blob: 2390c5541e71fb09c3353004224d76b4c81bddc8 [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 Axboe1429d7c2014-05-19 09:23:55 -060044 for (i = 0; i < hctx->ctx_map.map_size; i++)
45 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
Jens Axboe320ae512013-10-24 09:20:05 +010080static int blk_mq_queue_enter(struct request_queue *q)
81{
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
Tejun Heoadd703f2014-07-01 10:34:38 -060088 ret = wait_event_interruptible(q->mq_freeze_wq,
89 !q->mq_freeze_depth || blk_queue_dying(q));
90 if (blk_queue_dying(q))
91 return -ENODEV;
92 if (ret)
93 return ret;
94 }
Jens Axboe320ae512013-10-24 09:20:05 +010095}
96
97static void blk_mq_queue_exit(struct request_queue *q)
98{
Tejun Heoadd703f2014-07-01 10:34:38 -060099 percpu_ref_put(&q->mq_usage_counter);
100}
101
102static void blk_mq_usage_counter_release(struct percpu_ref *ref)
103{
104 struct request_queue *q =
105 container_of(ref, struct request_queue, mq_usage_counter);
106
107 wake_up_all(&q->mq_freeze_wq);
Jens Axboe320ae512013-10-24 09:20:05 +0100108}
109
Keith Buschb4c6a022014-12-19 17:54:14 -0700110void blk_mq_freeze_queue_start(struct request_queue *q)
Ming Lei43a5e4e2013-12-26 21:31:35 +0800111{
Tejun Heocddd5d12014-08-16 08:02:24 -0400112 bool freeze;
113
Tejun Heo72d6f022014-07-01 10:33:02 -0600114 spin_lock_irq(q->queue_lock);
Tejun Heocddd5d12014-08-16 08:02:24 -0400115 freeze = !q->mq_freeze_depth++;
Tejun Heo72d6f022014-07-01 10:33:02 -0600116 spin_unlock_irq(q->queue_lock);
117
Tejun Heocddd5d12014-08-16 08:02:24 -0400118 if (freeze) {
Tejun Heo9eca8042014-09-24 13:07:33 -0400119 percpu_ref_kill(&q->mq_usage_counter);
Tejun Heocddd5d12014-08-16 08:02:24 -0400120 blk_mq_run_queues(q, false);
121 }
Tejun Heof3af0202014-11-04 13:52:27 -0500122}
Keith Buschb4c6a022014-12-19 17:54:14 -0700123EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
Tejun Heof3af0202014-11-04 13:52:27 -0500124
125static void blk_mq_freeze_queue_wait(struct request_queue *q)
126{
Tejun Heoadd703f2014-07-01 10:34:38 -0600127 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
Ming Lei43a5e4e2013-12-26 21:31:35 +0800128}
129
Tejun Heof3af0202014-11-04 13:52:27 -0500130/*
131 * Guarantee no request is in use, so we can change any data structure of
132 * the queue afterward.
133 */
134void blk_mq_freeze_queue(struct request_queue *q)
135{
136 blk_mq_freeze_queue_start(q);
137 blk_mq_freeze_queue_wait(q);
138}
139
Keith Buschb4c6a022014-12-19 17:54:14 -0700140void blk_mq_unfreeze_queue(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +0100141{
Tejun Heocddd5d12014-08-16 08:02:24 -0400142 bool wake;
Jens Axboe320ae512013-10-24 09:20:05 +0100143
144 spin_lock_irq(q->queue_lock);
Tejun Heo780db202014-07-01 10:31:13 -0600145 wake = !--q->mq_freeze_depth;
146 WARN_ON_ONCE(q->mq_freeze_depth < 0);
Jens Axboe320ae512013-10-24 09:20:05 +0100147 spin_unlock_irq(q->queue_lock);
Tejun Heoadd703f2014-07-01 10:34:38 -0600148 if (wake) {
149 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
Joe Lawrencea492f072014-08-28 08:15:21 -0600258 ret = blk_mq_queue_enter(q);
259 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 {
676 queue_for_each_hw_ctx(q, hctx, i)
677 blk_mq_tag_idle(hctx);
678 }
Jens Axboe320ae512013-10-24 09:20:05 +0100679}
680
681/*
682 * Reverse check our software queue for entries that we could potentially
683 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
684 * too much time checking for merges.
685 */
686static bool blk_mq_attempt_merge(struct request_queue *q,
687 struct blk_mq_ctx *ctx, struct bio *bio)
688{
689 struct request *rq;
690 int checked = 8;
691
692 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
693 int el_ret;
694
695 if (!checked--)
696 break;
697
698 if (!blk_rq_merge_ok(rq, bio))
699 continue;
700
701 el_ret = blk_try_merge(rq, bio);
702 if (el_ret == ELEVATOR_BACK_MERGE) {
703 if (bio_attempt_back_merge(q, rq, bio)) {
704 ctx->rq_merged++;
705 return true;
706 }
707 break;
708 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
709 if (bio_attempt_front_merge(q, rq, bio)) {
710 ctx->rq_merged++;
711 return true;
712 }
713 break;
714 }
715 }
716
717 return false;
718}
719
Jens Axboe320ae512013-10-24 09:20:05 +0100720/*
Jens Axboe1429d7c2014-05-19 09:23:55 -0600721 * Process software queues that have been marked busy, splicing them
722 * to the for-dispatch
723 */
724static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
725{
726 struct blk_mq_ctx *ctx;
727 int i;
728
729 for (i = 0; i < hctx->ctx_map.map_size; i++) {
730 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
731 unsigned int off, bit;
732
733 if (!bm->word)
734 continue;
735
736 bit = 0;
737 off = i * hctx->ctx_map.bits_per_word;
738 do {
739 bit = find_next_bit(&bm->word, bm->depth, bit);
740 if (bit >= bm->depth)
741 break;
742
743 ctx = hctx->ctxs[bit + off];
744 clear_bit(bit, &bm->word);
745 spin_lock(&ctx->lock);
746 list_splice_tail_init(&ctx->rq_list, list);
747 spin_unlock(&ctx->lock);
748
749 bit++;
750 } while (1);
751 }
752}
753
754/*
Jens Axboe320ae512013-10-24 09:20:05 +0100755 * Run this hardware queue, pulling any software queues mapped to it in.
756 * Note that this function currently has various problems around ordering
757 * of IO. In particular, we'd like FIFO behaviour on handling existing
758 * items on the hctx->dispatch list. Ignore that for now.
759 */
760static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
761{
762 struct request_queue *q = hctx->queue;
Jens Axboe320ae512013-10-24 09:20:05 +0100763 struct request *rq;
764 LIST_HEAD(rq_list);
Jens Axboe74c45052014-10-29 11:14:52 -0600765 LIST_HEAD(driver_list);
766 struct list_head *dptr;
Jens Axboe1429d7c2014-05-19 09:23:55 -0600767 int queued;
Jens Axboe320ae512013-10-24 09:20:05 +0100768
Jens Axboefd1270d2014-04-16 09:23:48 -0600769 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
Jens Axboee4043dc2014-04-09 10:18:23 -0600770
Jens Axboe5d12f902014-03-19 15:25:02 -0600771 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
Jens Axboe320ae512013-10-24 09:20:05 +0100772 return;
773
774 hctx->run++;
775
776 /*
777 * Touch any software queue that has pending entries.
778 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600779 flush_busy_ctxs(hctx, &rq_list);
Jens Axboe320ae512013-10-24 09:20:05 +0100780
781 /*
782 * If we have previous entries on our dispatch list, grab them
783 * and stuff them at the front for more fair dispatch.
784 */
785 if (!list_empty_careful(&hctx->dispatch)) {
786 spin_lock(&hctx->lock);
787 if (!list_empty(&hctx->dispatch))
788 list_splice_init(&hctx->dispatch, &rq_list);
789 spin_unlock(&hctx->lock);
790 }
791
792 /*
Jens Axboe74c45052014-10-29 11:14:52 -0600793 * Start off with dptr being NULL, so we start the first request
794 * immediately, even if we have more pending.
795 */
796 dptr = NULL;
797
798 /*
Jens Axboe320ae512013-10-24 09:20:05 +0100799 * Now process all the entries, sending them to the driver.
