blob: 4c0622fae41383d0f5577ea9b8b127d93df33bb6 [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{
Ming Lei0048b482015-08-09 03:41:51 -0400547 return tags->rqs[tag];
Christoph Hellwig24d2f902014-04-15 14:14:00 -0600548}
549EXPORT_SYMBOL(blk_mq_tag_to_rq);
550
Jens Axboe320ae512013-10-24 09:20:05 +0100551struct blk_mq_timeout_data {
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700552 unsigned long next;
553 unsigned int next_set;
Jens Axboe320ae512013-10-24 09:20:05 +0100554};
555
Christoph Hellwig90415832014-09-22 10:21:48 -0600556void blk_mq_rq_timed_out(struct request *req, bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100557{
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700558 struct blk_mq_ops *ops = req->q->mq_ops;
559 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
Jens Axboe87ee7b12014-04-24 08:51:47 -0600560
561 /*
562 * We know that complete is set at this point. If STARTED isn't set
563 * anymore, then the request isn't active and the "timeout" should
564 * just be ignored. This can happen due to the bitflag ordering.
565 * Timeout first checks if STARTED is set, and if it is, assumes
566 * the request is active. But if we race with completion, then
567 * we both flags will get cleared. So check here again, and ignore
568 * a timeout event with a request that isn't active.
569 */
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700570 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
571 return;
Jens Axboe87ee7b12014-04-24 08:51:47 -0600572
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700573 if (ops->timeout)
Christoph Hellwig0152fb62014-09-13 16:40:13 -0700574 ret = ops->timeout(req, reserved);
Jens Axboe87ee7b12014-04-24 08:51:47 -0600575
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700576 switch (ret) {
577 case BLK_EH_HANDLED:
578 __blk_mq_complete_request(req);
579 break;
580 case BLK_EH_RESET_TIMER:
581 blk_add_timer(req);
582 blk_clear_rq_complete(req);
583 break;
584 case BLK_EH_NOT_HANDLED:
585 break;
586 default:
587 printk(KERN_ERR "block: bad eh return: %d\n", ret);
588 break;
589 }
Jens Axboe87ee7b12014-04-24 08:51:47 -0600590}
Keith Busch5b3f25f2015-01-07 18:55:46 -0700591
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700592static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
593 struct request *rq, void *priv, bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100594{
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700595 struct blk_mq_timeout_data *data = priv;
596
Keith Buscheb130db2015-01-08 08:59:53 -0700597 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
598 /*
599 * If a request wasn't started before the queue was
600 * marked dying, kill it here or it'll go unnoticed.
601 */
Christoph Hellwigf4829a92015-09-27 21:01:50 +0200602 if (unlikely(blk_queue_dying(rq->q)))
603 blk_mq_complete_request(rq, -EIO);
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700604 return;
Keith Buscheb130db2015-01-08 08:59:53 -0700605 }
Jens Axboe320ae512013-10-24 09:20:05 +0100606
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700607 if (time_after_eq(jiffies, rq->deadline)) {
608 if (!blk_mark_rq_complete(rq))
Christoph Hellwig0152fb62014-09-13 16:40:13 -0700609 blk_mq_rq_timed_out(rq, reserved);
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700610 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
611 data->next = rq->deadline;
612 data->next_set = 1;
613 }
Jens Axboe320ae512013-10-24 09:20:05 +0100614}
615
Christoph Hellwig287922e2015-10-30 20:57:30 +0800616static void blk_mq_timeout_work(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +0100617{
Christoph Hellwig287922e2015-10-30 20:57:30 +0800618 struct request_queue *q =
619 container_of(work, struct request_queue, timeout_work);
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700620 struct blk_mq_timeout_data data = {
621 .next = 0,
622 .next_set = 0,
623 };
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700624 int i;
Jens Axboe320ae512013-10-24 09:20:05 +0100625
Christoph Hellwig287922e2015-10-30 20:57:30 +0800626 if (blk_queue_enter(q, true))
627 return;
628
Christoph Hellwig0bf6cd52015-09-27 21:01:51 +0200629 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
Jens Axboe320ae512013-10-24 09:20:05 +0100630
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700631 if (data.next_set) {
632 data.next = blk_rq_timeout(round_jiffies_up(data.next));
633 mod_timer(&q->timeout, data.next);
Jens Axboe0d2602c2014-05-13 15:10:52 -0600634 } else {
Christoph Hellwig0bf6cd52015-09-27 21:01:51 +0200635 struct blk_mq_hw_ctx *hctx;
636
Ming Leif054b562015-04-21 10:00:19 +0800637 queue_for_each_hw_ctx(q, hctx, i) {
638 /* the hctx may be unmapped, so check it here */
639 if (blk_mq_hw_queue_mapped(hctx))
640 blk_mq_tag_idle(hctx);
641 }
Jens Axboe0d2602c2014-05-13 15:10:52 -0600642 }
Christoph Hellwig287922e2015-10-30 20:57:30 +0800643 blk_queue_exit(q);
Jens Axboe320ae512013-10-24 09:20:05 +0100644}
645
646/*
647 * Reverse check our software queue for entries that we could potentially
648 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
649 * too much time checking for merges.
650 */
651static bool blk_mq_attempt_merge(struct request_queue *q,
652 struct blk_mq_ctx *ctx, struct bio *bio)
653{
654 struct request *rq;
655 int checked = 8;
656
657 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
658 int el_ret;
659
660 if (!checked--)
661 break;
662
663 if (!blk_rq_merge_ok(rq, bio))
664 continue;
665
666 el_ret = blk_try_merge(rq, bio);
667 if (el_ret == ELEVATOR_BACK_MERGE) {
668 if (bio_attempt_back_merge(q, rq, bio)) {
669 ctx->rq_merged++;
670 return true;
671 }
672 break;
673 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
674 if (bio_attempt_front_merge(q, rq, bio)) {
675 ctx->rq_merged++;
676 return true;
677 }
678 break;
679 }
680 }
681
682 return false;
683}
684
Jens Axboe320ae512013-10-24 09:20:05 +0100685/*
Jens Axboe1429d7c2014-05-19 09:23:55 -0600686 * Process software queues that have been marked busy, splicing them
687 * to the for-dispatch
688 */
689static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
690{
691 struct blk_mq_ctx *ctx;
692 int i;
693
Jens Axboe569fd0c2015-04-17 08:28:50 -0600694 for (i = 0; i < hctx->ctx_map.size; i++) {
Jens Axboe1429d7c2014-05-19 09:23:55 -0600695 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
696 unsigned int off, bit;
697
698 if (!bm->word)
699 continue;
700
701 bit = 0;
702 off = i * hctx->ctx_map.bits_per_word;
703 do {
704 bit = find_next_bit(&bm->word, bm->depth, bit);
705 if (bit >= bm->depth)
706 break;
707
708 ctx = hctx->ctxs[bit + off];
709 clear_bit(bit, &bm->word);
710 spin_lock(&ctx->lock);
711 list_splice_tail_init(&ctx->rq_list, list);
712 spin_unlock(&ctx->lock);
713
714 bit++;
715 } while (1);
716 }
717}
718
719/*
Jens Axboe320ae512013-10-24 09:20:05 +0100720 * Run this hardware queue, pulling any software queues mapped to it in.
721 * Note that this function currently has various problems around ordering
722 * of IO. In particular, we'd like FIFO behaviour on handling existing
723 * items on the hctx->dispatch list. Ignore that for now.
724 */
725static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
726{
727 struct request_queue *q = hctx->queue;
Jens Axboe320ae512013-10-24 09:20:05 +0100728 struct request *rq;
729 LIST_HEAD(rq_list);
Jens Axboe74c45052014-10-29 11:14:52 -0600730 LIST_HEAD(driver_list);
731 struct list_head *dptr;
Jens Axboe1429d7c2014-05-19 09:23:55 -0600732 int queued;
Jens Axboe320ae512013-10-24 09:20:05 +0100733
Jens Axboefd1270d2014-04-16 09:23:48 -0600734 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
Jens Axboee4043dc2014-04-09 10:18:23 -0600735
Jens Axboe5d12f902014-03-19 15:25:02 -0600736 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
Jens Axboe320ae512013-10-24 09:20:05 +0100737 return;
738
739 hctx->run++;
740
741 /*
742 * Touch any software queue that has pending entries.
743 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600744 flush_busy_ctxs(hctx, &rq_list);
Jens Axboe320ae512013-10-24 09:20:05 +0100745
746 /*
747 * If we have previous entries on our dispatch list, grab them
748 * and stuff them at the front for more fair dispatch.
749 */
750 if (!list_empty_careful(&hctx->dispatch)) {
751 spin_lock(&hctx->lock);
752 if (!list_empty(&hctx->dispatch))
753 list_splice_init(&hctx->dispatch, &rq_list);
754 spin_unlock(&hctx->lock);
755 }
756
757 /*
Jens Axboe74c45052014-10-29 11:14:52 -0600758 * Start off with dptr being NULL, so we start the first request
759 * immediately, even if we have more pending.
760 */
761 dptr = NULL;
762
763 /*
Jens Axboe320ae512013-10-24 09:20:05 +0100764 * Now process all the entries, sending them to the driver.
765 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600766 queued = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100767 while (!list_empty(&rq_list)) {
Jens Axboe74c45052014-10-29 11:14:52 -0600768 struct blk_mq_queue_data bd;
Jens Axboe320ae512013-10-24 09:20:05 +0100769 int ret;
770
771 rq = list_first_entry(&rq_list, struct request, queuelist);
772 list_del_init(&rq->queuelist);
Jens Axboe320ae512013-10-24 09:20:05 +0100773
Jens Axboe74c45052014-10-29 11:14:52 -0600774 bd.rq = rq;
775 bd.list = dptr;
776 bd.last = list_empty(&rq_list);
777
778 ret = q->mq_ops->queue_rq(hctx, &bd);
Jens Axboe320ae512013-10-24 09:20:05 +0100779 switch (ret) {
780 case BLK_MQ_RQ_QUEUE_OK:
781 queued++;
782 continue;
783 case BLK_MQ_RQ_QUEUE_BUSY:
Jens Axboe320ae512013-10-24 09:20:05 +0100784 list_add(&rq->queuelist, &rq_list);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200785 __blk_mq_requeue_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100786 break;
787 default:
788 pr_err("blk-mq: bad return on queue: %d\n", ret);
Jens Axboe320ae512013-10-24 09:20:05 +0100789 case BLK_MQ_RQ_QUEUE_ERROR:
Christoph Hellwig1e93b8c2014-02-11 08:27:13 -0800790 rq->errors = -EIO;
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700791 blk_mq_end_request(rq, rq->errors);
Jens Axboe320ae512013-10-24 09:20:05 +0100792 break;
793 }
794
795 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
796 break;
Jens Axboe74c45052014-10-29 11:14:52 -0600797
798 /*
799 * We've done the first request. If we have more than 1
800 * left in the list, set dptr to defer issue.
