blob: 7d842db59699692dd0ebce5daf6a71443f0e4314 [file] [log] [blame]
Jens Axboe75bb4622014-05-28 10:15:41 -06001/*
2 * Block multiqueue core code
3 *
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
6 */
Jens Axboe320ae512013-10-24 09:20:05 +01007#include <linux/kernel.h>
8#include <linux/module.h>
9#include <linux/backing-dev.h>
10#include <linux/bio.h>
11#include <linux/blkdev.h>
12#include <linux/mm.h>
13#include <linux/init.h>
14#include <linux/slab.h>
15#include <linux/workqueue.h>
16#include <linux/smp.h>
17#include <linux/llist.h>
18#include <linux/list_sort.h>
19#include <linux/cpu.h>
20#include <linux/cache.h>
21#include <linux/sched/sysctl.h>
22#include <linux/delay.h>
Jens Axboeaedcd722014-09-17 08:27:03 -060023#include <linux/crash_dump.h>
Jens Axboe320ae512013-10-24 09:20:05 +010024
25#include <trace/events/block.h>
26
27#include <linux/blk-mq.h>
28#include "blk.h"
29#include "blk-mq.h"
30#include "blk-mq-tag.h"
31
32static DEFINE_MUTEX(all_q_mutex);
33static LIST_HEAD(all_q_list);
34
35static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
36
Jens Axboe320ae512013-10-24 09:20:05 +010037/*
38 * Check if any of the ctx's have pending work in this hardware queue
39 */
40static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
41{
42 unsigned int i;
43
Jens Axboe569fd0c2015-04-17 08:28:50 -060044 for (i = 0; i < hctx->ctx_map.size; i++)
Jens Axboe1429d7c2014-05-19 09:23:55 -060045 if (hctx->ctx_map.map[i].word)
Jens Axboe320ae512013-10-24 09:20:05 +010046 return true;
47
48 return false;
49}
50
Jens Axboe1429d7c2014-05-19 09:23:55 -060051static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
52 struct blk_mq_ctx *ctx)
53{
54 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
55}
56
57#define CTX_TO_BIT(hctx, ctx) \
58 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
59
Jens Axboe320ae512013-10-24 09:20:05 +010060/*
61 * Mark this ctx as having pending work in this hardware queue
62 */
63static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
64 struct blk_mq_ctx *ctx)
65{
Jens Axboe1429d7c2014-05-19 09:23:55 -060066 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
67
68 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
69 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
70}
71
72static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
73 struct blk_mq_ctx *ctx)
74{
75 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
76
77 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
Jens Axboe320ae512013-10-24 09:20:05 +010078}
79
Keith Buschbfd343a2015-03-11 23:56:39 -040080static int blk_mq_queue_enter(struct request_queue *q, gfp_t gfp)
Jens Axboe320ae512013-10-24 09:20:05 +010081{
Tejun Heoadd703f2014-07-01 10:34:38 -060082 while (true) {
83 int ret;
Jens Axboe320ae512013-10-24 09:20:05 +010084
Tejun Heoadd703f2014-07-01 10:34:38 -060085 if (percpu_ref_tryget_live(&q->mq_usage_counter))
86 return 0;
Keith Busch3b632cf2014-06-06 10:22:07 -060087
Keith Buschbfd343a2015-03-11 23:56:39 -040088 if (!(gfp & __GFP_WAIT))
89 return -EBUSY;
90
Tejun Heoadd703f2014-07-01 10:34:38 -060091 ret = wait_event_interruptible(q->mq_freeze_wq,
Christoph Hellwig4ecd4fe2015-05-07 09:38:13 +020092 !atomic_read(&q->mq_freeze_depth) ||
93 blk_queue_dying(q));
Tejun Heoadd703f2014-07-01 10:34:38 -060094 if (blk_queue_dying(q))
95 return -ENODEV;
96 if (ret)
97 return ret;
98 }
Jens Axboe320ae512013-10-24 09:20:05 +010099}
100
101static void blk_mq_queue_exit(struct request_queue *q)
102{
Tejun Heoadd703f2014-07-01 10:34:38 -0600103 percpu_ref_put(&q->mq_usage_counter);
104}
105
106static void blk_mq_usage_counter_release(struct percpu_ref *ref)
107{
108 struct request_queue *q =
109 container_of(ref, struct request_queue, mq_usage_counter);
110
111 wake_up_all(&q->mq_freeze_wq);
Jens Axboe320ae512013-10-24 09:20:05 +0100112}
113
Keith Buschb4c6a022014-12-19 17:54:14 -0700114void blk_mq_freeze_queue_start(struct request_queue *q)
Ming Lei43a5e4e2013-12-26 21:31:35 +0800115{
Christoph Hellwig4ecd4fe2015-05-07 09:38:13 +0200116 int freeze_depth;
Tejun Heocddd5d12014-08-16 08:02:24 -0400117
Christoph Hellwig4ecd4fe2015-05-07 09:38:13 +0200118 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
119 if (freeze_depth == 1) {
Tejun Heo9eca8042014-09-24 13:07:33 -0400120 percpu_ref_kill(&q->mq_usage_counter);
Mike Snitzerb94ec292015-03-11 23:56:38 -0400121 blk_mq_run_hw_queues(q, false);
Tejun Heocddd5d12014-08-16 08:02:24 -0400122 }
Tejun Heof3af0202014-11-04 13:52:27 -0500123}
Keith Buschb4c6a022014-12-19 17:54:14 -0700124EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
Tejun Heof3af0202014-11-04 13:52:27 -0500125
126static void blk_mq_freeze_queue_wait(struct request_queue *q)
127{
Tejun Heoadd703f2014-07-01 10:34:38 -0600128 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
Ming Lei43a5e4e2013-12-26 21:31:35 +0800129}
130
Tejun Heof3af0202014-11-04 13:52:27 -0500131/*
132 * Guarantee no request is in use, so we can change any data structure of
133 * the queue afterward.
134 */
135void blk_mq_freeze_queue(struct request_queue *q)
136{
137 blk_mq_freeze_queue_start(q);
138 blk_mq_freeze_queue_wait(q);
139}
Jens Axboec761d962015-01-02 15:05:12 -0700140EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
Tejun Heof3af0202014-11-04 13:52:27 -0500141
Keith Buschb4c6a022014-12-19 17:54:14 -0700142void blk_mq_unfreeze_queue(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +0100143{
Christoph Hellwig4ecd4fe2015-05-07 09:38:13 +0200144 int freeze_depth;
Jens Axboe320ae512013-10-24 09:20:05 +0100145
Christoph Hellwig4ecd4fe2015-05-07 09:38:13 +0200146 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
147 WARN_ON_ONCE(freeze_depth < 0);
148 if (!freeze_depth) {
Tejun Heoadd703f2014-07-01 10:34:38 -0600149 percpu_ref_reinit(&q->mq_usage_counter);
Jens Axboe320ae512013-10-24 09:20:05 +0100150 wake_up_all(&q->mq_freeze_wq);
Tejun Heoadd703f2014-07-01 10:34:38 -0600151 }
Jens Axboe320ae512013-10-24 09:20:05 +0100152}
Keith Buschb4c6a022014-12-19 17:54:14 -0700153EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
Jens Axboe320ae512013-10-24 09:20:05 +0100154
Jens Axboeaed3ea92014-12-22 14:04:42 -0700155void blk_mq_wake_waiters(struct request_queue *q)
156{
157 struct blk_mq_hw_ctx *hctx;
158 unsigned int i;
159
160 queue_for_each_hw_ctx(q, hctx, i)
161 if (blk_mq_hw_queue_mapped(hctx))
162 blk_mq_tag_wakeup_all(hctx->tags, true);
Keith Busch3fd59402015-01-08 08:53:56 -0700163
164 /*
165 * If we are called because the queue has now been marked as
166 * dying, we need to ensure that processes currently waiting on
167 * the queue are notified as well.
168 */
169 wake_up_all(&q->mq_freeze_wq);
Jens Axboeaed3ea92014-12-22 14:04:42 -0700170}
171
Jens Axboe320ae512013-10-24 09:20:05 +0100172bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
173{
174 return blk_mq_has_free_tags(hctx->tags);
175}
176EXPORT_SYMBOL(blk_mq_can_queue);
177
Jens Axboe94eddfb2013-11-19 09:25:07 -0700178static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
179 struct request *rq, unsigned int rw_flags)
Jens Axboe320ae512013-10-24 09:20:05 +0100180{
Jens Axboe94eddfb2013-11-19 09:25:07 -0700181 if (blk_queue_io_stat(q))
182 rw_flags |= REQ_IO_STAT;
183
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200184 INIT_LIST_HEAD(&rq->queuelist);
185 /* csd/requeue_work/fifo_time is initialized before use */
186 rq->q = q;
Jens Axboe320ae512013-10-24 09:20:05 +0100187 rq->mq_ctx = ctx;
Jens Axboe0d2602c2014-05-13 15:10:52 -0600188 rq->cmd_flags |= rw_flags;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200189 /* do not touch atomic flags, it needs atomic ops against the timer */
190 rq->cpu = -1;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200191 INIT_HLIST_NODE(&rq->hash);
192 RB_CLEAR_NODE(&rq->rb_node);
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200193 rq->rq_disk = NULL;
194 rq->part = NULL;
Jens Axboe3ee32372014-06-09 09:36:53 -0600195 rq->start_time = jiffies;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200196#ifdef CONFIG_BLK_CGROUP
197 rq->rl = NULL;
Ming Lei0fec08b2014-01-03 10:00:08 -0700198 set_start_time_ns(rq);
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200199 rq->io_start_time_ns = 0;
200#endif
201 rq->nr_phys_segments = 0;
202#if defined(CONFIG_BLK_DEV_INTEGRITY)
203 rq->nr_integrity_segments = 0;
204#endif
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200205 rq->special = NULL;
206 /* tag was already set */
207 rq->errors = 0;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200208
Tony Battersby6f4a16262014-08-22 15:53:39 -0400209 rq->cmd = rq->__cmd;
210
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200211 rq->extra_len = 0;
212 rq->sense_len = 0;
213 rq->resid_len = 0;
214 rq->sense = NULL;
215
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200216 INIT_LIST_HEAD(&rq->timeout_list);
Jens Axboef6be4fb2014-06-06 11:03:48 -0600217 rq->timeout = 0;
218
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200219 rq->end_io = NULL;
220 rq->end_io_data = NULL;
221 rq->next_rq = NULL;
222
Jens Axboe320ae512013-10-24 09:20:05 +0100223 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
224}
225
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200226static struct request *
Ming Leicb96a422014-06-01 00:43:37 +0800227__blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200228{
229 struct request *rq;
230 unsigned int tag;
231
Ming Leicb96a422014-06-01 00:43:37 +0800232 tag = blk_mq_get_tag(data);
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200233 if (tag != BLK_MQ_TAG_FAIL) {
Ming Leicb96a422014-06-01 00:43:37 +0800234 rq = data->hctx->tags->rqs[tag];
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200235
Ming Leicb96a422014-06-01 00:43:37 +0800236 if (blk_mq_tag_busy(data->hctx)) {
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200237 rq->cmd_flags = REQ_MQ_INFLIGHT;
Ming Leicb96a422014-06-01 00:43:37 +0800238 atomic_inc(&data->hctx->nr_active);
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200239 }
240
241 rq->tag = tag;
Ming Leicb96a422014-06-01 00:43:37 +0800242 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200243 return rq;
244 }
245
246 return NULL;
247}
248
Christoph Hellwig4ce01dd2014-05-27 20:59:46 +0200249struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
250 bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100251{
Christoph Hellwigd8525642014-05-27 20:59:50 +0200252 struct blk_mq_ctx *ctx;
253 struct blk_mq_hw_ctx *hctx;
Jens Axboe320ae512013-10-24 09:20:05 +0100254 struct request *rq;
Ming Leicb96a422014-06-01 00:43:37 +0800255 struct blk_mq_alloc_data alloc_data;
Joe Lawrencea492f072014-08-28 08:15:21 -0600256 int ret;
Jens Axboe320ae512013-10-24 09:20:05 +0100257
Keith Buschbfd343a2015-03-11 23:56:39 -0400258 ret = blk_mq_queue_enter(q, gfp);
Joe Lawrencea492f072014-08-28 08:15:21 -0600259 if (ret)
260 return ERR_PTR(ret);
Jens Axboe320ae512013-10-24 09:20:05 +0100261
Christoph Hellwigd8525642014-05-27 20:59:50 +0200262 ctx = blk_mq_get_ctx(q);
263 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Ming Leicb96a422014-06-01 00:43:37 +0800264 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
265 reserved, ctx, hctx);
Christoph Hellwigd8525642014-05-27 20:59:50 +0200266
Ming Leicb96a422014-06-01 00:43:37 +0800267 rq = __blk_mq_alloc_request(&alloc_data, rw);
Christoph Hellwigd8525642014-05-27 20:59:50 +0200268 if (!rq && (gfp & __GFP_WAIT)) {
