /* is_sync */
1314 old_rq = rq;
1315 blk_mq_put_ctx(data.ctx);
1316 if (!old_rq)
Jens Axboe7b371632015-11-05 10:41:40 -07001317 goto done;
1318 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1319 goto done;
Shaohua Lif984df12015-05-08 10:51:32 -07001320 blk_mq_insert_request(old_rq, false, true, true);
Jens Axboe7b371632015-11-05 10:41:40 -07001321 goto done;
Jens Axboe07068d52014-05-22 10:40:51 -06001322 }
1323
1324 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1325 /*
1326 * For a SYNC request, send it to the hardware immediately. For
1327 * an ASYNC request, just ensure that we run it later on. The
1328 * latter allows for merging opportunities and more efficient
1329 * dispatching.
1330 */
1331run_queue:
1332 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1333 }
Jens Axboe07068d52014-05-22 10:40:51 -06001334 blk_mq_put_ctx(data.ctx);
Jens Axboe7b371632015-11-05 10:41:40 -07001335done:
1336 return cookie;
Jens Axboe07068d52014-05-22 10:40:51 -06001337}
1338
1339/*
1340 * Single hardware queue variant. This will attempt to use any per-process
1341 * plug for merging and IO deferral.
1342 */
Jens Axboedece1632015-11-05 10:41:16 -07001343static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
Jens Axboe07068d52014-05-22 10:40:51 -06001344{
1345 const int is_sync = rw_is_sync(bio->bi_rw);
1346 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
Jeff Moyere6c44382015-05-08 10:51:30 -07001347 struct blk_plug *plug;
1348 unsigned int request_count = 0;
Jens Axboe07068d52014-05-22 10:40:51 -06001349 struct blk_map_ctx data;
1350 struct request *rq;
Jens Axboe7b371632015-11-05 10:41:40 -07001351 blk_qc_t cookie;
Jens Axboe07068d52014-05-22 10:40:51 -06001352
Jens Axboe07068d52014-05-22 10:40:51 -06001353 blk_queue_bounce(q, &bio);
1354
1355 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001356 bio_io_error(bio);
Jens Axboedece1632015-11-05 10:41:16 -07001357 return BLK_QC_T_NONE;
Jens Axboe07068d52014-05-22 10:40:51 -06001358 }
1359
Kent Overstreet54efd502015-04-23 22:37:18 -07001360 blk_queue_split(q, &bio, q->bio_split);
1361
Jeff Moyere6c44382015-05-08 10:51:30 -07001362 if (!is_flush_fua && !blk_queue_nomerges(q) &&
Shaohua Li5b3f3412015-05-08 10:51:33 -07001363 blk_attempt_plug_merge(q, bio, &request_count, NULL))
Jens Axboedece1632015-11-05 10:41:16 -07001364 return BLK_QC_T_NONE;
Jens Axboe07068d52014-05-22 10:40:51 -06001365
1366 rq = blk_mq_map_request(q, bio, &data);
Jens Axboeff87bce2014-06-03 11:59:49 -06001367 if (unlikely(!rq))
Jens Axboedece1632015-11-05 10:41:16 -07001368 return BLK_QC_T_NONE;
Jens Axboe320ae512013-10-24 09:20:05 +01001369
Jens Axboe7b371632015-11-05 10:41:40 -07001370 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
Jens Axboe320ae512013-10-24 09:20:05 +01001371
1372 if (unlikely(is_flush_fua)) {
1373 blk_mq_bio_to_request(rq, bio);
Jens Axboe320ae512013-10-24 09:20:05 +01001374 blk_insert_flush(rq);
1375 goto run_queue;
1376 }
1377
1378 /*
1379 * A task plug currently exists. Since this is completely lockless,
1380 * utilize that to temporarily store requests until the task is
1381 * either done or scheduled away.
1382 */
Jeff Moyere6c44382015-05-08 10:51:30 -07001383 plug = current->plug;
1384 if (plug) {
1385 blk_mq_bio_to_request(rq, bio);
Ming Lei676d0602015-10-20 23:13:56 +08001386 if (!request_count)
Jeff Moyere6c44382015-05-08 10:51:30 -07001387 trace_block_plug(q);
Jens Axboeb094f892015-11-20 20:29:45 -07001388
1389 blk_mq_put_ctx(data.ctx);
1390
1391 if (request_count >= BLK_MAX_REQUEST_COUNT) {
Jeff Moyere6c44382015-05-08 10:51:30 -07001392 blk_flush_plug_list(plug, false);
1393 trace_block_plug(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001394 }
Jens Axboeb094f892015-11-20 20:29:45 -07001395
Jeff Moyere6c44382015-05-08 10:51:30 -07001396 list_add_tail(&rq->queuelist, &plug->mq_list);
Jens Axboe7b371632015-11-05 10:41:40 -07001397 return cookie;
Jens Axboe320ae512013-10-24 09:20:05 +01001398 }
1399
Jens Axboe07068d52014-05-22 10:40:51 -06001400 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1401 /*
1402 * For a SYNC request, send it to the hardware immediately. For
1403 * an ASYNC request, just ensure that we run it later on. The
1404 * latter allows for merging opportunities and more efficient
1405 * dispatching.
1406 */
1407run_queue:
1408 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
Jens Axboe320ae512013-10-24 09:20:05 +01001409 }
1410
Jens Axboe07068d52014-05-22 10:40:51 -06001411 blk_mq_put_ctx(data.ctx);
Jens Axboe7b371632015-11-05 10:41:40 -07001412 return cookie;
Jens Axboe320ae512013-10-24 09:20:05 +01001413}
1414
1415/*
1416 * Default mapping to a software queue, since we use one per CPU.
1417 */
1418struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1419{
1420 return q->queue_hw_ctx[q->mq_map[cpu]];
1421}
1422EXPORT_SYMBOL(blk_mq_map_queue);
1423
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001424static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1425 struct blk_mq_tags *tags, unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01001426{
1427 struct page *page;
1428
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001429 if (tags->rqs && set->ops->exit_request) {
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001430 int i;
1431
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001432 for (i = 0; i < tags->nr_tags; i++) {
1433 if (!tags->rqs[i])
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001434 continue;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001435 set->ops->exit_request(set->driver_data, tags->rqs[i],
1436 hctx_idx, i);
Jens Axboea5164402014-09-10 09:02:03 -06001437 tags->rqs[i] = NULL;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001438 }
1439 }
1440
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001441 while (!list_empty(&tags->page_list)) {
1442 page = list_first_entry(&tags->page_list, struct page, lru);
Dave Hansen67534712014-01-08 20:17:46 -07001443 list_del_init(&page->lru);
Catalin Marinasf75782e2015-09-14 18:16:02 +01001444 /*
1445 * Remove kmemleak object previously allocated in
1446 * blk_mq_init_rq_map().
