blob: 33c428530193548e38e0ec4ce3608e08be8b1d8d [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);
Jens Axboe201f2012015-02-10 13:31:34 -070036static void blk_mq_run_queues(struct request_queue *q);
Jens Axboe320ae512013-10-24 09:20:05 +010037
Jens Axboe320ae512013-10-24 09:20:05 +010038/*
39 * Check if any of the ctx's have pending work in this hardware queue
40 */
41static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
42{
43 unsigned int i;
44
Jens Axboe1429d7c2014-05-19 09:23:55 -060045 for (i = 0; i < hctx->ctx_map.map_size; i++)
46 if (hctx->ctx_map.map[i].word)
Jens Axboe320ae512013-10-24 09:20:05 +010047 return true;
48
49 return false;
50}
51
Jens Axboe1429d7c2014-05-19 09:23:55 -060052static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
53 struct blk_mq_ctx *ctx)
54{
55 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
56}
57
58#define CTX_TO_BIT(hctx, ctx) \
59 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
60
Jens Axboe320ae512013-10-24 09:20:05 +010061/*
62 * Mark this ctx as having pending work in this hardware queue
63 */
64static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
65 struct blk_mq_ctx *ctx)
66{
Jens Axboe1429d7c2014-05-19 09:23:55 -060067 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
68
69 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
70 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
71}
72
73static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
74 struct blk_mq_ctx *ctx)
75{
76 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
77
78 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
Jens Axboe320ae512013-10-24 09:20:05 +010079}
80
Jens Axboe320ae512013-10-24 09:20:05 +010081static int blk_mq_queue_enter(struct request_queue *q)
82{
Tejun Heoadd703f2014-07-01 10:34:38 -060083 while (true) {
84 int ret;
Jens Axboe320ae512013-10-24 09:20:05 +010085
Tejun Heoadd703f2014-07-01 10:34:38 -060086 if (percpu_ref_tryget_live(&q->mq_usage_counter))
87 return 0;
Keith Busch3b632cf2014-06-06 10:22:07 -060088
Tejun Heoadd703f2014-07-01 10:34:38 -060089 ret = wait_event_interruptible(q->mq_freeze_wq,
90 !q->mq_freeze_depth || blk_queue_dying(q));
91 if (blk_queue_dying(q))
92 return -ENODEV;
93 if (ret)
94 return ret;
95 }
Jens Axboe320ae512013-10-24 09:20:05 +010096}
97
98static void blk_mq_queue_exit(struct request_queue *q)
99{
Tejun Heoadd703f2014-07-01 10:34:38 -0600100 percpu_ref_put(&q->mq_usage_counter);
101}
102
103static void blk_mq_usage_counter_release(struct percpu_ref *ref)
104{
105 struct request_queue *q =
106 container_of(ref, struct request_queue, mq_usage_counter);
107
108 wake_up_all(&q->mq_freeze_wq);
Jens Axboe320ae512013-10-24 09:20:05 +0100109}
110
Keith Buschb4c6a022014-12-19 17:54:14 -0700111void blk_mq_freeze_queue_start(struct request_queue *q)
Ming Lei43a5e4e2013-12-26 21:31:35 +0800112{
Tejun Heocddd5d12014-08-16 08:02:24 -0400113 bool freeze;
114
Tejun Heo72d6f022014-07-01 10:33:02 -0600115 spin_lock_irq(q->queue_lock);
Tejun Heocddd5d12014-08-16 08:02:24 -0400116 freeze = !q->mq_freeze_depth++;
Tejun Heo72d6f022014-07-01 10:33:02 -0600117 spin_unlock_irq(q->queue_lock);
118
Tejun Heocddd5d12014-08-16 08:02:24 -0400119 if (freeze) {
Tejun Heo9eca8042014-09-24 13:07:33 -0400120 percpu_ref_kill(&q->mq_usage_counter);
Jens Axboe201f2012015-02-10 13:31:34 -0700121 blk_mq_run_queues(q);
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{
Tejun Heocddd5d12014-08-16 08:02:24 -0400144 bool wake;
Jens Axboe320ae512013-10-24 09:20:05 +0100145
146 spin_lock_irq(q->queue_lock);
Tejun Heo780db202014-07-01 10:31:13 -0600147 wake = !--q->mq_freeze_depth;
148 WARN_ON_ONCE(q->mq_freeze_depth < 0);
Jens Axboe320ae512013-10-24 09:20:05 +0100149 spin_unlock_irq(q->queue_lock);
Tejun Heoadd703f2014-07-01 10:34:38 -0600150 if (wake) {
151 percpu_ref_reinit(&q->mq_usage_counter);
Jens Axboe320ae512013-10-24 09:20:05 +0100152 wake_up_all(&q->mq_freeze_wq);
Tejun Heoadd703f2014-07-01 10:34:38 -0600153 }
Jens Axboe320ae512013-10-24 09:20:05 +0100154}
Keith Buschb4c6a022014-12-19 17:54:14 -0700155EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
Jens Axboe320ae512013-10-24 09:20:05 +0100156
Jens Axboeaed3ea92014-12-22 14:04:42 -0700157void blk_mq_wake_waiters(struct request_queue *q)
158{
159 struct blk_mq_hw_ctx *hctx;
160 unsigned int i;
161
162 queue_for_each_hw_ctx(q, hctx, i)
163 if (blk_mq_hw_queue_mapped(hctx))
164 blk_mq_tag_wakeup_all(hctx->tags, true);
Keith Busch3fd59402015-01-08 08:53:56 -0700165
166 /*
167 * If we are called because the queue has now been marked as
168 * dying, we need to ensure that processes currently waiting on
169 * the queue are notified as well.
170 */
171 wake_up_all(&q->mq_freeze_wq);
Jens Axboeaed3ea92014-12-22 14:04:42 -0700172}
173
Jens Axboe320ae512013-10-24 09:20:05 +0100174bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
175{
176 return blk_mq_has_free_tags(hctx->tags);
177}
178EXPORT_SYMBOL(blk_mq_can_queue);
179
Jens Axboe94eddfb2013-11-19 09:25:07 -0700180static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
181 struct request *rq, unsigned int rw_flags)
Jens Axboe320ae512013-10-24 09:20:05 +0100182{
Jens Axboe94eddfb2013-11-19 09:25:07 -0700183 if (blk_queue_io_stat(q))
184 rw_flags |= REQ_IO_STAT;
185
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200186 INIT_LIST_HEAD(&rq->queuelist);
187 /* csd/requeue_work/fifo_time is initialized before use */
188 rq->q = q;
Jens Axboe320ae512013-10-24 09:20:05 +0100189 rq->mq_ctx = ctx;
Jens Axboe0d2602c2014-05-13 15:10:52 -0600190 rq->cmd_flags |= rw_flags;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200191 /* do not touch atomic flags, it needs atomic ops against the timer */
192 rq->cpu = -1;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200193 INIT_HLIST_NODE(&rq->hash);
194 RB_CLEAR_NODE(&rq->rb_node);
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200195 rq->rq_disk = NULL;
196 rq->part = NULL;
Jens Axboe3ee32372014-06-09 09:36:53 -0600197 rq->start_time = jiffies;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200198#ifdef CONFIG_BLK_CGROUP
199 rq->rl = NULL;
Ming Lei0fec08b2014-01-03 10:00:08 -0700200 set_start_time_ns(rq);
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200201 rq->io_start_time_ns = 0;
202#endif
203 rq->nr_phys_segments = 0;
204#if defined(CONFIG_BLK_DEV_INTEGRITY)
205 rq->nr_integrity_segments = 0;
206#endif
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200207 rq->special = NULL;
208 /* tag was already set */
209 rq->errors = 0;
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200210
Tony Battersby6f4a16262014-08-22 15:53:39 -0400211 rq->cmd = rq->__cmd;
212
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200213 rq->extra_len = 0;
214 rq->sense_len = 0;
215 rq->resid_len = 0;
216 rq->sense = NULL;
217
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200218 INIT_LIST_HEAD(&rq->timeout_list);
Jens Axboef6be4fb2014-06-06 11:03:48 -0600219 rq->timeout = 0;
220
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200221 rq->end_io = NULL;
222 rq->end_io_data = NULL;
223 rq->next_rq = NULL;
224
Jens Axboe320ae512013-10-24 09:20:05 +0100225 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
226}
227
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200228static struct request *
Ming Leicb96a422014-06-01 00:43:37 +0800229__blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200230{
231 struct request *rq;
232 unsigned int tag;
233
Ming Leicb96a422014-06-01 00:43:37 +0800234 tag = blk_mq_get_tag(data);
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200235 if (tag != BLK_MQ_TAG_FAIL) {
Ming Leicb96a422014-06-01 00:43:37 +0800236 rq = data->hctx->tags->rqs[tag];
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200237
Ming Leicb96a422014-06-01 00:43:37 +0800238 if (blk_mq_tag_busy(data->hctx)) {
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200239 rq->cmd_flags = REQ_MQ_INFLIGHT;
Ming Leicb96a422014-06-01 00:43:37 +0800240 atomic_inc(&data->hctx->nr_active);
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200241 }
242
243 rq->tag = tag;
Ming Leicb96a422014-06-01 00:43:37 +0800244 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
Christoph Hellwig5dee8572014-05-27 20:59:47 +0200245 return rq;
246 }
247
248 return NULL;
249}
250
Christoph Hellwig4ce01dd2014-05-27 20:59:46 +0200251struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
252 bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100253{
Christoph Hellwigd8525642014-05-27 20:59:50 +0200254 struct blk_mq_ctx *ctx;
255 struct blk_mq_hw_ctx *hctx;
Jens Axboe320ae512013-10-24 09:20:05 +0100256 struct request *rq;