269 __blk_mq_run_hw_queue(hctx);
270 blk_mq_put_ctx(ctx);
271
272 ctx = blk_mq_get_ctx(q);
273 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Ming Leicb96a422014-06-01 00:43:37 +0800274 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
275 hctx);
276 rq = __blk_mq_alloc_request(&alloc_data, rw);
277 ctx = alloc_data.ctx;
Christoph Hellwigd8525642014-05-27 20:59:50 +0200278 }
279 blk_mq_put_ctx(ctx);
Keith Buschc76541a2014-12-19 17:54:13 -0700280 if (!rq) {
281 blk_mq_queue_exit(q);
Joe Lawrencea492f072014-08-28 08:15:21 -0600282 return ERR_PTR(-EWOULDBLOCK);
Keith Buschc76541a2014-12-19 17:54:13 -0700283 }
Jens Axboe320ae512013-10-24 09:20:05 +0100284 return rq;
285}
Jens Axboe4bb659b2014-05-09 09:36:49 -0600286EXPORT_SYMBOL(blk_mq_alloc_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100287
Jens Axboe320ae512013-10-24 09:20:05 +0100288static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
289 struct blk_mq_ctx *ctx, struct request *rq)
290{
291 const int tag = rq->tag;
292 struct request_queue *q = rq->q;
293
Jens Axboe0d2602c2014-05-13 15:10:52 -0600294 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
295 atomic_dec(&hctx->nr_active);
David Hildenbrand683d0e12014-09-18 11:04:31 +0200296 rq->cmd_flags = 0;
Jens Axboe0d2602c2014-05-13 15:10:52 -0600297
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200298 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
Jens Axboe0d2602c2014-05-13 15:10:52 -0600299 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
Jens Axboe320ae512013-10-24 09:20:05 +0100300 blk_mq_queue_exit(q);
301}
302
Jens Axboe7c7f2f22014-11-17 10:41:57 -0700303void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100304{
305 struct blk_mq_ctx *ctx = rq->mq_ctx;
Jens Axboe7c7f2f22014-11-17 10:41:57 -0700306
307 ctx->rq_completed[rq_is_sync(rq)]++;
308 __blk_mq_free_request(hctx, ctx, rq);
309
310}
311EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
312
313void blk_mq_free_request(struct request *rq)
314{
Jens Axboe320ae512013-10-24 09:20:05 +0100315 struct blk_mq_hw_ctx *hctx;
316 struct request_queue *q = rq->q;
317
Jens Axboe7c7f2f22014-11-17 10:41:57 -0700318 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
319 blk_mq_free_hctx_request(hctx, rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100320}
Jens Axboe1a3b5952014-11-17 10:40:48 -0700321EXPORT_SYMBOL_GPL(blk_mq_free_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100322
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700323inline void __blk_mq_end_request(struct request *rq, int error)
Jens Axboe320ae512013-10-24 09:20:05 +0100324{
Ming Lei0d11e6a2013-12-05 10:50:39 -0700325 blk_account_io_done(rq);
326
Christoph Hellwig91b63632014-04-16 09:44:53 +0200327 if (rq->end_io) {
Jens Axboe320ae512013-10-24 09:20:05 +0100328 rq->end_io(rq, error);
Christoph Hellwig91b63632014-04-16 09:44:53 +0200329 } else {
330 if (unlikely(blk_bidi_rq(rq)))
331 blk_mq_free_request(rq->next_rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100332 blk_mq_free_request(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +0200333 }
Jens Axboe320ae512013-10-24 09:20:05 +0100334}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700335EXPORT_SYMBOL(__blk_mq_end_request);
Christoph Hellwig63151a42014-04-16 09:44:52 +0200336
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700337void blk_mq_end_request(struct request *rq, int error)
Christoph Hellwig63151a42014-04-16 09:44:52 +0200338{
339 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
340 BUG();
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700341 __blk_mq_end_request(rq, error);
Christoph Hellwig63151a42014-04-16 09:44:52 +0200342}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700343EXPORT_SYMBOL(blk_mq_end_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100344
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800345static void __blk_mq_complete_request_remote(void *data)
Jens Axboe320ae512013-10-24 09:20:05 +0100346{
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800347 struct request *rq = data;
Jens Axboe320ae512013-10-24 09:20:05 +0100348
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800349 rq->q->softirq_done_fn(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100350}
351
Jens Axboeed851862014-05-30 21:20:50 -0600352static void blk_mq_ipi_complete_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100353{
354 struct blk_mq_ctx *ctx = rq->mq_ctx;
Christoph Hellwig38535202014-04-25 02:32:53 -0700355 bool shared = false;
Jens Axboe320ae512013-10-24 09:20:05 +0100356 int cpu;
357
Christoph Hellwig38535202014-04-25 02:32:53 -0700358 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800359 rq->q->softirq_done_fn(rq);
360 return;
361 }
Jens Axboe320ae512013-10-24 09:20:05 +0100362
363 cpu = get_cpu();
Christoph Hellwig38535202014-04-25 02:32:53 -0700364 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
365 shared = cpus_share_cache(cpu, ctx->cpu);
366
367 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800368 rq->csd.func = __blk_mq_complete_request_remote;
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800369 rq->csd.info = rq;
370 rq->csd.flags = 0;
Frederic Weisbeckerc46fff22014-02-24 16:40:02 +0100371 smp_call_function_single_async(ctx->cpu, &rq->csd);
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800372 } else {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800373 rq->q->softirq_done_fn(rq);
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800374 }
Jens Axboe320ae512013-10-24 09:20:05 +0100375 put_cpu();
376}
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800377
Jens Axboeed851862014-05-30 21:20:50 -0600378void __blk_mq_complete_request(struct request *rq)
379{
380 struct request_queue *q = rq->q;
381
382 if (!q->softirq_done_fn)
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700383 blk_mq_end_request(rq, rq->errors);
Jens Axboeed851862014-05-30 21:20:50 -0600384 else
385 blk_mq_ipi_complete_request(rq);
386}
387
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800388/**
389 * blk_mq_complete_request - end I/O on a request
390 * @rq: the request being processed
391 *
392 * Description:
393 * Ends all I/O on a request. It does not handle partial completions.
394 * The actual completion happens out-of-order, through a IPI handler.
395 **/
396void blk_mq_complete_request(struct request *rq)
397{
Jens Axboe95f09682014-05-27 17:46:48 -0600398 struct request_queue *q = rq->q;
399
400 if (unlikely(blk_should_fake_timeout(q)))
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800401 return;
Jens Axboeed851862014-05-30 21:20:50 -0600402 if (!blk_mark_rq_complete(rq))
403 __blk_mq_complete_request(rq);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800404}
405EXPORT_SYMBOL(blk_mq_complete_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100406
Keith Busch973c0192015-01-07 18:55:43 -0700407int blk_mq_request_started(struct request *rq)
408{
409 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
410}
411EXPORT_SYMBOL_GPL(blk_mq_request_started);
412
Christoph Hellwige2490072014-09-13 16:40:09 -0700413void blk_mq_start_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100414{
415 struct request_queue *q = rq->q;
416
417 trace_block_rq_issue(q, rq);
418
Christoph Hellwig742ee692014-04-14 10:30:06 +0200419 rq->resid_len = blk_rq_bytes(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +0200420 if (unlikely(blk_bidi_rq(rq)))
421 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
Christoph Hellwig742ee692014-04-14 10:30:06 +0200422
Ming Lei2b8393b2014-06-10 00:16:41 +0800423 blk_add_timer(rq);
Jens Axboe87ee7b12014-04-24 08:51:47 -0600424
425 /*
Jens Axboe538b7532014-09-16 10:37:37 -0600426 * Ensure that ->deadline is visible before set the started
427 * flag and clear the completed flag.
428 */
429 smp_mb__before_atomic();
430
431 /*
Jens Axboe87ee7b12014-04-24 08:51:47 -0600432 * Mark us as started and clear complete. Complete might have been
433 * set if requeue raced with timeout, which then marked it as
434 * complete. So be sure to clear complete again when we start
435 * the request, otherwise we'll ignore the completion event.
436 */
Jens Axboe4b570522014-05-29 11:00:11 -0600437 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
438 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
439 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
440 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800441
442 if (q->dma_drain_size && blk_rq_bytes(rq)) {
443 /*
444 * Make sure space for the drain appears. We know we can do
445 * this because max_hw_segments has been adjusted to be one
446 * fewer than the device can handle.
447 */
448 rq->nr_phys_segments++;
449 }
Jens Axboe320ae512013-10-24 09:20:05 +0100450}
Christoph Hellwige2490072014-09-13 16:40:09 -0700451EXPORT_SYMBOL(blk_mq_start_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100452
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200453static void __blk_mq_requeue_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100454{
455 struct request_queue *q = rq->q;
456
457 trace_block_rq_requeue(q, rq);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800458
Christoph Hellwige2490072014-09-13 16:40:09 -0700459 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
460 if (q->dma_drain_size && blk_rq_bytes(rq))
461 rq->nr_phys_segments--;
462 }
Jens Axboe320ae512013-10-24 09:20:05 +0100463}
464
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200465void blk_mq_requeue_request(struct request *rq)
466{
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200467 __blk_mq_requeue_request(rq);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200468
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200469 BUG_ON(blk_queued_rq(rq));
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600470 blk_mq_add_to_requeue_list(rq, true);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200471}
472EXPORT_SYMBOL(blk_mq_requeue_request);
473
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600474static void blk_mq_requeue_work(struct work_struct *work)
475{
476 struct request_queue *q =
477 container_of(work, struct request_queue, requeue_work);
478 LIST_HEAD(rq_list);
479 struct request *rq, *next;
480 unsigned long flags;
481
482 spin_lock_irqsave(&q->requeue_lock, flags);
483 list_splice_init(&q->requeue_list, &rq_list);
484 spin_unlock_irqrestore(&q->requeue_lock, flags);
485
486 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
487 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
488 continue;
489
490 rq->cmd_flags &= ~REQ_SOFTBARRIER;
491 list_del_init(&rq->queuelist);
492 blk_mq_insert_request(rq, true, false, false);
493 }
494
495 while (!list_empty(&rq_list)) {
496 rq = list_entry(rq_list.next, struct request, queuelist);
497 list_del_init(&rq->queuelist);
498 blk_mq_insert_request(rq, false, false, false);
499 }
500
Jens Axboe8b957412014-09-19 13:10:29 -0600501 /*
502 * Use the start variant of queue running here, so that running
503 * the requeue work will kick stopped queues.
504 */
505 blk_mq_start_hw_queues(q);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600506}
507
508void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
509{
510 struct request_queue *q = rq->q;
511 unsigned long flags;
512
513 /*
514 * We abuse this flag that is otherwise used by the I/O scheduler to
515 * request head insertation from the workqueue.