800 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600801 queued = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100802 while (!list_empty(&rq_list)) {
Jens Axboe74c45052014-10-29 11:14:52 -0600803 struct blk_mq_queue_data bd;
Jens Axboe320ae512013-10-24 09:20:05 +0100804 int ret;
805
806 rq = list_first_entry(&rq_list, struct request, queuelist);
807 list_del_init(&rq->queuelist);
Jens Axboe320ae512013-10-24 09:20:05 +0100808
Jens Axboe74c45052014-10-29 11:14:52 -0600809 bd.rq = rq;
810 bd.list = dptr;
811 bd.last = list_empty(&rq_list);
812
813 ret = q->mq_ops->queue_rq(hctx, &bd);
Jens Axboe320ae512013-10-24 09:20:05 +0100814 switch (ret) {
815 case BLK_MQ_RQ_QUEUE_OK:
816 queued++;
817 continue;
818 case BLK_MQ_RQ_QUEUE_BUSY:
Jens Axboe320ae512013-10-24 09:20:05 +0100819 list_add(&rq->queuelist, &rq_list);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200820 __blk_mq_requeue_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100821 break;
822 default:
823 pr_err("blk-mq: bad return on queue: %d\n", ret);
Jens Axboe320ae512013-10-24 09:20:05 +0100824 case BLK_MQ_RQ_QUEUE_ERROR:
Christoph Hellwig1e93b8c2014-02-11 08:27:13 -0800825 rq->errors = -EIO;
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700826 blk_mq_end_request(rq, rq->errors);
Jens Axboe320ae512013-10-24 09:20:05 +0100827 break;
828 }
829
830 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
831 break;
Jens Axboe74c45052014-10-29 11:14:52 -0600832
833 /*
834 * We've done the first request. If we have more than 1
835 * left in the list, set dptr to defer issue.
836 */
837 if (!dptr && rq_list.next != rq_list.prev)
838 dptr = &driver_list;
Jens Axboe320ae512013-10-24 09:20:05 +0100839 }
840
841 if (!queued)
842 hctx->dispatched[0]++;
843 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
844 hctx->dispatched[ilog2(queued) + 1]++;
845
846 /*
847 * Any items that need requeuing? Stuff them into hctx->dispatch,
848 * that is where we will continue on next queue run.
849 */
850 if (!list_empty(&rq_list)) {
851 spin_lock(&hctx->lock);
852 list_splice(&rq_list, &hctx->dispatch);
853 spin_unlock(&hctx->lock);
854 }
855}
856
Jens Axboe506e9312014-05-07 10:26:44 -0600857/*
858 * It'd be great if the workqueue API had a way to pass
859 * in a mask and had some smarts for more clever placement.
860 * For now we just round-robin here, switching for every
861 * BLK_MQ_CPU_WORK_BATCH queued items.
862 */
863static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
864{
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100865 if (hctx->queue->nr_hw_queues == 1)
866 return WORK_CPU_UNBOUND;
Jens Axboe506e9312014-05-07 10:26:44 -0600867
868 if (--hctx->next_cpu_batch <= 0) {
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100869 int cpu = hctx->next_cpu, next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600870
871 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
872 if (next_cpu >= nr_cpu_ids)
873 next_cpu = cpumask_first(hctx->cpumask);
874
875 hctx->next_cpu = next_cpu;
876 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100877
878 return cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600879 }
880
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100881 return hctx->next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600882}
883
Jens Axboe320ae512013-10-24 09:20:05 +0100884void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
885{
Ming Lei19c66e52014-12-03 19:38:04 +0800886 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
887 !blk_mq_hw_queue_mapped(hctx)))
Jens Axboe320ae512013-10-24 09:20:05 +0100888 return;
889
Paolo Bonzini398205b2014-11-07 23:03:59 +0100890 if (!async) {
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100891 int cpu = get_cpu();
892 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
Paolo Bonzini398205b2014-11-07 23:03:59 +0100893 __blk_mq_run_hw_queue(hctx);
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100894 put_cpu();
Paolo Bonzini398205b2014-11-07 23:03:59 +0100895 return;
896 }
Jens Axboee4043dc2014-04-09 10:18:23 -0600897
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100898 put_cpu();
Jens Axboee4043dc2014-04-09 10:18:23 -0600899 }
Paolo Bonzini398205b2014-11-07 23:03:59 +0100900
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100901 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
902 &hctx->run_work, 0);
Jens Axboe320ae512013-10-24 09:20:05 +0100903}
904
905void blk_mq_run_queues(struct request_queue *q, bool async)
906{
907 struct blk_mq_hw_ctx *hctx;
908 int i;
909
910 queue_for_each_hw_ctx(q, hctx, i) {
911 if ((!blk_mq_hctx_has_pending(hctx) &&
912 list_empty_careful(&hctx->dispatch)) ||
Jens Axboe5d12f902014-03-19 15:25:02 -0600913 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
Jens Axboe320ae512013-10-24 09:20:05 +0100914 continue;
915
916 blk_mq_run_hw_queue(hctx, async);
917 }
918}
919EXPORT_SYMBOL(blk_mq_run_queues);
920
921void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
922{
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600923 cancel_delayed_work(&hctx->run_work);
924 cancel_delayed_work(&hctx->delay_work);
Jens Axboe320ae512013-10-24 09:20:05 +0100925 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
926}
927EXPORT_SYMBOL(blk_mq_stop_hw_queue);
928
Christoph Hellwig280d45f2013-10-25 14:45:58 +0100929void blk_mq_stop_hw_queues(struct request_queue *q)
930{
931 struct blk_mq_hw_ctx *hctx;
932 int i;
933
934 queue_for_each_hw_ctx(q, hctx, i)
935 blk_mq_stop_hw_queue(hctx);
936}
937EXPORT_SYMBOL(blk_mq_stop_hw_queues);
938
Jens Axboe320ae512013-10-24 09:20:05 +0100939void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
940{
941 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboee4043dc2014-04-09 10:18:23 -0600942
Jens Axboe0ffbce82014-06-25 08:22:34 -0600943 blk_mq_run_hw_queue(hctx, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100944}
945EXPORT_SYMBOL(blk_mq_start_hw_queue);
946
Christoph Hellwig2f268552014-04-16 09:44:56 +0200947void blk_mq_start_hw_queues(struct request_queue *q)
948{
949 struct blk_mq_hw_ctx *hctx;
950 int i;
951
952 queue_for_each_hw_ctx(q, hctx, i)
953 blk_mq_start_hw_queue(hctx);
954}
955EXPORT_SYMBOL(blk_mq_start_hw_queues);
956
957
Christoph Hellwig1b4a3252014-04-16 09:44:54 +0200958void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
Jens Axboe320ae512013-10-24 09:20:05 +0100959{
960 struct blk_mq_hw_ctx *hctx;
961 int i;
962
963 queue_for_each_hw_ctx(q, hctx, i) {
964 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
965 continue;
966
967 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Christoph Hellwig1b4a3252014-04-16 09:44:54 +0200968 blk_mq_run_hw_queue(hctx, async);
Jens Axboe320ae512013-10-24 09:20:05 +0100969 }
970}
971EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
972
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600973static void blk_mq_run_work_fn(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +0100974{
975 struct blk_mq_hw_ctx *hctx;
976
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600977 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
Jens Axboee4043dc2014-04-09 10:18:23 -0600978
Jens Axboe320ae512013-10-24 09:20:05 +0100979 __blk_mq_run_hw_queue(hctx);
980}
981
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600982static void blk_mq_delay_work_fn(struct work_struct *work)
983{
984 struct blk_mq_hw_ctx *hctx;
985
986 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
987
988 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
989 __blk_mq_run_hw_queue(hctx);
990}
991
992void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
993{
Ming Lei19c66e52014-12-03 19:38:04 +0800994 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
995 return;
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600996
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100997 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
998 &hctx->delay_work, msecs_to_jiffies(msecs));
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600999}
1000EXPORT_SYMBOL(blk_mq_delay_queue);
1001
Jens Axboe320ae512013-10-24 09:20:05 +01001002static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig72a0a362014-02-07 10:22:36 -08001003 struct request *rq, bool at_head)