801 */
802 if (!dptr && rq_list.next != rq_list.prev)
803 dptr = &driver_list;
Jens Axboe320ae512013-10-24 09:20:05 +0100804 }
805
806 if (!queued)
807 hctx->dispatched[0]++;
808 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
809 hctx->dispatched[ilog2(queued) + 1]++;
810
811 /*
812 * Any items that need requeuing? Stuff them into hctx->dispatch,
813 * that is where we will continue on next queue run.
814 */
815 if (!list_empty(&rq_list)) {
816 spin_lock(&hctx->lock);
817 list_splice(&rq_list, &hctx->dispatch);
818 spin_unlock(&hctx->lock);
Shaohua Li9ba52e52015-05-04 14:32:48 -0600819 /*
820 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
821 * it's possible the queue is stopped and restarted again
822 * before this. Queue restart will dispatch requests. And since
823 * requests in rq_list aren't added into hctx->dispatch yet,
824 * the requests in rq_list might get lost.
825 *
826 * blk_mq_run_hw_queue() already checks the STOPPED bit
827 **/
828 blk_mq_run_hw_queue(hctx, true);
Jens Axboe320ae512013-10-24 09:20:05 +0100829 }
830}
831
Jens Axboe506e9312014-05-07 10:26:44 -0600832/*
833 * It'd be great if the workqueue API had a way to pass
834 * in a mask and had some smarts for more clever placement.
835 * For now we just round-robin here, switching for every
836 * BLK_MQ_CPU_WORK_BATCH queued items.
837 */
838static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
839{
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100840 if (hctx->queue->nr_hw_queues == 1)
841 return WORK_CPU_UNBOUND;
Jens Axboe506e9312014-05-07 10:26:44 -0600842
843 if (--hctx->next_cpu_batch <= 0) {
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100844 int cpu = hctx->next_cpu, next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600845
846 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
847 if (next_cpu >= nr_cpu_ids)
848 next_cpu = cpumask_first(hctx->cpumask);
849
850 hctx->next_cpu = next_cpu;
851 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100852
853 return cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600854 }
855
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100856 return hctx->next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600857}
858
Jens Axboe320ae512013-10-24 09:20:05 +0100859void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
860{
Ming Lei19c66e52014-12-03 19:38:04 +0800861 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
862 !blk_mq_hw_queue_mapped(hctx)))
Jens Axboe320ae512013-10-24 09:20:05 +0100863 return;
864
Paolo Bonzini398205b2014-11-07 23:03:59 +0100865 if (!async) {
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100866 int cpu = get_cpu();
867 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
Paolo Bonzini398205b2014-11-07 23:03:59 +0100868 __blk_mq_run_hw_queue(hctx);
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100869 put_cpu();
Paolo Bonzini398205b2014-11-07 23:03:59 +0100870 return;
871 }
Jens Axboee4043dc2014-04-09 10:18:23 -0600872
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100873 put_cpu();
Jens Axboee4043dc2014-04-09 10:18:23 -0600874 }
Paolo Bonzini398205b2014-11-07 23:03:59 +0100875
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100876 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
877 &hctx->run_work, 0);
Jens Axboe320ae512013-10-24 09:20:05 +0100878}
879
Mike Snitzerb94ec292015-03-11 23:56:38 -0400880void blk_mq_run_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +0100881{
882 struct blk_mq_hw_ctx *hctx;
883 int i;
884
885 queue_for_each_hw_ctx(q, hctx, i) {
886 if ((!blk_mq_hctx_has_pending(hctx) &&
887 list_empty_careful(&hctx->dispatch)) ||
Jens Axboe5d12f902014-03-19 15:25:02 -0600888 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
Jens Axboe320ae512013-10-24 09:20:05 +0100889 continue;
890
Mike Snitzerb94ec292015-03-11 23:56:38 -0400891 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +0100892 }
893}
Mike Snitzerb94ec292015-03-11 23:56:38 -0400894EXPORT_SYMBOL(blk_mq_run_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +0100895
896void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
897{
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600898 cancel_delayed_work(&hctx->run_work);
899 cancel_delayed_work(&hctx->delay_work);
Jens Axboe320ae512013-10-24 09:20:05 +0100900 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
901}
902EXPORT_SYMBOL(blk_mq_stop_hw_queue);
903
Christoph Hellwig280d45f2013-10-25 14:45:58 +0100904void blk_mq_stop_hw_queues(struct request_queue *q)
905{
906 struct blk_mq_hw_ctx *hctx;
907 int i;
908
909 queue_for_each_hw_ctx(q, hctx, i)
910 blk_mq_stop_hw_queue(hctx);
911}
912EXPORT_SYMBOL(blk_mq_stop_hw_queues);
913
Jens Axboe320ae512013-10-24 09:20:05 +0100914void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
915{
916 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboee4043dc2014-04-09 10:18:23 -0600917
Jens Axboe0ffbce82014-06-25 08:22:34 -0600918 blk_mq_run_hw_queue(hctx, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100919}
920EXPORT_SYMBOL(blk_mq_start_hw_queue);
921
Christoph Hellwig2f268552014-04-16 09:44:56 +0200922void blk_mq_start_hw_queues(struct request_queue *q)
923{
924 struct blk_mq_hw_ctx *hctx;
925 int i;
926
927 queue_for_each_hw_ctx(q, hctx, i)
928 blk_mq_start_hw_queue(hctx);
929}
930EXPORT_SYMBOL(blk_mq_start_hw_queues);
931
Christoph Hellwig1b4a3252014-04-16 09:44:54 +0200932void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +0100933{
934 struct blk_mq_hw_ctx *hctx;
935 int i;
936
937 queue_for_each_hw_ctx(q, hctx, i) {
938 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
939 continue;
940
941 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Christoph Hellwig1b4a3252014-04-16 09:44:54 +0200942 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +0100943 }
944}
945EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
946
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600947static void blk_mq_run_work_fn(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +0100948{
949 struct blk_mq_hw_ctx *hctx;
950
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600951 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
Jens Axboee4043dc2014-04-09 10:18:23 -0600952
Jens Axboe320ae512013-10-24 09:20:05 +0100953 __blk_mq_run_hw_queue(hctx);
954}
955
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600956static void blk_mq_delay_work_fn(struct work_struct *work)
957{
958 struct blk_mq_hw_ctx *hctx;
959
960 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
961
962 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
963 __blk_mq_run_hw_queue(hctx);
964}
965
966void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
967{
Ming Lei19c66e52014-12-03 19:38:04 +0800968 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
969 return;
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600970
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100971 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
972 &hctx->delay_work, msecs_to_jiffies(msecs));
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600973}
974EXPORT_SYMBOL(blk_mq_delay_queue);
975
Ming Leicfd0c552015-10-20 23:13:57 +0800976static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
977 struct blk_mq_ctx *ctx,
978 struct request *rq,
979 bool at_head)
Jens Axboe320ae512013-10-24 09:20:05 +0100980{
Jens Axboe01b983c2013-11-19 18:59:10 -0700981 trace_block_rq_insert(hctx->queue, rq);
982
Christoph Hellwig72a0a362014-02-07 10:22:36 -0800983 if (at_head)
984 list_add(&rq->queuelist, &ctx->rq_list);
985 else
986 list_add_tail(&rq->queuelist, &ctx->rq_list);
Ming Leicfd0c552015-10-20 23:13:57 +0800987}
Jens Axboe4bb659b2014-05-09 09:36:49 -0600988
Ming Leicfd0c552015-10-20 23:13:57 +0800989static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
990 struct request *rq, bool at_head)
991{
992 struct blk_mq_ctx *ctx = rq->mq_ctx;
993
994 __blk_mq_insert_req_list(hctx, ctx, rq, at_head);
Jens Axboe320ae512013-10-24 09:20:05 +0100995 blk_mq_hctx_mark_pending(hctx, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +0100996}
997
Christoph Hellwigeeabc852014-03-21 08:57:37 -0600998void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
999 bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01001000{
1001 struct request_queue *q = rq->q;
1002 struct blk_mq_hw_ctx *hctx;
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001003 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001004
1005 current_ctx = blk_mq_get_ctx(q);
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001006 if (!cpu_online(ctx->cpu))
1007 rq->mq_ctx = ctx = current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001008
Jens Axboe320ae512013-10-24 09:20:05 +01001009 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1010
Christoph Hellwiga57a1782014-09-16 14:44:07 -07001011 spin_lock(&ctx->lock);
1012 __blk_mq_insert_request(hctx, rq, at_head);
1013 spin_unlock(&ctx->lock);
Jens Axboe320ae512013-10-24 09:20:05 +01001014
Jens Axboe320ae512013-10-24 09:20:05 +01001015 if (run_queue)
1016 blk_mq_run_hw_queue(hctx, async);
Jens Axboee4043dc2014-04-09 10:18:23 -06001017
1018 blk_mq_put_ctx(current_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001019}
1020
1021static void blk_mq_insert_requests(struct request_queue *q,
1022 struct blk_mq_ctx *ctx,
1023 struct list_head *list,
1024 int depth,
1025 bool from_schedule)
1026
1027{
1028 struct blk_mq_hw_ctx *hctx;
1029 struct blk_mq_ctx *current_ctx;
1030
1031 trace_block_unplug(q, depth, !from_schedule);
1032
1033 current_ctx = blk_mq_get_ctx(q);
1034
1035 if (!cpu_online(ctx->cpu))
1036 ctx = current_ctx;
1037 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1038
1039 /*
1040 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1041 * offline now
1042 */
1043 spin_lock(&ctx->lock);
1044 while (!list_empty(list)) {
1045 struct request *rq;
1046
1047 rq = list_first_entry(list, struct request, queuelist);
1048 list_del_init(&rq->queuelist);
1049 rq->mq_ctx = ctx;
Ming Leicfd0c552015-10-20 23:13:57 +08001050 __blk_mq_insert_req_list(hctx, ctx, rq, false);
Jens Axboe320ae512013-10-24 09:20:05 +01001051 }
Ming Leicfd0c552015-10-20 23:13:57 +08001052 blk_mq_hctx_mark_pending(hctx, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001053 spin_unlock(&ctx->lock);
1054
Jens Axboe320ae512013-10-24 09:20:05 +01001055 blk_mq_run_hw_queue(hctx, from_schedule);
Jens Axboee4043dc2014-04-09 10:18:23 -06001056 blk_mq_put_ctx(current_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001057}
1058
1059static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1060{
1061 struct request *rqa = container_of(a, struct request, queuelist);
1062 struct request *rqb = container_of(b, struct request, queuelist);
1063
1064 return !(rqa->mq_ctx < rqb->mq_ctx ||
1065 (rqa->mq_ctx == rqb->mq_ctx &&
1066 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1067}
1068
1069void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1070{
1071 struct blk_mq_ctx *this_ctx;
1072 struct request_queue *this_q;
1073 struct request *rq;
1074 LIST_HEAD(list);
1075 LIST_HEAD(ctx_list);
1076 unsigned int depth;
1077
1078 list_splice_init(&plug->mq_list, &list);
1079
1080 list_sort(NULL, &list, plug_ctx_cmp);
1081
1082 this_q = NULL;
1083 this_ctx = NULL;
1084 depth = 0;
1085
1086 while (!list_empty(&list)) {
1087 rq = list_entry_rq(list.next);
1088 list_del_init(&rq->queuelist);
1089 BUG_ON(!rq->q);
1090 if (rq->mq_ctx != this_ctx) {
1091 if (this_ctx) {
1092 blk_mq_insert_requests(this_q, this_ctx,
1093 &ctx_list, depth,
1094 from_schedule);
1095 }
1096
1097 this_ctx = rq->mq_ctx;
1098 this_q = rq->q;
1099 depth = 0;
1100 }
1101
1102 depth++;
1103 list_add_tail(&rq->queuelist, &ctx_list);
1104 }
1105
1106 /*
1107 * If 'this_ctx' is set, we know we have entries to complete
1108 * on 'ctx_list'. Do those.