516 */
517 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
518
519 spin_lock_irqsave(&q->requeue_lock, flags);
520 if (at_head) {
521 rq->cmd_flags |= REQ_SOFTBARRIER;
522 list_add(&rq->queuelist, &q->requeue_list);
523 } else {
524 list_add_tail(&rq->queuelist, &q->requeue_list);
525 }
526 spin_unlock_irqrestore(&q->requeue_lock, flags);
527}
528EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
529
Keith Buschc68ed592015-01-07 18:55:44 -0700530void blk_mq_cancel_requeue_work(struct request_queue *q)
531{
532 cancel_work_sync(&q->requeue_work);
533}
534EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
535
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600536void blk_mq_kick_requeue_list(struct request_queue *q)
537{
538 kblockd_schedule_work(&q->requeue_work);
539}
540EXPORT_SYMBOL(blk_mq_kick_requeue_list);
541
Jens Axboe1885b242015-01-07 18:55:45 -0700542void blk_mq_abort_requeue_list(struct request_queue *q)
543{
544 unsigned long flags;
545 LIST_HEAD(rq_list);
546
547 spin_lock_irqsave(&q->requeue_lock, flags);
548 list_splice_init(&q->requeue_list, &rq_list);
549 spin_unlock_irqrestore(&q->requeue_lock, flags);
550
551 while (!list_empty(&rq_list)) {
552 struct request *rq;
553
554 rq = list_first_entry(&rq_list, struct request, queuelist);
555 list_del_init(&rq->queuelist);
556 rq->errors = -EIO;
557 blk_mq_end_request(rq, rq->errors);
558 }
559}
560EXPORT_SYMBOL(blk_mq_abort_requeue_list);
561
Ming Lei7c94e1c2014-09-25 23:23:43 +0800562static inline bool is_flush_request(struct request *rq,
563 struct blk_flush_queue *fq, unsigned int tag)
Christoph Hellwig24d2f902014-04-15 14:14:00 -0600564{
Jens Axboe0e62f512014-06-04 10:23:49 -0600565 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
Ming Lei7c94e1c2014-09-25 23:23:43 +0800566 fq->flush_rq->tag == tag);
Jens Axboe0e62f512014-06-04 10:23:49 -0600567}
Shaohua Li22302372014-05-30 08:06:42 -0600568
Jens Axboe0e62f512014-06-04 10:23:49 -0600569struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
570{
571 struct request *rq = tags->rqs[tag];
Ming Leie97c2932014-09-25 23:23:46 +0800572 /* mq_ctx of flush rq is always cloned from the corresponding req */
573 struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
Shaohua Li22302372014-05-30 08:06:42 -0600574
Ming Lei7c94e1c2014-09-25 23:23:43 +0800575 if (!is_flush_request(rq, fq, tag))
Jens Axboe0e62f512014-06-04 10:23:49 -0600576 return rq;
577
Ming Lei7c94e1c2014-09-25 23:23:43 +0800578 return fq->flush_rq;
Christoph Hellwig24d2f902014-04-15 14:14:00 -0600579}
580EXPORT_SYMBOL(blk_mq_tag_to_rq);
581
Jens Axboe320ae512013-10-24 09:20:05 +0100582struct blk_mq_timeout_data {
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700583 unsigned long next;
584 unsigned int next_set;
Jens Axboe320ae512013-10-24 09:20:05 +0100585};
586
Christoph Hellwig90415832014-09-22 10:21:48 -0600587void blk_mq_rq_timed_out(struct request *req, bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100588{
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700589 struct blk_mq_ops *ops = req->q->mq_ops;
590 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
Jens Axboe87ee7b12014-04-24 08:51:47 -0600591
592 /*
593 * We know that complete is set at this point. If STARTED isn't set
594 * anymore, then the request isn't active and the "timeout" should
595 * just be ignored. This can happen due to the bitflag ordering.
596 * Timeout first checks if STARTED is set, and if it is, assumes
597 * the request is active. But if we race with completion, then
598 * we both flags will get cleared. So check here again, and ignore
599 * a timeout event with a request that isn't active.
600 */
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700601 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
602 return;
Jens Axboe87ee7b12014-04-24 08:51:47 -0600603
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700604 if (ops->timeout)
Christoph Hellwig0152fb62014-09-13 16:40:13 -0700605 ret = ops->timeout(req, reserved);
Jens Axboe87ee7b12014-04-24 08:51:47 -0600606
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700607 switch (ret) {
608 case BLK_EH_HANDLED:
609 __blk_mq_complete_request(req);
610 break;
611 case BLK_EH_RESET_TIMER:
612 blk_add_timer(req);
613 blk_clear_rq_complete(req);
614 break;
615 case BLK_EH_NOT_HANDLED:
616 break;
617 default:
618 printk(KERN_ERR "block: bad eh return: %d\n", ret);
619 break;
620 }
Jens Axboe87ee7b12014-04-24 08:51:47 -0600621}
Keith Busch5b3f25f2015-01-07 18:55:46 -0700622
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700623static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
624 struct request *rq, void *priv, bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100625{
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700626 struct blk_mq_timeout_data *data = priv;
627
Keith Buscheb130db2015-01-08 08:59:53 -0700628 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
629 /*
630 * If a request wasn't started before the queue was
631 * marked dying, kill it here or it'll go unnoticed.
632 */
633 if (unlikely(blk_queue_dying(rq->q))) {
634 rq->errors = -EIO;
635 blk_mq_complete_request(rq);
636 }
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700637 return;
Keith Buscheb130db2015-01-08 08:59:53 -0700638 }
Keith Busch5b3f25f2015-01-07 18:55:46 -0700639 if (rq->cmd_flags & REQ_NO_TIMEOUT)
640 return;
Jens Axboe320ae512013-10-24 09:20:05 +0100641
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700642 if (time_after_eq(jiffies, rq->deadline)) {
643 if (!blk_mark_rq_complete(rq))
Christoph Hellwig0152fb62014-09-13 16:40:13 -0700644 blk_mq_rq_timed_out(rq, reserved);
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700645 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
646 data->next = rq->deadline;
647 data->next_set = 1;
648 }
Jens Axboe320ae512013-10-24 09:20:05 +0100649}
650
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700651static void blk_mq_rq_timer(unsigned long priv)
Jens Axboe320ae512013-10-24 09:20:05 +0100652{
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700653 struct request_queue *q = (struct request_queue *)priv;
654 struct blk_mq_timeout_data data = {
655 .next = 0,
656 .next_set = 0,
657 };
Jens Axboe320ae512013-10-24 09:20:05 +0100658 struct blk_mq_hw_ctx *hctx;
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700659 int i;
Jens Axboe320ae512013-10-24 09:20:05 +0100660
Jens Axboe484b4062014-05-21 14:01:15 -0600661 queue_for_each_hw_ctx(q, hctx, i) {
662 /*
663 * If not software queues are currently mapped to this
664 * hardware queue, there's nothing to check
665 */
Ming Lei19c66e52014-12-03 19:38:04 +0800666 if (!blk_mq_hw_queue_mapped(hctx))
Jens Axboe484b4062014-05-21 14:01:15 -0600667 continue;
668
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700669 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
Jens Axboe484b4062014-05-21 14:01:15 -0600670 }
Jens Axboe320ae512013-10-24 09:20:05 +0100671
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700672 if (data.next_set) {
673 data.next = blk_rq_timeout(round_jiffies_up(data.next));
674 mod_timer(&q->timeout, data.next);
Jens Axboe0d2602c2014-05-13 15:10:52 -0600675 } else {
Ming Leif054b562015-04-21 10:00:19 +0800676 queue_for_each_hw_ctx(q, hctx, i) {
677 /* the hctx may be unmapped, so check it here */
678 if (blk_mq_hw_queue_mapped(hctx))
679 blk_mq_tag_idle(hctx);
680 }
Jens Axboe0d2602c2014-05-13 15:10:52 -0600681 }
Jens Axboe320ae512013-10-24 09:20:05 +0100682}
683
684/*
685 * Reverse check our software queue for entries that we could potentially
686 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
687 * too much time checking for merges.
688 */
689static bool blk_mq_attempt_merge(struct request_queue *q,
690 struct blk_mq_ctx *ctx, struct bio *bio)
691{
692 struct request *rq;
693 int checked = 8;
694
695 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
696 int el_ret;
697
698 if (!checked--)
699 break;
700
701 if (!blk_rq_merge_ok(rq, bio))
702 continue;
703
704 el_ret = blk_try_merge(rq, bio);
705 if (el_ret == ELEVATOR_BACK_MERGE) {
706 if (bio_attempt_back_merge(q, rq, bio)) {
707 ctx->rq_merged++;
708 return true;
709 }
710 break;
711 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
712 if (bio_attempt_front_merge(q, rq, bio)) {
713 ctx->rq_merged++;
714 return true;
715 }
716 break;
717 }
718 }
719
720 return false;
721}
722
Jens Axboe320ae512013-10-24 09:20:05 +0100723/*
Jens Axboe1429d7c2014-05-19 09:23:55 -0600724 * Process software queues that have been marked busy, splicing them
725 * to the for-dispatch
726 */
727static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
728{
729 struct blk_mq_ctx *ctx;
730 int i;
731
Jens Axboe569fd0c2015-04-17 08:28:50 -0600732 for (i = 0; i < hctx->ctx_map.size; i++) {
Jens Axboe1429d7c2014-05-19 09:23:55 -0600733 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
734 unsigned int off, bit;
735
736 if (!bm->word)
737 continue;
738
739 bit = 0;
740 off = i * hctx->ctx_map.bits_per_word;
741 do {
742 bit = find_next_bit(&bm->word, bm->depth, bit);
743 if (bit >= bm->depth)
744 break;
745
746 ctx = hctx->ctxs[bit + off];
747 clear_bit(bit, &bm->word);
748 spin_lock(&ctx->lock);
749 list_splice_tail_init(&ctx->rq_list, list);
750 spin_unlock(&ctx->lock);
751
752 bit++;
753 } while (1);
754 }
755}
756
757/*
Jens Axboe320ae512013-10-24 09:20:05 +0100758 * Run this hardware queue, pulling any software queues mapped to it in.
759 * Note that this function currently has various problems around ordering
760 * of IO. In particular, we'd like FIFO behaviour on handling existing
761 * items on the hctx->dispatch list. Ignore that for now.
762 */
763static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
764{
765 struct request_queue *q = hctx->queue;
Jens Axboe320ae512013-10-24 09:20:05 +0100766 struct request *rq;
767 LIST_HEAD(rq_list);
Jens Axboe74c45052014-10-29 11:14:52 -0600768 LIST_HEAD(driver_list);
769 struct list_head *dptr;
Jens Axboe1429d7c2014-05-19 09:23:55 -0600770 int queued;
Jens Axboe320ae512013-10-24 09:20:05 +0100771
Jens Axboefd1270d2014-04-16 09:23:48 -0600772 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
Jens Axboee4043dc2014-04-09 10:18:23 -0600773
Jens Axboe5d12f902014-03-19 15:25:02 -0600774 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
Jens Axboe320ae512013-10-24 09:20:05 +0100775 return;
776
777 hctx->run++;
778
779 /*
780 * Touch any software queue that has pending entries.
781 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600782 flush_busy_ctxs(hctx, &rq_list);
Jens Axboe320ae512013-10-24 09:20:05 +0100783
784 /*
785 * If we have previous entries on our dispatch list, grab them
786 * and stuff them at the front for more fair dispatch.
787 */
788 if (!list_empty_careful(&hctx->dispatch)) {
789 spin_lock(&hctx->lock);
790 if (!list_empty(&hctx->dispatch))
791 list_splice_init(&hctx->dispatch, &rq_list);
792 spin_unlock(&hctx->lock);
793 }
794
795 /*
Jens Axboe74c45052014-10-29 11:14:52 -0600796 * Start off with dptr being NULL, so we start the first request
797 * immediately, even if we have more pending.
798 */
799 dptr = NULL;
800
801 /*
Jens Axboe320ae512013-10-24 09:20:05 +0100802 * Now process all the entries, sending them to the driver.
803 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600804 queued = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100805 while (!list_empty(&rq_list)) {
Jens Axboe74c45052014-10-29 11:14:52 -0600806 struct blk_mq_queue_data bd;
Jens Axboe320ae512013-10-24 09:20:05 +0100807 int ret;
808
809 rq = list_first_entry(&rq_list, struct request, queuelist);
810 list_del_init(&rq->queuelist);
Jens Axboe320ae512013-10-24 09:20:05 +0100811
Jens Axboe74c45052014-10-29 11:14:52 -0600812 bd.rq = rq;
813 bd.list = dptr;
814 bd.last = list_empty(&rq_list);
815
816 ret = q->mq_ops->queue_rq(hctx, &bd);
Jens Axboe320ae512013-10-24 09:20:05 +0100817 switch (ret) {
818 case BLK_MQ_RQ_QUEUE_OK:
819 queued++;
820 continue;
821 case BLK_MQ_RQ_QUEUE_BUSY:
Jens Axboe320ae512013-10-24 09:20:05 +0100822 list_add(&rq->queuelist, &rq_list);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200823 __blk_mq_requeue_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100824 break;
825 default:
826 pr_err("blk-mq: bad return on queue: %d\n", ret);
Jens Axboe320ae512013-10-24 09:20:05 +0100827 case BLK_MQ_RQ_QUEUE_ERROR:
Christoph Hellwig1e93b8c2014-02-11 08:27:13 -0800828 rq->errors = -EIO;
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700829 blk_mq_end_request(rq, rq->errors);
Jens Axboe320ae512013-10-24 09:20:05 +0100830 break;
831 }
832
833 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
834 break;
Jens Axboe74c45052014-10-29 11:14:52 -0600835
836 /*
837 * We've done the first request. If we have more than 1
838 * left in the list, set dptr to defer issue.