Ming Leicb96a422014-06-01 00:43:37 +0800257 struct blk_mq_alloc_data alloc_data;
Joe Lawrencea492f072014-08-28 08:15:21 -0600258 int ret;
Jens Axboe320ae512013-10-24 09:20:05 +0100259
Joe Lawrencea492f072014-08-28 08:15:21 -0600260 ret = blk_mq_queue_enter(q);
261 if (ret)
262 return ERR_PTR(ret);
Jens Axboe320ae512013-10-24 09:20:05 +0100263
Christoph Hellwigd8525642014-05-27 20:59:50 +0200264 ctx = blk_mq_get_ctx(q);
265 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Ming Leicb96a422014-06-01 00:43:37 +0800266 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
267 reserved, ctx, hctx);
Christoph Hellwigd8525642014-05-27 20:59:50 +0200268
Ming Leicb96a422014-06-01 00:43:37 +0800269 rq = __blk_mq_alloc_request(&alloc_data, rw);
Christoph Hellwigd8525642014-05-27 20:59:50 +0200270 if (!rq && (gfp & __GFP_WAIT)) {
271 __blk_mq_run_hw_queue(hctx);
272 blk_mq_put_ctx(ctx);
273
274 ctx = blk_mq_get_ctx(q);
275 hctx = q->mq_ops->map_queue(q, ctx->cpu);
Ming Leicb96a422014-06-01 00:43:37 +0800276 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
277 hctx);
278 rq = __blk_mq_alloc_request(&alloc_data, rw);
279 ctx = alloc_data.ctx;
Christoph Hellwigd8525642014-05-27 20:59:50 +0200280 }
281 blk_mq_put_ctx(ctx);
Keith Buschc76541a2014-12-19 17:54:13 -0700282 if (!rq) {
283 blk_mq_queue_exit(q);
Joe Lawrencea492f072014-08-28 08:15:21 -0600284 return ERR_PTR(-EWOULDBLOCK);
Keith Buschc76541a2014-12-19 17:54:13 -0700285 }
Jens Axboe320ae512013-10-24 09:20:05 +0100286 return rq;
287}
Jens Axboe4bb659b2014-05-09 09:36:49 -0600288EXPORT_SYMBOL(blk_mq_alloc_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100289
Jens Axboe320ae512013-10-24 09:20:05 +0100290static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
291 struct blk_mq_ctx *ctx, struct request *rq)
292{
293 const int tag = rq->tag;
294 struct request_queue *q = rq->q;
295
Jens Axboe0d2602c2014-05-13 15:10:52 -0600296 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
297 atomic_dec(&hctx->nr_active);
David Hildenbrand683d0e12014-09-18 11:04:31 +0200298 rq->cmd_flags = 0;
Jens Axboe0d2602c2014-05-13 15:10:52 -0600299
Christoph Hellwigaf76e552014-05-06 12:12:45 +0200300 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
Jens Axboe0d2602c2014-05-13 15:10:52 -0600301 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
Jens Axboe320ae512013-10-24 09:20:05 +0100302 blk_mq_queue_exit(q);
303}
304
Jens Axboe7c7f2f22014-11-17 10:41:57 -0700305void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100306{
307 struct blk_mq_ctx *ctx = rq->mq_ctx;
Jens Axboe7c7f2f22014-11-17 10:41:57 -0700308
309 ctx->rq_completed[rq_is_sync(rq)]++;
310 __blk_mq_free_request(hctx, ctx, rq);
311
312}
313EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
314
315void blk_mq_free_request(struct request *rq)
316{
Jens Axboe320ae512013-10-24 09:20:05 +0100317 struct blk_mq_hw_ctx *hctx;
318 struct request_queue *q = rq->q;
319
Jens Axboe7c7f2f22014-11-17 10:41:57 -0700320 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
321 blk_mq_free_hctx_request(hctx, rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100322}
Jens Axboe1a3b5952014-11-17 10:40:48 -0700323EXPORT_SYMBOL_GPL(blk_mq_free_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100324
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700325inline void __blk_mq_end_request(struct request *rq, int error)
Jens Axboe320ae512013-10-24 09:20:05 +0100326{
Ming Lei0d11e6a2013-12-05 10:50:39 -0700327 blk_account_io_done(rq);
328
Christoph Hellwig91b63632014-04-16 09:44:53 +0200329 if (rq->end_io) {
Jens Axboe320ae512013-10-24 09:20:05 +0100330 rq->end_io(rq, error);
Christoph Hellwig91b63632014-04-16 09:44:53 +0200331 } else {
332 if (unlikely(blk_bidi_rq(rq)))
333 blk_mq_free_request(rq->next_rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100334 blk_mq_free_request(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +0200335 }
Jens Axboe320ae512013-10-24 09:20:05 +0100336}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700337EXPORT_SYMBOL(__blk_mq_end_request);
Christoph Hellwig63151a42014-04-16 09:44:52 +0200338
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700339void blk_mq_end_request(struct request *rq, int error)
Christoph Hellwig63151a42014-04-16 09:44:52 +0200340{
341 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
342 BUG();
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700343 __blk_mq_end_request(rq, error);
Christoph Hellwig63151a42014-04-16 09:44:52 +0200344}
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700345EXPORT_SYMBOL(blk_mq_end_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100346
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800347static void __blk_mq_complete_request_remote(void *data)
Jens Axboe320ae512013-10-24 09:20:05 +0100348{
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800349 struct request *rq = data;
Jens Axboe320ae512013-10-24 09:20:05 +0100350
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800351 rq->q->softirq_done_fn(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100352}
353
Jens Axboeed851862014-05-30 21:20:50 -0600354static void blk_mq_ipi_complete_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100355{
356 struct blk_mq_ctx *ctx = rq->mq_ctx;
Christoph Hellwig38535202014-04-25 02:32:53 -0700357 bool shared = false;
Jens Axboe320ae512013-10-24 09:20:05 +0100358 int cpu;
359
Christoph Hellwig38535202014-04-25 02:32:53 -0700360 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800361 rq->q->softirq_done_fn(rq);
362 return;
363 }
Jens Axboe320ae512013-10-24 09:20:05 +0100364
365 cpu = get_cpu();
Christoph Hellwig38535202014-04-25 02:32:53 -0700366 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
367 shared = cpus_share_cache(cpu, ctx->cpu);
368
369 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800370 rq->csd.func = __blk_mq_complete_request_remote;
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800371 rq->csd.info = rq;
372 rq->csd.flags = 0;
Frederic Weisbeckerc46fff22014-02-24 16:40:02 +0100373 smp_call_function_single_async(ctx->cpu, &rq->csd);
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800374 } else {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800375 rq->q->softirq_done_fn(rq);
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800376 }
Jens Axboe320ae512013-10-24 09:20:05 +0100377 put_cpu();
378}
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800379
Jens Axboeed851862014-05-30 21:20:50 -0600380void __blk_mq_complete_request(struct request *rq)
381{
382 struct request_queue *q = rq->q;
383
384 if (!q->softirq_done_fn)
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700385 blk_mq_end_request(rq, rq->errors);
Jens Axboeed851862014-05-30 21:20:50 -0600386 else
387 blk_mq_ipi_complete_request(rq);
388}
389
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800390/**
391 * blk_mq_complete_request - end I/O on a request
392 * @rq: the request being processed
393 *
394 * Description:
395 * Ends all I/O on a request. It does not handle partial completions.
396 * The actual completion happens out-of-order, through a IPI handler.
397 **/
398void blk_mq_complete_request(struct request *rq)
399{
Jens Axboe95f09682014-05-27 17:46:48 -0600400 struct request_queue *q = rq->q;
401
402 if (unlikely(blk_should_fake_timeout(q)))
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800403 return;
Jens Axboeed851862014-05-30 21:20:50 -0600404 if (!blk_mark_rq_complete(rq))
405 __blk_mq_complete_request(rq);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800406}
407EXPORT_SYMBOL(blk_mq_complete_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100408
Keith Busch973c0192015-01-07 18:55:43 -0700409int blk_mq_request_started(struct request *rq)
410{
411 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
412}
413EXPORT_SYMBOL_GPL(blk_mq_request_started);
414
Christoph Hellwige2490072014-09-13 16:40:09 -0700415void blk_mq_start_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100416{
417 struct request_queue *q = rq->q;
418
419 trace_block_rq_issue(q, rq);
420
Christoph Hellwig742ee692014-04-14 10:30:06 +0200421 rq->resid_len = blk_rq_bytes(rq);
Christoph Hellwig91b63632014-04-16 09:44:53 +0200422 if (unlikely(blk_bidi_rq(rq)))
423 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
Christoph Hellwig742ee692014-04-14 10:30:06 +0200424
Ming Lei2b8393b2014-06-10 00:16:41 +0800425 blk_add_timer(rq);
Jens Axboe87ee7b12014-04-24 08:51:47 -0600426
427 /*
Jens Axboe538b7532014-09-16 10:37:37 -0600428 * Ensure that ->deadline is visible before set the started
429 * flag and clear the completed flag.