1447 */
1448 kmemleak_free(page_address(page));
Jens Axboe320ae512013-10-24 09:20:05 +01001449 __free_pages(page, page->private);
1450 }
1451
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001452 kfree(tags->rqs);
Jens Axboe320ae512013-10-24 09:20:05 +01001453
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001454 blk_mq_free_tags(tags);
Jens Axboe320ae512013-10-24 09:20:05 +01001455}
1456
1457static size_t order_to_size(unsigned int order)
1458{
Ming Lei4ca08502014-04-19 18:00:18 +08001459 return (size_t)PAGE_SIZE << order;
Jens Axboe320ae512013-10-24 09:20:05 +01001460}
1461
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001462static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1463 unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01001464{
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001465 struct blk_mq_tags *tags;
Jens Axboe320ae512013-10-24 09:20:05 +01001466 unsigned int i, j, entries_per_page, max_order = 4;
1467 size_t rq_size, left;
1468
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001469 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
Shaohua Li24391c02015-01-23 14:18:00 -07001470 set->numa_node,
1471 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001472 if (!tags)
1473 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001474
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001475 INIT_LIST_HEAD(&tags->page_list);
1476
Jens Axboea5164402014-09-10 09:02:03 -06001477 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1478 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1479 set->numa_node);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001480 if (!tags->rqs) {
1481 blk_mq_free_tags(tags);
1482 return NULL;
1483 }
Jens Axboe320ae512013-10-24 09:20:05 +01001484
1485 /*
1486 * rq_size is the size of the request plus driver payload, rounded
1487 * to the cacheline size
1488 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001489 rq_size = round_up(sizeof(struct request) + set->cmd_size,
Jens Axboe320ae512013-10-24 09:20:05 +01001490 cache_line_size());
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001491 left = rq_size * set->queue_depth;
Jens Axboe320ae512013-10-24 09:20:05 +01001492
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001493 for (i = 0; i < set->queue_depth; ) {
Jens Axboe320ae512013-10-24 09:20:05 +01001494 int this_order = max_order;
1495 struct page *page;
1496 int to_do;
1497 void *p;
1498
1499 while (left < order_to_size(this_order - 1) && this_order)
1500 this_order--;
1501
1502 do {
Jens Axboea5164402014-09-10 09:02:03 -06001503 page = alloc_pages_node(set->numa_node,
Linus Torvaldsac211172015-04-09 14:12:22 -07001504 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
Jens Axboea5164402014-09-10 09:02:03 -06001505 this_order);
Jens Axboe320ae512013-10-24 09:20:05 +01001506 if (page)
1507 break;
1508 if (!this_order--)
1509 break;
1510 if (order_to_size(this_order) < rq_size)
1511 break;
1512 } while (1);
1513
1514 if (!page)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001515 goto fail;
Jens Axboe320ae512013-10-24 09:20:05 +01001516
1517 page->private = this_order;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001518 list_add_tail(&page->lru, &tags->page_list);
Jens Axboe320ae512013-10-24 09:20:05 +01001519
1520 p = page_address(page);
Catalin Marinasf75782e2015-09-14 18:16:02 +01001521 /*
1522 * Allow kmemleak to scan these pages as they contain pointers
1523 * to additional allocations like via ops->init_request().
1524 */
1525 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
Jens Axboe320ae512013-10-24 09:20:05 +01001526 entries_per_page = order_to_size(this_order) / rq_size;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001527 to_do = min(entries_per_page, set->queue_depth - i);
Jens Axboe320ae512013-10-24 09:20:05 +01001528 left -= to_do * rq_size;
1529 for (j = 0; j < to_do; j++) {
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001530 tags->rqs[i] = p;
1531 if (set->ops->init_request) {
1532 if (set->ops->init_request(set->driver_data,
1533 tags->rqs[i], hctx_idx, i,
Jens Axboea5164402014-09-10 09:02:03 -06001534 set->numa_node)) {
1535 tags->rqs[i] = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001536 goto fail;
Jens Axboea5164402014-09-10 09:02:03 -06001537 }
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001538 }
1539
Jens Axboe320ae512013-10-24 09:20:05 +01001540 p += rq_size;
1541 i++;
1542 }
1543 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001544 return tags;
Jens Axboe320ae512013-10-24 09:20:05 +01001545
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001546fail:
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001547 blk_mq_free_rq_map(set, tags, hctx_idx);
1548 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001549}
1550
Jens Axboe1429d7c2014-05-19 09:23:55 -06001551static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1552{
1553 kfree(bitmap->map);
1554}
1555
1556static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1557{
1558 unsigned int bpw = 8, total, num_maps, i;
1559
1560 bitmap->bits_per_word = bpw;
1561
1562 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1563 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1564 GFP_KERNEL, node);
1565 if (!bitmap->map)
1566 return -ENOMEM;
1567
Jens Axboe1429d7c2014-05-19 09:23:55 -06001568 total = nr_cpu_ids;
1569 for (i = 0; i < num_maps; i++) {
1570 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1571 total -= bitmap->map[i].depth;
1572 }
1573
1574 return 0;
1575}
1576
Jens Axboe484b4062014-05-21 14:01:15 -06001577static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1578{
1579 struct request_queue *q = hctx->queue;
1580 struct blk_mq_ctx *ctx;
1581 LIST_HEAD(tmp);
1582
1583 /*
1584 * Move ctx entries to new CPU, if this one is going away.