839 */
840 if (!dptr && rq_list.next != rq_list.prev)
841 dptr = &driver_list;
Jens Axboe320ae512013-10-24 09:20:05 +0100842 }
843
844 if (!queued)
845 hctx->dispatched[0]++;
846 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
847 hctx->dispatched[ilog2(queued) + 1]++;
848
849 /*
850 * Any items that need requeuing? Stuff them into hctx->dispatch,
851 * that is where we will continue on next queue run.
852 */
853 if (!list_empty(&rq_list)) {
854 spin_lock(&hctx->lock);
855 list_splice(&rq_list, &hctx->dispatch);
856 spin_unlock(&hctx->lock);
Shaohua Li9ba52e52015-05-04 14:32:48 -0600857 /*
858 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
859 * it's possible the queue is stopped and restarted again
860 * before this. Queue restart will dispatch requests. And since
861 * requests in rq_list aren't added into hctx->dispatch yet,
862 * the requests in rq_list might get lost.
863 *
864 * blk_mq_run_hw_queue() already checks the STOPPED bit
865 **/
866 blk_mq_run_hw_queue(hctx, true);
Jens Axboe320ae512013-10-24 09:20:05 +0100867 }
868}
869
Jens Axboe506e9312014-05-07 10:26:44 -0600870/*
871 * It'd be great if the workqueue API had a way to pass
872 * in a mask and had some smarts for more clever placement.
873 * For now we just round-robin here, switching for every
874 * BLK_MQ_CPU_WORK_BATCH queued items.
875 */
876static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
877{
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100878 if (hctx->queue->nr_hw_queues == 1)
879 return WORK_CPU_UNBOUND;
Jens Axboe506e9312014-05-07 10:26:44 -0600880
881 if (--hctx->next_cpu_batch <= 0) {
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100882 int cpu = hctx->next_cpu, next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600883
884 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
885 if (next_cpu >= nr_cpu_ids)
886 next_cpu = cpumask_first(hctx->cpumask);
887
888 hctx->next_cpu = next_cpu;
889 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100890
891 return cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600892 }
893
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100894 return hctx->next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600895}
896
Jens Axboe320ae512013-10-24 09:20:05 +0100897void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
898{
Ming Lei19c66e52014-12-03 19:38:04 +0800899 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
900 !blk_mq_hw_queue_mapped(hctx)))
Jens Axboe320ae512013-10-24 09:20:05 +0100901 return;
902
Paolo Bonzini398205b2014-11-07 23:03:59 +0100903 if (!async) {
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100904 int cpu = get_cpu();
905 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
Paolo Bonzini398205b2014-11-07 23:03:59 +0100906 __blk_mq_run_hw_queue(hctx);
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100907 put_cpu();
Paolo Bonzini398205b2014-11-07 23:03:59 +0100908 return;
909 }
Jens Axboee4043dc2014-04-09 10:18:23 -0600910
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100911 put_cpu();
Jens Axboee4043dc2014-04-09 10:18:23 -0600912 }
Paolo Bonzini398205b2014-11-07 23:03:59 +0100913
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100914 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
915 &hctx->run_work, 0);
Jens Axboe320ae512013-10-24 09:20:05 +0100916}
917
Mike Snitzerb94ec292015-03-11 23:56:38 -0400918void blk_mq_run_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +0100919{
920 struct blk_mq_hw_ctx *hctx;
921 int i;
922
923 queue_for_each_hw_ctx(q, hctx, i) {
924 if ((!blk_mq_hctx_has_pending(hctx) &&
925 list_empty_careful(&hctx->dispatch)) ||
Jens Axboe5d12f902014-03-19 15:25:02 -0600926 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
Jens Axboe320ae512013-10-24 09:20:05 +0100927 continue;
928
Mike Snitzerb94ec292015-03-11 23:56:38 -0400929 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +0100930 }
931}
Mike Snitzerb94ec292015-03-11 23:56:38 -0400932EXPORT_SYMBOL(blk_mq_run_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +0100933
934void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
935{
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600936 cancel_delayed_work(&hctx->run_work);
937 cancel_delayed_work(&hctx->delay_work);
Jens Axboe320ae512013-10-24 09:20:05 +0100938 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
939}
940EXPORT_SYMBOL(blk_mq_stop_hw_queue);
941
Christoph Hellwig280d45f2013-10-25 14:45:58 +0100942void blk_mq_stop_hw_queues(struct request_queue *q)
943{
944 struct blk_mq_hw_ctx *hctx;
945 int i;
946
947 queue_for_each_hw_ctx(q, hctx, i)
948 blk_mq_stop_hw_queue(hctx);
949}
950EXPORT_SYMBOL(blk_mq_stop_hw_queues);
951
Jens Axboe320ae512013-10-24 09:20:05 +0100952void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
953{
954 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboee4043dc2014-04-09 10:18:23 -0600955
Jens Axboe0ffbce82014-06-25 08:22:34 -0600956 blk_mq_run_hw_queue(hctx, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100957}
958EXPORT_SYMBOL(blk_mq_start_hw_queue);
959
Christoph Hellwig2f268552014-04-16 09:44:56 +0200960void blk_mq_start_hw_queues(struct request_queue *q)
961{
962 struct blk_mq_hw_ctx *hctx;
963 int i;
964
965 queue_for_each_hw_ctx(q, hctx, i)
966 blk_mq_start_hw_queue(hctx);
967}
968EXPORT_SYMBOL(blk_mq_start_hw_queues);
969
Christoph Hellwig1b4a3252014-04-16 09:44:54 +0200970void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +0100971{
972 struct blk_mq_hw_ctx *hctx;
973 int i;
974
975 queue_for_each_hw_ctx(q, hctx, i) {
976 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
977 continue;
978
979 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Christoph Hellwig1b4a3252014-04-16 09:44:54 +0200980 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +0100981 }
982}
983EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
984
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600985static void blk_mq_run_work_fn(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +0100986{
987 struct blk_mq_hw_ctx *hctx;
988
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600989 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
Jens Axboee4043dc2014-04-09 10:18:23 -0600990
Jens Axboe320ae512013-10-24 09:20:05 +0100991 __blk_mq_run_hw_queue(hctx);
992}
993
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600994static void blk_mq_delay_work_fn(struct work_struct *work)
995{
996 struct blk_mq_hw_ctx *hctx;
997
998 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
999
1000 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1001 __blk_mq_run_hw_queue(hctx);
1002}
1003
1004void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1005{
Ming Lei19c66e52014-12-03 19:38:04 +08001006 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1007 return;
Christoph Hellwig70f4db62014-04-16 10:48:08 -06001008
Christoph Hellwigb657d7e2014-11-24 09:27:23 +01001009 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1010 &hctx->delay_work, msecs_to_jiffies(msecs));
Christoph Hellwig70f4db62014-04-16 10:48:08 -06001011}
1012EXPORT_SYMBOL(blk_mq_delay_queue);
1013
Jens Axboe320ae512013-10-24 09:20:05 +01001014static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig72a0a362014-02-07 10:22:36 -08001015 struct request *rq, bool at_head)
Jens Axboe320ae512013-10-24 09:20:05 +01001016{
1017 struct blk_mq_ctx *ctx = rq->mq_ctx;
1018
Jens Axboe01b983c2013-11-19 18:59:10 -07001019 trace_block_rq_insert(hctx->queue, rq);
1020
Christoph Hellwig72a0a362014-02-07 10:22:36 -08001021 if (at_head)
1022 list_add(&rq->queuelist, &ctx->rq_list);
1023 else
1024 list_add_tail(&rq->queuelist, &ctx->rq_list);
Jens Axboe4bb659b2014-05-09 09:36:49 -06001025
Jens Axboe320ae512013-10-24 09:20:05 +01001026 blk_mq_hctx_mark_pending(hctx, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001027}
1028
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001029void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1030 bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01001031{
1032 struct request_queue *q = rq->q;
1033 struct blk_mq_hw_ctx *hctx;
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001034 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001035
1036 current_ctx = blk_mq_get_ctx(q);
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001037 if (!cpu_online(ctx->cpu))
1038 rq->mq_ctx = ctx = current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001039
Jens Axboe320ae512013-10-24 09:20:05 +01001040 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1041
Christoph Hellwiga57a1782014-09-16 14:44:07 -07001042 spin_lock(&ctx->lock);
1043 __blk_mq_insert_request(hctx, rq, at_head);
1044 spin_unlock(&ctx->lock);
Jens Axboe320ae512013-10-24 09:20:05 +01001045
Jens Axboe320ae512013-10-24 09:20:05 +01001046 if (run_queue)
1047 blk_mq_run_hw_queue(hctx, async);
Jens Axboee4043dc2014-04-09 10:18:23 -06001048
1049 blk_mq_put_ctx(current_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001050}
1051
1052static void blk_mq_insert_requests(struct request_queue *q,
1053 struct blk_mq_ctx *ctx,
1054 struct list_head *list,
1055 int depth,
1056 bool from_schedule)
1057
1058{
1059 struct blk_mq_hw_ctx *hctx;
1060 struct blk_mq_ctx *current_ctx;
1061
1062 trace_block_unplug(q, depth, !from_schedule);
1063
1064 current_ctx = blk_mq_get_ctx(q);
1065
1066 if (!cpu_online(ctx->cpu))
1067 ctx = current_ctx;
1068 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1069
1070 /*
1071 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1072 * offline now
1073 */
1074 spin_lock(&ctx->lock);
1075 while (!list_empty(list)) {
1076 struct request *rq;
1077
1078 rq = list_first_entry(list, struct request, queuelist);
1079 list_del_init(&rq->queuelist);
1080 rq->mq_ctx = ctx;
Christoph Hellwig72a0a362014-02-07 10:22:36 -08001081 __blk_mq_insert_request(hctx, rq, false);
Jens Axboe320ae512013-10-24 09:20:05 +01001082 }
1083 spin_unlock(&ctx->lock);
1084
Jens Axboe320ae512013-10-24 09:20:05 +01001085 blk_mq_run_hw_queue(hctx, from_schedule);
Jens Axboee4043dc2014-04-09 10:18:23 -06001086 blk_mq_put_ctx(current_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001087}
1088
1089static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1090{
1091 struct request *rqa = container_of(a, struct request, queuelist);
1092 struct request *rqb = container_of(b, struct request, queuelist);
1093
1094 return !(rqa->mq_ctx < rqb->mq_ctx ||
1095 (rqa->mq_ctx == rqb->mq_ctx &&
1096 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1097}
1098
1099void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1100{
1101 struct blk_mq_ctx *this_ctx;
1102 struct request_queue *this_q;
1103 struct request *rq;
1104 LIST_HEAD(list);
1105 LIST_HEAD(ctx_list);
1106 unsigned int depth;
1107
1108 list_splice_init(&plug->mq_list, &list);
1109
1110 list_sort(NULL, &list, plug_ctx_cmp);
1111
1112 this_q = NULL;
1113 this_ctx = NULL;
1114 depth = 0;
1115
1116 while (!list_empty(&list)) {
1117 rq = list_entry_rq(list.next);
1118 list_del_init(&rq->queuelist);
1119 BUG_ON(!rq->q);
1120 if (rq->mq_ctx != this_ctx) {
1121 if (this_ctx) {
1122 blk_mq_insert_requests(this_q, this_ctx,
1123 &ctx_list, depth,
1124 from_schedule);
1125 }
1126
1127 this_ctx = rq->mq_ctx;
1128 this_q = rq->q;
1129 depth = 0;
1130 }
1131
1132 depth++;
1133 list_add_tail(&rq->queuelist, &ctx_list);
1134 }
1135
1136 /*
1137 * If 'this_ctx' is set, we know we have entries to complete
1138 * on 'ctx_list'. Do those.