430 */
431 smp_mb__before_atomic();
432
433 /*
Jens Axboe87ee7b12014-04-24 08:51:47 -0600434 * Mark us as started and clear complete. Complete might have been
435 * set if requeue raced with timeout, which then marked it as
436 * complete. So be sure to clear complete again when we start
437 * the request, otherwise we'll ignore the completion event.
438 */
Jens Axboe4b570522014-05-29 11:00:11 -0600439 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
440 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
441 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
442 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800443
444 if (q->dma_drain_size && blk_rq_bytes(rq)) {
445 /*
446 * Make sure space for the drain appears. We know we can do
447 * this because max_hw_segments has been adjusted to be one
448 * fewer than the device can handle.
449 */
450 rq->nr_phys_segments++;
451 }
Jens Axboe320ae512013-10-24 09:20:05 +0100452}
Christoph Hellwige2490072014-09-13 16:40:09 -0700453EXPORT_SYMBOL(blk_mq_start_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100454
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200455static void __blk_mq_requeue_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100456{
457 struct request_queue *q = rq->q;
458
459 trace_block_rq_requeue(q, rq);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800460
Christoph Hellwige2490072014-09-13 16:40:09 -0700461 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
462 if (q->dma_drain_size && blk_rq_bytes(rq))
463 rq->nr_phys_segments--;
464 }
Jens Axboe320ae512013-10-24 09:20:05 +0100465}
466
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200467void blk_mq_requeue_request(struct request *rq)
468{
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200469 __blk_mq_requeue_request(rq);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200470
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200471 BUG_ON(blk_queued_rq(rq));
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600472 blk_mq_add_to_requeue_list(rq, true);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200473}
474EXPORT_SYMBOL(blk_mq_requeue_request);
475
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600476static void blk_mq_requeue_work(struct work_struct *work)
477{
478 struct request_queue *q =
479 container_of(work, struct request_queue, requeue_work);
480 LIST_HEAD(rq_list);
481 struct request *rq, *next;
482 unsigned long flags;
483
484 spin_lock_irqsave(&q->requeue_lock, flags);
485 list_splice_init(&q->requeue_list, &rq_list);
486 spin_unlock_irqrestore(&q->requeue_lock, flags);
487
488 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
489 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
490 continue;
491
492 rq->cmd_flags &= ~REQ_SOFTBARRIER;
493 list_del_init(&rq->queuelist);
494 blk_mq_insert_request(rq, true, false, false);
495 }
496
497 while (!list_empty(&rq_list)) {
498 rq = list_entry(rq_list.next, struct request, queuelist);
499 list_del_init(&rq->queuelist);
500 blk_mq_insert_request(rq, false, false, false);
501 }
502
Jens Axboe8b957412014-09-19 13:10:29 -0600503 /*
504 * Use the start variant of queue running here, so that running
505 * the requeue work will kick stopped queues.
506 */
507 blk_mq_start_hw_queues(q);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600508}
509
510void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
511{
512 struct request_queue *q = rq->q;
513 unsigned long flags;
514
515 /*
516 * We abuse this flag that is otherwise used by the I/O scheduler to
517 * request head insertation from the workqueue.
518 */
519 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
520
521 spin_lock_irqsave(&q->requeue_lock, flags);
522 if (at_head) {
523 rq->cmd_flags |= REQ_SOFTBARRIER;
524 list_add(&rq->queuelist, &q->requeue_list);
525 } else {
526 list_add_tail(&rq->queuelist, &q->requeue_list);
527 }
528 spin_unlock_irqrestore(&q->requeue_lock, flags);
529}
530EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
531
Keith Buschc68ed592015-01-07 18:55:44 -0700532void blk_mq_cancel_requeue_work(struct request_queue *q)
533{
534 cancel_work_sync(&q->requeue_work);
535}
536EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
537
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600538void blk_mq_kick_requeue_list(struct request_queue *q)
539{
540 kblockd_schedule_work(&q->requeue_work);
541}
542EXPORT_SYMBOL(blk_mq_kick_requeue_list);
543
Jens Axboe1885b242015-01-07 18:55:45 -0700544void blk_mq_abort_requeue_list(struct request_queue *q)
545{
546 unsigned long flags;
547 LIST_HEAD(rq_list);
548
549 spin_lock_irqsave(&q->requeue_lock, flags);
550 list_splice_init(&q->requeue_list, &rq_list);
551 spin_unlock_irqrestore(&q->requeue_lock, flags);
552
553 while (!list_empty(&rq_list)) {
554 struct request *rq;
555
556 rq = list_first_entry(&rq_list, struct request, queuelist);
557 list_del_init(&rq->queuelist);
558 rq->errors = -EIO;
559 blk_mq_end_request(rq, rq->errors);
560 }
561}
562EXPORT_SYMBOL(blk_mq_abort_requeue_list);
563
Ming Lei7c94e1c2014-09-25 23:23:43 +0800564static inline bool is_flush_request(struct request *rq,
565 struct blk_flush_queue *fq, unsigned int tag)
Christoph Hellwig24d2f902014-04-15 14:14:00 -0600566{
Jens Axboe0e62f512014-06-04 10:23:49 -0600567 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
Ming Lei7c94e1c2014-09-25 23:23:43 +0800568 fq->flush_rq->tag == tag);
Jens Axboe0e62f512014-06-04 10:23:49 -0600569}
Shaohua Li22302372014-05-30 08:06:42 -0600570
Jens Axboe0e62f512014-06-04 10:23:49 -0600571struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
572{
573 struct request *rq = tags->rqs[tag];
Ming Leie97c2932014-09-25 23:23:46 +0800574 /* mq_ctx of flush rq is always cloned from the corresponding req */
575 struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
Shaohua Li22302372014-05-30 08:06:42 -0600576
Ming Lei7c94e1c2014-09-25 23:23:43 +0800577 if (!is_flush_request(rq, fq, tag))
Jens Axboe0e62f512014-06-04 10:23:49 -0600578 return rq;
579
Ming Lei7c94e1c2014-09-25 23:23:43 +0800580 return fq->flush_rq;
Christoph Hellwig24d2f902014-04-15 14:14:00 -0600581}
582EXPORT_SYMBOL(blk_mq_tag_to_rq);
583
Jens Axboe320ae512013-10-24 09:20:05 +0100584struct blk_mq_timeout_data {
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700585 unsigned long next;
586 unsigned int next_set;
Jens Axboe320ae512013-10-24 09:20:05 +0100587};
588
Christoph Hellwig90415832014-09-22 10:21:48 -0600589void blk_mq_rq_timed_out(struct request *req, bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100590{
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700591 struct blk_mq_ops *ops = req->q->mq_ops;
592 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
Jens Axboe87ee7b12014-04-24 08:51:47 -0600593
594 /*
595 * We know that complete is set at this point. If STARTED isn't set
596 * anymore, then the request isn't active and the "timeout" should
597 * just be ignored. This can happen due to the bitflag ordering.
598 * Timeout first checks if STARTED is set, and if it is, assumes
599 * the request is active. But if we race with completion, then
600 * we both flags will get cleared. So check here again, and ignore
601 * a timeout event with a request that isn't active.
602 */
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700603 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
604 return;
Jens Axboe87ee7b12014-04-24 08:51:47 -0600605
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700606 if (ops->timeout)
Christoph Hellwig0152fb62014-09-13 16:40:13 -0700607 ret = ops->timeout(req, reserved);
Jens Axboe87ee7b12014-04-24 08:51:47 -0600608
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700609 switch (ret) {
610 case BLK_EH_HANDLED:
611 __blk_mq_complete_request(req);
612 break;
613 case BLK_EH_RESET_TIMER:
614 blk_add_timer(req);
615 blk_clear_rq_complete(req);
616 break;
617 case BLK_EH_NOT_HANDLED:
618 break;
619 default:
620 printk(KERN_ERR "block: bad eh return: %d\n", ret);
621 break;
622 }
Jens Axboe87ee7b12014-04-24 08:51:47 -0600623}
Keith Busch5b3f25f2015-01-07 18:55:46 -0700624
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700625static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
626 struct request *rq, void *priv, bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100627{
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700628 struct blk_mq_timeout_data *data = priv;
629
Keith Buscheb130db2015-01-08 08:59:53 -0700630 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
631 /*
632 * If a request wasn't started before the queue was
633 * marked dying, kill it here or it'll go unnoticed.