Jens Axboe320ae512013-10-24 09:20:05 +01001004{
1005 struct blk_mq_ctx *ctx = rq->mq_ctx;
1006
Jens Axboe01b983c2013-11-19 18:59:10 -07001007 trace_block_rq_insert(hctx->queue, rq);
1008
Christoph Hellwig72a0a362014-02-07 10:22:36 -08001009 if (at_head)
1010 list_add(&rq->queuelist, &ctx->rq_list);
1011 else
1012 list_add_tail(&rq->queuelist, &ctx->rq_list);
Jens Axboe4bb659b2014-05-09 09:36:49 -06001013
Jens Axboe320ae512013-10-24 09:20:05 +01001014 blk_mq_hctx_mark_pending(hctx, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001015}
1016
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001017void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1018 bool async)
Jens Axboe320ae512013-10-24 09:20:05 +01001019{
1020 struct request_queue *q = rq->q;
1021 struct blk_mq_hw_ctx *hctx;
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001022 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001023
1024 current_ctx = blk_mq_get_ctx(q);
Christoph Hellwigeeabc852014-03-21 08:57:37 -06001025 if (!cpu_online(ctx->cpu))
1026 rq->mq_ctx = ctx = current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001027
Jens Axboe320ae512013-10-24 09:20:05 +01001028 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1029
Christoph Hellwiga57a1782014-09-16 14:44:07 -07001030 spin_lock(&ctx->lock);
1031 __blk_mq_insert_request(hctx, rq, at_head);
1032 spin_unlock(&ctx->lock);
Jens Axboe320ae512013-10-24 09:20:05 +01001033
Jens Axboe320ae512013-10-24 09:20:05 +01001034 if (run_queue)
1035 blk_mq_run_hw_queue(hctx, async);
Jens Axboee4043dc2014-04-09 10:18:23 -06001036
1037 blk_mq_put_ctx(current_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001038}
1039
1040static void blk_mq_insert_requests(struct request_queue *q,
1041 struct blk_mq_ctx *ctx,
1042 struct list_head *list,
1043 int depth,
1044 bool from_schedule)
1045
1046{
1047 struct blk_mq_hw_ctx *hctx;
1048 struct blk_mq_ctx *current_ctx;
1049
1050 trace_block_unplug(q, depth, !from_schedule);
1051
1052 current_ctx = blk_mq_get_ctx(q);
1053
1054 if (!cpu_online(ctx->cpu))
1055 ctx = current_ctx;
1056 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1057
1058 /*
1059 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1060 * offline now
1061 */
1062 spin_lock(&ctx->lock);
1063 while (!list_empty(list)) {
1064 struct request *rq;
1065
1066 rq = list_first_entry(list, struct request, queuelist);
1067 list_del_init(&rq->queuelist);
1068 rq->mq_ctx = ctx;
Christoph Hellwig72a0a362014-02-07 10:22:36 -08001069 __blk_mq_insert_request(hctx, rq, false);
Jens Axboe320ae512013-10-24 09:20:05 +01001070 }
1071 spin_unlock(&ctx->lock);
1072
Jens Axboe320ae512013-10-24 09:20:05 +01001073 blk_mq_run_hw_queue(hctx, from_schedule);
Jens Axboee4043dc2014-04-09 10:18:23 -06001074 blk_mq_put_ctx(current_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001075}
1076
1077static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1078{
1079 struct request *rqa = container_of(a, struct request, queuelist);
1080 struct request *rqb = container_of(b, struct request, queuelist);
1081
1082 return !(rqa->mq_ctx < rqb->mq_ctx ||
1083 (rqa->mq_ctx == rqb->mq_ctx &&
1084 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1085}
1086
1087void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1088{
1089 struct blk_mq_ctx *this_ctx;
1090 struct request_queue *this_q;
1091 struct request *rq;
1092 LIST_HEAD(list);
1093 LIST_HEAD(ctx_list);
1094 unsigned int depth;
1095
1096 list_splice_init(&plug->mq_list, &list);
1097
1098 list_sort(NULL, &list, plug_ctx_cmp);
1099
1100 this_q = NULL;
1101 this_ctx = NULL;
1102 depth = 0;
1103
1104 while (!list_empty(&list)) {
1105 rq = list_entry_rq(list.next);
1106 list_del_init(&rq->queuelist);
1107 BUG_ON(!rq->q);
1108 if (rq->mq_ctx != this_ctx) {
1109 if (this_ctx) {
1110 blk_mq_insert_requests(this_q, this_ctx,
1111 &ctx_list, depth,
1112 from_schedule);
1113 }
1114
1115 this_ctx = rq->mq_ctx;
1116 this_q = rq->q;
1117 depth = 0;
1118 }
1119
1120 depth++;
1121 list_add_tail(&rq->queuelist, &ctx_list);
1122 }
1123
1124 /*
1125 * If 'this_ctx' is set, we know we have entries to complete
1126 * on 'ctx_list'. Do those.
1127 */
1128 if (this_ctx) {
1129 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1130 from_schedule);
1131 }
1132}
1133
1134static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1135{
1136 init_request_from_bio(rq, bio);
Jens Axboe4b570522014-05-29 11:00:11 -06001137
Jens Axboe3ee32372014-06-09 09:36:53 -06001138 if (blk_do_io_stat(rq))
Jens Axboe4b570522014-05-29 11:00:11 -06001139 blk_account_io_start(rq, 1);
Jens Axboe320ae512013-10-24 09:20:05 +01001140}
1141
Jens Axboe274a5842014-08-15 12:44:08 -06001142static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1143{
1144 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1145 !blk_queue_nomerges(hctx->queue);
1146}
1147
Jens Axboe07068d52014-05-22 10:40:51 -06001148static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1149 struct blk_mq_ctx *ctx,
1150 struct request *rq, struct bio *bio)
1151{
Jens Axboe274a5842014-08-15 12:44:08 -06001152 if (!hctx_allow_merges(hctx)) {
Jens Axboe07068d52014-05-22 10:40:51 -06001153 blk_mq_bio_to_request(rq, bio);
1154 spin_lock(&ctx->lock);
1155insert_rq:
1156 __blk_mq_insert_request(hctx, rq, false);
1157 spin_unlock(&ctx->lock);
1158 return false;
1159 } else {
Jens Axboe274a5842014-08-15 12:44:08 -06001160 struct request_queue *q = hctx->queue;
1161
Jens Axboe07068d52014-05-22 10:40:51 -06001162 spin_lock(&ctx->lock);
1163 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1164 blk_mq_bio_to_request(rq, bio);
1165 goto insert_rq;
1166 }
1167
1168 spin_unlock(&ctx->lock);
1169 __blk_mq_free_request(hctx, ctx, rq);
1170 return true;
1171 }
1172}
1173
1174struct blk_map_ctx {
1175 struct blk_mq_hw_ctx *hctx;
1176 struct blk_mq_ctx *ctx;
1177};
1178
1179static struct request *blk_mq_map_request(struct request_queue *q,
1180 struct bio *bio,
1181 struct blk_map_ctx *data)
Jens Axboe320ae512013-10-24 09:20:05 +01001182{
1183 struct blk_mq_hw_ctx *hctx;
1184 struct blk_mq_ctx *ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001185 struct request *rq;
Jens Axboe07068d52014-05-22 10:40:51 -06001186 int rw = bio_data_dir(bio);
Ming Leicb96a422014-06-01 00:43:37 +08001187 struct blk_mq_alloc_data alloc_data;
Jens Axboe320ae512013-10-24 09:20:05 +01001188
Jens Axboe07068d52014-05-22 10:40:51 -06001189 if (unlikely(blk_mq_queue_enter(q))) {
Nicholas Bellinger14ec77f2014-02-07 13:45:39 -07001190 bio_endio(bio, -EIO);
Jens Axboe07068d52014-05-22 10:40:51 -06001191 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001192 }
1193
1194 ctx = blk_mq_get_ctx(q);
1195 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1196
Jens Axboe07068d52014-05-22 10:40:51 -06001197 if (rw_is_sync(bio->bi_rw))
Shaohua Li27fbf4e2014-02-19 20:20:21 +08001198 rw |= REQ_SYNC;
Jens Axboe07068d52014-05-22 10:40:51 -06001199
Jens Axboe320ae512013-10-24 09:20:05 +01001200 trace_block_getrq(q, bio, rw);
Ming Leicb96a422014-06-01 00:43:37 +08001201 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1202 hctx);
1203 rq = __blk_mq_alloc_request(&alloc_data, rw);
Christoph Hellwig5dee8572014-05-27 20:59:47 +02001204 if (unlikely(!rq)) {
Christoph Hellwig793597a2014-05-27 20:59:49 +02001205 __blk_mq_run_hw_queue(hctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001206 blk_mq_put_ctx(ctx);
1207 trace_block_sleeprq(q, bio, rw);
Christoph Hellwig793597a2014-05-27 20:59:49 +02001208
1209 ctx = blk_mq_get_ctx(q);
Jens Axboe320ae512013-10-24 09:20:05 +01001210 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Ming Leicb96a422014-06-01 00:43:37 +08001211 blk_mq_set_alloc_data(&alloc_data, q,
1212 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1213 rq = __blk_mq_alloc_request(&alloc_data, rw);
1214 ctx = alloc_data.ctx;
1215 hctx = alloc_data.hctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001216 }
1217
1218 hctx->queued++;
Jens Axboe07068d52014-05-22 10:40:51 -06001219 data->hctx = hctx;
1220 data->ctx = ctx;
1221 return rq;
1222}
1223
1224/*
1225 * Multiple hardware queue variant. This will not use per-process plugs,
1226 * but will attempt to bypass the hctx queueing if we can go straight to
1227 * hardware for SYNC IO.