1139 */
1140 if (this_ctx) {
1141 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1142 from_schedule);
1143 }
1144}
1145
1146static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1147{
1148 init_request_from_bio(rq, bio);
Jens Axboe4b570522014-05-29 11:00:11 -06001149
Jens Axboe3ee32372014-06-09 09:36:53 -06001150 if (blk_do_io_stat(rq))
Jens Axboe4b570522014-05-29 11:00:11 -06001151 blk_account_io_start(rq, 1);
Jens Axboe320ae512013-10-24 09:20:05 +01001152}
1153
Jens Axboe274a5842014-08-15 12:44:08 -06001154static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1155{
1156 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1157 !blk_queue_nomerges(hctx->queue);
1158}
1159
Jens Axboe07068d52014-05-22 10:40:51 -06001160static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1161 struct blk_mq_ctx *ctx,
1162 struct request *rq, struct bio *bio)
1163{
Jens Axboe274a5842014-08-15 12:44:08 -06001164 if (!hctx_allow_merges(hctx)) {
Jens Axboe07068d52014-05-22 10:40:51 -06001165 blk_mq_bio_to_request(rq, bio);
1166 spin_lock(&ctx->lock);
1167insert_rq:
1168 __blk_mq_insert_request(hctx, rq, false);
1169 spin_unlock(&ctx->lock);
1170 return false;
1171 } else {
Jens Axboe274a5842014-08-15 12:44:08 -06001172 struct request_queue *q = hctx->queue;
1173
Jens Axboe07068d52014-05-22 10:40:51 -06001174 spin_lock(&ctx->lock);
1175 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1176 blk_mq_bio_to_request(rq, bio);
1177 goto insert_rq;
1178 }
1179
1180 spin_unlock(&ctx->lock);
1181 __blk_mq_free_request(hctx, ctx, rq);
1182 return true;
1183 }
1184}
1185
1186struct blk_map_ctx {
1187 struct blk_mq_hw_ctx *hctx;
1188 struct blk_mq_ctx *ctx;
1189};
1190
1191static struct request *blk_mq_map_request(struct request_queue *q,
1192 struct bio *bio,
1193 struct blk_map_ctx *data)
Jens Axboe320ae512013-10-24 09:20:05 +01001194{
1195 struct blk_mq_hw_ctx *hctx;
1196 struct blk_mq_ctx *ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001197 struct request *rq;
Jens Axboe07068d52014-05-22 10:40:51 -06001198 int rw = bio_data_dir(bio);
Ming Leicb96a422014-06-01 00:43:37 +08001199 struct blk_mq_alloc_data alloc_data;
Jens Axboe320ae512013-10-24 09:20:05 +01001200
Keith Buschbfd343a2015-03-11 23:56:39 -04001201 if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
Nicholas Bellinger14ec77f2014-02-07 13:45:39 -07001202 bio_endio(bio, -EIO);
Jens Axboe07068d52014-05-22 10:40:51 -06001203 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001204 }
1205
1206 ctx = blk_mq_get_ctx(q);
1207 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1208
Jens Axboe07068d52014-05-22 10:40:51 -06001209 if (rw_is_sync(bio->bi_rw))
Shaohua Li27fbf4e2014-02-19 20:20:21 +08001210 rw |= REQ_SYNC;
Jens Axboe07068d52014-05-22 10:40:51 -06001211
Jens Axboe320ae512013-10-24 09:20:05 +01001212 trace_block_getrq(q, bio, rw);
Ming Leicb96a422014-06-01 00:43:37 +08001213 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1214 hctx);
1215 rq = __blk_mq_alloc_request(&alloc_data, rw);
Christoph Hellwig5dee8572014-05-27 20:59:47 +02001216 if (unlikely(!rq)) {
Christoph Hellwig793597a2014-05-27 20:59:49 +02001217 __blk_mq_run_hw_queue(hctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001218 blk_mq_put_ctx(ctx);
1219 trace_block_sleeprq(q, bio, rw);
Christoph Hellwig793597a2014-05-27 20:59:49 +02001220
1221 ctx = blk_mq_get_ctx(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001222 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Ming Leicb96a422014-06-01 00:43:37 +08001223 blk_mq_set_alloc_data(&alloc_data, q,
1224 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1225 rq = __blk_mq_alloc_request(&alloc_data, rw);
1226 ctx = alloc_data.ctx;
1227 hctx = alloc_data.hctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001228 }
1229
1230 hctx->queued++;
Jens Axboe07068d52014-05-22 10:40:51 -06001231 data->hctx = hctx;
1232 data->ctx = ctx;
1233 return rq;
1234}
1235
Shaohua Lif984df12015-05-08 10:51:32 -07001236static int blk_mq_direct_issue_request(struct request *rq)
1237{
1238 int ret;
1239 struct request_queue *q = rq->q;
1240 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1241 rq->mq_ctx->cpu);
1242 struct blk_mq_queue_data bd = {
1243 .rq = rq,
1244 .list = NULL,
1245 .last = 1
1246 };
1247
1248 /*
1249 * For OK queue, we are done. For error, kill it. Any other
1250 * error (busy), just add it to our list as we previously
1251 * would have done
1252 */
1253 ret = q->mq_ops->queue_rq(hctx, &bd);
1254 if (ret == BLK_MQ_RQ_QUEUE_OK)
1255 return 0;
1256 else {
1257 __blk_mq_requeue_request(rq);
1258
1259 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1260 rq->errors = -EIO;
1261 blk_mq_end_request(rq, rq->errors);
1262 return 0;
1263 }
1264 return -1;
1265 }
1266}
1267
Jens Axboe07068d52014-05-22 10:40:51 -06001268/*
1269 * Multiple hardware queue variant. This will not use per-process plugs,
1270 * but will attempt to bypass the hctx queueing if we can go straight to
1271 * hardware for SYNC IO.
1272 */
1273static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1274{
1275 const int is_sync = rw_is_sync(bio->bi_rw);
1276 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1277 struct blk_map_ctx data;
1278 struct request *rq;
Shaohua Lif984df12015-05-08 10:51:32 -07001279 unsigned int request_count = 0;
1280 struct blk_plug *plug;
Shaohua Li5b3f3412015-05-08 10:51:33 -07001281 struct request *same_queue_rq = NULL;
Jens Axboe07068d52014-05-22 10:40:51 -06001282
1283 blk_queue_bounce(q, &bio);
1284
1285 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1286 bio_endio(bio, -EIO);
1287 return;
1288 }
1289
Shaohua Lif984df12015-05-08 10:51:32 -07001290 if (!is_flush_fua && !blk_queue_nomerges(q) &&
Shaohua Li5b3f3412015-05-08 10:51:33 -07001291 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
Shaohua Lif984df12015-05-08 10:51:32 -07001292 return;
1293
Jens Axboe07068d52014-05-22 10:40:51 -06001294 rq = blk_mq_map_request(q, bio, &data);
1295 if (unlikely(!rq))
1296 return;
1297
1298 if (unlikely(is_flush_fua)) {
1299 blk_mq_bio_to_request(rq, bio);
1300 blk_insert_flush(rq);
1301 goto run_queue;
1302 }
1303
Shaohua Lif984df12015-05-08 10:51:32 -07001304 plug = current->plug;
Jens Axboee167dfb2014-10-29 11:18:26 -06001305 /*
1306 * If the driver supports defer issued based on 'last', then
1307 * queue it up like normal since we can potentially save some
1308 * CPU this way.
1309 */
Shaohua Lif984df12015-05-08 10:51:32 -07001310 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1311 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1312 struct request *old_rq = NULL;
Jens Axboe07068d52014-05-22 10:40:51 -06001313
1314 blk_mq_bio_to_request(rq, bio);
Jens Axboe07068d52014-05-22 10:40:51 -06001315
1316 /*
Shaohua Lif984df12015-05-08 10:51:32 -07001317 * we do limited pluging. If bio can be merged, do merge.
1318 * Otherwise the existing request in the plug list will be
1319 * issued. So the plug list will have one request at most
Jens Axboe07068d52014-05-22 10:40:51 -06001320 */
Shaohua Lif984df12015-05-08 10:51:32 -07001321 if (plug) {
Shaohua Li5b3f3412015-05-08 10:51:33 -07001322 /*
1323 * The plug list might get flushed before this. If that
1324 * happens, same_queue_rq is invalid and plug list is empty
1325 **/
1326 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1327 old_rq = same_queue_rq;
Shaohua Lif984df12015-05-08 10:51:32 -07001328 list_del_init(&old_rq->queuelist);
Jens Axboe07068d52014-05-22 10:40:51 -06001329 }
Shaohua Lif984df12015-05-08 10:51:32 -07001330 list_add_tail(&rq->queuelist, &plug->mq_list);
1331 } else /* is_sync */
1332 old_rq = rq;
1333 blk_mq_put_ctx(data.ctx);
1334 if (!old_rq)
Shaohua Li239ad212015-05-08 10:51:31 -07001335 return;
Shaohua Lif984df12015-05-08 10:51:32 -07001336 if (!blk_mq_direct_issue_request(old_rq))
1337 return;
1338 blk_mq_insert_request(old_rq, false, true, true);
1339 return;
Jens Axboe07068d52014-05-22 10:40:51 -06001340 }
1341
1342 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1343 /*
1344 * For a SYNC request, send it to the hardware immediately. For
1345 * an ASYNC request, just ensure that we run it later on. The
1346 * latter allows for merging opportunities and more efficient
1347 * dispatching.
1348 */
1349run_queue:
1350 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1351 }
Jens Axboe07068d52014-05-22 10:40:51 -06001352 blk_mq_put_ctx(data.ctx);
1353}
1354
1355/*
1356 * Single hardware queue variant. This will attempt to use any per-process
1357 * plug for merging and IO deferral.
1358 */
1359static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1360{
1361 const int is_sync = rw_is_sync(bio->bi_rw);
1362 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
Jeff Moyere6c44382015-05-08 10:51:30 -07001363 struct blk_plug *plug;
1364 unsigned int request_count = 0;
Jens Axboe07068d52014-05-22 10:40:51 -06001365 struct blk_map_ctx data;
1366 struct request *rq;
1367
Jens Axboe07068d52014-05-22 10:40:51 -06001368 blk_queue_bounce(q, &bio);
1369
1370 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1371 bio_endio(bio, -EIO);
1372 return;
1373 }
1374
Jeff Moyere6c44382015-05-08 10:51:30 -07001375 if (!is_flush_fua && !blk_queue_nomerges(q) &&
Shaohua Li5b3f3412015-05-08 10:51:33 -07001376 blk_attempt_plug_merge(q, bio, &request_count, NULL))
Jens Axboe07068d52014-05-22 10:40:51 -06001377 return;
1378
1379 rq = blk_mq_map_request(q, bio, &data);
Jens Axboeff87bce2014-06-03 11:59:49 -06001380 if (unlikely(!rq))
1381 return;
Jens Axboe320ae512013-10-24 09:20:05 +01001382
1383 if (unlikely(is_flush_fua)) {
1384 blk_mq_bio_to_request(rq, bio);
Jens Axboe320ae512013-10-24 09:20:05 +01001385 blk_insert_flush(rq);
1386 goto run_queue;
1387 }
1388
1389 /*
1390 * A task plug currently exists. Since this is completely lockless,
1391 * utilize that to temporarily store requests until the task is
1392 * either done or scheduled away.
1393 */
Jeff Moyere6c44382015-05-08 10:51:30 -07001394 plug = current->plug;
1395 if (plug) {
1396 blk_mq_bio_to_request(rq, bio);
1397 if (list_empty(&plug->mq_list))
1398 trace_block_plug(q);
1399 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1400 blk_flush_plug_list(plug, false);
1401 trace_block_plug(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001402 }
Jeff Moyere6c44382015-05-08 10:51:30 -07001403 list_add_tail(&rq->queuelist, &plug->mq_list);
1404 blk_mq_put_ctx(data.ctx);
1405 return;
Jens Axboe320ae512013-10-24 09:20:05 +01001406 }
1407
Jens Axboe07068d52014-05-22 10:40:51 -06001408 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1409 /*
1410 * For a SYNC request, send it to the hardware immediately. For
1411 * an ASYNC request, just ensure that we run it later on. The
1412 * latter allows for merging opportunities and more efficient
1413 * dispatching.
1414 */
1415run_queue:
1416 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
Jens Axboe320ae512013-10-24 09:20:05 +01001417 }
1418
Jens Axboe07068d52014-05-22 10:40:51 -06001419 blk_mq_put_ctx(data.ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001420}
1421
1422/*
1423 * Default mapping to a software queue, since we use one per CPU.