634 */
635 if (unlikely(blk_queue_dying(rq->q))) {
636 rq->errors = -EIO;
637 blk_mq_complete_request(rq);
638 }
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700639 return;
Keith Buscheb130db2015-01-08 08:59:53 -0700640 }
Keith Busch5b3f25f2015-01-07 18:55:46 -0700641 if (rq->cmd_flags & REQ_NO_TIMEOUT)
642 return;
Jens Axboe320ae512013-10-24 09:20:05 +0100643
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700644 if (time_after_eq(jiffies, rq->deadline)) {
645 if (!blk_mark_rq_complete(rq))
Christoph Hellwig0152fb62014-09-13 16:40:13 -0700646 blk_mq_rq_timed_out(rq, reserved);
Christoph Hellwig46f92d42014-09-13 16:40:12 -0700647 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
648 data->next = rq->deadline;
649 data->next_set = 1;
650 }
Jens Axboe320ae512013-10-24 09:20:05 +0100651}
652
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700653static void blk_mq_rq_timer(unsigned long priv)
Jens Axboe320ae512013-10-24 09:20:05 +0100654{
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700655 struct request_queue *q = (struct request_queue *)priv;
656 struct blk_mq_timeout_data data = {
657 .next = 0,
658 .next_set = 0,
659 };
Jens Axboe320ae512013-10-24 09:20:05 +0100660 struct blk_mq_hw_ctx *hctx;
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700661 int i;
Jens Axboe320ae512013-10-24 09:20:05 +0100662
Jens Axboe484b4062014-05-21 14:01:15 -0600663 queue_for_each_hw_ctx(q, hctx, i) {
664 /*
665 * If not software queues are currently mapped to this
666 * hardware queue, there's nothing to check
667 */
Ming Lei19c66e52014-12-03 19:38:04 +0800668 if (!blk_mq_hw_queue_mapped(hctx))
Jens Axboe484b4062014-05-21 14:01:15 -0600669 continue;
670
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700671 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
Jens Axboe484b4062014-05-21 14:01:15 -0600672 }
Jens Axboe320ae512013-10-24 09:20:05 +0100673
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700674 if (data.next_set) {
675 data.next = blk_rq_timeout(round_jiffies_up(data.next));
676 mod_timer(&q->timeout, data.next);
Jens Axboe0d2602c2014-05-13 15:10:52 -0600677 } else {
678 queue_for_each_hw_ctx(q, hctx, i)
679 blk_mq_tag_idle(hctx);
680 }
Jens Axboe320ae512013-10-24 09:20:05 +0100681}
682
683/*
684 * Reverse check our software queue for entries that we could potentially
685 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
686 * too much time checking for merges.
687 */
688static bool blk_mq_attempt_merge(struct request_queue *q,
689 struct blk_mq_ctx *ctx, struct bio *bio)
690{
691 struct request *rq;
692 int checked = 8;
693
694 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
695 int el_ret;
696
697 if (!checked--)
698 break;
699
700 if (!blk_rq_merge_ok(rq, bio))
701 continue;
702
703 el_ret = blk_try_merge(rq, bio);
704 if (el_ret == ELEVATOR_BACK_MERGE) {
705 if (bio_attempt_back_merge(q, rq, bio)) {
706 ctx->rq_merged++;
707 return true;
708 }
709 break;
710 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
711 if (bio_attempt_front_merge(q, rq, bio)) {
712 ctx->rq_merged++;
713 return true;
714 }
715 break;
716 }
717 }
718
719 return false;
720}
721
Jens Axboe320ae512013-10-24 09:20:05 +0100722/*
Jens Axboe1429d7c2014-05-19 09:23:55 -0600723 * Process software queues that have been marked busy, splicing them
724 * to the for-dispatch
725 */
726static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
727{
728 struct blk_mq_ctx *ctx;
729 int i;
730
731 for (i = 0; i < hctx->ctx_map.map_size; i++) {
732 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
733 unsigned int off, bit;
734
735 if (!bm->word)
736 continue;
737
738 bit = 0;
739 off = i * hctx->ctx_map.bits_per_word;
740 do {
741 bit = find_next_bit(&bm->word, bm->depth, bit);
742 if (bit >= bm->depth)
743 break;
744
745 ctx = hctx->ctxs[bit + off];
746 clear_bit(bit, &bm->word);
747 spin_lock(&ctx->lock);
748 list_splice_tail_init(&ctx->rq_list, list);
749 spin_unlock(&ctx->lock);
750
751 bit++;
752 } while (1);
753 }
754}
755
756/*
Jens Axboe320ae512013-10-24 09:20:05 +0100757 * Run this hardware queue, pulling any software queues mapped to it in.
758 * Note that this function currently has various problems around ordering
759 * of IO. In particular, we'd like FIFO behaviour on handling existing
760 * items on the hctx->dispatch list. Ignore that for now.
761 */
762static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
763{
764 struct request_queue *q = hctx->queue;
Jens Axboe320ae512013-10-24 09:20:05 +0100765 struct request *rq;
766 LIST_HEAD(rq_list);
Jens Axboe74c45052014-10-29 11:14:52 -0600767 LIST_HEAD(driver_list);
768 struct list_head *dptr;
Jens Axboe1429d7c2014-05-19 09:23:55 -0600769 int queued;
Jens Axboe320ae512013-10-24 09:20:05 +0100770
Jens Axboefd1270d2014-04-16 09:23:48 -0600771 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
Jens Axboee4043dc2014-04-09 10:18:23 -0600772
Jens Axboe5d12f902014-03-19 15:25:02 -0600773 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
Jens Axboe320ae512013-10-24 09:20:05 +0100774 return;
775
776 hctx->run++;
777
778 /*
779 * Touch any software queue that has pending entries.
780 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600781 flush_busy_ctxs(hctx, &rq_list);
Jens Axboe320ae512013-10-24 09:20:05 +0100782
783 /*
784 * If we have previous entries on our dispatch list, grab them
785 * and stuff them at the front for more fair dispatch.
786 */
787 if (!list_empty_careful(&hctx->dispatch)) {
788 spin_lock(&hctx->lock);
789 if (!list_empty(&hctx->dispatch))
790 list_splice_init(&hctx->dispatch, &rq_list);
791 spin_unlock(&hctx->lock);
792 }
793
794 /*
Jens Axboe74c45052014-10-29 11:14:52 -0600795 * Start off with dptr being NULL, so we start the first request
796 * immediately, even if we have more pending.
797 */
798 dptr = NULL;
799
800 /*
Jens Axboe320ae512013-10-24 09:20:05 +0100801 * Now process all the entries, sending them to the driver.
802 */
Jens Axboe1429d7c2014-05-19 09:23:55 -0600803 queued = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100804 while (!list_empty(&rq_list)) {
Jens Axboe74c45052014-10-29 11:14:52 -0600805 struct blk_mq_queue_data bd;
Jens Axboe320ae512013-10-24 09:20:05 +0100806 int ret;
807
808 rq = list_first_entry(&rq_list, struct request, queuelist);
809 list_del_init(&rq->queuelist);
Jens Axboe320ae512013-10-24 09:20:05 +0100810
Jens Axboe74c45052014-10-29 11:14:52 -0600811 bd.rq = rq;
812 bd.list = dptr;
813 bd.last = list_empty(&rq_list);
814
815 ret = q->mq_ops->queue_rq(hctx, &bd);
Jens Axboe320ae512013-10-24 09:20:05 +0100816 switch (ret) {
817 case BLK_MQ_RQ_QUEUE_OK:
818 queued++;
819 continue;
820 case BLK_MQ_RQ_QUEUE_BUSY:
Jens Axboe320ae512013-10-24 09:20:05 +0100821 list_add(&rq->queuelist, &rq_list);
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200822 __blk_mq_requeue_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100823 break;
824 default:
825 pr_err("blk-mq: bad return on queue: %d\n", ret);
Jens Axboe320ae512013-10-24 09:20:05 +0100826 case BLK_MQ_RQ_QUEUE_ERROR:
Christoph Hellwig1e93b8c2014-02-11 08:27:13 -0800827 rq->errors = -EIO;
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700828 blk_mq_end_request(rq, rq->errors);
Jens Axboe320ae512013-10-24 09:20:05 +0100829 break;
830 }
831
832 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
833 break;
Jens Axboe74c45052014-10-29 11:14:52 -0600834
835 /*
836 * We've done the first request. If we have more than 1
837 * left in the list, set dptr to defer issue.
838 */
839 if (!dptr && rq_list.next != rq_list.prev)
840 dptr = &driver_list;
Jens Axboe320ae512013-10-24 09:20:05 +0100841 }
842
843 if (!queued)
844 hctx->dispatched[0]++;
845 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
846 hctx->dispatched[ilog2(queued) + 1]++;
847
848 /*
849 * Any items that need requeuing? Stuff them into hctx->dispatch,
850 * that is where we will continue on next queue run.
851 */
852 if (!list_empty(&rq_list)) {
853 spin_lock(&hctx->lock);
854 list_splice(&rq_list, &hctx->dispatch);
855 spin_unlock(&hctx->lock);
856 }
857}
858
Jens Axboe506e9312014-05-07 10:26:44 -0600859/*
860 * It'd be great if the workqueue API had a way to pass
861 * in a mask and had some smarts for more clever placement.
862 * For now we just round-robin here, switching for every
863 * BLK_MQ_CPU_WORK_BATCH queued items.