1585 */
1586 ctx = __blk_mq_get_ctx(q, cpu);
1587
1588 spin_lock(&ctx->lock);
1589 if (!list_empty(&ctx->rq_list)) {
1590 list_splice_init(&ctx->rq_list, &tmp);
1591 blk_mq_hctx_clear_pending(hctx, ctx);
1592 }
1593 spin_unlock(&ctx->lock);
1594
1595 if (list_empty(&tmp))
1596 return NOTIFY_OK;
1597
1598 ctx = blk_mq_get_ctx(q);
1599 spin_lock(&ctx->lock);
1600
1601 while (!list_empty(&tmp)) {
1602 struct request *rq;
1603
1604 rq = list_first_entry(&tmp, struct request, queuelist);
1605 rq->mq_ctx = ctx;
1606 list_move_tail(&rq->queuelist, &ctx->rq_list);
1607 }
1608
1609 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1610 blk_mq_hctx_mark_pending(hctx, ctx);
1611
1612 spin_unlock(&ctx->lock);
1613
1614 blk_mq_run_hw_queue(hctx, true);
1615 blk_mq_put_ctx(ctx);
1616 return NOTIFY_OK;
1617}
1618
Jens Axboe484b4062014-05-21 14:01:15 -06001619static int blk_mq_hctx_notify(void *data, unsigned long action,
1620 unsigned int cpu)
1621{
1622 struct blk_mq_hw_ctx *hctx = data;
1623
1624 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1625 return blk_mq_hctx_cpu_offline(hctx, cpu);
Ming Lei2a34c082015-04-21 10:00:20 +08001626
1627 /*
1628 * In case of CPU online, tags may be reallocated
1629 * in blk_mq_map_swqueue() after mapping is updated.
1630 */
Jens Axboe484b4062014-05-21 14:01:15 -06001631
1632 return NOTIFY_OK;
1633}
1634
Ming Leic3b4afc2015-06-04 22:25:04 +08001635/* hctx->ctxs will be freed in queue's release handler */
Ming Lei08e98fc2014-09-25 23:23:38 +08001636static void blk_mq_exit_hctx(struct request_queue *q,
1637 struct blk_mq_tag_set *set,
1638 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1639{
Ming Leif70ced02014-09-25 23:23:47 +08001640 unsigned flush_start_tag = set->queue_depth;
1641
Ming Lei08e98fc2014-09-25 23:23:38 +08001642 blk_mq_tag_idle(hctx);
1643
Ming Leif70ced02014-09-25 23:23:47 +08001644 if (set->ops->exit_request)
1645 set->ops->exit_request(set->driver_data,
1646 hctx->fq->flush_rq, hctx_idx,
1647 flush_start_tag + hctx_idx);
1648
Ming Lei08e98fc2014-09-25 23:23:38 +08001649 if (set->ops->exit_hctx)
1650 set->ops->exit_hctx(hctx, hctx_idx);
1651
1652 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
Ming Leif70ced02014-09-25 23:23:47 +08001653 blk_free_flush_queue(hctx->fq);
Ming Lei08e98fc2014-09-25 23:23:38 +08001654 blk_mq_free_bitmap(&hctx->ctx_map);
1655}
1656
Ming Lei624dbe42014-05-27 23:35:13 +08001657static void blk_mq_exit_hw_queues(struct request_queue *q,
1658 struct blk_mq_tag_set *set, int nr_queue)
1659{
1660 struct blk_mq_hw_ctx *hctx;
1661 unsigned int i;
1662
1663 queue_for_each_hw_ctx(q, hctx, i) {
1664 if (i == nr_queue)
1665 break;
Ming Lei08e98fc2014-09-25 23:23:38 +08001666 blk_mq_exit_hctx(q, set, hctx, i);
Ming Lei624dbe42014-05-27 23:35:13 +08001667 }
Ming Lei624dbe42014-05-27 23:35:13 +08001668}
1669
1670static void blk_mq_free_hw_queues(struct request_queue *q,
1671 struct blk_mq_tag_set *set)
1672{
1673 struct blk_mq_hw_ctx *hctx;
1674 unsigned int i;
1675
Ming Leie09aae72015-01-29 20:17:27 +08001676 queue_for_each_hw_ctx(q, hctx, i)
Ming Lei624dbe42014-05-27 23:35:13 +08001677 free_cpumask_var(hctx->cpumask);
Ming Lei624dbe42014-05-27 23:35:13 +08001678}
1679
Ming Lei08e98fc2014-09-25 23:23:38 +08001680static int blk_mq_init_hctx(struct request_queue *q,
1681 struct blk_mq_tag_set *set,
1682 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1683{
1684 int node;
Ming Leif70ced02014-09-25 23:23:47 +08001685 unsigned flush_start_tag = set->queue_depth;
Ming Lei08e98fc2014-09-25 23:23:38 +08001686
1687 node = hctx->numa_node;
1688 if (node == NUMA_NO_NODE)
1689 node = hctx->numa_node = set->numa_node;
1690
1691 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1692 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1693 spin_lock_init(&hctx->lock);
1694 INIT_LIST_HEAD(&hctx->dispatch);
1695 hctx->queue = q;
1696 hctx->queue_num = hctx_idx;
Jeff Moyer2404e602015-11-03 10:40:06 -05001697 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
Ming Lei08e98fc2014-09-25 23:23:38 +08001698
1699 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1700 blk_mq_hctx_notify, hctx);
1701 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1702
1703 hctx->tags = set->tags[hctx_idx];
1704
1705 /*
1706 * Allocate space for all possible cpus to avoid allocation at
1707 * runtime
1708 */
1709 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1710 GFP_KERNEL, node);
1711 if (!hctx->ctxs)
1712 goto unregister_cpu_notifier;
1713
1714 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1715 goto free_ctxs;
1716
1717 hctx->nr_ctx = 0;
1718
1719 if (set->ops->init_hctx &&
1720 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1721 goto free_bitmap;
1722
Ming Leif70ced02014-09-25 23:23:47 +08001723 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1724 if (!hctx->fq)