1424 */
1425struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1426{
1427 return q->queue_hw_ctx[q->mq_map[cpu]];
1428}
1429EXPORT_SYMBOL(blk_mq_map_queue);
1430
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001431static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1432 struct blk_mq_tags *tags, unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01001433{
1434 struct page *page;
1435
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001436 if (tags->rqs && set->ops->exit_request) {
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001437 int i;
1438
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001439 for (i = 0; i < tags->nr_tags; i++) {
1440 if (!tags->rqs[i])
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001441 continue;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001442 set->ops->exit_request(set->driver_data, tags->rqs[i],
1443 hctx_idx, i);
Jens Axboea5164402014-09-10 09:02:03 -06001444 tags->rqs[i] = NULL;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001445 }
1446 }
1447
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001448 while (!list_empty(&tags->page_list)) {
1449 page = list_first_entry(&tags->page_list, struct page, lru);
Dave Hansen67534712014-01-08 20:17:46 -07001450 list_del_init(&page->lru);
Jens Axboe320ae512013-10-24 09:20:05 +01001451 __free_pages(page, page->private);
1452 }
1453
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001454 kfree(tags->rqs);
Jens Axboe320ae512013-10-24 09:20:05 +01001455
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001456 blk_mq_free_tags(tags);
Jens Axboe320ae512013-10-24 09:20:05 +01001457}
1458
1459static size_t order_to_size(unsigned int order)
1460{
Ming Lei4ca08502014-04-19 18:00:18 +08001461 return (size_t)PAGE_SIZE << order;
Jens Axboe320ae512013-10-24 09:20:05 +01001462}
1463
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001464static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1465 unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01001466{
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001467 struct blk_mq_tags *tags;
Jens Axboe320ae512013-10-24 09:20:05 +01001468 unsigned int i, j, entries_per_page, max_order = 4;
1469 size_t rq_size, left;
1470
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001471 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
Shaohua Li24391c02015-01-23 14:18:00 -07001472 set->numa_node,
1473 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001474 if (!tags)
1475 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001476
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001477 INIT_LIST_HEAD(&tags->page_list);
1478
Jens Axboea5164402014-09-10 09:02:03 -06001479 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1480 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1481 set->numa_node);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001482 if (!tags->rqs) {
1483 blk_mq_free_tags(tags);
1484 return NULL;
1485 }
Jens Axboe320ae512013-10-24 09:20:05 +01001486
1487 /*
1488 * rq_size is the size of the request plus driver payload, rounded
1489 * to the cacheline size
1490 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001491 rq_size = round_up(sizeof(struct request) + set->cmd_size,
Jens Axboe320ae512013-10-24 09:20:05 +01001492 cache_line_size());
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001493 left = rq_size * set->queue_depth;
Jens Axboe320ae512013-10-24 09:20:05 +01001494
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001495 for (i = 0; i < set->queue_depth; ) {
Jens Axboe320ae512013-10-24 09:20:05 +01001496 int this_order = max_order;
1497 struct page *page;
1498 int to_do;
1499 void *p;
1500
1501 while (left < order_to_size(this_order - 1) && this_order)
1502 this_order--;
1503
1504 do {
Jens Axboea5164402014-09-10 09:02:03 -06001505 page = alloc_pages_node(set->numa_node,
Linus Torvaldsac211172015-04-09 14:12:22 -07001506 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
Jens Axboea5164402014-09-10 09:02:03 -06001507 this_order);
Jens Axboe320ae512013-10-24 09:20:05 +01001508 if (page)
1509 break;
1510 if (!this_order--)
1511 break;
1512 if (order_to_size(this_order) < rq_size)
1513 break;
1514 } while (1);
1515
1516 if (!page)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001517 goto fail;
Jens Axboe320ae512013-10-24 09:20:05 +01001518
1519 page->private = this_order;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001520 list_add_tail(&page->lru, &tags->page_list);
Jens Axboe320ae512013-10-24 09:20:05 +01001521
1522 p = page_address(page);
1523 entries_per_page = order_to_size(this_order) / rq_size;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001524 to_do = min(entries_per_page, set->queue_depth - i);
Jens Axboe320ae512013-10-24 09:20:05 +01001525 left -= to_do * rq_size;
1526 for (j = 0; j < to_do; j++) {
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001527 tags->rqs[i] = p;
1528 if (set->ops->init_request) {
1529 if (set->ops->init_request(set->driver_data,
1530 tags->rqs[i], hctx_idx, i,
Jens Axboea5164402014-09-10 09:02:03 -06001531 set->numa_node)) {
1532 tags->rqs[i] = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001533 goto fail;
Jens Axboea5164402014-09-10 09:02:03 -06001534 }
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001535 }
1536
Jens Axboe320ae512013-10-24 09:20:05 +01001537 p += rq_size;
1538 i++;
1539 }
1540 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001541 return tags;
Jens Axboe320ae512013-10-24 09:20:05 +01001542
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001543fail:
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001544 blk_mq_free_rq_map(set, tags, hctx_idx);
1545 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001546}
1547
Jens Axboe1429d7c2014-05-19 09:23:55 -06001548static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1549{
1550 kfree(bitmap->map);
1551}
1552
1553static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1554{
1555 unsigned int bpw = 8, total, num_maps, i;
1556
1557 bitmap->bits_per_word = bpw;
1558
1559 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1560 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1561 GFP_KERNEL, node);
1562 if (!bitmap->map)
1563 return -ENOMEM;
1564
Jens Axboe1429d7c2014-05-19 09:23:55 -06001565 total = nr_cpu_ids;
1566 for (i = 0; i < num_maps; i++) {
1567 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1568 total -= bitmap->map[i].depth;
1569 }
1570
1571 return 0;
1572}
1573
Jens Axboe484b4062014-05-21 14:01:15 -06001574static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1575{
1576 struct request_queue *q = hctx->queue;
1577 struct blk_mq_ctx *ctx;
1578 LIST_HEAD(tmp);
1579
1580 /*
1581 * Move ctx entries to new CPU, if this one is going away.
1582 */
1583 ctx = __blk_mq_get_ctx(q, cpu);
1584
1585 spin_lock(&ctx->lock);
1586 if (!list_empty(&ctx->rq_list)) {
1587 list_splice_init(&ctx->rq_list, &tmp);
1588 blk_mq_hctx_clear_pending(hctx, ctx);
1589 }
1590 spin_unlock(&ctx->lock);
1591
1592 if (list_empty(&tmp))
1593 return NOTIFY_OK;
1594
1595 ctx = blk_mq_get_ctx(q);
1596 spin_lock(&ctx->lock);
1597
1598 while (!list_empty(&tmp)) {
1599 struct request *rq;
1600
1601 rq = list_first_entry(&tmp, struct request, queuelist);
1602 rq->mq_ctx = ctx;
1603 list_move_tail(&rq->queuelist, &ctx->rq_list);
1604 }
1605
1606 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1607 blk_mq_hctx_mark_pending(hctx, ctx);
1608
1609 spin_unlock(&ctx->lock);
1610
1611 blk_mq_run_hw_queue(hctx, true);
1612 blk_mq_put_ctx(ctx);
1613 return NOTIFY_OK;
1614}
1615
Jens Axboe484b4062014-05-21 14:01:15 -06001616static int blk_mq_hctx_notify(void *data, unsigned long action,
1617 unsigned int cpu)
1618{
1619 struct blk_mq_hw_ctx *hctx = data;
1620
1621 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1622 return blk_mq_hctx_cpu_offline(hctx, cpu);
Ming Lei2a34c082015-04-21 10:00:20 +08001623
1624 /*
1625 * In case of CPU online, tags may be reallocated
1626 * in blk_mq_map_swqueue() after mapping is updated.
1627 */
Jens Axboe484b4062014-05-21 14:01:15 -06001628
1629 return NOTIFY_OK;
1630}
1631
Ming Leic3b4afc2015-06-04 22:25:04 +08001632/* hctx->ctxs will be freed in queue's release handler */
Ming Lei08e98fc2014-09-25 23:23:38 +08001633static void blk_mq_exit_hctx(struct request_queue *q,
1634 struct blk_mq_tag_set *set,
1635 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1636{
Ming Leif70ced02014-09-25 23:23:47 +08001637 unsigned flush_start_tag = set->queue_depth;
1638
Ming Lei08e98fc2014-09-25 23:23:38 +08001639 blk_mq_tag_idle(hctx);
1640
Ming Leif70ced02014-09-25 23:23:47 +08001641 if (set->ops->exit_request)
1642 set->ops->exit_request(set->driver_data,
1643 hctx->fq->flush_rq, hctx_idx,
1644 flush_start_tag + hctx_idx);
1645
Ming Lei08e98fc2014-09-25 23:23:38 +08001646 if (set->ops->exit_hctx)
1647 set->ops->exit_hctx(hctx, hctx_idx);
1648
1649 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
Ming Leif70ced02014-09-25 23:23:47 +08001650 blk_free_flush_queue(hctx->fq);
Ming Lei08e98fc2014-09-25 23:23:38 +08001651 blk_mq_free_bitmap(&hctx->ctx_map);
1652}
1653
Ming Lei624dbe42014-05-27 23:35:13 +08001654static void blk_mq_exit_hw_queues(struct request_queue *q,
1655 struct blk_mq_tag_set *set, int nr_queue)
1656{
1657 struct blk_mq_hw_ctx *hctx;
1658 unsigned int i;
1659
1660 queue_for_each_hw_ctx(q, hctx, i) {
1661 if (i == nr_queue)
1662 break;
Ming Lei08e98fc2014-09-25 23:23:38 +08001663 blk_mq_exit_hctx(q, set, hctx, i);
Ming Lei624dbe42014-05-27 23:35:13 +08001664 }
Ming Lei624dbe42014-05-27 23:35:13 +08001665}
1666
1667static void blk_mq_free_hw_queues(struct request_queue *q,
1668 struct blk_mq_tag_set *set)
1669{
1670 struct blk_mq_hw_ctx *hctx;
1671 unsigned int i;
1672
Ming Leie09aae72015-01-29 20:17:27 +08001673 queue_for_each_hw_ctx(q, hctx, i)
Ming Lei624dbe42014-05-27 23:35:13 +08001674 free_cpumask_var(hctx->cpumask);
Ming Lei624dbe42014-05-27 23:35:13 +08001675}
1676
Ming Lei08e98fc2014-09-25 23:23:38 +08001677static int blk_mq_init_hctx(struct request_queue *q,
1678 struct blk_mq_tag_set *set,
1679 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1680{
1681 int node;
Ming Leif70ced02014-09-25 23:23:47 +08001682 unsigned flush_start_tag = set->queue_depth;
Ming Lei08e98fc2014-09-25 23:23:38 +08001683
1684 node = hctx->numa_node;
1685 if (node == NUMA_NO_NODE)
1686 node = hctx->numa_node = set->numa_node;
1687
1688 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1689 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1690 spin_lock_init(&hctx->lock);
1691 INIT_LIST_HEAD(&hctx->dispatch);
1692 hctx->queue = q;
1693 hctx->queue_num = hctx_idx;
1694 hctx->flags = set->flags;
Ming Lei08e98fc2014-09-25 23:23:38 +08001695
1696 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1697 blk_mq_hctx_notify, hctx);
1698 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1699
1700 hctx->tags = set->tags[hctx_idx];
1701
1702 /*
1703 * Allocate space for all possible cpus to avoid allocation at
1704 * runtime
1705 */
1706 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1707 GFP_KERNEL, node);
1708 if (!hctx->ctxs)
1709 goto unregister_cpu_notifier;
1710
1711 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1712 goto free_ctxs;
1713
1714 hctx->nr_ctx = 0;
1715
1716 if (set->ops->init_hctx &&
1717 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1718 goto free_bitmap;
1719
Ming Leif70ced02014-09-25 23:23:47 +08001720 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1721 if (!hctx->fq)
1722 goto exit_hctx;
1723
1724 if (set->ops->init_request &&
1725 set->ops->init_request(set->driver_data,
1726 hctx->fq->flush_rq, hctx_idx,
1727 flush_start_tag + hctx_idx, node))
1728 goto free_fq;
1729
Ming Lei08e98fc2014-09-25 23:23:38 +08001730 return 0;
1731
Ming Leif70ced02014-09-25 23:23:47 +08001732 free_fq:
1733 kfree(hctx->fq);
1734 exit_hctx:
1735 if (set->ops->exit_hctx)
1736 set->ops->exit_hctx(hctx, hctx_idx);
Ming Lei08e98fc2014-09-25 23:23:38 +08001737 free_bitmap:
1738 blk_mq_free_bitmap(&hctx->ctx_map);
1739 free_ctxs:
1740 kfree(hctx->ctxs);
1741 unregister_cpu_notifier:
1742 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1743
1744 return -1;
1745}
1746
Jens Axboe320ae512013-10-24 09:20:05 +01001747static int blk_mq_init_hw_queues(struct request_queue *q,
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001748 struct blk_mq_tag_set *set)
Jens Axboe320ae512013-10-24 09:20:05 +01001749{
1750 struct blk_mq_hw_ctx *hctx;
Ming Lei624dbe42014-05-27 23:35:13 +08001751 unsigned int i;
Jens Axboe320ae512013-10-24 09:20:05 +01001752
1753 /*
1754 * Initialize hardware queues
1755 */
1756 queue_for_each_hw_ctx(q, hctx, i) {
Ming Lei08e98fc2014-09-25 23:23:38 +08001757 if (blk_mq_init_hctx(q, set, hctx, i))
Jens Axboe320ae512013-10-24 09:20:05 +01001758 break;
1759 }
1760
1761 if (i == q->nr_hw_queues)
1762 return 0;
1763
1764 /*
1765 * Init failed
1766 */
Ming Lei624dbe42014-05-27 23:35:13 +08001767 blk_mq_exit_hw_queues(q, set, i);
Jens Axboe320ae512013-10-24 09:20:05 +01001768
1769 return 1;
1770}
1771
1772static void blk_mq_init_cpu_queues(struct request_queue *q,
1773 unsigned int nr_hw_queues)
1774{
1775 unsigned int i;
1776
1777 for_each_possible_cpu(i) {
1778 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1779 struct blk_mq_hw_ctx *hctx;
1780
1781 memset(__ctx, 0, sizeof(*__ctx));
1782 __ctx->cpu = i;
1783 spin_lock_init(&__ctx->lock);
1784 INIT_LIST_HEAD(&__ctx->rq_list);
1785 __ctx->queue = q;
1786
1787 /* If the cpu isn't online, the cpu is mapped to first hctx */
Jens Axboe320ae512013-10-24 09:20:05 +01001788 if (!cpu_online(i))
1789 continue;
1790
Jens Axboee4043dc2014-04-09 10:18:23 -06001791 hctx = q->mq_ops->map_queue(q, i);
Jens Axboee4043dc2014-04-09 10:18:23 -06001792
Jens Axboe320ae512013-10-24 09:20:05 +01001793 /*
1794 * Set local node, IFF we have more than one hw queue. If
1795 * not, we remain on the home node of the device
1796 */
1797 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1798 hctx->numa_node = cpu_to_node(i);
1799 }
1800}
1801
1802static void blk_mq_map_swqueue(struct request_queue *q)
1803{
1804 unsigned int i;
1805 struct blk_mq_hw_ctx *hctx;
1806 struct blk_mq_ctx *ctx;
Ming Lei2a34c082015-04-21 10:00:20 +08001807 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01001808
1809 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboee4043dc2014-04-09 10:18:23 -06001810 cpumask_clear(hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001811 hctx->nr_ctx = 0;
1812 }
1813
1814 /*
1815 * Map software to hardware queues
1816 */
1817 queue_for_each_ctx(q, ctx, i) {
1818 /* If the cpu isn't online, the cpu is mapped to first hctx */
Jens Axboee4043dc2014-04-09 10:18:23 -06001819 if (!cpu_online(i))
1820 continue;
1821
Jens Axboe320ae512013-10-24 09:20:05 +01001822 hctx = q->mq_ops->map_queue(q, i);
Jens Axboee4043dc2014-04-09 10:18:23 -06001823 cpumask_set_cpu(i, hctx->cpumask);
Keith Buschf26cdc82015-06-01 09:29:53 -06001824 cpumask_set_cpu(i, hctx->tags->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001825 ctx->index_hw = hctx->nr_ctx;
1826 hctx->ctxs[hctx->nr_ctx++] = ctx;
1827 }
Jens Axboe506e9312014-05-07 10:26:44 -06001828
1829 queue_for_each_hw_ctx(q, hctx, i) {
Chong Yuan889fa312015-04-15 11:39:29 -06001830 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1831
Jens Axboe484b4062014-05-21 14:01:15 -06001832 /*
Jens Axboea68aafa2014-08-15 13:19:15 -06001833 * If no software queues are mapped to this hardware queue,
1834 * disable it and free the request entries.