1228 */
1229static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1230{
1231 const int is_sync = rw_is_sync(bio->bi_rw);
1232 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1233 struct blk_map_ctx data;
1234 struct request *rq;
1235
1236 blk_queue_bounce(q, &bio);
1237
1238 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1239 bio_endio(bio, -EIO);
1240 return;
1241 }
1242
1243 rq = blk_mq_map_request(q, bio, &data);
1244 if (unlikely(!rq))
1245 return;
1246
1247 if (unlikely(is_flush_fua)) {
1248 blk_mq_bio_to_request(rq, bio);
1249 blk_insert_flush(rq);
1250 goto run_queue;
1251 }
1252
Jens Axboee167dfb2014-10-29 11:18:26 -06001253 /*
1254 * If the driver supports defer issued based on 'last', then
1255 * queue it up like normal since we can potentially save some
1256 * CPU this way.
1257 */
1258 if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
Jens Axboe74c45052014-10-29 11:14:52 -06001259 struct blk_mq_queue_data bd = {
1260 .rq = rq,
1261 .list = NULL,
1262 .last = 1
1263 };
Jens Axboe07068d52014-05-22 10:40:51 -06001264 int ret;
1265
1266 blk_mq_bio_to_request(rq, bio);
Jens Axboe07068d52014-05-22 10:40:51 -06001267
1268 /*
1269 * For OK queue, we are done. For error, kill it. Any other
1270 * error (busy), just add it to our list as we previously
1271 * would have done
1272 */
Jens Axboe74c45052014-10-29 11:14:52 -06001273 ret = q->mq_ops->queue_rq(data.hctx, &bd);
Jens Axboe07068d52014-05-22 10:40:51 -06001274 if (ret == BLK_MQ_RQ_QUEUE_OK)
1275 goto done;
1276 else {
1277 __blk_mq_requeue_request(rq);
1278
1279 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1280 rq->errors = -EIO;
Christoph Hellwigc8a446a2014-09-13 16:40:10 -07001281 blk_mq_end_request(rq, rq->errors);
Jens Axboe07068d52014-05-22 10:40:51 -06001282 goto done;
1283 }
1284 }
1285 }
1286
1287 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1288 /*
1289 * For a SYNC request, send it to the hardware immediately. For
1290 * an ASYNC request, just ensure that we run it later on. The
1291 * latter allows for merging opportunities and more efficient
1292 * dispatching.
1293 */
1294run_queue:
1295 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1296 }
1297done:
1298 blk_mq_put_ctx(data.ctx);
1299}
1300
1301/*
1302 * Single hardware queue variant. This will attempt to use any per-process
1303 * plug for merging and IO deferral.
1304 */
1305static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1306{
1307 const int is_sync = rw_is_sync(bio->bi_rw);
1308 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1309 unsigned int use_plug, request_count = 0;
1310 struct blk_map_ctx data;
1311 struct request *rq;
1312
1313 /*
1314 * If we have multiple hardware queues, just go directly to
1315 * one of those for sync IO.
1316 */
1317 use_plug = !is_flush_fua && !is_sync;
1318
1319 blk_queue_bounce(q, &bio);
1320
1321 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1322 bio_endio(bio, -EIO);
1323 return;
1324 }
1325
1326 if (use_plug && !blk_queue_nomerges(q) &&
1327 blk_attempt_plug_merge(q, bio, &request_count))
1328 return;
1329
1330 rq = blk_mq_map_request(q, bio, &data);
Jens Axboeff87bce2014-06-03 11:59:49 -06001331 if (unlikely(!rq))
1332 return;
Jens Axboe320ae512013-10-24 09:20:05 +01001333
1334 if (unlikely(is_flush_fua)) {
1335 blk_mq_bio_to_request(rq, bio);
Jens Axboe320ae512013-10-24 09:20:05 +01001336 blk_insert_flush(rq);
1337 goto run_queue;
1338 }
1339
1340 /*
1341 * A task plug currently exists. Since this is completely lockless,
1342 * utilize that to temporarily store requests until the task is
1343 * either done or scheduled away.
1344 */
1345 if (use_plug) {
1346 struct blk_plug *plug = current->plug;
1347
1348 if (plug) {
1349 blk_mq_bio_to_request(rq, bio);
Shaohua Li92f399c2013-10-29 12:01:03 -06001350 if (list_empty(&plug->mq_list))
Jens Axboe320ae512013-10-24 09:20:05 +01001351 trace_block_plug(q);
1352 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1353 blk_flush_plug_list(plug, false);
1354 trace_block_plug(q);
1355 }
1356 list_add_tail(&rq->queuelist, &plug->mq_list);
Jens Axboe07068d52014-05-22 10:40:51 -06001357 blk_mq_put_ctx(data.ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001358 return;
1359 }
1360 }
1361
Jens Axboe07068d52014-05-22 10:40:51 -06001362 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1363 /*
1364 * For a SYNC request, send it to the hardware immediately. For
1365 * an ASYNC request, just ensure that we run it later on. The
1366 * latter allows for merging opportunities and more efficient
1367 * dispatching.
1368 */
1369run_queue:
1370 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
Jens Axboe320ae512013-10-24 09:20:05 +01001371 }
1372
Jens Axboe07068d52014-05-22 10:40:51 -06001373 blk_mq_put_ctx(data.ctx);
Jens Axboe320ae512013-10-24 09:20:05 +01001374}
1375
1376/*
1377 * Default mapping to a software queue, since we use one per CPU.