1725 goto exit_hctx;
1726
1727 if (set->ops->init_request &&
1728 set->ops->init_request(set->driver_data,
1729 hctx->fq->flush_rq, hctx_idx,
1730 flush_start_tag + hctx_idx, node))
1731 goto free_fq;
1732
Ming Lei08e98fc2014-09-25 23:23:38 +08001733 return 0;
1734
Ming Leif70ced02014-09-25 23:23:47 +08001735 free_fq:
1736 kfree(hctx->fq);
1737 exit_hctx:
1738 if (set->ops->exit_hctx)
1739 set->ops->exit_hctx(hctx, hctx_idx);
Ming Lei08e98fc2014-09-25 23:23:38 +08001740 free_bitmap:
1741 blk_mq_free_bitmap(&hctx->ctx_map);
1742 free_ctxs:
1743 kfree(hctx->ctxs);
1744 unregister_cpu_notifier:
1745 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1746
1747 return -1;
1748}
1749
Jens Axboe320ae512013-10-24 09:20:05 +01001750static void blk_mq_init_cpu_queues(struct request_queue *q,
1751 unsigned int nr_hw_queues)
1752{
1753 unsigned int i;
1754
1755 for_each_possible_cpu(i) {
1756 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1757 struct blk_mq_hw_ctx *hctx;
1758
1759 memset(__ctx, 0, sizeof(*__ctx));
1760 __ctx->cpu = i;
1761 spin_lock_init(&__ctx->lock);
1762 INIT_LIST_HEAD(&__ctx->rq_list);
1763 __ctx->queue = q;
1764
1765 /* If the cpu isn't online, the cpu is mapped to first hctx */
Jens Axboe320ae512013-10-24 09:20:05 +01001766 if (!cpu_online(i))
1767 continue;
1768
Jens Axboee4043dc2014-04-09 10:18:23 -06001769 hctx = q->mq_ops->map_queue(q, i);
Jens Axboee4043dc2014-04-09 10:18:23 -06001770
Jens Axboe320ae512013-10-24 09:20:05 +01001771 /*
1772 * Set local node, IFF we have more than one hw queue. If
1773 * not, we remain on the home node of the device
1774 */
1775 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
Raghavendra K Tbffed452015-12-02 16:59:05 +05301776 hctx->numa_node = local_memory_node(cpu_to_node(i));
Jens Axboe320ae512013-10-24 09:20:05 +01001777 }
1778}
1779
Akinobu Mita57783222015-09-27 02:09:23 +09001780static void blk_mq_map_swqueue(struct request_queue *q,
1781 const struct cpumask *online_mask)
Jens Axboe320ae512013-10-24 09:20:05 +01001782{
1783 unsigned int i;
1784 struct blk_mq_hw_ctx *hctx;
1785 struct blk_mq_ctx *ctx;
Ming Lei2a34c082015-04-21 10:00:20 +08001786 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01001787
Akinobu Mita60de0742015-09-27 02:09:25 +09001788 /*
1789 * Avoid others reading imcomplete hctx->cpumask through sysfs
1790 */
1791 mutex_lock(&q->sysfs_lock);
1792
Jens Axboe320ae512013-10-24 09:20:05 +01001793 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboee4043dc2014-04-09 10:18:23 -06001794 cpumask_clear(hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001795 hctx->nr_ctx = 0;
1796 }
1797
1798 /*
1799 * Map software to hardware queues
1800 */
1801 queue_for_each_ctx(q, ctx, i) {
1802 /* If the cpu isn't online, the cpu is mapped to first hctx */
Akinobu Mita57783222015-09-27 02:09:23 +09001803 if (!cpumask_test_cpu(i, online_mask))
Jens Axboee4043dc2014-04-09 10:18:23 -06001804 continue;
1805
Jens Axboe320ae512013-10-24 09:20:05 +01001806 hctx = q->mq_ops->map_queue(q, i);
Keith Busch868f2f02015-12-17 17:08:14 -07001807
Jens Axboee4043dc2014-04-09 10:18:23 -06001808 cpumask_set_cpu(i, hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001809 ctx->index_hw = hctx->nr_ctx;
1810 hctx->ctxs[hctx->nr_ctx++] = ctx;
1811 }
Jens Axboe506e9312014-05-07 10:26:44 -06001812
Akinobu Mita60de0742015-09-27 02:09:25 +09001813 mutex_unlock(&q->sysfs_lock);
1814
Jens Axboe506e9312014-05-07 10:26:44 -06001815 queue_for_each_hw_ctx(q, hctx, i) {
Chong Yuan889fa312015-04-15 11:39:29 -06001816 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1817
Jens Axboe484b4062014-05-21 14:01:15 -06001818 /*
Jens Axboea68aafa2014-08-15 13:19:15 -06001819 * If no software queues are mapped to this hardware queue,
1820 * disable it and free the request entries.
Jens Axboe484b4062014-05-21 14:01:15 -06001821 */
1822 if (!hctx->nr_ctx) {
Jens Axboe484b4062014-05-21 14:01:15 -06001823 if (set->tags[i]) {
1824 blk_mq_free_rq_map(set, set->tags[i], i);
1825 set->tags[i] = NULL;
Jens Axboe484b4062014-05-21 14:01:15 -06001826 }
Ming Lei2a34c082015-04-21 10:00:20 +08001827 hctx->tags = NULL;
Jens Axboe484b4062014-05-21 14:01:15 -06001828 continue;
1829 }
1830
Ming Lei2a34c082015-04-21 10:00:20 +08001831 /* unmapped hw queue can be remapped after CPU topo changed */
1832 if (!set->tags[i])
1833 set->tags[i] = blk_mq_init_rq_map(set, i);
1834 hctx->tags = set->tags[i];
1835 WARN_ON(!hctx->tags);
1836
Raghavendra K Te0e827b2015-12-02 16:57:06 +05301837 cpumask_copy(hctx->tags->cpumask, hctx->cpumask);
Jens Axboe484b4062014-05-21 14:01:15 -06001838 /*
Chong Yuan889fa312015-04-15 11:39:29 -06001839 * Set the map size to the number of mapped software queues.
1840 * This is more accurate and more efficient than looping
1841 * over all possibly mapped software queues.