Jens Axboe484b4062014-05-21 14:01:15 -06001835 */
1836 if (!hctx->nr_ctx) {
Jens Axboe484b4062014-05-21 14:01:15 -06001837 if (set->tags[i]) {
1838 blk_mq_free_rq_map(set, set->tags[i], i);
1839 set->tags[i] = NULL;
Jens Axboe484b4062014-05-21 14:01:15 -06001840 }
Ming Lei2a34c082015-04-21 10:00:20 +08001841 hctx->tags = NULL;
Jens Axboe484b4062014-05-21 14:01:15 -06001842 continue;
1843 }
1844
Ming Lei2a34c082015-04-21 10:00:20 +08001845 /* unmapped hw queue can be remapped after CPU topo changed */
1846 if (!set->tags[i])
1847 set->tags[i] = blk_mq_init_rq_map(set, i);
1848 hctx->tags = set->tags[i];
1849 WARN_ON(!hctx->tags);
1850
Jens Axboe484b4062014-05-21 14:01:15 -06001851 /*
Chong Yuan889fa312015-04-15 11:39:29 -06001852 * Set the map size to the number of mapped software queues.
1853 * This is more accurate and more efficient than looping
1854 * over all possibly mapped software queues.
1855 */
Jens Axboe569fd0c2015-04-17 08:28:50 -06001856 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
Chong Yuan889fa312015-04-15 11:39:29 -06001857
1858 /*
Jens Axboe484b4062014-05-21 14:01:15 -06001859 * Initialize batch roundrobin counts
1860 */
Jens Axboe506e9312014-05-07 10:26:44 -06001861 hctx->next_cpu = cpumask_first(hctx->cpumask);
1862 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1863 }
Jens Axboe320ae512013-10-24 09:20:05 +01001864}
1865
Jens Axboe0d2602c2014-05-13 15:10:52 -06001866static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1867{
1868 struct blk_mq_hw_ctx *hctx;
1869 struct request_queue *q;
1870 bool shared;
1871 int i;
1872
1873 if (set->tag_list.next == set->tag_list.prev)
1874 shared = false;
1875 else
1876 shared = true;
1877
1878 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1879 blk_mq_freeze_queue(q);
1880
1881 queue_for_each_hw_ctx(q, hctx, i) {
1882 if (shared)
1883 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1884 else
1885 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1886 }
1887 blk_mq_unfreeze_queue(q);
1888 }
1889}
1890
1891static void blk_mq_del_queue_tag_set(struct request_queue *q)
1892{
1893 struct blk_mq_tag_set *set = q->tag_set;
1894
Jens Axboe0d2602c2014-05-13 15:10:52 -06001895 mutex_lock(&set->tag_list_lock);
1896 list_del_init(&q->tag_set_list);
1897 blk_mq_update_tag_set_depth(set);
1898 mutex_unlock(&set->tag_list_lock);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001899}
1900
1901static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1902 struct request_queue *q)
1903{
1904 q->tag_set = set;
1905
1906 mutex_lock(&set->tag_list_lock);
1907 list_add_tail(&q->tag_set_list, &set->tag_list);
1908 blk_mq_update_tag_set_depth(set);
1909 mutex_unlock(&set->tag_list_lock);
1910}
1911
Ming Leie09aae72015-01-29 20:17:27 +08001912/*
1913 * It is the actual release handler for mq, but we do it from
1914 * request queue's release handler for avoiding use-after-free
1915 * and headache because q->mq_kobj shouldn't have been introduced,
1916 * but we can't group ctx/kctx kobj without it.
1917 */
1918void blk_mq_release(struct request_queue *q)
1919{
1920 struct blk_mq_hw_ctx *hctx;
1921 unsigned int i;
1922
1923 /* hctx kobj stays in hctx */
Ming Leic3b4afc2015-06-04 22:25:04 +08001924 queue_for_each_hw_ctx(q, hctx, i) {
1925 if (!hctx)
1926 continue;
1927 kfree(hctx->ctxs);
Ming Leie09aae72015-01-29 20:17:27 +08001928 kfree(hctx);
Ming Leic3b4afc2015-06-04 22:25:04 +08001929 }
Ming Leie09aae72015-01-29 20:17:27 +08001930
1931 kfree(q->queue_hw_ctx);
1932
1933 /* ctx kobj stays in queue_ctx */
1934 free_percpu(q->queue_ctx);
1935}
1936
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001937struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
Jens Axboe320ae512013-10-24 09:20:05 +01001938{
Mike Snitzerb62c21b2015-03-12 23:56:02 -04001939 struct request_queue *uninit_q, *q;
1940
1941 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1942 if (!uninit_q)
1943 return ERR_PTR(-ENOMEM);
1944
1945 q = blk_mq_init_allocated_queue(set, uninit_q);
1946 if (IS_ERR(q))
1947 blk_cleanup_queue(uninit_q);
1948
1949 return q;
1950}
1951EXPORT_SYMBOL(blk_mq_init_queue);
1952
1953struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1954 struct request_queue *q)
1955{
Jens Axboe320ae512013-10-24 09:20:05 +01001956 struct blk_mq_hw_ctx **hctxs;
Ming Leie6cdb092014-06-03 11:24:06 +08001957 struct blk_mq_ctx __percpu *ctx;
Jens Axboef14bbe72014-05-27 12:06:53 -06001958 unsigned int *map;
Jens Axboe320ae512013-10-24 09:20:05 +01001959 int i;
1960
Jens Axboe320ae512013-10-24 09:20:05 +01001961 ctx = alloc_percpu(struct blk_mq_ctx);
1962 if (!ctx)
1963 return ERR_PTR(-ENOMEM);
1964
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001965 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1966 set->numa_node);
Jens Axboe320ae512013-10-24 09:20:05 +01001967
1968 if (!hctxs)
1969 goto err_percpu;
1970
Jens Axboef14bbe72014-05-27 12:06:53 -06001971 map = blk_mq_make_queue_map(set);
1972 if (!map)
1973 goto err_map;
1974
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001975 for (i = 0; i < set->nr_hw_queues; i++) {
Jens Axboef14bbe72014-05-27 12:06:53 -06001976 int node = blk_mq_hw_queue_to_node(map, i);
1977
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02001978 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1979 GFP_KERNEL, node);
Jens Axboe320ae512013-10-24 09:20:05 +01001980 if (!hctxs[i])
1981 goto err_hctxs;
1982
Jens Axboea86073e2014-10-13 15:41:54 -06001983 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1984 node))
Jens Axboee4043dc2014-04-09 10:18:23 -06001985 goto err_hctxs;
1986
Jens Axboe0d2602c2014-05-13 15:10:52 -06001987 atomic_set(&hctxs[i]->nr_active, 0);
Jens Axboef14bbe72014-05-27 12:06:53 -06001988 hctxs[i]->numa_node = node;
Jens Axboe320ae512013-10-24 09:20:05 +01001989 hctxs[i]->queue_num = i;
1990 }
1991
Tejun Heo17497ac2014-09-24 13:31:50 -04001992 /*
1993 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1994 * See blk_register_queue() for details.
1995 */
Tejun Heoa34375e2014-09-08 09:51:30 +09001996 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
Tejun Heo17497ac2014-09-24 13:31:50 -04001997 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
Mike Snitzerb62c21b2015-03-12 23:56:02 -04001998 goto err_hctxs;
Ming Lei3d2936f2014-05-27 23:35:14 +08001999
Jens Axboe320ae512013-10-24 09:20:05 +01002000 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
Ming Leie56f6982015-07-16 19:53:22 +08002001 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
Jens Axboe320ae512013-10-24 09:20:05 +01002002
2003 q->nr_queues = nr_cpu_ids;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002004 q->nr_hw_queues = set->nr_hw_queues;
Jens Axboef14bbe72014-05-27 12:06:53 -06002005 q->mq_map = map;
Jens Axboe320ae512013-10-24 09:20:05 +01002006
2007 q->queue_ctx = ctx;
2008 q->queue_hw_ctx = hctxs;
2009
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002010 q->mq_ops = set->ops;
Jens Axboe94eddfb2013-11-19 09:25:07 -07002011 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
Jens Axboe320ae512013-10-24 09:20:05 +01002012
Jens Axboe05f1dd52014-05-29 09:53:32 -06002013 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2014 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2015
Christoph Hellwig1be036e2014-02-07 10:22:39 -08002016 q->sg_reserved_size = INT_MAX;
2017
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06002018 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2019 INIT_LIST_HEAD(&q->requeue_list);
2020 spin_lock_init(&q->requeue_lock);
2021
Jens Axboe07068d52014-05-22 10:40:51 -06002022 if (q->nr_hw_queues > 1)
2023 blk_queue_make_request(q, blk_mq_make_request);
2024 else
2025 blk_queue_make_request(q, blk_sq_make_request);
2026
Jens Axboeeba71762014-05-20 15:17:27 -06002027 /*
2028 * Do this after blk_queue_make_request() overrides it...