1378 */
1379struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1380{
1381 return q->queue_hw_ctx[q->mq_map[cpu]];
1382}
1383EXPORT_SYMBOL(blk_mq_map_queue);
1384
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001385static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1386 struct blk_mq_tags *tags, unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01001387{
1388 struct page *page;
1389
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001390 if (tags->rqs && set->ops->exit_request) {
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001391 int i;
1392
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001393 for (i = 0; i < tags->nr_tags; i++) {
1394 if (!tags->rqs[i])
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001395 continue;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001396 set->ops->exit_request(set->driver_data, tags->rqs[i],
1397 hctx_idx, i);
Jens Axboea5164402014-09-10 09:02:03 -06001398 tags->rqs[i] = NULL;
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001399 }
1400 }
1401
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001402 while (!list_empty(&tags->page_list)) {
1403 page = list_first_entry(&tags->page_list, struct page, lru);
Dave Hansen67534712014-01-08 20:17:46 -07001404 list_del_init(&page->lru);
Jens Axboe320ae512013-10-24 09:20:05 +01001405 __free_pages(page, page->private);
1406 }
1407
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001408 kfree(tags->rqs);
Jens Axboe320ae512013-10-24 09:20:05 +01001409
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001410 blk_mq_free_tags(tags);
Jens Axboe320ae512013-10-24 09:20:05 +01001411}
1412
1413static size_t order_to_size(unsigned int order)
1414{
Ming Lei4ca08502014-04-19 18:00:18 +08001415 return (size_t)PAGE_SIZE << order;
Jens Axboe320ae512013-10-24 09:20:05 +01001416}
1417
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001418static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1419 unsigned int hctx_idx)
Jens Axboe320ae512013-10-24 09:20:05 +01001420{
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001421 struct blk_mq_tags *tags;
Jens Axboe320ae512013-10-24 09:20:05 +01001422 unsigned int i, j, entries_per_page, max_order = 4;
1423 size_t rq_size, left;
1424
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001425 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1426 set->numa_node);
1427 if (!tags)
1428 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001429
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001430 INIT_LIST_HEAD(&tags->page_list);
1431
Jens Axboea5164402014-09-10 09:02:03 -06001432 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1433 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1434 set->numa_node);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001435 if (!tags->rqs) {
1436 blk_mq_free_tags(tags);
1437 return NULL;
1438 }
Jens Axboe320ae512013-10-24 09:20:05 +01001439
1440 /*
1441 * rq_size is the size of the request plus driver payload, rounded
1442 * to the cacheline size
1443 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001444 rq_size = round_up(sizeof(struct request) + set->cmd_size,
Jens Axboe320ae512013-10-24 09:20:05 +01001445 cache_line_size());
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001446 left = rq_size * set->queue_depth;
Jens Axboe320ae512013-10-24 09:20:05 +01001447
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001448 for (i = 0; i < set->queue_depth; ) {
Jens Axboe320ae512013-10-24 09:20:05 +01001449 int this_order = max_order;
1450 struct page *page;
1451 int to_do;
1452 void *p;
1453
1454 while (left < order_to_size(this_order - 1) && this_order)
1455 this_order--;
1456
1457 do {
Jens Axboea5164402014-09-10 09:02:03 -06001458 page = alloc_pages_node(set->numa_node,
1459 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1460 this_order);
Jens Axboe320ae512013-10-24 09:20:05 +01001461 if (page)
1462 break;
1463 if (!this_order--)
1464 break;
1465 if (order_to_size(this_order) < rq_size)
1466 break;
1467 } while (1);
1468
1469 if (!page)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001470 goto fail;
Jens Axboe320ae512013-10-24 09:20:05 +01001471
1472 page->private = this_order;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001473 list_add_tail(&page->lru, &tags->page_list);
Jens Axboe320ae512013-10-24 09:20:05 +01001474
1475 p = page_address(page);
1476 entries_per_page = order_to_size(this_order) / rq_size;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001477 to_do = min(entries_per_page, set->queue_depth - i);
Jens Axboe320ae512013-10-24 09:20:05 +01001478 left -= to_do * rq_size;
1479 for (j = 0; j < to_do; j++) {
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001480 tags->rqs[i] = p;
David Hildenbrand683d0e12014-09-18 11:04:31 +02001481 tags->rqs[i]->atomic_flags = 0;
1482 tags->rqs[i]->cmd_flags = 0;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001483 if (set->ops->init_request) {
1484 if (set->ops->init_request(set->driver_data,
1485 tags->rqs[i], hctx_idx, i,
Jens Axboea5164402014-09-10 09:02:03 -06001486 set->numa_node)) {
1487 tags->rqs[i] = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001488 goto fail;
Jens Axboea5164402014-09-10 09:02:03 -06001489 }
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001490 }
1491
Jens Axboe320ae512013-10-24 09:20:05 +01001492 p += rq_size;
1493 i++;
1494 }
1495 }
1496
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001497 return tags;
Jens Axboe320ae512013-10-24 09:20:05 +01001498
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001499fail:
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001500 blk_mq_free_rq_map(set, tags, hctx_idx);
1501 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001502}
1503
Jens Axboe1429d7c2014-05-19 09:23:55 -06001504static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1505{
1506 kfree(bitmap->map);
1507}
1508
1509static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1510{
1511 unsigned int bpw = 8, total, num_maps, i;
1512
1513 bitmap->bits_per_word = bpw;
1514
1515 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1516 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1517 GFP_KERNEL, node);
1518 if (!bitmap->map)
1519 return -ENOMEM;
1520
1521 bitmap->map_size = num_maps;
1522
1523 total = nr_cpu_ids;
1524 for (i = 0; i < num_maps; i++) {
1525 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1526 total -= bitmap->map[i].depth;
1527 }
1528
1529 return 0;
1530}
1531
Jens Axboe484b4062014-05-21 14:01:15 -06001532static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1533{
1534 struct request_queue *q = hctx->queue;
1535 struct blk_mq_ctx *ctx;
1536 LIST_HEAD(tmp);
1537
1538 /*
1539 * Move ctx entries to new CPU, if this one is going away.
1540 */
1541 ctx = __blk_mq_get_ctx(q, cpu);
1542
1543 spin_lock(&ctx->lock);
1544 if (!list_empty(&ctx->rq_list)) {
1545 list_splice_init(&ctx->rq_list, &tmp);
1546 blk_mq_hctx_clear_pending(hctx, ctx);
1547 }
1548 spin_unlock(&ctx->lock);
1549
1550 if (list_empty(&tmp))
1551 return NOTIFY_OK;
1552
1553 ctx = blk_mq_get_ctx(q);
1554 spin_lock(&ctx->lock);
1555
1556 while (!list_empty(&tmp)) {
1557 struct request *rq;
1558
1559 rq = list_first_entry(&tmp, struct request, queuelist);
1560 rq->mq_ctx = ctx;
1561 list_move_tail(&rq->queuelist, &ctx->rq_list);
1562 }
1563
1564 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1565 blk_mq_hctx_mark_pending(hctx, ctx);
1566
1567 spin_unlock(&ctx->lock);
1568
1569 blk_mq_run_hw_queue(hctx, true);
1570 blk_mq_put_ctx(ctx);
1571 return NOTIFY_OK;
1572}
1573
1574static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1575{
1576 struct request_queue *q = hctx->queue;
1577 struct blk_mq_tag_set *set = q->tag_set;
1578
1579 if (set->tags[hctx->queue_num])
1580 return NOTIFY_OK;
1581
1582 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1583 if (!set->tags[hctx->queue_num])
1584 return NOTIFY_STOP;
1585
1586 hctx->tags = set->tags[hctx->queue_num];
1587 return NOTIFY_OK;
1588}
1589
1590static int blk_mq_hctx_notify(void *data, unsigned long action,
1591 unsigned int cpu)
1592{
1593 struct blk_mq_hw_ctx *hctx = data;
1594
1595 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1596 return blk_mq_hctx_cpu_offline(hctx, cpu);
1597 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1598 return blk_mq_hctx_cpu_online(hctx, cpu);
1599
1600 return NOTIFY_OK;
1601}
1602
Ming Lei08e98fc2014-09-25 23:23:38 +08001603static void blk_mq_exit_hctx(struct request_queue *q,
1604 struct blk_mq_tag_set *set,
1605 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1606{
Ming Leif70ced02014-09-25 23:23:47 +08001607 unsigned flush_start_tag = set->queue_depth;
1608
Ming Lei08e98fc2014-09-25 23:23:38 +08001609 blk_mq_tag_idle(hctx);
1610
Ming Leif70ced02014-09-25 23:23:47 +08001611 if (set->ops->exit_request)