1842 */
Jens Axboe569fd0c2015-04-17 08:28:50 -06001843 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
Chong Yuan889fa312015-04-15 11:39:29 -06001844
1845 /*
Jens Axboe484b4062014-05-21 14:01:15 -06001846 * Initialize batch roundrobin counts
1847 */
Jens Axboe506e9312014-05-07 10:26:44 -06001848 hctx->next_cpu = cpumask_first(hctx->cpumask);
1849 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1850 }
Jens Axboe320ae512013-10-24 09:20:05 +01001851}
1852
Jeff Moyer2404e602015-11-03 10:40:06 -05001853static void queue_set_hctx_shared(struct request_queue *q, bool shared)
Jens Axboe0d2602c2014-05-13 15:10:52 -06001854{
1855 struct blk_mq_hw_ctx *hctx;
Jens Axboe0d2602c2014-05-13 15:10:52 -06001856 int i;
1857
Jeff Moyer2404e602015-11-03 10:40:06 -05001858 queue_for_each_hw_ctx(q, hctx, i) {
1859 if (shared)
1860 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1861 else
1862 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1863 }
1864}
1865
1866static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1867{
1868 struct request_queue *q;
Jens Axboe0d2602c2014-05-13 15:10:52 -06001869
1870 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1871 blk_mq_freeze_queue(q);
Jeff Moyer2404e602015-11-03 10:40:06 -05001872 queue_set_hctx_shared(q, shared);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001873 blk_mq_unfreeze_queue(q);
1874 }
1875}
1876
1877static void blk_mq_del_queue_tag_set(struct request_queue *q)
1878{
1879 struct blk_mq_tag_set *set = q->tag_set;
1880
Jens Axboe0d2602c2014-05-13 15:10:52 -06001881 mutex_lock(&set->tag_list_lock);
1882 list_del_init(&q->tag_set_list);
Jeff Moyer2404e602015-11-03 10:40:06 -05001883 if (list_is_singular(&set->tag_list)) {
1884 /* just transitioned to unshared */
1885 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1886 /* update existing queue */
1887 blk_mq_update_tag_set_depth(set, false);
1888 }
Jens Axboe0d2602c2014-05-13 15:10:52 -06001889 mutex_unlock(&set->tag_list_lock);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001890}
1891
1892static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1893 struct request_queue *q)
1894{
1895 q->tag_set = set;
1896
1897 mutex_lock(&set->tag_list_lock);
Jeff Moyer2404e602015-11-03 10:40:06 -05001898
1899 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1900 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1901 set->flags |= BLK_MQ_F_TAG_SHARED;
1902 /* update existing queue */
1903 blk_mq_update_tag_set_depth(set, true);
1904 }
1905 if (set->flags & BLK_MQ_F_TAG_SHARED)
1906 queue_set_hctx_shared(q, true);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001907 list_add_tail(&q->tag_set_list, &set->tag_list);
Jeff Moyer2404e602015-11-03 10:40:06 -05001908
Jens Axboe0d2602c2014-05-13 15:10:52 -06001909 mutex_unlock(&set->tag_list_lock);
1910}
1911
Ming Leie09aae72015-01-29 20:17:27 +08001912/*
1913 * It is the actual release handler for mq, but we do it from
1914 * request queue's release handler for avoiding use-after-free
1915 * and headache because q->mq_kobj shouldn't have been introduced,
1916 * but we can't group ctx/kctx kobj without it.
1917 */
1918void blk_mq_release(struct request_queue *q)
1919{
1920 struct blk_mq_hw_ctx *hctx;
1921 unsigned int i;
1922
1923 /* hctx kobj stays in hctx */
Ming Leic3b4afc2015-06-04 22:25:04 +08001924 queue_for_each_hw_ctx(q, hctx, i) {
1925 if (!hctx)
1926 continue;
1927 kfree(hctx->ctxs);
Ming Leie09aae72015-01-29 20:17:27 +08001928 kfree(hctx);
Ming Leic3b4afc2015-06-04 22:25:04 +08001929 }
Ming Leie09aae72015-01-29 20:17:27 +08001930
Akinobu Mitaa723bab2015-09-27 02:09:21 +09001931 kfree(q->mq_map);
1932 q->mq_map = NULL;
1933
Ming Leie09aae72015-01-29 20:17:27 +08001934 kfree(q->queue_hw_ctx);
1935
1936 /* ctx kobj stays in queue_ctx */
1937 free_percpu(q->queue_ctx);
1938}
1939
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001940struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
Jens Axboe320ae512013-10-24 09:20:05 +01001941{
Mike Snitzerb62c21b2015-03-12 23:56:02 -04001942 struct request_queue *uninit_q, *q;
1943
1944 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1945 if (!uninit_q)
1946 return ERR_PTR(-ENOMEM);
1947
1948 q = blk_mq_init_allocated_queue(set, uninit_q);
1949 if (IS_ERR(q))
1950 blk_cleanup_queue(uninit_q);
1951
1952 return q;
1953}
1954EXPORT_SYMBOL(blk_mq_init_queue);
1955
Keith Busch868f2f02015-12-17 17:08:14 -07001956static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
1957 struct request_queue *q)
Mike Snitzerb62c21b2015-03-12 23:56:02 -04001958{
Keith Busch868f2f02015-12-17 17:08:14 -07001959 int i, j;
1960 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001961
Keith Busch868f2f02015-12-17 17:08:14 -07001962 blk_mq_sysfs_unregister(q);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001963 for (i = 0; i < set->nr_hw_queues; i++) {
Keith Busch868f2f02015-12-17 17:08:14 -07001964 int node;
Jens Axboef14bbe72014-05-27 12:06:53 -06001965
Keith Busch868f2f02015-12-17 17:08:14 -07001966 if (hctxs[i])
1967 continue;
1968
1969 node = blk_mq_hw_queue_to_node(q->mq_map, i);
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02001970 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1971 GFP_KERNEL, node);
Jens Axboe320ae512013-10-24 09:20:05 +01001972 if (!hctxs[i])
Keith Busch868f2f02015-12-17 17:08:14 -07001973 break;
Jens Axboe320ae512013-10-24 09:20:05 +01001974
Jens Axboea86073e2014-10-13 15:41:54 -06001975 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
Keith Busch868f2f02015-12-17 17:08:14 -07001976 node)) {
1977 kfree(hctxs[i]);
1978 hctxs[i] = NULL;
1979 break;
1980 }
Jens Axboee4043dc2014-04-09 10:18:23 -06001981
Jens Axboe0d2602c2014-05-13 15:10:52 -06001982 atomic_set(&hctxs[i]->nr_active, 0);
Jens Axboef14bbe72014-05-27 12:06:53 -06001983 hctxs[i]->numa_node = node;
Jens Axboe320ae512013-10-24 09:20:05 +01001984 hctxs[i]->queue_num = i;
Keith Busch868f2f02015-12-17 17:08:14 -07001985