864 */
865static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
866{
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100867 if (hctx->queue->nr_hw_queues == 1)
868 return WORK_CPU_UNBOUND;
Jens Axboe506e9312014-05-07 10:26:44 -0600869
870 if (--hctx->next_cpu_batch <= 0) {
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100871 int cpu = hctx->next_cpu, next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600872
873 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
874 if (next_cpu >= nr_cpu_ids)
875 next_cpu = cpumask_first(hctx->cpumask);
876
877 hctx->next_cpu = next_cpu;
878 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100879
880 return cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600881 }
882
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100883 return hctx->next_cpu;
Jens Axboe506e9312014-05-07 10:26:44 -0600884}
885
Jens Axboe320ae512013-10-24 09:20:05 +0100886void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
887{
Ming Lei19c66e52014-12-03 19:38:04 +0800888 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
889 !blk_mq_hw_queue_mapped(hctx)))
Jens Axboe320ae512013-10-24 09:20:05 +0100890 return;
891
Paolo Bonzini398205b2014-11-07 23:03:59 +0100892 if (!async) {
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100893 int cpu = get_cpu();
894 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
Paolo Bonzini398205b2014-11-07 23:03:59 +0100895 __blk_mq_run_hw_queue(hctx);
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100896 put_cpu();
Paolo Bonzini398205b2014-11-07 23:03:59 +0100897 return;
898 }
Jens Axboee4043dc2014-04-09 10:18:23 -0600899
Paolo Bonzini2a90d4a2014-11-07 23:04:00 +0100900 put_cpu();
Jens Axboee4043dc2014-04-09 10:18:23 -0600901 }
Paolo Bonzini398205b2014-11-07 23:03:59 +0100902
Christoph Hellwigb657d7e2014-11-24 09:27:23 +0100903 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
904 &hctx->run_work, 0);
Jens Axboe320ae512013-10-24 09:20:05 +0100905}
906
Jens Axboe201f2012015-02-10 13:31:34 -0700907static void blk_mq_run_queues(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +0100908{
909 struct blk_mq_hw_ctx *hctx;
910 int i;
911
912 queue_for_each_hw_ctx(q, hctx, i) {
913 if ((!blk_mq_hctx_has_pending(hctx) &&
914 list_empty_careful(&hctx->dispatch)) ||
Jens Axboe5d12f902014-03-19 15:25:02 -0600915 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
Jens Axboe320ae512013-10-24 09:20:05 +0100916 continue;
917
Jens Axboe201f2012015-02-10 13:31:34 -0700918 blk_mq_run_hw_queue(hctx, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100919 }
920}
Jens Axboe320ae512013-10-24 09:20:05 +0100921
922void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
923{
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600924 cancel_delayed_work(&hctx->run_work);
925 cancel_delayed_work(&hctx->delay_work);
Jens Axboe320ae512013-10-24 09:20:05 +0100926 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
927}
928EXPORT_SYMBOL(blk_mq_stop_hw_queue);
929
Christoph Hellwig280d45f2013-10-25 14:45:58 +0100930void blk_mq_stop_hw_queues(struct request_queue *q)
931{
932 struct blk_mq_hw_ctx *hctx;
933 int i;
934
935 queue_for_each_hw_ctx(q, hctx, i)
936 blk_mq_stop_hw_queue(hctx);
937}
938EXPORT_SYMBOL(blk_mq_stop_hw_queues);
939
Jens Axboe320ae512013-10-24 09:20:05 +0100940void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
941{
942 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
Jens Axboee4043dc2014-04-09 10:18:23 -0600943
Jens Axboe0ffbce82014-06-25 08:22:34 -0600944 blk_mq_run_hw_queue(hctx, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100945}
946EXPORT_SYMBOL(blk_mq_start_hw_queue);
947
Christoph Hellwig2f268552014-04-16 09:44:56 +0200948void blk_mq_start_hw_queues(struct request_queue *q)
949{
950 struct blk_mq_hw_ctx *hctx;
951 int i;
952
953 queue_for_each_hw_ctx(q, hctx, i)
954 blk_mq_start_hw_queue(hctx);
955}
956EXPORT_SYMBOL(blk_mq_start_hw_queues);
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,
Shaohua Li24391c02015-01-23 14:18:00 -07001426 set->numa_node,
1427 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001428 if (!tags)
1429 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001430
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001431 INIT_LIST_HEAD(&tags->page_list);
1432
Jens Axboea5164402014-09-10 09:02:03 -06001433 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1434 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1435 set->numa_node);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001436 if (!tags->rqs) {
1437 blk_mq_free_tags(tags);
1438 return NULL;
1439 }
Jens Axboe320ae512013-10-24 09:20:05 +01001440
1441 /*
1442 * rq_size is the size of the request plus driver payload, rounded
1443 * to the cacheline size
1444 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001445 rq_size = round_up(sizeof(struct request) + set->cmd_size,
Jens Axboe320ae512013-10-24 09:20:05 +01001446 cache_line_size());
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001447 left = rq_size * set->queue_depth;
Jens Axboe320ae512013-10-24 09:20:05 +01001448
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001449 for (i = 0; i < set->queue_depth; ) {
Jens Axboe320ae512013-10-24 09:20:05 +01001450 int this_order = max_order;
1451 struct page *page;
1452 int to_do;
1453 void *p;
1454
1455 while (left < order_to_size(this_order - 1) && this_order)
1456 this_order--;
1457
1458 do {
Jens Axboea5164402014-09-10 09:02:03 -06001459 page = alloc_pages_node(set->numa_node,
Linus Torvaldsac211172015-04-09 14:12:22 -07001460 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
Jens Axboea5164402014-09-10 09:02:03 -06001461 this_order);
Jens Axboe320ae512013-10-24 09:20:05 +01001462 if (page)
1463 break;
1464 if (!this_order--)
1465 break;
1466 if (order_to_size(this_order) < rq_size)
1467 break;
1468 } while (1);
1469
1470 if (!page)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001471 goto fail;
Jens Axboe320ae512013-10-24 09:20:05 +01001472
1473 page->private = this_order;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001474 list_add_tail(&page->lru, &tags->page_list);
Jens Axboe320ae512013-10-24 09:20:05 +01001475
1476 p = page_address(page);
1477 entries_per_page = order_to_size(this_order) / rq_size;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001478 to_do = min(entries_per_page, set->queue_depth - i);
Jens Axboe320ae512013-10-24 09:20:05 +01001479 left -= to_do * rq_size;
1480 for (j = 0; j < to_do; j++) {
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001481 tags->rqs[i] = p;
1482 if (set->ops->init_request) {
1483 if (set->ops->init_request(set->driver_data,
1484 tags->rqs[i], hctx_idx, i,
Jens Axboea5164402014-09-10 09:02:03 -06001485 set->numa_node)) {
1486 tags->rqs[i] = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001487 goto fail;
Jens Axboea5164402014-09-10 09:02:03 -06001488 }
Christoph Hellwige9b267d2014-04-15 13:59:10 -06001489 }
1490
Jens Axboe320ae512013-10-24 09:20:05 +01001491 p += rq_size;
1492 i++;
1493 }
1494 }
1495
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001496 return tags;
Jens Axboe320ae512013-10-24 09:20:05 +01001497
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001498fail:
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001499 blk_mq_free_rq_map(set, tags, hctx_idx);
1500 return NULL;
Jens Axboe320ae512013-10-24 09:20:05 +01001501}
1502
Jens Axboe1429d7c2014-05-19 09:23:55 -06001503static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1504{
1505 kfree(bitmap->map);
1506}
1507
1508static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1509{
1510 unsigned int bpw = 8, total, num_maps, i;
1511
1512 bitmap->bits_per_word = bpw;
1513
1514 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1515 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1516 GFP_KERNEL, node);
1517 if (!bitmap->map)
1518 return -ENOMEM;
1519
1520 bitmap->map_size = num_maps;
1521
1522 total = nr_cpu_ids;
1523 for (i = 0; i < num_maps; i++) {
1524 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1525 total -= bitmap->map[i].depth;
1526 }
1527
1528 return 0;
1529}
1530
Jens Axboe484b4062014-05-21 14:01:15 -06001531static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1532{
1533 struct request_queue *q = hctx->queue;
1534 struct blk_mq_ctx *ctx;
1535 LIST_HEAD(tmp);
1536
1537 /*
1538 * Move ctx entries to new CPU, if this one is going away.