1612 set->ops->exit_request(set->driver_data,
1613 hctx->fq->flush_rq, hctx_idx,
1614 flush_start_tag + hctx_idx);
1615
Ming Lei08e98fc2014-09-25 23:23:38 +08001616 if (set->ops->exit_hctx)
1617 set->ops->exit_hctx(hctx, hctx_idx);
1618
1619 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
Ming Leif70ced02014-09-25 23:23:47 +08001620 blk_free_flush_queue(hctx->fq);
Ming Lei08e98fc2014-09-25 23:23:38 +08001621 kfree(hctx->ctxs);
1622 blk_mq_free_bitmap(&hctx->ctx_map);
1623}
1624
Ming Lei624dbe42014-05-27 23:35:13 +08001625static void blk_mq_exit_hw_queues(struct request_queue *q,
1626 struct blk_mq_tag_set *set, int nr_queue)
1627{
1628 struct blk_mq_hw_ctx *hctx;
1629 unsigned int i;
1630
1631 queue_for_each_hw_ctx(q, hctx, i) {
1632 if (i == nr_queue)
1633 break;
Ming Lei08e98fc2014-09-25 23:23:38 +08001634 blk_mq_exit_hctx(q, set, hctx, i);
Ming Lei624dbe42014-05-27 23:35:13 +08001635 }
Ming Lei624dbe42014-05-27 23:35:13 +08001636}
1637
1638static void blk_mq_free_hw_queues(struct request_queue *q,
1639 struct blk_mq_tag_set *set)
1640{
1641 struct blk_mq_hw_ctx *hctx;
1642 unsigned int i;
1643
Ming Leie09aae72015-01-29 20:17:27 +08001644 queue_for_each_hw_ctx(q, hctx, i)
Ming Lei624dbe42014-05-27 23:35:13 +08001645 free_cpumask_var(hctx->cpumask);
Ming Lei624dbe42014-05-27 23:35:13 +08001646}
1647
Ming Lei08e98fc2014-09-25 23:23:38 +08001648static int blk_mq_init_hctx(struct request_queue *q,
1649 struct blk_mq_tag_set *set,
1650 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1651{
1652 int node;
Ming Leif70ced02014-09-25 23:23:47 +08001653 unsigned flush_start_tag = set->queue_depth;
Ming Lei08e98fc2014-09-25 23:23:38 +08001654
1655 node = hctx->numa_node;
1656 if (node == NUMA_NO_NODE)
1657 node = hctx->numa_node = set->numa_node;
1658
1659 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1660 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1661 spin_lock_init(&hctx->lock);
1662 INIT_LIST_HEAD(&hctx->dispatch);
1663 hctx->queue = q;
1664 hctx->queue_num = hctx_idx;
1665 hctx->flags = set->flags;
Ming Lei08e98fc2014-09-25 23:23:38 +08001666
1667 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1668 blk_mq_hctx_notify, hctx);
1669 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1670
1671 hctx->tags = set->tags[hctx_idx];
1672
1673 /*
1674 * Allocate space for all possible cpus to avoid allocation at
1675 * runtime
1676 */
1677 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1678 GFP_KERNEL, node);
1679 if (!hctx->ctxs)
1680 goto unregister_cpu_notifier;
1681
1682 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1683 goto free_ctxs;
1684
1685 hctx->nr_ctx = 0;
1686
1687 if (set->ops->init_hctx &&
1688 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1689 goto free_bitmap;
1690
Ming Leif70ced02014-09-25 23:23:47 +08001691 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1692 if (!hctx->fq)
1693 goto exit_hctx;
1694
1695 if (set->ops->init_request &&
1696 set->ops->init_request(set->driver_data,
1697 hctx->fq->flush_rq, hctx_idx,
1698 flush_start_tag + hctx_idx, node))
1699 goto free_fq;
1700
Ming Lei08e98fc2014-09-25 23:23:38 +08001701 return 0;
1702
Ming Leif70ced02014-09-25 23:23:47 +08001703 free_fq:
1704 kfree(hctx->fq);
1705 exit_hctx:
1706 if (set->ops->exit_hctx)
1707 set->ops->exit_hctx(hctx, hctx_idx);
Ming Lei08e98fc2014-09-25 23:23:38 +08001708 free_bitmap:
1709 blk_mq_free_bitmap(&hctx->ctx_map);
1710 free_ctxs:
1711 kfree(hctx->ctxs);
1712 unregister_cpu_notifier:
1713 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1714
1715 return -1;
1716}
1717
Jens Axboe320ae512013-10-24 09:20:05 +01001718static int blk_mq_init_hw_queues(struct request_queue *q,
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001719 struct blk_mq_tag_set *set)
Jens Axboe320ae512013-10-24 09:20:05 +01001720{
1721 struct blk_mq_hw_ctx *hctx;
Ming Lei624dbe42014-05-27 23:35:13 +08001722 unsigned int i;
Jens Axboe320ae512013-10-24 09:20:05 +01001723
1724 /*
1725 * Initialize hardware queues
1726 */
1727 queue_for_each_hw_ctx(q, hctx, i) {
Ming Lei08e98fc2014-09-25 23:23:38 +08001728 if (blk_mq_init_hctx(q, set, hctx, i))
Jens Axboe320ae512013-10-24 09:20:05 +01001729 break;
1730 }
1731
1732 if (i == q->nr_hw_queues)
1733 return 0;
1734
1735 /*
1736 * Init failed
1737 */
Ming Lei624dbe42014-05-27 23:35:13 +08001738 blk_mq_exit_hw_queues(q, set, i);
Jens Axboe320ae512013-10-24 09:20:05 +01001739
1740 return 1;
1741}
1742
1743static void blk_mq_init_cpu_queues(struct request_queue *q,
1744 unsigned int nr_hw_queues)
1745{
1746 unsigned int i;
1747
1748 for_each_possible_cpu(i) {
1749 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1750 struct blk_mq_hw_ctx *hctx;
1751
1752 memset(__ctx, 0, sizeof(*__ctx));
1753 __ctx->cpu = i;
1754 spin_lock_init(&__ctx->lock);
1755 INIT_LIST_HEAD(&__ctx->rq_list);
1756 __ctx->queue = q;
1757
1758 /* If the cpu isn't online, the cpu is mapped to first hctx */
Jens Axboe320ae512013-10-24 09:20:05 +01001759 if (!cpu_online(i))
1760 continue;
1761
Jens Axboee4043dc2014-04-09 10:18:23 -06001762 hctx = q->mq_ops->map_queue(q, i);
1763 cpumask_set_cpu(i, hctx->cpumask);
1764 hctx->nr_ctx++;
1765
Jens Axboe320ae512013-10-24 09:20:05 +01001766 /*
1767 * Set local node, IFF we have more than one hw queue. If
1768 * not, we remain on the home node of the device
1769 */
1770 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1771 hctx->numa_node = cpu_to_node(i);
1772 }
1773}
1774
1775static void blk_mq_map_swqueue(struct request_queue *q)
1776{
1777 unsigned int i;
1778 struct blk_mq_hw_ctx *hctx;
1779 struct blk_mq_ctx *ctx;
1780
1781 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboee4043dc2014-04-09 10:18:23 -06001782 cpumask_clear(hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001783 hctx->nr_ctx = 0;
1784 }
1785
1786 /*
1787 * Map software to hardware queues
1788 */
1789 queue_for_each_ctx(q, ctx, i) {
1790 /* If the cpu isn't online, the cpu is mapped to first hctx */
Jens Axboee4043dc2014-04-09 10:18:23 -06001791 if (!cpu_online(i))
1792 continue;
1793
Jens Axboe320ae512013-10-24 09:20:05 +01001794 hctx = q->mq_ops->map_queue(q, i);
Jens Axboee4043dc2014-04-09 10:18:23 -06001795 cpumask_set_cpu(i, hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001796 ctx->index_hw = hctx->nr_ctx;
1797 hctx->ctxs[hctx->nr_ctx++] = ctx;
1798 }
Jens Axboe506e9312014-05-07 10:26:44 -06001799
1800 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboe484b4062014-05-21 14:01:15 -06001801 /*
Jens Axboea68aafa2014-08-15 13:19:15 -06001802 * If no software queues are mapped to this hardware queue,
1803 * disable it and free the request entries.
Jens Axboe484b4062014-05-21 14:01:15 -06001804 */
1805 if (!hctx->nr_ctx) {
1806 struct blk_mq_tag_set *set = q->tag_set;
1807
1808 if (set->tags[i]) {
1809 blk_mq_free_rq_map(set, set->tags[i], i);
1810 set->tags[i] = NULL;
1811 hctx->tags = NULL;
1812 }
1813 continue;
1814 }
1815
1816 /*
1817 * Initialize batch roundrobin counts
1818 */
Jens Axboe506e9312014-05-07 10:26:44 -06001819 hctx->next_cpu = cpumask_first(hctx->cpumask);
1820 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1821 }
Jens Axboe320ae512013-10-24 09:20:05 +01001822}
1823
Jens Axboe0d2602c2014-05-13 15:10:52 -06001824static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1825{
1826 struct blk_mq_hw_ctx *hctx;
1827 struct request_queue *q;
1828 bool shared;
1829 int i;
1830
1831 if (set->tag_list.next == set->tag_list.prev)
1832 shared = false;
1833 else
1834 shared = true;
1835
1836 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1837 blk_mq_freeze_queue(q);
1838
1839 queue_for_each_hw_ctx(q, hctx, i) {
1840 if (shared)
1841 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1842 else
1843 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1844 }
1845 blk_mq_unfreeze_queue(q);
1846 }
1847}
1848
1849static void blk_mq_del_queue_tag_set(struct request_queue *q)
1850{
1851 struct blk_mq_tag_set *set = q->tag_set;
1852
Jens Axboe0d2602c2014-05-13 15:10:52 -06001853 mutex_lock(&set->tag_list_lock);
1854 list_del_init(&q->tag_set_list);
1855 blk_mq_update_tag_set_depth(set);
1856 mutex_unlock(&set->tag_list_lock);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001857}
1858
1859static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1860 struct request_queue *q)
1861{
1862 q->tag_set = set;
1863
1864 mutex_lock(&set->tag_list_lock);
1865 list_add_tail(&q->tag_set_list, &set->tag_list);
1866 blk_mq_update_tag_set_depth(set);
1867 mutex_unlock(&set->tag_list_lock);
1868}
1869
Ming Leie09aae72015-01-29 20:17:27 +08001870/*
1871 * It is the actual release handler for mq, but we do it from
1872 * request queue's release handler for avoiding use-after-free
1873 * and headache because q->mq_kobj shouldn't have been introduced,
1874 * but we can't group ctx/kctx kobj without it.