1986 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
1987 free_cpumask_var(hctxs[i]->cpumask);
1988 kfree(hctxs[i]);
1989 hctxs[i] = NULL;
1990 break;
1991 }
1992 blk_mq_hctx_kobj_init(hctxs[i]);
Jens Axboe320ae512013-10-24 09:20:05 +01001993 }
Keith Busch868f2f02015-12-17 17:08:14 -07001994 for (j = i; j < q->nr_hw_queues; j++) {
1995 struct blk_mq_hw_ctx *hctx = hctxs[j];
1996
1997 if (hctx) {
1998 if (hctx->tags) {
1999 blk_mq_free_rq_map(set, hctx->tags, j);
2000 set->tags[j] = NULL;
2001 }
2002 blk_mq_exit_hctx(q, set, hctx, j);
2003 free_cpumask_var(hctx->cpumask);
2004 kobject_put(&hctx->kobj);
2005 kfree(hctx->ctxs);
2006 kfree(hctx);
2007 hctxs[j] = NULL;
2008
2009 }
2010 }
2011 q->nr_hw_queues = i;
2012 blk_mq_sysfs_register(q);
2013}
2014
2015struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2016 struct request_queue *q)
2017{
Ming Lei66841672016-02-12 15:27:00 +08002018 /* mark the queue as mq asap */
2019 q->mq_ops = set->ops;
2020
Keith Busch868f2f02015-12-17 17:08:14 -07002021 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2022 if (!q->queue_ctx)
2023 return ERR_PTR(-ENOMEM);
2024
2025 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2026 GFP_KERNEL, set->numa_node);
2027 if (!q->queue_hw_ctx)
2028 goto err_percpu;
2029
2030 q->mq_map = blk_mq_make_queue_map(set);
2031 if (!q->mq_map)
2032 goto err_map;
2033
2034 blk_mq_realloc_hw_ctxs(set, q);
2035 if (!q->nr_hw_queues)
2036 goto err_hctxs;
Jens Axboe320ae512013-10-24 09:20:05 +01002037
Christoph Hellwig287922e2015-10-30 20:57:30 +08002038 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
Ming Leie56f6982015-07-16 19:53:22 +08002039 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
Jens Axboe320ae512013-10-24 09:20:05 +01002040
2041 q->nr_queues = nr_cpu_ids;
Jens Axboe320ae512013-10-24 09:20:05 +01002042
Jens Axboe94eddfb2013-11-19 09:25:07 -07002043 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
Jens Axboe320ae512013-10-24 09:20:05 +01002044
Jens Axboe05f1dd52014-05-29 09:53:32 -06002045 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2046 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2047
Christoph Hellwig1be036e2014-02-07 10:22:39 -08002048 q->sg_reserved_size = INT_MAX;
2049
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06002050 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2051 INIT_LIST_HEAD(&q->requeue_list);
2052 spin_lock_init(&q->requeue_lock);
2053
Jens Axboe07068d52014-05-22 10:40:51 -06002054 if (q->nr_hw_queues > 1)
2055 blk_queue_make_request(q, blk_mq_make_request);
2056 else
2057 blk_queue_make_request(q, blk_sq_make_request);
2058
Jens Axboeeba71762014-05-20 15:17:27 -06002059 /*
2060 * Do this after blk_queue_make_request() overrides it...
2061 */
2062 q->nr_requests = set->queue_depth;
2063
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002064 if (set->ops->complete)
2065 blk_queue_softirq_done(q, set->ops->complete);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08002066
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002067 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +01002068
Akinobu Mita57783222015-09-27 02:09:23 +09002069 get_online_cpus();
Jens Axboe320ae512013-10-24 09:20:05 +01002070 mutex_lock(&all_q_mutex);
Akinobu Mita4593fdb2015-09-27 02:09:20 +09002071
Jens Axboe320ae512013-10-24 09:20:05 +01002072 list_add_tail(&q->all_q_node, &all_q_list);
Jens Axboe0d2602c2014-05-13 15:10:52 -06002073 blk_mq_add_queue_tag_set(set, q);
Akinobu Mita57783222015-09-27 02:09:23 +09002074 blk_mq_map_swqueue(q, cpu_online_mask);
Jens Axboe484b4062014-05-21 14:01:15 -06002075
Akinobu Mita4593fdb2015-09-27 02:09:20 +09002076 mutex_unlock(&all_q_mutex);
Akinobu Mita57783222015-09-27 02:09:23 +09002077 put_online_cpus();
Akinobu Mita4593fdb2015-09-27 02:09:20 +09002078
Jens Axboe320ae512013-10-24 09:20:05 +01002079 return q;
Christoph Hellwig18741982014-02-10 09:29:00 -07002080
Jens Axboe320ae512013-10-24 09:20:05 +01002081err_hctxs:
Keith Busch868f2f02015-12-17 17:08:14 -07002082 kfree(q->mq_map);
Jens Axboef14bbe72014-05-27 12:06:53 -06002083err_map:
Keith Busch868f2f02015-12-17 17:08:14 -07002084 kfree(q->queue_hw_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01002085err_percpu:
Keith Busch868f2f02015-12-17 17:08:14 -07002086 free_percpu(q->queue_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01002087 return ERR_PTR(-ENOMEM);
2088}
Mike Snitzerb62c21b2015-03-12 23:56:02 -04002089EXPORT_SYMBOL(blk_mq_init_allocated_queue);
Jens Axboe320ae512013-10-24 09:20:05 +01002090
2091void blk_mq_free_queue(struct request_queue *q)
2092{
Ming Lei624dbe42014-05-27 23:35:13 +08002093 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01002094
Akinobu Mita0e626362015-09-27 02:09:22 +09002095 mutex_lock(&all_q_mutex);
2096 list_del_init(&q->all_q_node);
2097 mutex_unlock(&all_q_mutex);
2098
Jens Axboe0d2602c2014-05-13 15:10:52 -06002099 blk_mq_del_queue_tag_set(q);
2100
Ming Lei624dbe42014-05-27 23:35:13 +08002101 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2102 blk_mq_free_hw_queues(q, set);
Jens Axboe320ae512013-10-24 09:20:05 +01002103}
Jens Axboe320ae512013-10-24 09:20:05 +01002104
2105/* Basically redo blk_mq_init_queue with queue frozen */
Akinobu Mita57783222015-09-27 02:09:23 +09002106static void blk_mq_queue_reinit(struct request_queue *q,
2107 const struct cpumask *online_mask)
Jens Axboe320ae512013-10-24 09:20:05 +01002108{
Christoph Hellwig4ecd4fe2015-05-07 09:38:13 +02002109 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
Jens Axboe320ae512013-10-24 09:20:05 +01002110
Jens Axboe67aec142014-05-30 08:25:36 -06002111 blk_mq_sysfs_unregister(q);
2112
Akinobu Mita57783222015-09-27 02:09:23 +09002113 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
Jens Axboe320ae512013-10-24 09:20:05 +01002114
2115 /*
2116 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2117 * we should change hctx numa_node according to new topology (this
2118 * involves free and re-allocate memory, worthy doing?)