1109 */
1110 if (this_ctx) {
1111 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1112 from_schedule);
1113 }
1114}
1115
1116static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1117{
1118 init_request_from_bio(rq, bio);
Jens Axboe4b570522014-05-29 11:00:11 -06001119
Jens Axboe3ee32372014-06-09 09:36:53 -06001120 if (blk_do_io_stat(rq))
Jens Axboe4b570522014-05-29 11:00:11 -06001121 blk_account_io_start(rq, 1);
Jens Axboe320ae512013-10-24 09:20:05 +01001122}
1123
Jens Axboe274a5842014-08-15 12:44:08 -06001124static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1125{
1126 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1127 !blk_queue_nomerges(hctx->queue);
1128}
1129
Jens Axboe07068d52014-05-22 10:40:51 -06001130static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1131 struct blk_mq_ctx *ctx,
1132 struct request *rq, struct bio *bio)
1133{
Ming Leie18378a2015-10-20 23:13:54 +08001134 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
Jens Axboe07068d52014-05-22 10:40:51 -06001135 blk_mq_bio_to_request(rq, bio);
1136 spin_lock(&ctx->lock);
1137insert_rq:
1138 __blk_mq_insert_request(hctx, rq, false);
1139 spin_unlock(&ctx->lock);
1140 return false;
1141 } else {
Jens Axboe274a5842014-08-15 12:44:08 -06001142 struct request_queue *q = hctx->queue;
1143
Jens Axboe07068d52014-05-22 10:40:51 -06001144 spin_lock(&ctx->lock);
1145 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1146 blk_mq_bio_to_request(rq, bio);
1147 goto insert_rq;
1148 }
1149
1150 spin_unlock(&ctx->lock);
1151 __blk_mq_free_request(hctx, ctx, rq);
1152 return true;
1153 }
1154}
1155
1156struct blk_map_ctx {
1157 struct blk_mq_hw_ctx *hctx;
1158 struct blk_mq_ctx *ctx;
1159};
1160
1161static struct request *blk_mq_map_request(struct request_queue *q,
1162 struct bio *bio,
1163 struct blk_map_ctx *data)
Jens Axboe320ae512013-10-24 09:20:05 +01001164{
1165 struct blk_mq_hw_ctx *hctx;
1166 struct blk_mq_ctx *ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001167 struct request *rq;
Jens Axboe07068d52014-05-22 10:40:51 -06001168 int rw = bio_data_dir(bio);
Ming Leicb96a422014-06-01 00:43:37 +08001169 struct blk_mq_alloc_data alloc_data;
Jens Axboe320ae512013-10-24 09:20:05 +01001170
Dan Williams3ef28e82015-10-21 13:20:12 -04001171 blk_queue_enter_live(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001172 ctx = blk_mq_get_ctx(q);
1173 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1174
Jens Axboe07068d52014-05-22 10:40:51 -06001175 if (rw_is_sync(bio->bi_rw))
Shaohua Li27fbf4e2014-02-19 20:20:21 +08001176 rw |= REQ_SYNC;
Jens Axboe07068d52014-05-22 10:40:51 -06001177
Jens Axboe320ae512013-10-24 09:20:05 +01001178 trace_block_getrq(q, bio, rw);
Christoph Hellwig6f3b0e82015-11-26 09:13:05 +01001179 blk_mq_set_alloc_data(&alloc_data, q, BLK_MQ_REQ_NOWAIT, ctx, hctx);
Ming Leicb96a422014-06-01 00:43:37 +08001180 rq = __blk_mq_alloc_request(&alloc_data, rw);
Christoph Hellwig5dee8572014-05-27 20:59:47 +02001181 if (unlikely(!rq)) {
Christoph Hellwig793597a2014-05-27 20:59:49 +02001182 __blk_mq_run_hw_queue(hctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001183 blk_mq_put_ctx(ctx);
1184 trace_block_sleeprq(q, bio, rw);
Christoph Hellwig793597a2014-05-27 20:59:49 +02001185
1186 ctx = blk_mq_get_ctx(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001187 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Christoph Hellwig6f3b0e82015-11-26 09:13:05 +01001188 blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
Ming Leicb96a422014-06-01 00:43:37 +08001189 rq = __blk_mq_alloc_request(&alloc_data, rw);
1190 ctx = alloc_data.ctx;
1191 hctx = alloc_data.hctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001192 }
1193
1194 hctx->queued++;
Jens Axboe07068d52014-05-22 10:40:51 -06001195 data->hctx = hctx;
1196 data->ctx = ctx;
1197 return rq;
1198}
1199
Jens Axboe7b371632015-11-05 10:41:40 -07001200static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
Shaohua Lif984df12015-05-08 10:51:32 -07001201{
1202 int ret;
1203 struct request_queue *q = rq->q;
1204 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1205 rq->mq_ctx->cpu);
1206 struct blk_mq_queue_data bd = {
1207 .rq = rq,
1208 .list = NULL,
1209 .last = 1
1210 };
Jens Axboe7b371632015-11-05 10:41:40 -07001211 blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
Shaohua Lif984df12015-05-08 10:51:32 -07001212
1213 /*
1214 * For OK queue, we are done. For error, kill it. Any other
1215 * error (busy), just add it to our list as we previously
1216 * would have done
1217 */
1218 ret = q->mq_ops->queue_rq(hctx, &bd);
Jens Axboe7b371632015-11-05 10:41:40 -07001219 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1220 *cookie = new_cookie;
Shaohua Lif984df12015-05-08 10:51:32 -07001221 return 0;
Shaohua Lif984df12015-05-08 10:51:32 -07001222 }
Jens Axboe7b371632015-11-05 10:41:40 -07001223
1224 __blk_mq_requeue_request(rq);
1225
1226 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1227 *cookie = BLK_QC_T_NONE;
1228 rq->errors = -EIO;
1229 blk_mq_end_request(rq, rq->errors);
1230 return 0;
1231 }
1232
1233 return -1;
Shaohua Lif984df12015-05-08 10:51:32 -07001234}
1235
Jens Axboe07068d52014-05-22 10:40:51 -06001236/*
1237 * Multiple hardware queue variant. This will not use per-process plugs,
1238 * but will attempt to bypass the hctx queueing if we can go straight to
1239 * hardware for SYNC IO.
1240 */
Jens Axboedece1632015-11-05 10:41:16 -07001241static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
Jens Axboe07068d52014-05-22 10:40:51 -06001242{
1243 const int is_sync = rw_is_sync(bio->bi_rw);
1244 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1245 struct blk_map_ctx data;
1246 struct request *rq;
Shaohua Lif984df12015-05-08 10:51:32 -07001247 unsigned int request_count = 0;
1248 struct blk_plug *plug;
Shaohua Li5b3f3412015-05-08 10:51:33 -07001249 struct request *same_queue_rq = NULL;
Jens Axboe7b371632015-11-05 10:41:40 -07001250 blk_qc_t cookie;
Jens Axboe07068d52014-05-22 10:40:51 -06001251
1252 blk_queue_bounce(q, &bio);
1253
1254 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001255 bio_io_error(bio);
Jens Axboedece1632015-11-05 10:41:16 -07001256 return BLK_QC_T_NONE;
Jens Axboe07068d52014-05-22 10:40:51 -06001257 }
1258
Kent Overstreet54efd502015-04-23 22:37:18 -07001259 blk_queue_split(q, &bio, q->bio_split);
1260
Jeff Moyer0809e3a2015-10-20 23:13:51 +08001261 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1262 if (blk_attempt_plug_merge(q, bio, &request_count,
1263 &same_queue_rq))
Jens Axboedece1632015-11-05 10:41:16 -07001264 return BLK_QC_T_NONE;
Jeff Moyer0809e3a2015-10-20 23:13:51 +08001265 } else
1266 request_count = blk_plug_queued_count(q);
Shaohua Lif984df12015-05-08 10:51:32 -07001267
Jens Axboe07068d52014-05-22 10:40:51 -06001268 rq = blk_mq_map_request(q, bio, &data);
1269 if (unlikely(!rq))
Jens Axboedece1632015-11-05 10:41:16 -07001270 return BLK_QC_T_NONE;
Jens Axboe07068d52014-05-22 10:40:51 -06001271
Jens Axboe7b371632015-11-05 10:41:40 -07001272 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
Jens Axboe07068d52014-05-22 10:40:51 -06001273
1274 if (unlikely(is_flush_fua)) {
1275 blk_mq_bio_to_request(rq, bio);
1276 blk_insert_flush(rq);
1277 goto run_queue;
1278 }
1279
Shaohua Lif984df12015-05-08 10:51:32 -07001280 plug = current->plug;
Jens Axboee167dfb2014-10-29 11:18:26 -06001281 /*
1282 * If the driver supports defer issued based on 'last', then
1283 * queue it up like normal since we can potentially save some
1284 * CPU this way.