1875 */
1876void blk_mq_release(struct request_queue *q)
1877{
1878 struct blk_mq_hw_ctx *hctx;
1879 unsigned int i;
1880
1881 /* hctx kobj stays in hctx */
1882 queue_for_each_hw_ctx(q, hctx, i)
1883 kfree(hctx);
1884
1885 kfree(q->queue_hw_ctx);
1886
1887 /* ctx kobj stays in queue_ctx */
1888 free_percpu(q->queue_ctx);
1889}
1890
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001891struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
Jens Axboe320ae512013-10-24 09:20:05 +01001892{
1893 struct blk_mq_hw_ctx **hctxs;
Ming Leie6cdb092014-06-03 11:24:06 +08001894 struct blk_mq_ctx __percpu *ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001895 struct request_queue *q;
Jens Axboef14bbe72014-05-27 12:06:53 -06001896 unsigned int *map;
Jens Axboe320ae512013-10-24 09:20:05 +01001897 int i;
1898
Jens Axboe320ae512013-10-24 09:20:05 +01001899 ctx = alloc_percpu(struct blk_mq_ctx);
1900 if (!ctx)
1901 return ERR_PTR(-ENOMEM);
1902
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001903 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1904 set->numa_node);
Jens Axboe320ae512013-10-24 09:20:05 +01001905
1906 if (!hctxs)
1907 goto err_percpu;
1908
Jens Axboef14bbe72014-05-27 12:06:53 -06001909 map = blk_mq_make_queue_map(set);
1910 if (!map)
1911 goto err_map;
1912
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001913 for (i = 0; i < set->nr_hw_queues; i++) {
Jens Axboef14bbe72014-05-27 12:06:53 -06001914 int node = blk_mq_hw_queue_to_node(map, i);
1915
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02001916 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1917 GFP_KERNEL, node);
Jens Axboe320ae512013-10-24 09:20:05 +01001918 if (!hctxs[i])
1919 goto err_hctxs;
1920
Jens Axboea86073e2014-10-13 15:41:54 -06001921 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1922 node))
Jens Axboee4043dc2014-04-09 10:18:23 -06001923 goto err_hctxs;
1924
Jens Axboe0d2602c2014-05-13 15:10:52 -06001925 atomic_set(&hctxs[i]->nr_active, 0);
Jens Axboef14bbe72014-05-27 12:06:53 -06001926 hctxs[i]->numa_node = node;
Jens Axboe320ae512013-10-24 09:20:05 +01001927 hctxs[i]->queue_num = i;
1928 }
1929
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001930 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
Jens Axboe320ae512013-10-24 09:20:05 +01001931 if (!q)
1932 goto err_hctxs;
1933
Tejun Heo17497ac2014-09-24 13:31:50 -04001934 /*
1935 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1936 * See blk_register_queue() for details.
1937 */
Tejun Heoa34375e2014-09-08 09:51:30 +09001938 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
Tejun Heo17497ac2014-09-24 13:31:50 -04001939 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
Ming Lei3d2936f2014-05-27 23:35:14 +08001940 goto err_map;
1941
Jens Axboe320ae512013-10-24 09:20:05 +01001942 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1943 blk_queue_rq_timeout(q, 30000);
1944
1945 q->nr_queues = nr_cpu_ids;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001946 q->nr_hw_queues = set->nr_hw_queues;
Jens Axboef14bbe72014-05-27 12:06:53 -06001947 q->mq_map = map;
Jens Axboe320ae512013-10-24 09:20:05 +01001948
1949 q->queue_ctx = ctx;
1950 q->queue_hw_ctx = hctxs;
1951
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001952 q->mq_ops = set->ops;
Jens Axboe94eddfb2013-11-19 09:25:07 -07001953 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
Jens Axboe320ae512013-10-24 09:20:05 +01001954
Jens Axboe05f1dd52014-05-29 09:53:32 -06001955 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1956 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1957
Christoph Hellwig1be036e2014-02-07 10:22:39 -08001958 q->sg_reserved_size = INT_MAX;
1959
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001960 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1961 INIT_LIST_HEAD(&q->requeue_list);
1962 spin_lock_init(&q->requeue_lock);
1963
Jens Axboe07068d52014-05-22 10:40:51 -06001964 if (q->nr_hw_queues > 1)
1965 blk_queue_make_request(q, blk_mq_make_request);
1966 else
1967 blk_queue_make_request(q, blk_sq_make_request);
1968
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001969 if (set->timeout)
1970 blk_queue_rq_timeout(q, set->timeout);
Jens Axboe320ae512013-10-24 09:20:05 +01001971
Jens Axboeeba71762014-05-20 15:17:27 -06001972 /*
1973 * Do this after blk_queue_make_request() overrides it...
1974 */
1975 q->nr_requests = set->queue_depth;
1976
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001977 if (set->ops->complete)
1978 blk_queue_softirq_done(q, set->ops->complete);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08001979
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001980 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +01001981
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001982 if (blk_mq_init_hw_queues(q, set))
Ming Lei1bcb1ea2014-09-25 23:23:39 +08001983 goto err_hw;
Christoph Hellwig18741982014-02-10 09:29:00 -07001984
Jens Axboe320ae512013-10-24 09:20:05 +01001985 mutex_lock(&all_q_mutex);
1986 list_add_tail(&q->all_q_node, &all_q_list);
1987 mutex_unlock(&all_q_mutex);
1988
Jens Axboe0d2602c2014-05-13 15:10:52 -06001989 blk_mq_add_queue_tag_set(set, q);
1990
Jens Axboe484b4062014-05-21 14:01:15 -06001991 blk_mq_map_swqueue(q);
1992
Jens Axboe320ae512013-10-24 09:20:05 +01001993 return q;
Christoph Hellwig18741982014-02-10 09:29:00 -07001994
Jens Axboe320ae512013-10-24 09:20:05 +01001995err_hw:
Jens Axboe320ae512013-10-24 09:20:05 +01001996 blk_cleanup_queue(q);
1997err_hctxs:
Jens Axboef14bbe72014-05-27 12:06:53 -06001998 kfree(map);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001999 for (i = 0; i < set->nr_hw_queues; i++) {
Jens Axboe320ae512013-10-24 09:20:05 +01002000 if (!hctxs[i])
2001 break;
Jens Axboee4043dc2014-04-09 10:18:23 -06002002 free_cpumask_var(hctxs[i]->cpumask);
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02002003 kfree(hctxs[i]);
Jens Axboe320ae512013-10-24 09:20:05 +01002004 }
Jens Axboef14bbe72014-05-27 12:06:53 -06002005err_map:
Jens Axboe320ae512013-10-24 09:20:05 +01002006 kfree(hctxs);
2007err_percpu:
2008 free_percpu(ctx);
2009 return ERR_PTR(-ENOMEM);
2010}
2011EXPORT_SYMBOL(blk_mq_init_queue);
2012
2013void blk_mq_free_queue(struct request_queue *q)
2014{
Ming Lei624dbe42014-05-27 23:35:13 +08002015 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01002016
Jens Axboe0d2602c2014-05-13 15:10:52 -06002017 blk_mq_del_queue_tag_set(q);
2018
Ming Lei624dbe42014-05-27 23:35:13 +08002019 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2020 blk_mq_free_hw_queues(q, set);
Jens Axboe320ae512013-10-24 09:20:05 +01002021
Tejun Heoadd703f2014-07-01 10:34:38 -06002022 percpu_ref_exit(&q->mq_usage_counter);
Ming Lei3d2936f2014-05-27 23:35:14 +08002023
Jens Axboe320ae512013-10-24 09:20:05 +01002024 kfree(q->mq_map);
2025
Jens Axboe320ae512013-10-24 09:20:05 +01002026 q->mq_map = NULL;
2027
2028 mutex_lock(&all_q_mutex);
2029 list_del_init(&q->all_q_node);
2030 mutex_unlock(&all_q_mutex);
2031}
Jens Axboe320ae512013-10-24 09:20:05 +01002032
2033/* Basically redo blk_mq_init_queue with queue frozen */
Paul Gortmakerf618ef72013-11-14 08:26:02 -07002034static void blk_mq_queue_reinit(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01002035{
Tejun Heof3af0202014-11-04 13:52:27 -05002036 WARN_ON_ONCE(!q->mq_freeze_depth);
Jens Axboe320ae512013-10-24 09:20:05 +01002037
Jens Axboe67aec142014-05-30 08:25:36 -06002038 blk_mq_sysfs_unregister(q);
2039
Jens Axboe320ae512013-10-24 09:20:05 +01002040 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2041
2042 /*
2043 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2044 * we should change hctx numa_node according to new topology (this
2045 * involves free and re-allocate memory, worthy doing?)