2119 */
2120
Akinobu Mita57783222015-09-27 02:09:23 +09002121 blk_mq_map_swqueue(q, online_mask);
Jens Axboe320ae512013-10-24 09:20:05 +01002122
Jens Axboe67aec142014-05-30 08:25:36 -06002123 blk_mq_sysfs_register(q);
Jens Axboe320ae512013-10-24 09:20:05 +01002124}
2125
Paul Gortmakerf618ef72013-11-14 08:26:02 -07002126static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2127 unsigned long action, void *hcpu)
Jens Axboe320ae512013-10-24 09:20:05 +01002128{
2129 struct request_queue *q;
Akinobu Mita57783222015-09-27 02:09:23 +09002130 int cpu = (unsigned long)hcpu;
2131 /*
2132 * New online cpumask which is going to be set in this hotplug event.
2133 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2134 * one-by-one and dynamically allocating this could result in a failure.
2135 */
2136 static struct cpumask online_new;
Jens Axboe320ae512013-10-24 09:20:05 +01002137
2138 /*
Akinobu Mita57783222015-09-27 02:09:23 +09002139 * Before hotadded cpu starts handling requests, new mappings must
2140 * be established. Otherwise, these requests in hw queue might
2141 * never be dispatched.
2142 *
2143 * For example, there is a single hw queue (hctx) and two CPU queues
2144 * (ctx0 for CPU0, and ctx1 for CPU1).
2145 *
2146 * Now CPU1 is just onlined and a request is inserted into
2147 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2148 * still zero.
2149 *
2150 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2151 * set in pending bitmap and tries to retrieve requests in
2152 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2153 * so the request in ctx1->rq_list is ignored.
Jens Axboe320ae512013-10-24 09:20:05 +01002154 */
Akinobu Mita57783222015-09-27 02:09:23 +09002155 switch (action & ~CPU_TASKS_FROZEN) {
2156 case CPU_DEAD:
2157 case CPU_UP_CANCELED:
2158 cpumask_copy(&online_new, cpu_online_mask);
2159 break;
2160 case CPU_UP_PREPARE:
2161 cpumask_copy(&online_new, cpu_online_mask);
2162 cpumask_set_cpu(cpu, &online_new);
2163 break;
2164 default:
Jens Axboe320ae512013-10-24 09:20:05 +01002165 return NOTIFY_OK;
Akinobu Mita57783222015-09-27 02:09:23 +09002166 }
Jens Axboe320ae512013-10-24 09:20:05 +01002167
2168 mutex_lock(&all_q_mutex);
Tejun Heof3af0202014-11-04 13:52:27 -05002169
2170 /*
2171 * We need to freeze and reinit all existing queues. Freezing
2172 * involves synchronous wait for an RCU grace period and doing it
2173 * one by one may take a long time. Start freezing all queues in
2174 * one swoop and then wait for the completions so that freezing can
2175 * take place in parallel.
2176 */
2177 list_for_each_entry(q, &all_q_list, all_q_node)
2178 blk_mq_freeze_queue_start(q);
Ming Leif054b562015-04-21 10:00:19 +08002179 list_for_each_entry(q, &all_q_list, all_q_node) {
Tejun Heof3af0202014-11-04 13:52:27 -05002180 blk_mq_freeze_queue_wait(q);
2181
Ming Leif054b562015-04-21 10:00:19 +08002182 /*
2183 * timeout handler can't touch hw queue during the
2184 * reinitialization
2185 */
2186 del_timer_sync(&q->timeout);
2187 }
2188
Jens Axboe320ae512013-10-24 09:20:05 +01002189 list_for_each_entry(q, &all_q_list, all_q_node)
Akinobu Mita57783222015-09-27 02:09:23 +09002190 blk_mq_queue_reinit(q, &online_new);
Tejun Heof3af0202014-11-04 13:52:27 -05002191
2192 list_for_each_entry(q, &all_q_list, all_q_node)
2193 blk_mq_unfreeze_queue(q);
2194
Jens Axboe320ae512013-10-24 09:20:05 +01002195 mutex_unlock(&all_q_mutex);
2196 return NOTIFY_OK;
2197}
2198
Jens Axboea5164402014-09-10 09:02:03 -06002199static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2200{
2201 int i;
2202
2203 for (i = 0; i < set->nr_hw_queues; i++) {
2204 set->tags[i] = blk_mq_init_rq_map(set, i);
2205 if (!set->tags[i])
2206 goto out_unwind;
2207 }
2208
2209 return 0;
2210
2211out_unwind:
2212 while (--i >= 0)
2213 blk_mq_free_rq_map(set, set->tags[i], i);
2214
Jens Axboea5164402014-09-10 09:02:03 -06002215 return -ENOMEM;
2216}
2217
2218/*
2219 * Allocate the request maps associated with this tag_set. Note that this
2220 * may reduce the depth asked for, if memory is tight. set->queue_depth
2221 * will be updated to reflect the allocated depth.
2222 */
2223static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2224{
2225 unsigned int depth;
2226 int err;
2227
2228 depth = set->queue_depth;
2229 do {
2230 err = __blk_mq_alloc_rq_maps(set);
2231 if (!err)
2232 break;
2233
2234 set->queue_depth >>= 1;
2235 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2236 err = -ENOMEM;
2237 break;
2238 }
2239 } while (set->queue_depth);
2240
2241 if (!set->queue_depth || err) {
2242 pr_err("blk-mq: failed to allocate request map\n");
2243 return -ENOMEM;
2244 }
2245
2246 if (depth != set->queue_depth)
2247 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2248 depth, set->queue_depth);
2249
2250 return 0;
2251}
2252
Keith Buschf26cdc82015-06-01 09:29:53 -06002253struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2254{
2255 return tags->cpumask;
2256}
2257EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2258
Jens Axboea4391c62014-06-05 15:21:56 -06002259/*
2260 * Alloc a tag set to be associated with one or more request queues.