1285 */
Shaohua Lif984df12015-05-08 10:51:32 -07001286 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1287 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1288 struct request *old_rq = NULL;
Jens Axboe07068d52014-05-22 10:40:51 -06001289
1290 blk_mq_bio_to_request(rq, bio);
Jens Axboe07068d52014-05-22 10:40:51 -06001291
1292 /*
Jens Axboeb094f892015-11-20 20:29:45 -07001293 * We do limited pluging. If the bio can be merged, do that.
Shaohua Lif984df12015-05-08 10:51:32 -07001294 * Otherwise the existing request in the plug list will be
1295 * issued. So the plug list will have one request at most
Jens Axboe07068d52014-05-22 10:40:51 -06001296 */
Shaohua Lif984df12015-05-08 10:51:32 -07001297 if (plug) {
Shaohua Li5b3f3412015-05-08 10:51:33 -07001298 /*
1299 * The plug list might get flushed before this. If that
Jens Axboeb094f892015-11-20 20:29:45 -07001300 * happens, same_queue_rq is invalid and plug list is
1301 * empty
1302 */
Shaohua Li5b3f3412015-05-08 10:51:33 -07001303 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1304 old_rq = same_queue_rq;
Shaohua Lif984df12015-05-08 10:51:32 -07001305 list_del_init(&old_rq->queuelist);
Jens Axboe07068d52014-05-22 10:40:51 -06001306 }
Shaohua Lif984df12015-05-08 10:51:32 -07001307 list_add_tail(&rq->queuelist, &plug->mq_list);
1308 } else /* is_sync */
1309 old_rq = rq;
1310 blk_mq_put_ctx(data.ctx);
1311 if (!old_rq)
Jens Axboe7b371632015-11-05 10:41:40 -07001312 goto done;
1313 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1314 goto done;
Shaohua Lif984df12015-05-08 10:51:32 -07001315 blk_mq_insert_request(old_rq, false, true, true);
Jens Axboe7b371632015-11-05 10:41:40 -07001316 goto done;
Jens Axboe07068d52014-05-22 10:40:51 -06001317 }
1318
1319 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1320 /*
1321 * For a SYNC request, send it to the hardware immediately. For
1322 * an ASYNC request, just ensure that we run it later on. The
1323 * latter allows for merging opportunities and more efficient
1324 * dispatching.
1325 */
1326run_queue:
1327 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1328 }
Jens Axboe07068d52014-05-22 10:40:51 -06001329 blk_mq_put_ctx(data.ctx);
Jens Axboe7b371632015-11-05 10:41:40 -07001330done:
1331 return cookie;
Jens Axboe07068d52014-05-22 10:40:51 -06001332}
1333
1334/*
1335 * Single hardware queue variant. This will attempt to use any per-process
1336 * plug for merging and IO deferral.
1337 */
Jens Axboedece1632015-11-05 10:41:16 -07001338static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
Jens Axboe07068d52014-05-22 10:40:51 -06001339{
1340 const int is_sync = rw_is_sync(bio->bi_rw);
1341 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
Jeff Moyere6c44382015-05-08 10:51:30 -07001342 struct blk_plug *plug;
1343 unsigned int request_count = 0;
Jens Axboe07068d52014-05-22 10:40:51 -06001344 struct blk_map_ctx data;
1345 struct request *rq;
Jens Axboe7b371632015-11-05 10:41:40 -07001346 blk_qc_t cookie;
Jens Axboe07068d52014-05-22 10:40:51 -06001347
Jens Axboe07068d52014-05-22 10:40:51 -06001348 blk_queue_bounce(q, &bio);
1349
1350 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001351 bio_io_error(bio);
Jens Axboedece1632015-11-05 10:41:16 -07001352 return BLK_QC_T_NONE;
Jens Axboe07068d52014-05-22 10:40:51 -06001353 }
1354
Kent Overstreet54efd502015-04-23 22:37:18 -07001355 blk_queue_split(q, &bio, q->bio_split);
1356
Jeff Moyere6c44382015-05-08 10:51:30 -07001357 if (!is_flush_fua && !blk_queue_nomerges(q) &&
Shaohua Li5b3f3412015-05-08 10:51:33 -07001358 blk_attempt_plug_merge(q, bio, &request_count, NULL))
Jens Axboedece1632015-11-05 10:41:16 -07001359 return BLK_QC_T_NONE;
Jens Axboe07068d52014-05-22 10:40:51 -06001360
1361 rq = blk_mq_map_request(q, bio, &data);
Jens Axboeff87bce2014-06-03 11:59:49 -06001362 if (unlikely(!rq))
Jens Axboedece1632015-11-05 10:41:16 -07001363 return BLK_QC_T_NONE;
Jens Axboe320ae512013-10-24 09:20:05 +01001364
Jens Axboe7b371632015-11-05 10:41:40 -07001365 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
Jens Axboe320ae512013-10-24 09:20:05 +01001366
1367 if (unlikely(is_flush_fua)) {
1368 blk_mq_bio_to_request(rq, bio);
Jens Axboe320ae512013-10-24 09:20:05 +01001369 blk_insert_flush(rq);
1370 goto run_queue;
1371 }
1372
1373 /*
1374 * A task plug currently exists. Since this is completely lockless,
1375 * utilize that to temporarily store requests until the task is
1376 * either done or scheduled away.
1377 */
Jeff Moyere6c44382015-05-08 10:51:30 -07001378 plug = current->plug;
1379 if (plug) {
1380 blk_mq_bio_to_request(rq, bio);
Ming Lei676d0602015-10-20 23:13:56 +08001381 if (!request_count)
Jeff Moyere6c44382015-05-08 10:51:30 -07001382 trace_block_plug(q);
Jens Axboeb094f892015-11-20 20:29:45 -07001383
1384 blk_mq_put_ctx(data.ctx);
1385
1386 if (request_count >= BLK_MAX_REQUEST_COUNT) {
Jeff Moyere6c44382015-05-08 10:51:30 -07001387 blk_flush_plug_list(plug, false);
1388 trace_block_plug(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001389 }
Jens Axboeb094f892015-11-20 20:29:45 -07001390
Jeff Moyere6c44382015-05-08 10:51:30 -07001391 list_add_tail(&rq->queuelist, &plug->mq_list);
Jens Axboe7b371632015-11-05 10:41:40 -07001392 return cookie;
Jens Axboe320ae512013-10-24 09:20:05 +01001393 }
1394
Jens Axboe07068d52014-05-22 10:40:51 -06001395 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1396 /*
1397 * For a SYNC request, send it to the hardware immediately. For
1398 * an ASYNC request, just ensure that we run it later on. The
1399 * latter allows for merging opportunities and more efficient
1400 * dispatching.
1401 */
1402run_queue:
1403 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
Jens Axboe320ae512013-10-24 09:20:05 +01001404 }
1405
Jens Axboe07068d52014-05-22 10:40:51 -06001406 blk_mq_put_ctx(data.ctx);
Jens Axboe7b371632015-11-05 10:41:40 -07001407 return cookie;
Jens Axboe320ae512013-10-24 09:20:05 +01001408}
1409
1410/*
1411 * Default mapping to a software queue, since we use one per CPU.
1412 */
1413struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1414{
1415 return q->queue_hw_ctx[q->mq_map[cpu]];
1416}
1417EXPORT_SYMBOL(blk_mq_map_queue);
1418
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001419static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1420 struct blk_mq_tags *tags, unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01001421{
1422 struct page *page;
1423
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001424 if (tags->rqs && set->ops->exit_request) {
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001425 int i;
1426
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001427 for (i = 0; i < tags->nr_tags; i++) {
1428 if (!tags->rqs[i])
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001429 continue;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001430 set->ops->exit_request(set->driver_data, tags->rqs[i],
1431 hctx_idx, i);
Jens Axboea5164402014-09-10 09:02:03 -06001432 tags->rqs[i] = NULL;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001433 }
1434 }
1435
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001436 while (!list_empty(&tags->page_list)) {
1437 page = list_first_entry(&tags->page_list, struct page, lru);
Dave Hansen67534712014-01-08 20:17:46 -07001438 list_del_init(&page->lru);
Catalin Marinasf75782e2015-09-14 18:16:02 +01001439 /*
1440 * Remove kmemleak object previously allocated in
1441 * blk_mq_init_rq_map().