1539 */
1540 ctx = __blk_mq_get_ctx(q, cpu);
1541
1542 spin_lock(&ctx->lock);
1543 if (!list_empty(&ctx->rq_list)) {
1544 list_splice_init(&ctx->rq_list, &tmp);
1545 blk_mq_hctx_clear_pending(hctx, ctx);
1546 }
1547 spin_unlock(&ctx->lock);
1548
1549 if (list_empty(&tmp))
1550 return NOTIFY_OK;
1551
1552 ctx = blk_mq_get_ctx(q);
1553 spin_lock(&ctx->lock);
1554
1555 while (!list_empty(&tmp)) {
1556 struct request *rq;
1557
1558 rq = list_first_entry(&tmp, struct request, queuelist);
1559 rq->mq_ctx = ctx;
1560 list_move_tail(&rq->queuelist, &ctx->rq_list);
1561 }
1562
1563 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1564 blk_mq_hctx_mark_pending(hctx, ctx);
1565
1566 spin_unlock(&ctx->lock);
1567
1568 blk_mq_run_hw_queue(hctx, true);
1569 blk_mq_put_ctx(ctx);
1570 return NOTIFY_OK;
1571}
1572
1573static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1574{
1575 struct request_queue *q = hctx->queue;
1576 struct blk_mq_tag_set *set = q->tag_set;
1577
1578 if (set->tags[hctx->queue_num])
1579 return NOTIFY_OK;
1580
1581 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1582 if (!set->tags[hctx->queue_num])
1583 return NOTIFY_STOP;
1584
1585 hctx->tags = set->tags[hctx->queue_num];
1586 return NOTIFY_OK;
1587}
1588
1589static int blk_mq_hctx_notify(void *data, unsigned long action,
1590 unsigned int cpu)
1591{
1592 struct blk_mq_hw_ctx *hctx = data;
1593
1594 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1595 return blk_mq_hctx_cpu_offline(hctx, cpu);
1596 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1597 return blk_mq_hctx_cpu_online(hctx, cpu);
1598
1599 return NOTIFY_OK;
1600}
1601
Ming Lei08e98fc2014-09-25 23:23:38 +08001602static void blk_mq_exit_hctx(struct request_queue *q,
1603 struct blk_mq_tag_set *set,
1604 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1605{
Ming Leif70ced02014-09-25 23:23:47 +08001606 unsigned flush_start_tag = set->queue_depth;
1607
Ming Lei08e98fc2014-09-25 23:23:38 +08001608 blk_mq_tag_idle(hctx);
1609
Ming Leif70ced02014-09-25 23:23:47 +08001610 if (set->ops->exit_request)
1611 set->ops->exit_request(set->driver_data,
1612 hctx->fq->flush_rq, hctx_idx,
1613 flush_start_tag + hctx_idx);
1614
Ming Lei08e98fc2014-09-25 23:23:38 +08001615 if (set->ops->exit_hctx)
1616 set->ops->exit_hctx(hctx, hctx_idx);
1617
1618 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
Ming Leif70ced02014-09-25 23:23:47 +08001619 blk_free_flush_queue(hctx->fq);
Ming Lei08e98fc2014-09-25 23:23:38 +08001620 kfree(hctx->ctxs);
1621 blk_mq_free_bitmap(&hctx->ctx_map);
1622}
1623
Ming Lei624dbe42014-05-27 23:35:13 +08001624static void blk_mq_exit_hw_queues(struct request_queue *q,
1625 struct blk_mq_tag_set *set, int nr_queue)
1626{
1627 struct blk_mq_hw_ctx *hctx;
1628 unsigned int i;
1629
1630 queue_for_each_hw_ctx(q, hctx, i) {
1631 if (i == nr_queue)
1632 break;
Ming Lei08e98fc2014-09-25 23:23:38 +08001633 blk_mq_exit_hctx(q, set, hctx, i);
Ming Lei624dbe42014-05-27 23:35:13 +08001634 }
Ming Lei624dbe42014-05-27 23:35:13 +08001635}
1636
1637static void blk_mq_free_hw_queues(struct request_queue *q,
1638 struct blk_mq_tag_set *set)
1639{
1640 struct blk_mq_hw_ctx *hctx;
1641 unsigned int i;
1642
Ming Leie09aae72015-01-29 20:17:27 +08001643 queue_for_each_hw_ctx(q, hctx, i)
Ming Lei624dbe42014-05-27 23:35:13 +08001644 free_cpumask_var(hctx->cpumask);
Ming Lei624dbe42014-05-27 23:35:13 +08001645}
1646
Ming Lei08e98fc2014-09-25 23:23:38 +08001647static int blk_mq_init_hctx(struct request_queue *q,
1648 struct blk_mq_tag_set *set,
1649 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1650{
1651 int node;
Ming Leif70ced02014-09-25 23:23:47 +08001652 unsigned flush_start_tag = set->queue_depth;
Ming Lei08e98fc2014-09-25 23:23:38 +08001653
1654 node = hctx->numa_node;
1655 if (node == NUMA_NO_NODE)
1656 node = hctx->numa_node = set->numa_node;
1657
1658 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1659 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1660 spin_lock_init(&hctx->lock);
1661 INIT_LIST_HEAD(&hctx->dispatch);
1662 hctx->queue = q;
1663 hctx->queue_num = hctx_idx;
1664 hctx->flags = set->flags;
Ming Lei08e98fc2014-09-25 23:23:38 +08001665
1666 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1667 blk_mq_hctx_notify, hctx);
1668 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1669
1670 hctx->tags = set->tags[hctx_idx];
1671
1672 /*
1673 * Allocate space for all possible cpus to avoid allocation at
1674 * runtime
1675 */
1676 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1677 GFP_KERNEL, node);
1678 if (!hctx->ctxs)
1679 goto unregister_cpu_notifier;
1680
1681 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1682 goto free_ctxs;
1683
1684 hctx->nr_ctx = 0;
1685
1686 if (set->ops->init_hctx &&
1687 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1688 goto free_bitmap;
1689
Ming Leif70ced02014-09-25 23:23:47 +08001690 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1691 if (!hctx->fq)
1692 goto exit_hctx;
1693
1694 if (set->ops->init_request &&
1695 set->ops->init_request(set->driver_data,
1696 hctx->fq->flush_rq, hctx_idx,
1697 flush_start_tag + hctx_idx, node))
1698 goto free_fq;
1699
Ming Lei08e98fc2014-09-25 23:23:38 +08001700 return 0;
1701
Ming Leif70ced02014-09-25 23:23:47 +08001702 free_fq:
1703 kfree(hctx->fq);
1704 exit_hctx:
1705 if (set->ops->exit_hctx)
1706 set->ops->exit_hctx(hctx, hctx_idx);
Ming Lei08e98fc2014-09-25 23:23:38 +08001707 free_bitmap:
1708 blk_mq_free_bitmap(&hctx->ctx_map);
1709 free_ctxs:
1710 kfree(hctx->ctxs);
1711 unregister_cpu_notifier:
1712 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1713
1714 return -1;
1715}
1716
Jens Axboe320ae512013-10-24 09:20:05 +01001717static int blk_mq_init_hw_queues(struct request_queue *q,
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001718 struct blk_mq_tag_set *set)
Jens Axboe320ae512013-10-24 09:20:05 +01001719{
1720 struct blk_mq_hw_ctx *hctx;
Ming Lei624dbe42014-05-27 23:35:13 +08001721 unsigned int i;
Jens Axboe320ae512013-10-24 09:20:05 +01001722
1723 /*
1724 * Initialize hardware queues
1725 */
1726 queue_for_each_hw_ctx(q, hctx, i) {
Ming Lei08e98fc2014-09-25 23:23:38 +08001727 if (blk_mq_init_hctx(q, set, hctx, i))
Jens Axboe320ae512013-10-24 09:20:05 +01001728 break;
1729 }
1730
1731 if (i == q->nr_hw_queues)
1732 return 0;
1733
1734 /*
1735 * Init failed
1736 */
Ming Lei624dbe42014-05-27 23:35:13 +08001737 blk_mq_exit_hw_queues(q, set, i);
Jens Axboe320ae512013-10-24 09:20:05 +01001738
1739 return 1;
1740}
1741
1742static void blk_mq_init_cpu_queues(struct request_queue *q,
1743 unsigned int nr_hw_queues)
1744{
1745 unsigned int i;
1746
1747 for_each_possible_cpu(i) {
1748 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1749 struct blk_mq_hw_ctx *hctx;
1750
1751 memset(__ctx, 0, sizeof(*__ctx));
1752 __ctx->cpu = i;
1753 spin_lock_init(&__ctx->lock);
1754 INIT_LIST_HEAD(&__ctx->rq_list);
1755 __ctx->queue = q;
1756
1757 /* If the cpu isn't online, the cpu is mapped to first hctx */
Jens Axboe320ae512013-10-24 09:20:05 +01001758 if (!cpu_online(i))
1759 continue;
1760
Jens Axboee4043dc2014-04-09 10:18:23 -06001761 hctx = q->mq_ops->map_queue(q, i);
1762 cpumask_set_cpu(i, hctx->cpumask);
1763 hctx->nr_ctx++;
1764
Jens Axboe320ae512013-10-24 09:20:05 +01001765 /*
1766 * Set local node, IFF we have more than one hw queue. If
1767 * not, we remain on the home node of the device
1768 */
1769 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1770 hctx->numa_node = cpu_to_node(i);
1771 }
1772}
1773
1774static void blk_mq_map_swqueue(struct request_queue *q)
1775{
1776 unsigned int i;
1777 struct blk_mq_hw_ctx *hctx;
1778 struct blk_mq_ctx *ctx;
1779
1780 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboee4043dc2014-04-09 10:18:23 -06001781 cpumask_clear(hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001782 hctx->nr_ctx = 0;
1783 }
1784
1785 /*
1786 * Map software to hardware queues
1787 */
1788 queue_for_each_ctx(q, ctx, i) {
1789 /* If the cpu isn't online, the cpu is mapped to first hctx */
Jens Axboee4043dc2014-04-09 10:18:23 -06001790 if (!cpu_online(i))
1791 continue;
1792
Jens Axboe320ae512013-10-24 09:20:05 +01001793 hctx = q->mq_ops->map_queue(q, i);
Jens Axboee4043dc2014-04-09 10:18:23 -06001794 cpumask_set_cpu(i, hctx->cpumask);
Jens Axboe320ae512013-10-24 09:20:05 +01001795 ctx->index_hw = hctx->nr_ctx;
1796 hctx->ctxs[hctx->nr_ctx++] = ctx;
1797 }
Jens Axboe506e9312014-05-07 10:26:44 -06001798
1799 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboe484b4062014-05-21 14:01:15 -06001800 /*
Jens Axboea68aafa2014-08-15 13:19:15 -06001801 * If no software queues are mapped to this hardware queue,
1802 * disable it and free the request entries.