2029 */
2030 q->nr_requests = set->queue_depth;
2031
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002032 if (set->ops->complete)
2033 blk_queue_softirq_done(q, set->ops->complete);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08002034
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002035 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +01002036
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002037 if (blk_mq_init_hw_queues(q, set))
Mike Snitzerb62c21b2015-03-12 23:56:02 -04002038 goto err_hctxs;
Christoph Hellwig18741982014-02-10 09:29:00 -07002039
Jens Axboe320ae512013-10-24 09:20:05 +01002040 mutex_lock(&all_q_mutex);
2041 list_add_tail(&q->all_q_node, &all_q_list);
2042 mutex_unlock(&all_q_mutex);
2043
Jens Axboe0d2602c2014-05-13 15:10:52 -06002044 blk_mq_add_queue_tag_set(set, q);
2045
Jens Axboe484b4062014-05-21 14:01:15 -06002046 blk_mq_map_swqueue(q);
2047
Jens Axboe320ae512013-10-24 09:20:05 +01002048 return q;
Christoph Hellwig18741982014-02-10 09:29:00 -07002049
Jens Axboe320ae512013-10-24 09:20:05 +01002050err_hctxs:
Jens Axboef14bbe72014-05-27 12:06:53 -06002051 kfree(map);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002052 for (i = 0; i < set->nr_hw_queues; i++) {
Jens Axboe320ae512013-10-24 09:20:05 +01002053 if (!hctxs[i])
2054 break;
Jens Axboee4043dc2014-04-09 10:18:23 -06002055 free_cpumask_var(hctxs[i]->cpumask);
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02002056 kfree(hctxs[i]);
Jens Axboe320ae512013-10-24 09:20:05 +01002057 }
Jens Axboef14bbe72014-05-27 12:06:53 -06002058err_map:
Jens Axboe320ae512013-10-24 09:20:05 +01002059 kfree(hctxs);
2060err_percpu:
2061 free_percpu(ctx);
2062 return ERR_PTR(-ENOMEM);
2063}
Mike Snitzerb62c21b2015-03-12 23:56:02 -04002064EXPORT_SYMBOL(blk_mq_init_allocated_queue);
Jens Axboe320ae512013-10-24 09:20:05 +01002065
2066void blk_mq_free_queue(struct request_queue *q)
2067{
Ming Lei624dbe42014-05-27 23:35:13 +08002068 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01002069
Jens Axboe0d2602c2014-05-13 15:10:52 -06002070 blk_mq_del_queue_tag_set(q);
2071
Ming Lei624dbe42014-05-27 23:35:13 +08002072 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2073 blk_mq_free_hw_queues(q, set);
Jens Axboe320ae512013-10-24 09:20:05 +01002074
Tejun Heoadd703f2014-07-01 10:34:38 -06002075 percpu_ref_exit(&q->mq_usage_counter);
Ming Lei3d2936f2014-05-27 23:35:14 +08002076
Jens Axboe320ae512013-10-24 09:20:05 +01002077 kfree(q->mq_map);
2078
Jens Axboe320ae512013-10-24 09:20:05 +01002079 q->mq_map = NULL;
2080
2081 mutex_lock(&all_q_mutex);
2082 list_del_init(&q->all_q_node);
2083 mutex_unlock(&all_q_mutex);
2084}
Jens Axboe320ae512013-10-24 09:20:05 +01002085
2086/* Basically redo blk_mq_init_queue with queue frozen */
Paul Gortmakerf618ef72013-11-14 08:26:02 -07002087static void blk_mq_queue_reinit(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01002088{
Christoph Hellwig4ecd4fe2015-05-07 09:38:13 +02002089 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
Jens Axboe320ae512013-10-24 09:20:05 +01002090
Jens Axboe67aec142014-05-30 08:25:36 -06002091 blk_mq_sysfs_unregister(q);
2092
Jens Axboe320ae512013-10-24 09:20:05 +01002093 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2094
2095 /*
2096 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2097 * we should change hctx numa_node according to new topology (this
2098 * involves free and re-allocate memory, worthy doing?)
2099 */
2100
2101 blk_mq_map_swqueue(q);
2102
Jens Axboe67aec142014-05-30 08:25:36 -06002103 blk_mq_sysfs_register(q);
Jens Axboe320ae512013-10-24 09:20:05 +01002104}
2105
Paul Gortmakerf618ef72013-11-14 08:26:02 -07002106static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2107 unsigned long action, void *hcpu)
Jens Axboe320ae512013-10-24 09:20:05 +01002108{
2109 struct request_queue *q;
2110
2111 /*
Jens Axboe9fccfed2014-05-08 14:50:19 -06002112 * Before new mappings are established, hotadded cpu might already
2113 * start handling requests. This doesn't break anything as we map
2114 * offline CPUs to first hardware queue. We will re-init the queue
2115 * below to get optimal settings.
Jens Axboe320ae512013-10-24 09:20:05 +01002116 */
2117 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2118 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2119 return NOTIFY_OK;
2120
2121 mutex_lock(&all_q_mutex);
Tejun Heof3af0202014-11-04 13:52:27 -05002122
2123 /*
2124 * We need to freeze and reinit all existing queues. Freezing
2125 * involves synchronous wait for an RCU grace period and doing it
2126 * one by one may take a long time. Start freezing all queues in
2127 * one swoop and then wait for the completions so that freezing can
2128 * take place in parallel.
2129 */
2130 list_for_each_entry(q, &all_q_list, all_q_node)
2131 blk_mq_freeze_queue_start(q);
Ming Leif054b562015-04-21 10:00:19 +08002132 list_for_each_entry(q, &all_q_list, all_q_node) {
Tejun Heof3af0202014-11-04 13:52:27 -05002133 blk_mq_freeze_queue_wait(q);
2134
Ming Leif054b562015-04-21 10:00:19 +08002135 /*
2136 * timeout handler can't touch hw queue during the
2137 * reinitialization
2138 */
2139 del_timer_sync(&q->timeout);
2140 }
2141
Jens Axboe320ae512013-10-24 09:20:05 +01002142 list_for_each_entry(q, &all_q_list, all_q_node)
2143 blk_mq_queue_reinit(q);
Tejun Heof3af0202014-11-04 13:52:27 -05002144
2145 list_for_each_entry(q, &all_q_list, all_q_node)
2146 blk_mq_unfreeze_queue(q);
2147
Jens Axboe320ae512013-10-24 09:20:05 +01002148 mutex_unlock(&all_q_mutex);
2149 return NOTIFY_OK;
2150}
2151
Jens Axboea5164402014-09-10 09:02:03 -06002152static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2153{
2154 int i;
2155
2156 for (i = 0; i < set->nr_hw_queues; i++) {
2157 set->tags[i] = blk_mq_init_rq_map(set, i);
2158 if (!set->tags[i])
2159 goto out_unwind;
2160 }
2161
2162 return 0;
2163
2164out_unwind:
2165 while (--i >= 0)
2166 blk_mq_free_rq_map(set, set->tags[i], i);
2167
Jens Axboea5164402014-09-10 09:02:03 -06002168 return -ENOMEM;
2169}
2170
2171/*
2172 * Allocate the request maps associated with this tag_set. Note that this
2173 * may reduce the depth asked for, if memory is tight. set->queue_depth
2174 * will be updated to reflect the allocated depth.
2175 */
2176static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2177{
2178 unsigned int depth;
2179 int err;
2180
2181 depth = set->queue_depth;
2182 do {
2183 err = __blk_mq_alloc_rq_maps(set);
2184 if (!err)
2185 break;
2186
2187 set->queue_depth >>= 1;
2188 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2189 err = -ENOMEM;
2190 break;
2191 }
2192 } while (set->queue_depth);
2193
2194 if (!set->queue_depth || err) {
2195 pr_err("blk-mq: failed to allocate request map\n");
2196 return -ENOMEM;
2197 }
2198
2199 if (depth != set->queue_depth)
2200 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2201 depth, set->queue_depth);
2202
2203 return 0;
2204}
2205
Keith Buschf26cdc82015-06-01 09:29:53 -06002206struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2207{
2208 return tags->cpumask;
2209}
2210EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2211
Jens Axboea4391c62014-06-05 15:21:56 -06002212/*
2213 * Alloc a tag set to be associated with one or more request queues.
2214 * May fail with EINVAL for various error conditions. May adjust the
2215 * requested depth down, if if it too large. In that case, the set
2216 * value will be stored in set->queue_depth.
2217 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002218int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2219{
Bart Van Assche205fb5f2014-10-30 14:45:11 +01002220 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2221
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002222 if (!set->nr_hw_queues)
2223 return -EINVAL;
Jens Axboea4391c62014-06-05 15:21:56 -06002224 if (!set->queue_depth)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002225 return -EINVAL;
2226 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2227 return -EINVAL;
2228
Xiaoguang Wangf9018ac2015-03-30 13:19:14 +08002229 if (!set->ops->queue_rq || !set->ops->map_queue)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002230 return -EINVAL;
2231
Jens Axboea4391c62014-06-05 15:21:56 -06002232 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2233 pr_info("blk-mq: reduced tag depth to %u\n",
2234 BLK_MQ_MAX_DEPTH);
2235 set->queue_depth = BLK_MQ_MAX_DEPTH;
2236 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002237
Shaohua Li6637fad2014-11-30 16:00:58 -08002238 /*
2239 * If a crashdump is active, then we are potentially in a very
2240 * memory constrained environment. Limit us to 1 queue and
2241 * 64 tags to prevent using too much memory.
2242 */
2243 if (is_kdump_kernel()) {
2244 set->nr_hw_queues = 1;
2245 set->queue_depth = min(64U, set->queue_depth);
2246 }
2247
Ming Lei48479002014-04-19 18:00:17 +08002248 set->tags = kmalloc_node(set->nr_hw_queues *
2249 sizeof(struct blk_mq_tags *),
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002250 GFP_KERNEL, set->numa_node);
2251 if (!set->tags)
Jens Axboea5164402014-09-10 09:02:03 -06002252 return -ENOMEM;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002253
Jens Axboea5164402014-09-10 09:02:03 -06002254 if (blk_mq_alloc_rq_maps(set))
2255 goto enomem;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002256
Jens Axboe0d2602c2014-05-13 15:10:52 -06002257 mutex_init(&set->tag_list_lock);
2258 INIT_LIST_HEAD(&set->tag_list);
2259
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002260 return 0;
Jens Axboea5164402014-09-10 09:02:03 -06002261enomem:
Robert Elliott5676e7b2014-09-02 11:38:44 -05002262 kfree(set->tags);
2263 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002264 return -ENOMEM;
2265}
2266EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2267
2268void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2269{
2270 int i;
2271
Jens Axboe484b4062014-05-21 14:01:15 -06002272 for (i = 0; i < set->nr_hw_queues; i++) {
Keith Buschf26cdc82015-06-01 09:29:53 -06002273 if (set->tags[i]) {
Jens Axboe484b4062014-05-21 14:01:15 -06002274 blk_mq_free_rq_map(set, set->tags[i], i);
Keith Buschf26cdc82015-06-01 09:29:53 -06002275 free_cpumask_var(set->tags[i]->cpumask);
2276 }
Jens Axboe484b4062014-05-21 14:01:15 -06002277 }
2278
Ming Lei981bd182014-04-24 00:07:34 +08002279 kfree(set->tags);
Robert Elliott5676e7b2014-09-02 11:38:44 -05002280 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002281}
2282EXPORT_SYMBOL(blk_mq_free_tag_set);
2283
Jens Axboee3a2b3f2014-05-20 11:49:02 -06002284int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2285{
2286 struct blk_mq_tag_set *set = q->tag_set;
2287 struct blk_mq_hw_ctx *hctx;
2288 int i, ret;
2289
2290 if (!set || nr > set->queue_depth)
2291 return -EINVAL;
2292
2293 ret = 0;
2294 queue_for_each_hw_ctx(q, hctx, i) {
2295 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2296 if (ret)
2297 break;
2298 }
2299
2300 if (!ret)
2301 q->nr_requests = nr;
2302
2303 return ret;
2304}
2305
Jens Axboe676141e2014-03-20 13:29:18 -06002306void blk_mq_disable_hotplug(void)
2307{
2308 mutex_lock(&all_q_mutex);
2309}
2310
2311void blk_mq_enable_hotplug(void)
2312{
2313 mutex_unlock(&all_q_mutex);
2314}
2315
Jens Axboe320ae512013-10-24 09:20:05 +01002316static int __init blk_mq_init(void)
2317{
Jens Axboe320ae512013-10-24 09:20:05 +01002318 blk_mq_cpu_init();
2319
Tejun Heoadd703f2014-07-01 10:34:38 -06002320 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
Jens Axboe320ae512013-10-24 09:20:05 +01002321
2322 return 0;
2323}
2324subsys_initcall(blk_mq_init);