1442 */
1443 kmemleak_free(page_address(page));
Jens Axboe320ae512013-10-24 09:20:05 +01001444 __free_pages(page, page->private);
1445 }
1446
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001447 kfree(tags->rqs);
Jens Axboe320ae512013-10-24 09:20:05 +01001448
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001449 blk_mq_free_tags(tags);
Jens Axboe320ae512013-10-24 09:20:05 +01001450}
1451
1452static size_t order_to_size(unsigned int order)
1453{
Ming Lei4ca08502014-04-19 18:00:18 +08001454 return (size_t)PAGE_SIZE << order;
Jens Axboe320ae512013-10-24 09:20:05 +01001455}
1456
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001457static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1458 unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01001459{
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001460 struct blk_mq_tags *tags;
Jens Axboe320ae512013-10-24 09:20:05 +01001461 unsigned int i, j, entries_per_page, max_order = 4;
1462 size_t rq_size, left;
1463
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001464 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
Shaohua Li24391c02015-01-23 14:18:00 -07001465 set->numa_node,
1466 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001467 if (!tags)
1468 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001469
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001470 INIT_LIST_HEAD(&tags->page_list);
1471
Jens Axboea5164402014-09-10 09:02:03 -06001472 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1473 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1474 set->numa_node);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001475 if (!tags->rqs) {
1476 blk_mq_free_tags(tags);
1477 return NULL;
1478 }
Jens Axboe320ae512013-10-24 09:20:05 +01001479
1480 /*
1481 * rq_size is the size of the request plus driver payload, rounded
1482 * to the cacheline size
1483 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001484 rq_size = round_up(sizeof(struct request) + set->cmd_size,
Jens Axboe320ae512013-10-24 09:20:05 +01001485 cache_line_size());
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001486 left = rq_size * set->queue_depth;
Jens Axboe320ae512013-10-24 09:20:05 +01001487
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001488 for (i = 0; i < set->queue_depth; ) {
Jens Axboe320ae512013-10-24 09:20:05 +01001489 int this_order = max_order;
1490 struct page *page;
1491 int to_do;
1492 void *p;
1493
1494 while (left < order_to_size(this_order - 1) && this_order)
1495 this_order--;
1496
1497 do {
Jens Axboea5164402014-09-10 09:02:03 -06001498 page = alloc_pages_node(set->numa_node,
Linus Torvaldsac211172015-04-09 14:12:22 -07001499 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
Jens Axboea5164402014-09-10 09:02:03 -06001500 this_order);
Jens Axboe320ae512013-10-24 09:20:05 +01001501 if (page)
1502 break;
1503 if (!this_order--)
1504 break;
1505 if (order_to_size(this_order) < rq_size)
1506 break;
1507 } while (1);
1508
1509 if (!page)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001510 goto fail;
Jens Axboe320ae512013-10-24 09:20:05 +01001511
1512 page->private = this_order;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001513 list_add_tail(&page->lru, &tags->page_list);
Jens Axboe320ae512013-10-24 09:20:05 +01001514
1515 p = page_address(page);
Catalin Marinasf75782e2015-09-14 18:16:02 +01001516 /*
1517 * Allow kmemleak to scan these pages as they contain pointers
1518 * to additional allocations like via ops->init_request().
1519 */
1520 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
Jens Axboe320ae512013-10-24 09:20:05 +01001521 entries_per_page = order_to_size(this_order) / rq_size;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001522 to_do = min(entries_per_page, set->queue_depth - i);
Jens Axboe320ae512013-10-24 09:20:05 +01001523 left -= to_do * rq_size;
1524 for (j = 0; j < to_do; j++) {
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001525 tags->rqs[i] = p;
1526 if (set->ops->init_request) {
1527 if (set->ops->init_request(set->driver_data,
1528 tags->rqs[i], hctx_idx, i,
Jens Axboea5164402014-09-10 09:02:03 -06001529 set->numa_node)) {
1530 tags->rqs[i] = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001531 goto fail;
Jens Axboea5164402014-09-10 09:02:03 -06001532 }
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001533 }
1534
Jens Axboe320ae512013-10-24 09:20:05 +01001535 p += rq_size;
1536 i++;
1537 }
1538 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001539 return tags;
Jens Axboe320ae512013-10-24 09:20:05 +01001540
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001541fail:
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001542 blk_mq_free_rq_map(set, tags, hctx_idx);
1543 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001544}
1545
Jens Axboe1429d7c2014-05-19 09:23:55 -06001546static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1547{
1548 kfree(bitmap->map);
1549}
1550
1551static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1552{
1553 unsigned int bpw = 8, total, num_maps, i;
1554
1555 bitmap->bits_per_word = bpw;
1556
1557 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1558 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1559 GFP_KERNEL, node);
1560 if (!bitmap->map)
1561 return -ENOMEM;
1562
Jens Axboe1429d7c2014-05-19 09:23:55 -06001563 total = nr_cpu_ids;
1564 for (i = 0; i < num_maps; i++) {
1565 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1566 total -= bitmap->map[i].depth;
1567 }
1568
1569 return 0;
1570}
1571
Jens Axboe484b4062014-05-21 14:01:15 -06001572static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1573{
1574 struct request_queue *q = hctx->queue;
1575 struct blk_mq_ctx *ctx;
1576 LIST_HEAD(tmp);
1577
1578 /*
1579 * Move ctx entries to new CPU, if this one is going away.
1580 */
1581 ctx = __blk_mq_get_ctx(q, cpu);
1582
1583 spin_lock(&ctx->lock);
1584 if (!list_empty(&ctx->rq_list)) {
1585 list_splice_init(&ctx->rq_list, &tmp);
1586 blk_mq_hctx_clear_pending(hctx, ctx);
1587 }
1588 spin_unlock(&ctx->lock);
1589
1590 if (list_empty(&tmp))
1591 return NOTIFY_OK;
1592
1593 ctx = blk_mq_get_ctx(q);
1594 spin_lock(&ctx->lock);
1595
1596 while (!list_empty(&tmp)) {
1597 struct request *rq;
1598
1599 rq = list_first_entry(&tmp, struct request, queuelist);
1600 rq->mq_ctx = ctx;
1601 list_move_tail(&rq->queuelist, &ctx->rq_list);
1602 }
1603
1604 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1605 blk_mq_hctx_mark_pending(hctx, ctx);
1606
1607 spin_unlock(&ctx->lock);
1608
1609 blk_mq_run_hw_queue(hctx, true);
1610 blk_mq_put_ctx(ctx);
1611 return NOTIFY_OK;
1612}
1613
Jens Axboe484b4062014-05-21 14:01:15 -06001614static int blk_mq_hctx_notify(void *data, unsigned long action,
1615 unsigned int cpu)
1616{
1617 struct blk_mq_hw_ctx *hctx = data;
1618
1619 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1620 return blk_mq_hctx_cpu_offline(hctx, cpu);
Ming Lei2a34c082015-04-21 10:00:20 +08001621
1622 /*
1623 * In case of CPU online, tags may be reallocated
1624 * in blk_mq_map_swqueue() after mapping is updated.
1625 */
Jens Axboe484b4062014-05-21 14:01:15 -06001626
1627 return NOTIFY_OK;
1628}
1629
Ming Leic3b4afc2015-06-04 22:25:04 +08001630/* hctx->ctxs will be freed in queue's release handler */
Ming Lei08e98fc2014-09-25 23:23:38 +08001631static void blk_mq_exit_hctx(struct request_queue *q,
1632 struct blk_mq_tag_set *set,
1633 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1634{
Ming Leif70ced02014-09-25 23:23:47 +08001635 unsigned flush_start_tag = set->queue_depth;
1636
Ming Lei08e98fc2014-09-25 23:23:38 +08001637 blk_mq_tag_idle(hctx);
1638
Ming Leif70ced02014-09-25 23:23:47 +08001639 if (set->ops->exit_request)
1640 set->ops->exit_request(set->driver_data,
1641 hctx->fq->flush_rq, hctx_idx,
1642 flush_start_tag + hctx_idx);
1643
Ming Lei08e98fc2014-09-25 23:23:38 +08001644 if (set->ops->exit_hctx)
1645 set->ops->exit_hctx(hctx, hctx_idx);
1646
1647 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
Ming Leif70ced02014-09-25 23:23:47 +08001648 blk_free_flush_queue(hctx->fq);
Ming Lei08e98fc2014-09-25 23:23:38 +08001649 blk_mq_free_bitmap(&hctx->ctx_map);
1650}
1651
Ming Lei624dbe42014-05-27 23:35:13 +08001652static void blk_mq_exit_hw_queues(struct request_queue *q,
1653 struct blk_mq_tag_set *set, int nr_queue)
1654{
1655 struct blk_mq_hw_ctx *hctx;
1656 unsigned int i;
1657
1658 queue_for_each_hw_ctx(q, hctx, i) {
1659 if (i == nr_queue)
1660 break;
Ming Lei08e98fc2014-09-25 23:23:38 +08001661 blk_mq_exit_hctx(q, set, hctx, i);
Ming Lei624dbe42014-05-27 23:35:13 +08001662 }
Ming Lei624dbe42014-05-27 23:35:13 +08001663}
1664
1665static void blk_mq_free_hw_queues(struct request_queue *q,
1666 struct blk_mq_tag_set *set)
1667{
1668 struct blk_mq_hw_ctx *hctx;
1669 unsigned int i;
1670
Ming Leie09aae72015-01-29 20:17:27 +08001671 queue_for_each_hw_ctx(q, hctx, i)
Ming Lei624dbe42014-05-27 23:35:13 +08001672 free_cpumask_var(hctx->cpumask);
Ming Lei624dbe42014-05-27 23:35:13 +08001673}
1674
Ming Lei08e98fc2014-09-25 23:23:38 +08001675static int blk_mq_init_hctx(struct request_queue *q,
1676 struct blk_mq_tag_set *set,
1677 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1678{
1679 int node;
Ming Leif70ced02014-09-25 23:23:47 +08001680 unsigned flush_start_tag = set->queue_depth;
Ming Lei08e98fc2014-09-25 23:23:38 +08001681
1682 node = hctx->numa_node;
1683 if (node == NUMA_NO_NODE)
1684 node = hctx->numa_node = set->numa_node;
1685
1686 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1687 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1688 spin_lock_init(&hctx->lock);
1689 INIT_LIST_HEAD(&hctx->dispatch);