Jens Axboe484b4062014-05-21 14:01:15 -06001803 */
1804 if (!hctx->nr_ctx) {
1805 struct blk_mq_tag_set *set = q->tag_set;
1806
1807 if (set->tags[i]) {
1808 blk_mq_free_rq_map(set, set->tags[i], i);
1809 set->tags[i] = NULL;
1810 hctx->tags = NULL;
1811 }
1812 continue;
1813 }
1814
1815 /*
1816 * Initialize batch roundrobin counts
1817 */
Jens Axboe506e9312014-05-07 10:26:44 -06001818 hctx->next_cpu = cpumask_first(hctx->cpumask);
1819 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1820 }
Jens Axboe320ae512013-10-24 09:20:05 +01001821}
1822
Jens Axboe0d2602c2014-05-13 15:10:52 -06001823static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1824{
1825 struct blk_mq_hw_ctx *hctx;
1826 struct request_queue *q;
1827 bool shared;
1828 int i;
1829
1830 if (set->tag_list.next == set->tag_list.prev)
1831 shared = false;
1832 else
1833 shared = true;
1834
1835 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1836 blk_mq_freeze_queue(q);
1837
1838 queue_for_each_hw_ctx(q, hctx, i) {
1839 if (shared)
1840 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1841 else
1842 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1843 }
1844 blk_mq_unfreeze_queue(q);
1845 }
1846}
1847
1848static void blk_mq_del_queue_tag_set(struct request_queue *q)
1849{
1850 struct blk_mq_tag_set *set = q->tag_set;
1851
Jens Axboe0d2602c2014-05-13 15:10:52 -06001852 mutex_lock(&set->tag_list_lock);
1853 list_del_init(&q->tag_set_list);
1854 blk_mq_update_tag_set_depth(set);
1855 mutex_unlock(&set->tag_list_lock);
Jens Axboe0d2602c2014-05-13 15:10:52 -06001856}
1857
1858static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1859 struct request_queue *q)
1860{
1861 q->tag_set = set;
1862
1863 mutex_lock(&set->tag_list_lock);
1864 list_add_tail(&q->tag_set_list, &set->tag_list);
1865 blk_mq_update_tag_set_depth(set);
1866 mutex_unlock(&set->tag_list_lock);
1867}
1868
Ming Leie09aae72015-01-29 20:17:27 +08001869/*
1870 * It is the actual release handler for mq, but we do it from
1871 * request queue's release handler for avoiding use-after-free
1872 * and headache because q->mq_kobj shouldn't have been introduced,
1873 * but we can't group ctx/kctx kobj without it.
1874 */
1875void blk_mq_release(struct request_queue *q)
1876{
1877 struct blk_mq_hw_ctx *hctx;
1878 unsigned int i;
1879
1880 /* hctx kobj stays in hctx */
1881 queue_for_each_hw_ctx(q, hctx, i)
1882 kfree(hctx);
1883
1884 kfree(q->queue_hw_ctx);
1885
1886 /* ctx kobj stays in queue_ctx */
1887 free_percpu(q->queue_ctx);
1888}
1889
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001890struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
Jens Axboe320ae512013-10-24 09:20:05 +01001891{
1892 struct blk_mq_hw_ctx **hctxs;
Ming Leie6cdb092014-06-03 11:24:06 +08001893 struct blk_mq_ctx __percpu *ctx;
Jens Axboe320ae512013-10-24 09:20:05 +01001894 struct request_queue *q;
Jens Axboef14bbe72014-05-27 12:06:53 -06001895 unsigned int *map;
Jens Axboe320ae512013-10-24 09:20:05 +01001896 int i;
1897
Jens Axboe320ae512013-10-24 09:20:05 +01001898 ctx = alloc_percpu(struct blk_mq_ctx);
1899 if (!ctx)
1900 return ERR_PTR(-ENOMEM);
1901
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001902 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1903 set->numa_node);
Jens Axboe320ae512013-10-24 09:20:05 +01001904
1905 if (!hctxs)
1906 goto err_percpu;
1907
Jens Axboef14bbe72014-05-27 12:06:53 -06001908 map = blk_mq_make_queue_map(set);
1909 if (!map)
1910 goto err_map;
1911
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001912 for (i = 0; i < set->nr_hw_queues; i++) {
Jens Axboef14bbe72014-05-27 12:06:53 -06001913 int node = blk_mq_hw_queue_to_node(map, i);
1914
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02001915 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1916 GFP_KERNEL, node);
Jens Axboe320ae512013-10-24 09:20:05 +01001917 if (!hctxs[i])
1918 goto err_hctxs;
1919
Jens Axboea86073e2014-10-13 15:41:54 -06001920 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1921 node))
Jens Axboee4043dc2014-04-09 10:18:23 -06001922 goto err_hctxs;
1923
Jens Axboe0d2602c2014-05-13 15:10:52 -06001924 atomic_set(&hctxs[i]->nr_active, 0);
Jens Axboef14bbe72014-05-27 12:06:53 -06001925 hctxs[i]->numa_node = node;
Jens Axboe320ae512013-10-24 09:20:05 +01001926 hctxs[i]->queue_num = i;
1927 }
1928
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001929 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
Jens Axboe320ae512013-10-24 09:20:05 +01001930 if (!q)
1931 goto err_hctxs;
1932
Tejun Heo17497ac2014-09-24 13:31:50 -04001933 /*
1934 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1935 * See blk_register_queue() for details.
1936 */
Tejun Heoa34375e2014-09-08 09:51:30 +09001937 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
Tejun Heo17497ac2014-09-24 13:31:50 -04001938 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
Mike Snitzer9a30b092015-03-12 23:53:26 -04001939 goto err_mq_usage;
Ming Lei3d2936f2014-05-27 23:35:14 +08001940
Jens Axboe320ae512013-10-24 09:20:05 +01001941 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1942 blk_queue_rq_timeout(q, 30000);
1943
1944 q->nr_queues = nr_cpu_ids;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001945 q->nr_hw_queues = set->nr_hw_queues;
Jens Axboef14bbe72014-05-27 12:06:53 -06001946 q->mq_map = map;
Jens Axboe320ae512013-10-24 09:20:05 +01001947
1948 q->queue_ctx = ctx;
1949 q->queue_hw_ctx = hctxs;
1950
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001951 q->mq_ops = set->ops;
Jens Axboe94eddfb2013-11-19 09:25:07 -07001952 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
Jens Axboe320ae512013-10-24 09:20:05 +01001953
Jens Axboe05f1dd52014-05-29 09:53:32 -06001954 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1955 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1956
Christoph Hellwig1be036e2014-02-07 10:22:39 -08001957 q->sg_reserved_size = INT_MAX;
1958
Christoph Hellwig6fca6a62014-05-28 08:08:02 -06001959 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1960 INIT_LIST_HEAD(&q->requeue_list);
1961 spin_lock_init(&q->requeue_lock);
1962
Jens Axboe07068d52014-05-22 10:40:51 -06001963 if (q->nr_hw_queues > 1)
1964 blk_queue_make_request(q, blk_mq_make_request);
1965 else
1966 blk_queue_make_request(q, blk_sq_make_request);
1967
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001968 if (set->timeout)
1969 blk_queue_rq_timeout(q, set->timeout);
Jens Axboe320ae512013-10-24 09:20:05 +01001970
Jens Axboeeba71762014-05-20 15:17:27 -06001971 /*
1972 * Do this after blk_queue_make_request() overrides it...