2046 */
2047
2048 blk_mq_map_swqueue(q);
2049
Jens Axboe67aec142014-05-30 08:25:36 -06002050 blk_mq_sysfs_register(q);
Jens Axboe320ae512013-10-24 09:20:05 +01002051}
2052
Paul Gortmakerf618ef72013-11-14 08:26:02 -07002053static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2054 unsigned long action, void *hcpu)
Jens Axboe320ae512013-10-24 09:20:05 +01002055{
2056 struct request_queue *q;
2057
2058 /*
Jens Axboe9fccfed2014-05-08 14:50:19 -06002059 * Before new mappings are established, hotadded cpu might already
2060 * start handling requests. This doesn't break anything as we map
2061 * offline CPUs to first hardware queue. We will re-init the queue
2062 * below to get optimal settings.
Jens Axboe320ae512013-10-24 09:20:05 +01002063 */
2064 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2065 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2066 return NOTIFY_OK;
2067
2068 mutex_lock(&all_q_mutex);
Tejun Heof3af0202014-11-04 13:52:27 -05002069
2070 /*
2071 * We need to freeze and reinit all existing queues. Freezing
2072 * involves synchronous wait for an RCU grace period and doing it
2073 * one by one may take a long time. Start freezing all queues in
2074 * one swoop and then wait for the completions so that freezing can
2075 * take place in parallel.
2076 */
2077 list_for_each_entry(q, &all_q_list, all_q_node)
2078 blk_mq_freeze_queue_start(q);
2079 list_for_each_entry(q, &all_q_list, all_q_node)
2080 blk_mq_freeze_queue_wait(q);
2081
Jens Axboe320ae512013-10-24 09:20:05 +01002082 list_for_each_entry(q, &all_q_list, all_q_node)
2083 blk_mq_queue_reinit(q);
Tejun Heof3af0202014-11-04 13:52:27 -05002084
2085 list_for_each_entry(q, &all_q_list, all_q_node)
2086 blk_mq_unfreeze_queue(q);
2087
Jens Axboe320ae512013-10-24 09:20:05 +01002088 mutex_unlock(&all_q_mutex);
2089 return NOTIFY_OK;
2090}
2091
Jens Axboea5164402014-09-10 09:02:03 -06002092static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2093{
2094 int i;
2095
2096 for (i = 0; i < set->nr_hw_queues; i++) {
2097 set->tags[i] = blk_mq_init_rq_map(set, i);
2098 if (!set->tags[i])
2099 goto out_unwind;
2100 }
2101
2102 return 0;
2103
2104out_unwind:
2105 while (--i >= 0)
2106 blk_mq_free_rq_map(set, set->tags[i], i);
2107
Jens Axboea5164402014-09-10 09:02:03 -06002108 return -ENOMEM;
2109}
2110
2111/*
2112 * Allocate the request maps associated with this tag_set. Note that this
2113 * may reduce the depth asked for, if memory is tight. set->queue_depth
2114 * will be updated to reflect the allocated depth.
2115 */
2116static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2117{
2118 unsigned int depth;
2119 int err;
2120
2121 depth = set->queue_depth;
2122 do {
2123 err = __blk_mq_alloc_rq_maps(set);
2124 if (!err)
2125 break;
2126
2127 set->queue_depth >>= 1;
2128 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2129 err = -ENOMEM;
2130 break;
2131 }
2132 } while (set->queue_depth);
2133
2134 if (!set->queue_depth || err) {
2135 pr_err("blk-mq: failed to allocate request map\n");
2136 return -ENOMEM;
2137 }
2138
2139 if (depth != set->queue_depth)
2140 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2141 depth, set->queue_depth);
2142
2143 return 0;
2144}
2145
Jens Axboea4391c62014-06-05 15:21:56 -06002146/*
2147 * Alloc a tag set to be associated with one or more request queues.
2148 * May fail with EINVAL for various error conditions. May adjust the
2149 * requested depth down, if if it too large. In that case, the set
2150 * value will be stored in set->queue_depth.
2151 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002152int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2153{
Bart Van Assche205fb5f2014-10-30 14:45:11 +01002154 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2155
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002156 if (!set->nr_hw_queues)
2157 return -EINVAL;
Jens Axboea4391c62014-06-05 15:21:56 -06002158 if (!set->queue_depth)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002159 return -EINVAL;
2160 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2161 return -EINVAL;
2162
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02002163 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002164 return -EINVAL;
2165
Jens Axboea4391c62014-06-05 15:21:56 -06002166 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2167 pr_info("blk-mq: reduced tag depth to %u\n",
2168 BLK_MQ_MAX_DEPTH);
2169 set->queue_depth = BLK_MQ_MAX_DEPTH;
2170 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002171
Shaohua Li6637fad2014-11-30 16:00:58 -08002172 /*
2173 * If a crashdump is active, then we are potentially in a very
2174 * memory constrained environment. Limit us to 1 queue and
2175 * 64 tags to prevent using too much memory.
2176 */
2177 if (is_kdump_kernel()) {
2178 set->nr_hw_queues = 1;
2179 set->queue_depth = min(64U, set->queue_depth);
2180 }
2181
Ming Lei48479002014-04-19 18:00:17 +08002182 set->tags = kmalloc_node(set->nr_hw_queues *
2183 sizeof(struct blk_mq_tags *),
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002184 GFP_KERNEL, set->numa_node);
2185 if (!set->tags)
Jens Axboea5164402014-09-10 09:02:03 -06002186 return -ENOMEM;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002187
Jens Axboea5164402014-09-10 09:02:03 -06002188 if (blk_mq_alloc_rq_maps(set))
2189 goto enomem;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002190
Jens Axboe0d2602c2014-05-13 15:10:52 -06002191 mutex_init(&set->tag_list_lock);
2192 INIT_LIST_HEAD(&set->tag_list);
2193
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002194 return 0;
Jens Axboea5164402014-09-10 09:02:03 -06002195enomem:
Robert Elliott5676e7b2014-09-02 11:38:44 -05002196 kfree(set->tags);
2197 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002198 return -ENOMEM;
2199}
2200EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2201
2202void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2203{
2204 int i;
2205
Jens Axboe484b4062014-05-21 14:01:15 -06002206 for (i = 0; i < set->nr_hw_queues; i++) {
2207 if (set->tags[i])
2208 blk_mq_free_rq_map(set, set->tags[i], i);
2209 }
2210
Ming Lei981bd182014-04-24 00:07:34 +08002211 kfree(set->tags);
Robert Elliott5676e7b2014-09-02 11:38:44 -05002212 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002213}
2214EXPORT_SYMBOL(blk_mq_free_tag_set);
2215
Jens Axboee3a2b3f2014-05-20 11:49:02 -06002216int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2217{
2218 struct blk_mq_tag_set *set = q->tag_set;
2219 struct blk_mq_hw_ctx *hctx;
2220 int i, ret;
2221
2222 if (!set || nr > set->queue_depth)
2223 return -EINVAL;
2224
2225 ret = 0;
2226 queue_for_each_hw_ctx(q, hctx, i) {
2227 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2228 if (ret)
2229 break;
2230 }
2231
2232 if (!ret)
2233 q->nr_requests = nr;
2234
2235 return ret;
2236}
2237
Jens Axboe676141e2014-03-20 13:29:18 -06002238void blk_mq_disable_hotplug(void)
2239{
2240 mutex_lock(&all_q_mutex);
2241}
2242
2243void blk_mq_enable_hotplug(void)
2244{
2245 mutex_unlock(&all_q_mutex);
2246}
2247
Jens Axboe320ae512013-10-24 09:20:05 +01002248static int __init blk_mq_init(void)
2249{
Jens Axboe320ae512013-10-24 09:20:05 +01002250 blk_mq_cpu_init();
2251
Tejun Heoadd703f2014-07-01 10:34:38 -06002252 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
Jens Axboe320ae512013-10-24 09:20:05 +01002253
2254 return 0;
2255}
2256subsys_initcall(blk_mq_init);