1690 hctx->queue = q;
1691 hctx->queue_num = hctx_idx;
Jeff Moyer2404e602015-11-03 10:40:06 -05001692 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
Ming Lei08e98fc2014-09-25 23:23:38 +08001693
1694 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1695 blk_mq_hctx_notify, hctx);
1696 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1697
1698 hctx->tags = set->tags[hctx_idx];
1699
1700 /*
1701 * Allocate space for all possible cpus to avoid allocation at
1702 * runtime
1703 */
1704 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1705 GFP_KERNEL, node);
1706 if (!hctx->ctxs)
1707 goto unregister_cpu_notifier;
1708
1709 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1710 goto free_ctxs;
1711
1712 hctx->nr_ctx = 0;
1713
1714 if (set->ops->init_hctx &&
1715 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1716 goto free_bitmap;
1717
Ming Leif70ced02014-09-25 23:23:47 +08001718 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1719 if (!hctx->fq)
1720 goto exit_hctx;
1721
1722 if (set->ops->init_request &&
1723 set->ops->init_request(set->driver_data,
1724 hctx->fq->flush_rq, hctx_idx,
1725 flush_start_tag + hctx_idx, node))
1726 goto free_fq;
1727
Ming Lei08e98fc2014-09-25 23:23:38 +08001728 return 0;
1729
Ming Leif70ced02014-09-25 23:23:47 +08001730 free_fq:
1731 kfree(hctx->fq);
1732 exit_hctx:
1733 if (set->ops->exit_hctx)
1734 set->ops->exit_hctx(hctx, hctx_idx);
Ming Lei08e98fc2014-09-25 23:23:38 +08001735 free_bitmap:
1736 blk_mq_free_bitmap(&hctx->ctx_map);
1737 free_ctxs:
1738 kfree(hctx->ctxs);
1739 unregister_cpu_notifier:
1740 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1741
1742 return -1;
1743}
1744
Jens Axboe320ae512013-10-24 09:20:05 +01001745static int blk_mq_init_hw_queues(struct request_queue *q,
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001746 struct blk_mq_tag_set *set)
Jens Axboe320ae512013-10-24 09:20:05 +01001747{
1748 struct blk_mq_hw_ctx *hctx;
Ming Lei624dbe42014-05-27 23:35:13 +08001749 unsigned int i;
Jens Axboe320ae512013-10-24 09:20:05 +01001750
1751 /*
1752 * Initialize hardware queues
1753 */
1754 queue_for_each_hw_ctx(q, hctx, i) {
Ming Lei08e98fc2014-09-25 23:23:38 +08001755 if (blk_mq_init_hctx(q, set, hctx, i))
Jens Axboe320ae512013-10-24 09:20:05 +01001756 break;
1757 }
1758
1759 if (i == q->nr_hw_queues)
1760 return 0;
1761
1762 /*
1763 * Init failed
1764 */
Ming Lei624dbe42014-05-27 23:35:13 +08001765 blk_mq_exit_hw_queues(q, set, i);
Jens Axboe320ae512013-10-24 09:20:05 +01001766
1767 return 1;
1768}
1769
1770static void blk_mq_init_cpu_queues(struct request_queue *q,
1771 unsigned int nr_hw_queues)
1772{
1773 unsigned int i;
1774
1775 for_each_possible_cpu(i) {
1776 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1777 struct blk_mq_hw_ctx *hctx;
1778
1779 memset(__ctx, 0, sizeof(*__ctx));
1780 __ctx->cpu = i;
1781 spin_lock_init(&__ctx->lock);
1782 INIT_LIST_HEAD(&__ctx->rq_list);
1783 __ctx->queue = q;
1784
1785 /* If the cpu isn't online, the cpu is mapped to first hctx */
Jens Axboe320ae512013-10-24 09:20:05 +01001786 if (!cpu_online(i))
1787 continue;
1788
Jens Axboee4043dc2014-04-09 10:18:23 -06001789 hctx = q->mq_ops->map_queue(q, i);
Jens Axboee4043dc2014-04-09 10:18:23 -06001790
Jens Axboe320ae512013-10-24 09:20:05 +01001791 /*
1792 * Set local node, IFF we have more than one hw queue. If
1793 * not, we remain on the home node of the device
1794 */
1795 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
Raghavendra K Tbffed452015-12-02 16:59:05 +05301796 hctx->numa_node = local_memory_node(cpu_to_node(i));
Jens Axboe320ae512013-10-24 09:20:05 +01001797 }
1798}
1799
Akinobu Mita57783222015-09-27 02:09:23 +09001800static void blk_mq_map_swqueue(struct request_queue *q,
1801 const struct cpumask *online_mask)
Jens Axboe320ae512013-10-24 09:20:05 +01001802{
1803 unsigned int i;
1804 struct blk_mq_hw_ctx *hctx;
1805 struct blk_mq_ctx *ctx;
Ming Lei2a34c082015-04-21 10:00:20 +08001806 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01001807
Akinobu Mita60de0742015-09-27 02:09:25 +09001808 /*
1809 * Avoid others reading imcomplete hctx->cpumask through sysfs
1810 */
1811 mutex_lock(&q->sysfs_lock);
1812
Jens Axboe320ae512013-10-24 09:20:05 +01001813 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboee4043dc2014-04-09 10:18:23 -06001814 cpumask_clear(hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001815 hctx->nr_ctx = 0;
1816 }
1817
1818 /*
1819 * Map software to hardware queues
1820 */
1821 queue_for_each_ctx(q, ctx, i) {
1822 /* If the cpu isn't online, the cpu is mapped to first hctx */
Akinobu Mita57783222015-09-27 02:09:23 +09001823 if (!cpumask_test_cpu(i, online_mask))
Jens Axboee4043dc2014-04-09 10:18:23 -06001824 continue;
1825
Jens Axboe320ae512013-10-24 09:20:05 +01001826 hctx = q->mq_ops->map_queue(q, i);
Jens Axboee4043dc2014-04-09 10:18:23 -06001827 cpumask_set_cpu(i, hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001828 ctx->index_hw = hctx->nr_ctx;
1829 hctx->ctxs[hctx->nr_ctx++] = ctx;
1830 }
Jens Axboe506e9312014-05-07 10:26:44 -06001831
Akinobu Mita60de0742015-09-27 02:09:25 +09001832 mutex_unlock(&q->sysfs_lock);
1833
Jens Axboe506e9312014-05-07 10:26:44 -06001834 queue_for_each_hw_ctx(q, hctx, i) {
Chong Yuan889fa312015-04-15 11:39:29 -06001835 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1836
Jens Axboe484b4062014-05-21 14:01:15 -06001837 /*
Jens Axboea68aafa2014-08-15 13:19:15 -06001838 * If no software queues are mapped to this hardware queue,
1839 * disable it and free the request entries.
Jens Axboe484b4062014-05-21 14:01:15 -06001840 */
1841 if (!hctx->nr_ctx) {
Jens Axboe484b4062014-05-21 14:01:15 -06001842 if (set->tags[i]) {
1843 blk_mq_free_rq_map(set, set->tags[i], i);
1844 set->tags[i] = NULL;
Jens Axboe484b4062014-05-21 14:01:15 -06001845 }
Ming Lei2a34c082015-04-21 10:00:20 +08001846 hctx->tags = NULL;
Jens Axboe484b4062014-05-21 14:01:15 -06001847 continue;
1848 }
1849
Ming Lei2a34c082015-04-21 10:00:20 +08001850 /* unmapped hw queue can be remapped after CPU topo changed */
1851 if (!set->tags[i])
1852 set->tags[i] = blk_mq_init_rq_map(set, i);
1853 hctx->tags = set->tags[i];
1854 WARN_ON(!hctx->tags);
1855
Raghavendra K Te0e827b2015-12-02 16:57:06 +05301856 cpumask_copy(hctx->tags->cpumask, hctx->cpumask);
Jens Axboe484b4062014-05-21 14:01:15 -06001857 /*
Chong Yuan889fa312015-04-15 11:39:29 -06001858 * Set the map size to the number of mapped software queues.
1859 * This is more accurate and more efficient than looping
1860 * over all possibly mapped software queues.
1861 */
Jens Axboe569fd0c2015-04-17 08:28:50 -06001862 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
Chong Yuan889fa312015-04-15 11:39:29 -06001863
1864 /*
Jens Axboe484b4062014-05-21 14:01:15 -06001865 * Initialize batch roundrobin counts
1866 */
Jens Axboe506e9312014-05-07 10:26:44 -06001867 hctx->next_cpu = cpumask_first(hctx->cpumask);
1868 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1869 }
Jens Axboe320ae512013-10-24 09:20:05 +01001870}
1871
Jeff Moyer2404e602015-11-03 10:40:06 -05001872static void queue_set_hctx_shared(struct request_queue *q, bool shared)
Jens Axboe0d2602c2014-05-13 15:10:52 -06001873{
1874 struct blk_mq_hw_ctx *hctx;
Jens Axboe0d2602c2014-05-13 15:10:52 -06001875 int i;
1876
Jeff Moyer2404e602015-11-03 10:40:06 -05001877 queue_for_each_hw_ctx(q, hctx, i) {
1878 if (shared)
1879 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1880 else
1881 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1882 }
1883}
1884
1885static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1886{
1887 struct request_queue *q;
Jens Axboe0d2602c2014-05-13 15:10:52 -06001888
1889 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1890 blk_mq_freeze_queue(q);
Jeff Moyer2404e602015-11-03 10:40:06 -05001891 queue_set_hctx_shared(q, shared);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001892 blk_mq_unfreeze_queue(q);
1893 }
1894}
1895
1896static void blk_mq_del_queue_tag_set(struct request_queue *q)
1897{
1898 struct blk_mq_tag_set *set = q->tag_set;
1899
Jens Axboe0d2602c2014-05-13 15:10:52 -06001900 mutex_lock(&set->tag_list_lock);
1901 list_del_init(&q->tag_set_list);
Jeff Moyer2404e602015-11-03 10:40:06 -05001902 if (list_is_singular(&set->tag_list)) {
1903 /* just transitioned to unshared */
1904 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1905 /* update existing queue */
1906 blk_mq_update_tag_set_depth(set, false);
1907 }
Jens Axboe0d2602c2014-05-13 15:10:52 -06001908 mutex_unlock(&set->tag_list_lock);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001909}
1910
1911static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1912 struct request_queue *q)
1913{
1914 q->tag_set = set;
1915
1916 mutex_lock(&set->tag_list_lock);
Jeff Moyer2404e602015-11-03 10:40:06 -05001917
1918 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1919 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1920 set->flags |= BLK_MQ_F_TAG_SHARED;
1921 /* update existing queue */
1922 blk_mq_update_tag_set_depth(set, true);
1923 }
1924 if (set->flags & BLK_MQ_F_TAG_SHARED)
1925 queue_set_hctx_shared(q, true);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001926 list_add_tail(&q->tag_set_list, &set->tag_list);
Jeff Moyer2404e602015-11-03 10:40:06 -05001927
Jens Axboe0d2602c2014-05-13 15:10:52 -06001928 mutex_unlock(&set->tag_list_lock);
1929}
1930
Ming Leie09aae72015-01-29 20:17:27 +08001931/*
1932 * It is the actual release handler for mq, but we do it from
1933 * request queue's release handler for avoiding use-after-free
1934 * and headache because q->mq_kobj shouldn't have been introduced,
1935 * but we can't group ctx/kctx kobj without it.