2261 * May fail with EINVAL for various error conditions. May adjust the
2262 * requested depth down, if if it too large. In that case, the set
2263 * value will be stored in set->queue_depth.
2264 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002265int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2266{
Bart Van Assche205fb5f2014-10-30 14:45:11 +01002267 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2268
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002269 if (!set->nr_hw_queues)
2270 return -EINVAL;
Jens Axboea4391c62014-06-05 15:21:56 -06002271 if (!set->queue_depth)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002272 return -EINVAL;
2273 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2274 return -EINVAL;
2275
Xiaoguang Wangf9018ac2015-03-30 13:19:14 +08002276 if (!set->ops->queue_rq || !set->ops->map_queue)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002277 return -EINVAL;
2278
Jens Axboea4391c62014-06-05 15:21:56 -06002279 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2280 pr_info("blk-mq: reduced tag depth to %u\n",
2281 BLK_MQ_MAX_DEPTH);
2282 set->queue_depth = BLK_MQ_MAX_DEPTH;
2283 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002284
Shaohua Li6637fad2014-11-30 16:00:58 -08002285 /*
2286 * If a crashdump is active, then we are potentially in a very
2287 * memory constrained environment. Limit us to 1 queue and
2288 * 64 tags to prevent using too much memory.
2289 */
2290 if (is_kdump_kernel()) {
2291 set->nr_hw_queues = 1;
2292 set->queue_depth = min(64U, set->queue_depth);
2293 }
Keith Busch868f2f02015-12-17 17:08:14 -07002294 /*
2295 * There is no use for more h/w queues than cpus.
2296 */
2297 if (set->nr_hw_queues > nr_cpu_ids)
2298 set->nr_hw_queues = nr_cpu_ids;
Shaohua Li6637fad2014-11-30 16:00:58 -08002299
Keith Busch868f2f02015-12-17 17:08:14 -07002300 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002301 GFP_KERNEL, set->numa_node);
2302 if (!set->tags)
Jens Axboea5164402014-09-10 09:02:03 -06002303 return -ENOMEM;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002304
Jens Axboea5164402014-09-10 09:02:03 -06002305 if (blk_mq_alloc_rq_maps(set))
2306 goto enomem;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002307
Jens Axboe0d2602c2014-05-13 15:10:52 -06002308 mutex_init(&set->tag_list_lock);
2309 INIT_LIST_HEAD(&set->tag_list);
2310
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002311 return 0;
Jens Axboea5164402014-09-10 09:02:03 -06002312enomem:
Robert Elliott5676e7b2014-09-02 11:38:44 -05002313 kfree(set->tags);
2314 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002315 return -ENOMEM;
2316}
2317EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2318
2319void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2320{
2321 int i;
2322
Keith Busch868f2f02015-12-17 17:08:14 -07002323 for (i = 0; i < nr_cpu_ids; i++) {
Junichi Nomuraf42d79a2015-10-14 05:02:15 +00002324 if (set->tags[i])
Jens Axboe484b4062014-05-21 14:01:15 -06002325 blk_mq_free_rq_map(set, set->tags[i], i);
2326 }
2327
Ming Lei981bd182014-04-24 00:07:34 +08002328 kfree(set->tags);
Robert Elliott5676e7b2014-09-02 11:38:44 -05002329 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002330}
2331EXPORT_SYMBOL(blk_mq_free_tag_set);
2332
Jens Axboee3a2b3f2014-05-20 11:49:02 -06002333int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2334{
2335 struct blk_mq_tag_set *set = q->tag_set;
2336 struct blk_mq_hw_ctx *hctx;
2337 int i, ret;
2338
2339 if (!set || nr > set->queue_depth)
2340 return -EINVAL;
2341
2342 ret = 0;
2343 queue_for_each_hw_ctx(q, hctx, i) {
Keith Busche9137d42016-02-18 14:56:35 -07002344 if (!hctx->tags)
2345 continue;
Jens Axboee3a2b3f2014-05-20 11:49:02 -06002346 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2347 if (ret)
2348 break;
2349 }
2350
2351 if (!ret)
2352 q->nr_requests = nr;
2353
2354 return ret;
2355}
2356
Keith Busch868f2f02015-12-17 17:08:14 -07002357void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2358{
2359 struct request_queue *q;
2360
2361 if (nr_hw_queues > nr_cpu_ids)
2362 nr_hw_queues = nr_cpu_ids;
2363 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2364 return;
2365
2366 list_for_each_entry(q, &set->tag_list, tag_set_list)
2367 blk_mq_freeze_queue(q);
2368
2369 set->nr_hw_queues = nr_hw_queues;
2370 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2371 blk_mq_realloc_hw_ctxs(set, q);
2372
2373 if (q->nr_hw_queues > 1)
2374 blk_queue_make_request(q, blk_mq_make_request);
2375 else
2376 blk_queue_make_request(q, blk_sq_make_request);
2377
2378 blk_mq_queue_reinit(q, cpu_online_mask);
2379 }
2380
2381 list_for_each_entry(q, &set->tag_list, tag_set_list)
2382 blk_mq_unfreeze_queue(q);
2383}
2384EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2385
Jens Axboe676141e2014-03-20 13:29:18 -06002386void blk_mq_disable_hotplug(void)
2387{
2388 mutex_lock(&all_q_mutex);
2389}
2390
2391void blk_mq_enable_hotplug(void)
2392{
2393 mutex_unlock(&all_q_mutex);
2394}
2395
Jens Axboe320ae512013-10-24 09:20:05 +01002396static int __init blk_mq_init(void)
2397{
Jens Axboe320ae512013-10-24 09:20:05 +01002398 blk_mq_cpu_init();
2399
Tejun Heoadd703f2014-07-01 10:34:38 -06002400 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
Jens Axboe320ae512013-10-24 09:20:05 +01002401
2402 return 0;
2403}
2404subsys_initcall(blk_mq_init);