1936 */
1937void blk_mq_release(struct request_queue *q)
1938{
1939 struct blk_mq_hw_ctx *hctx;
1940 unsigned int i;
1941
1942 /* hctx kobj stays in hctx */
Ming Leic3b4afc2015-06-04 22:25:04 +08001943 queue_for_each_hw_ctx(q, hctx, i) {
1944 if (!hctx)
1945 continue;
1946 kfree(hctx->ctxs);
Ming Leie09aae72015-01-29 20:17:27 +08001947 kfree(hctx);
Ming Leic3b4afc2015-06-04 22:25:04 +08001948 }
Ming Leie09aae72015-01-29 20:17:27 +08001949
Akinobu Mitaa723bab2015-09-27 02:09:21 +09001950 kfree(q->mq_map);
1951 q->mq_map = NULL;
1952
Ming Leie09aae72015-01-29 20:17:27 +08001953 kfree(q->queue_hw_ctx);
1954
1955 /* ctx kobj stays in queue_ctx */
1956 free_percpu(q->queue_ctx);
1957}
1958
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001959struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
Jens Axboe320ae512013-10-24 09:20:05 +01001960{
Mike Snitzerb62c21b2015-03-12 23:56:02 -04001961 struct request_queue *uninit_q, *q;
1962
1963 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1964 if (!uninit_q)
1965 return ERR_PTR(-ENOMEM);
1966
1967 q = blk_mq_init_allocated_queue(set, uninit_q);
1968 if (IS_ERR(q))
1969 blk_cleanup_queue(uninit_q);
1970
1971 return q;
1972}
1973EXPORT_SYMBOL(blk_mq_init_queue);
1974
1975struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1976 struct request_queue *q)
1977{
Jens Axboe320ae512013-10-24 09:20:05 +01001978 struct blk_mq_hw_ctx **hctxs;
Ming Leie6cdb092014-06-03 11:24:06 +08001979 struct blk_mq_ctx __percpu *ctx;
Jens Axboef14bbe72014-05-27 12:06:53 -06001980 unsigned int *map;
Jens Axboe320ae512013-10-24 09:20:05 +01001981 int i;
1982
Jens Axboe320ae512013-10-24 09:20:05 +01001983 ctx = alloc_percpu(struct blk_mq_ctx);
1984 if (!ctx)
1985 return ERR_PTR(-ENOMEM);
1986
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001987 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1988 set->numa_node);
Jens Axboe320ae512013-10-24 09:20:05 +01001989
1990 if (!hctxs)
1991 goto err_percpu;
1992
Jens Axboef14bbe72014-05-27 12:06:53 -06001993 map = blk_mq_make_queue_map(set);
1994 if (!map)
1995 goto err_map;
1996
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001997 for (i = 0; i < set->nr_hw_queues; i++) {
Jens Axboef14bbe72014-05-27 12:06:53 -06001998 int node = blk_mq_hw_queue_to_node(map, i);
1999
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02002000 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2001 GFP_KERNEL, node);
Jens Axboe320ae512013-10-24 09:20:05 +01002002 if (!hctxs[i])
2003 goto err_hctxs;
2004
Jens Axboea86073e2014-10-13 15:41:54 -06002005 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2006 node))
Jens Axboee4043dc2014-04-09 10:18:23 -06002007 goto err_hctxs;
2008
Jens Axboe0d2602c2014-05-13 15:10:52 -06002009 atomic_set(&hctxs[i]->nr_active, 0);
Jens Axboef14bbe72014-05-27 12:06:53 -06002010 hctxs[i]->numa_node = node;
Jens Axboe320ae512013-10-24 09:20:05 +01002011 hctxs[i]->queue_num = i;
2012 }
2013
Christoph Hellwig287922e2015-10-30 20:57:30 +08002014 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
Ming Leie56f6982015-07-16 19:53:22 +08002015 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
Jens Axboe320ae512013-10-24 09:20:05 +01002016
2017 q->nr_queues = nr_cpu_ids;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002018 q->nr_hw_queues = set->nr_hw_queues;
Jens Axboef14bbe72014-05-27 12:06:53 -06002019 q->mq_map = map;
Jens Axboe320ae512013-10-24 09:20:05 +01002020
2021 q->queue_ctx = ctx;
2022 q->queue_hw_ctx = hctxs;
2023
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002024 q->mq_ops = set->ops;
Jens Axboe94eddfb2013-11-19 09:25:07 -07002025 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
Jens Axboe320ae512013-10-24 09:20:05 +01002026
Jens Axboe05f1dd52014-05-29 09:53:32 -06002027 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2028 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2029
Christoph Hellwig1be036e2014-02-07 10:22:39 -08002030 q->sg_reserved_size = INT_MAX;
2031
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06002032 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2033 INIT_LIST_HEAD(&q->requeue_list);
2034 spin_lock_init(&q->requeue_lock);
2035
Jens Axboe07068d52014-05-22 10:40:51 -06002036 if (q->nr_hw_queues > 1)
2037 blk_queue_make_request(q, blk_mq_make_request);
2038 else
2039 blk_queue_make_request(q, blk_sq_make_request);
2040
Jens Axboeeba71762014-05-20 15:17:27 -06002041 /*
2042 * Do this after blk_queue_make_request() overrides it...
2043 */
2044 q->nr_requests = set->queue_depth;
2045
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002046 if (set->ops->complete)
2047 blk_queue_softirq_done(q, set->ops->complete);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08002048
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002049 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +01002050
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002051 if (blk_mq_init_hw_queues(q, set))
Mike Snitzerb62c21b2015-03-12 23:56:02 -04002052 goto err_hctxs;
Christoph Hellwig18741982014-02-10 09:29:00 -07002053
Akinobu Mita57783222015-09-27 02:09:23 +09002054 get_online_cpus();
Jens Axboe320ae512013-10-24 09:20:05 +01002055 mutex_lock(&all_q_mutex);
Akinobu Mita4593fdb2015-09-27 02:09:20 +09002056
Jens Axboe320ae512013-10-24 09:20:05 +01002057 list_add_tail(&q->all_q_node, &all_q_list);
Jens Axboe0d2602c2014-05-13 15:10:52 -06002058 blk_mq_add_queue_tag_set(set, q);
Akinobu Mita57783222015-09-27 02:09:23 +09002059 blk_mq_map_swqueue(q, cpu_online_mask);
Jens Axboe484b4062014-05-21 14:01:15 -06002060
Akinobu Mita4593fdb2015-09-27 02:09:20 +09002061 mutex_unlock(&all_q_mutex);
Akinobu Mita57783222015-09-27 02:09:23 +09002062 put_online_cpus();
Akinobu Mita4593fdb2015-09-27 02:09:20 +09002063
Jens Axboe320ae512013-10-24 09:20:05 +01002064 return q;
Christoph Hellwig18741982014-02-10 09:29:00 -07002065
Jens Axboe320ae512013-10-24 09:20:05 +01002066err_hctxs:
Jens Axboef14bbe72014-05-27 12:06:53 -06002067 kfree(map);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002068 for (i = 0; i < set->nr_hw_queues; i++) {
Jens Axboe320ae512013-10-24 09:20:05 +01002069 if (!hctxs[i])
2070 break;
Jens Axboee4043dc2014-04-09 10:18:23 -06002071 free_cpumask_var(hctxs[i]->cpumask);
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02002072 kfree(hctxs[i]);
Jens Axboe320ae512013-10-24 09:20:05 +01002073 }
Jens Axboef14bbe72014-05-27 12:06:53 -06002074err_map:
Jens Axboe320ae512013-10-24 09:20:05 +01002075 kfree(hctxs);
2076err_percpu:
2077 free_percpu(ctx);
2078 return ERR_PTR(-ENOMEM);
2079}
Mike Snitzerb62c21b2015-03-12 23:56:02 -04002080EXPORT_SYMBOL(blk_mq_init_allocated_queue);
Jens Axboe320ae512013-10-24 09:20:05 +01002081
2082void blk_mq_free_queue(struct request_queue *q)
2083{
Ming Lei624dbe42014-05-27 23:35:13 +08002084 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01002085
Akinobu Mita0e626362015-09-27 02:09:22 +09002086 mutex_lock(&all_q_mutex);
2087 list_del_init(&q->all_q_node);
2088 mutex_unlock(&all_q_mutex);
2089
Jens Axboe0d2602c2014-05-13 15:10:52 -06002090 blk_mq_del_queue_tag_set(q);
2091
Ming Lei624dbe42014-05-27 23:35:13 +08002092 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2093 blk_mq_free_hw_queues(q, set);
Jens Axboe320ae512013-10-24 09:20:05 +01002094}
Jens Axboe320ae512013-10-24 09:20:05 +01002095
2096/* Basically redo blk_mq_init_queue with queue frozen */
Akinobu Mita57783222015-09-27 02:09:23 +09002097static void blk_mq_queue_reinit(struct request_queue *q,
2098 const struct cpumask *online_mask)
Jens Axboe320ae512013-10-24 09:20:05 +01002099{
Christoph Hellwig4ecd4fe2015-05-07 09:38:13 +02002100 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
Jens Axboe320ae512013-10-24 09:20:05 +01002101
Jens Axboe67aec142014-05-30 08:25:36 -06002102 blk_mq_sysfs_unregister(q);
2103
Akinobu Mita57783222015-09-27 02:09:23 +09002104 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
Jens Axboe320ae512013-10-24 09:20:05 +01002105
2106 /*
2107 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2108 * we should change hctx numa_node according to new topology (this
2109 * involves free and re-allocate memory, worthy doing?)
2110 */
2111
Akinobu Mita57783222015-09-27 02:09:23 +09002112 blk_mq_map_swqueue(q, online_mask);
Jens Axboe320ae512013-10-24 09:20:05 +01002113
Jens Axboe67aec142014-05-30 08:25:36 -06002114 blk_mq_sysfs_register(q);
Jens Axboe320ae512013-10-24 09:20:05 +01002115}
2116
Paul Gortmakerf618ef72013-11-14 08:26:02 -07002117static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2118 unsigned long action, void *hcpu)
Jens Axboe320ae512013-10-24 09:20:05 +01002119{
2120 struct request_queue *q;
Akinobu Mita57783222015-09-27 02:09:23 +09002121 int cpu = (unsigned long)hcpu;
2122 /*
2123 * New online cpumask which is going to be set in this hotplug event.
2124 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2125 * one-by-one and dynamically allocating this could result in a failure.
2126 */
2127 static struct cpumask online_new;
Jens Axboe320ae512013-10-24 09:20:05 +01002128
2129 /*
Akinobu Mita57783222015-09-27 02:09:23 +09002130 * Before hotadded cpu starts handling requests, new mappings must
2131 * be established. Otherwise, these requests in hw queue might
2132 * never be dispatched.
2133 *
2134 * For example, there is a single hw queue (hctx) and two CPU queues
2135 * (ctx0 for CPU0, and ctx1 for CPU1).
2136 *
2137 * Now CPU1 is just onlined and a request is inserted into
2138 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2139 * still zero.
2140 *
2141 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2142 * set in pending bitmap and tries to retrieve requests in
2143 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2144 * so the request in ctx1->rq_list is ignored.
Jens Axboe320ae512013-10-24 09:20:05 +01002145 */
Akinobu Mita57783222015-09-27 02:09:23 +09002146 switch (action & ~CPU_TASKS_FROZEN) {
2147 case CPU_DEAD:
2148 case CPU_UP_CANCELED:
2149 cpumask_copy(&online_new, cpu_online_mask);
2150 break;
2151 case CPU_UP_PREPARE:
2152 cpumask_copy(&online_new, cpu_online_mask);
2153 cpumask_set_cpu(cpu, &online_new);
2154 break;
2155 default:
Jens Axboe320ae512013-10-24 09:20:05 +01002156 return NOTIFY_OK;
Akinobu Mita57783222015-09-27 02:09:23 +09002157 }
Jens Axboe320ae512013-10-24 09:20:05 +01002158
2159 mutex_lock(&all_q_mutex);
Tejun Heof3af0202014-11-04 13:52:27 -05002160
2161 /*
2162 * We need to freeze and reinit all existing queues. Freezing
2163 * involves synchronous wait for an RCU grace period and doing it
2164 * one by one may take a long time. Start freezing all queues in
2165 * one swoop and then wait for the completions so that freezing can
2166 * take place in parallel.
2167 */
2168 list_for_each_entry(q, &all_q_list, all_q_node)
2169 blk_mq_freeze_queue_start(q);
Ming Leif054b562015-04-21 10:00:19 +08002170 list_for_each_entry(q, &all_q_list, all_q_node) {
Tejun Heof3af0202014-11-04 13:52:27 -05002171 blk_mq_freeze_queue_wait(q);
2172
Ming Leif054b562015-04-21 10:00:19 +08002173 /*
2174 * timeout handler can't touch hw queue during the
2175 * reinitialization
2176 */
2177 del_timer_sync(&q->timeout);
2178 }
2179
Jens Axboe320ae512013-10-24 09:20:05 +01002180 list_for_each_entry(q, &all_q_list, all_q_node)
Akinobu Mita57783222015-09-27 02:09:23 +09002181 blk_mq_queue_reinit(q, &online_new);
Tejun Heof3af0202014-11-04 13:52:27 -05002182
2183 list_for_each_entry(q, &all_q_list, all_q_node)
2184 blk_mq_unfreeze_queue(q);
2185
Jens Axboe320ae512013-10-24 09:20:05 +01002186 mutex_unlock(&all_q_mutex);
2187 return NOTIFY_OK;
2188}
2189
Jens Axboea5164402014-09-10 09:02:03 -06002190static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2191{
2192 int i;
2193
2194 for (i = 0; i < set->nr_hw_queues; i++) {
2195 set->tags[i] = blk_mq_init_rq_map(set, i);
2196 if (!set->tags[i])
2197 goto out_unwind;
2198 }
2199
2200 return 0;
2201
2202out_unwind:
2203 while (--i >= 0)
2204 blk_mq_free_rq_map(set, set->tags[i], i);
2205
Jens Axboea5164402014-09-10 09:02:03 -06002206 return -ENOMEM;
2207}
2208
2209/*
2210 * Allocate the request maps associated with this tag_set. Note that this
2211 * may reduce the depth asked for, if memory is tight. set->queue_depth
2212 * will be updated to reflect the allocated depth.
2213 */
2214static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2215{
2216 unsigned int depth;
2217 int err;
2218
2219 depth = set->queue_depth;
2220 do {
2221 err = __blk_mq_alloc_rq_maps(set);
2222 if (!err)
2223 break;
2224
2225 set->queue_depth >>= 1;
2226 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2227 err = -ENOMEM;
2228 break;
2229 }
2230 } while (set->queue_depth);
2231
2232 if (!set->queue_depth || err) {
2233 pr_err("blk-mq: failed to allocate request map\n");
2234 return -ENOMEM;
2235 }
2236
2237 if (depth != set->queue_depth)
2238 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2239 depth, set->queue_depth);
2240
2241 return 0;
2242}
2243
Keith Buschf26cdc82015-06-01 09:29:53 -06002244struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2245{
2246 return tags->cpumask;
2247}
2248EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2249
Jens Axboea4391c62014-06-05 15:21:56 -06002250/*
2251 * Alloc a tag set to be associated with one or more request queues.
2252 * May fail with EINVAL for various error conditions. May adjust the
2253 * requested depth down, if if it too large. In that case, the set
2254 * value will be stored in set->queue_depth.
2255 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002256int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2257{
Bart Van Assche205fb5f2014-10-30 14:45:11 +01002258 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2259
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002260 if (!set->nr_hw_queues)
2261 return -EINVAL;
Jens Axboea4391c62014-06-05 15:21:56 -06002262 if (!set->queue_depth)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002263 return -EINVAL;
2264 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2265 return -EINVAL;
2266
Xiaoguang Wangf9018ac2015-03-30 13:19:14 +08002267 if (!set->ops->queue_rq || !set->ops->map_queue)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002268 return -EINVAL;
2269
Jens Axboea4391c62014-06-05 15:21:56 -06002270 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2271 pr_info("blk-mq: reduced tag depth to %u\n",
2272 BLK_MQ_MAX_DEPTH);
2273 set->queue_depth = BLK_MQ_MAX_DEPTH;
2274 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002275
Shaohua Li6637fad2014-11-30 16:00:58 -08002276 /*
2277 * If a crashdump is active, then we are potentially in a very
2278 * memory constrained environment. Limit us to 1 queue and
2279 * 64 tags to prevent using too much memory.
2280 */
2281 if (is_kdump_kernel()) {
2282 set->nr_hw_queues = 1;
2283 set->queue_depth = min(64U, set->queue_depth);
2284 }
2285
Ming Lei48479002014-04-19 18:00:17 +08002286 set->tags = kmalloc_node(set->nr_hw_queues *
2287 sizeof(struct blk_mq_tags *),
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002288 GFP_KERNEL, set->numa_node);
2289 if (!set->tags)
Jens Axboea5164402014-09-10 09:02:03 -06002290 return -ENOMEM;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002291
Jens Axboea5164402014-09-10 09:02:03 -06002292 if (blk_mq_alloc_rq_maps(set))
2293 goto enomem;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002294
Jens Axboe0d2602c2014-05-13 15:10:52 -06002295 mutex_init(&set->tag_list_lock);
2296 INIT_LIST_HEAD(&set->tag_list);
2297
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002298 return 0;
Jens Axboea5164402014-09-10 09:02:03 -06002299enomem:
Robert Elliott5676e7b2014-09-02 11:38:44 -05002300 kfree(set->tags);
2301 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002302 return -ENOMEM;
2303}
2304EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2305
2306void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2307{
2308 int i;
2309
Jens Axboe484b4062014-05-21 14:01:15 -06002310 for (i = 0; i < set->nr_hw_queues; i++) {
Junichi Nomuraf42d79a2015-10-14 05:02:15 +00002311 if (set->tags[i])
Jens Axboe484b4062014-05-21 14:01:15 -06002312 blk_mq_free_rq_map(set, set->tags[i], i);
2313 }
2314
Ming Lei981bd182014-04-24 00:07:34 +08002315 kfree(set->tags);
Robert Elliott5676e7b2014-09-02 11:38:44 -05002316 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002317}
2318EXPORT_SYMBOL(blk_mq_free_tag_set);
2319
Jens Axboee3a2b3f2014-05-20 11:49:02 -06002320int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2321{
2322 struct blk_mq_tag_set *set = q->tag_set;
2323 struct blk_mq_hw_ctx *hctx;
2324 int i, ret;
2325
2326 if (!set || nr > set->queue_depth)
2327 return -EINVAL;
2328
2329 ret = 0;
2330 queue_for_each_hw_ctx(q, hctx, i) {
2331 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2332 if (ret)
2333 break;
2334 }
2335
2336 if (!ret)
2337 q->nr_requests = nr;
2338
2339 return ret;
2340}
2341
Jens Axboe676141e2014-03-20 13:29:18 -06002342void blk_mq_disable_hotplug(void)
2343{
2344 mutex_lock(&all_q_mutex);
2345}
2346
2347void blk_mq_enable_hotplug(void)
2348{
2349 mutex_unlock(&all_q_mutex);
2350}
2351
Jens Axboe320ae512013-10-24 09:20:05 +01002352static int __init blk_mq_init(void)
2353{
Jens Axboe320ae512013-10-24 09:20:05 +01002354 blk_mq_cpu_init();
2355
Tejun Heoadd703f2014-07-01 10:34:38 -06002356 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
Jens Axboe320ae512013-10-24 09:20:05 +01002357
2358 return 0;
2359}
2360subsys_initcall(blk_mq_init);