1973 */
1974 q->nr_requests = set->queue_depth;
1975
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001976 if (set->ops->complete)
1977 blk_queue_softirq_done(q, set->ops->complete);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08001978
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001979 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
Jens Axboe320ae512013-10-24 09:20:05 +01001980
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001981 if (blk_mq_init_hw_queues(q, set))
Mike Snitzer9a30b092015-03-12 23:53:26 -04001982 goto err_mq_usage;
Christoph Hellwig18741982014-02-10 09:29:00 -07001983
Jens Axboe320ae512013-10-24 09:20:05 +01001984 mutex_lock(&all_q_mutex);
1985 list_add_tail(&q->all_q_node, &all_q_list);
1986 mutex_unlock(&all_q_mutex);
1987
Jens Axboe0d2602c2014-05-13 15:10:52 -06001988 blk_mq_add_queue_tag_set(set, q);
1989
Jens Axboe484b4062014-05-21 14:01:15 -06001990 blk_mq_map_swqueue(q);
1991
Jens Axboe320ae512013-10-24 09:20:05 +01001992 return q;
Christoph Hellwig18741982014-02-10 09:29:00 -07001993
Mike Snitzer9a30b092015-03-12 23:53:26 -04001994err_mq_usage:
Jens Axboe320ae512013-10-24 09:20:05 +01001995 blk_cleanup_queue(q);
1996err_hctxs:
Jens Axboef14bbe72014-05-27 12:06:53 -06001997 kfree(map);
Christoph Hellwig24d2f902014-04-15 14:14:00 -06001998 for (i = 0; i < set->nr_hw_queues; i++) {
Jens Axboe320ae512013-10-24 09:20:05 +01001999 if (!hctxs[i])
2000 break;
Jens Axboee4043dc2014-04-09 10:18:23 -06002001 free_cpumask_var(hctxs[i]->cpumask);
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02002002 kfree(hctxs[i]);
Jens Axboe320ae512013-10-24 09:20:05 +01002003 }
Jens Axboef14bbe72014-05-27 12:06:53 -06002004err_map:
Jens Axboe320ae512013-10-24 09:20:05 +01002005 kfree(hctxs);
2006err_percpu:
2007 free_percpu(ctx);
2008 return ERR_PTR(-ENOMEM);
2009}
2010EXPORT_SYMBOL(blk_mq_init_queue);
2011
2012void blk_mq_free_queue(struct request_queue *q)
2013{
Ming Lei624dbe42014-05-27 23:35:13 +08002014 struct blk_mq_tag_set *set = q->tag_set;
Jens Axboe320ae512013-10-24 09:20:05 +01002015
Jens Axboe0d2602c2014-05-13 15:10:52 -06002016 blk_mq_del_queue_tag_set(q);
2017
Ming Lei624dbe42014-05-27 23:35:13 +08002018 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2019 blk_mq_free_hw_queues(q, set);
Jens Axboe320ae512013-10-24 09:20:05 +01002020
Tejun Heoadd703f2014-07-01 10:34:38 -06002021 percpu_ref_exit(&q->mq_usage_counter);
Ming Lei3d2936f2014-05-27 23:35:14 +08002022
Jens Axboe320ae512013-10-24 09:20:05 +01002023 kfree(q->mq_map);
2024
Jens Axboe320ae512013-10-24 09:20:05 +01002025 q->mq_map = NULL;
2026
2027 mutex_lock(&all_q_mutex);
2028 list_del_init(&q->all_q_node);
2029 mutex_unlock(&all_q_mutex);
2030}
Jens Axboe320ae512013-10-24 09:20:05 +01002031
2032/* Basically redo blk_mq_init_queue with queue frozen */
Paul Gortmakerf618ef72013-11-14 08:26:02 -07002033static void blk_mq_queue_reinit(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01002034{
Tejun Heof3af0202014-11-04 13:52:27 -05002035 WARN_ON_ONCE(!q->mq_freeze_depth);
Jens Axboe320ae512013-10-24 09:20:05 +01002036
Jens Axboe67aec142014-05-30 08:25:36 -06002037 blk_mq_sysfs_unregister(q);
2038
Jens Axboe320ae512013-10-24 09:20:05 +01002039 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2040
2041 /*
2042 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2043 * we should change hctx numa_node according to new topology (this
2044 * involves free and re-allocate memory, worthy doing?)
2045 */
2046
2047 blk_mq_map_swqueue(q);
2048
Jens Axboe67aec142014-05-30 08:25:36 -06002049 blk_mq_sysfs_register(q);
Jens Axboe320ae512013-10-24 09:20:05 +01002050}
2051
Paul Gortmakerf618ef72013-11-14 08:26:02 -07002052static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2053 unsigned long action, void *hcpu)
Jens Axboe320ae512013-10-24 09:20:05 +01002054{
2055 struct request_queue *q;
2056
2057 /*
Jens Axboe9fccfed2014-05-08 14:50:19 -06002058 * Before new mappings are established, hotadded cpu might already
2059 * start handling requests. This doesn't break anything as we map
2060 * offline CPUs to first hardware queue. We will re-init the queue
2061 * below to get optimal settings.
Jens Axboe320ae512013-10-24 09:20:05 +01002062 */
2063 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2064 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2065 return NOTIFY_OK;
2066
2067 mutex_lock(&all_q_mutex);
Tejun Heof3af0202014-11-04 13:52:27 -05002068
2069 /*
2070 * We need to freeze and reinit all existing queues. Freezing
2071 * involves synchronous wait for an RCU grace period and doing it
2072 * one by one may take a long time. Start freezing all queues in
2073 * one swoop and then wait for the completions so that freezing can
2074 * take place in parallel.
2075 */
2076 list_for_each_entry(q, &all_q_list, all_q_node)
2077 blk_mq_freeze_queue_start(q);
2078 list_for_each_entry(q, &all_q_list, all_q_node)
2079 blk_mq_freeze_queue_wait(q);
2080
Jens Axboe320ae512013-10-24 09:20:05 +01002081 list_for_each_entry(q, &all_q_list, all_q_node)
2082 blk_mq_queue_reinit(q);
Tejun Heof3af0202014-11-04 13:52:27 -05002083
2084 list_for_each_entry(q, &all_q_list, all_q_node)
2085 blk_mq_unfreeze_queue(q);
2086
Jens Axboe320ae512013-10-24 09:20:05 +01002087 mutex_unlock(&all_q_mutex);
2088 return NOTIFY_OK;
2089}
2090
Jens Axboea5164402014-09-10 09:02:03 -06002091static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2092{
2093 int i;
2094
2095 for (i = 0; i < set->nr_hw_queues; i++) {
2096 set->tags[i] = blk_mq_init_rq_map(set, i);
2097 if (!set->tags[i])
2098 goto out_unwind;
2099 }
2100
2101 return 0;
2102
2103out_unwind:
2104 while (--i >= 0)
2105 blk_mq_free_rq_map(set, set->tags[i], i);
2106
Jens Axboea5164402014-09-10 09:02:03 -06002107 return -ENOMEM;
2108}
2109
2110/*
2111 * Allocate the request maps associated with this tag_set. Note that this
2112 * may reduce the depth asked for, if memory is tight. set->queue_depth
2113 * will be updated to reflect the allocated depth.
2114 */
2115static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2116{
2117 unsigned int depth;
2118 int err;
2119
2120 depth = set->queue_depth;
2121 do {
2122 err = __blk_mq_alloc_rq_maps(set);
2123 if (!err)
2124 break;
2125
2126 set->queue_depth >>= 1;
2127 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2128 err = -ENOMEM;
2129 break;
2130 }
2131 } while (set->queue_depth);
2132
2133 if (!set->queue_depth || err) {
2134 pr_err("blk-mq: failed to allocate request map\n");
2135 return -ENOMEM;
2136 }
2137
2138 if (depth != set->queue_depth)
2139 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2140 depth, set->queue_depth);
2141
2142 return 0;
2143}
2144
Jens Axboea4391c62014-06-05 15:21:56 -06002145/*
2146 * Alloc a tag set to be associated with one or more request queues.
2147 * May fail with EINVAL for various error conditions. May adjust the
2148 * requested depth down, if if it too large. In that case, the set
2149 * value will be stored in set->queue_depth.
2150 */
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002151int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2152{
Bart Van Assche205fb5f2014-10-30 14:45:11 +01002153 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2154
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002155 if (!set->nr_hw_queues)
2156 return -EINVAL;
Jens Axboea4391c62014-06-05 15:21:56 -06002157 if (!set->queue_depth)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002158 return -EINVAL;
2159 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2160 return -EINVAL;
2161
Christoph Hellwigcdef54d2014-05-28 18:11:06 +02002162 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002163 return -EINVAL;
2164
Jens Axboea4391c62014-06-05 15:21:56 -06002165 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2166 pr_info("blk-mq: reduced tag depth to %u\n",
2167 BLK_MQ_MAX_DEPTH);
2168 set->queue_depth = BLK_MQ_MAX_DEPTH;
2169 }
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002170
Shaohua Li6637fad2014-11-30 16:00:58 -08002171 /*
2172 * If a crashdump is active, then we are potentially in a very
2173 * memory constrained environment. Limit us to 1 queue and
2174 * 64 tags to prevent using too much memory.
2175 */
2176 if (is_kdump_kernel()) {
2177 set->nr_hw_queues = 1;
2178 set->queue_depth = min(64U, set->queue_depth);
2179 }
2180
Ming Lei48479002014-04-19 18:00:17 +08002181 set->tags = kmalloc_node(set->nr_hw_queues *
2182 sizeof(struct blk_mq_tags *),
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002183 GFP_KERNEL, set->numa_node);
2184 if (!set->tags)
Jens Axboea5164402014-09-10 09:02:03 -06002185 return -ENOMEM;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002186
Jens Axboea5164402014-09-10 09:02:03 -06002187 if (blk_mq_alloc_rq_maps(set))
2188 goto enomem;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002189
Jens Axboe0d2602c2014-05-13 15:10:52 -06002190 mutex_init(&set->tag_list_lock);
2191 INIT_LIST_HEAD(&set->tag_list);
2192
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002193 return 0;
Jens Axboea5164402014-09-10 09:02:03 -06002194enomem:
Robert Elliott5676e7b2014-09-02 11:38:44 -05002195 kfree(set->tags);
2196 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002197 return -ENOMEM;
2198}
2199EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2200
2201void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2202{
2203 int i;
2204
Jens Axboe484b4062014-05-21 14:01:15 -06002205 for (i = 0; i < set->nr_hw_queues; i++) {
2206 if (set->tags[i])
2207 blk_mq_free_rq_map(set, set->tags[i], i);
2208 }
2209
Ming Lei981bd182014-04-24 00:07:34 +08002210 kfree(set->tags);
Robert Elliott5676e7b2014-09-02 11:38:44 -05002211 set->tags = NULL;
Christoph Hellwig24d2f902014-04-15 14:14:00 -06002212}
2213EXPORT_SYMBOL(blk_mq_free_tag_set);
2214
Jens Axboee3a2b3f2014-05-20 11:49:02 -06002215int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2216{
2217 struct blk_mq_tag_set *set = q->tag_set;
2218 struct blk_mq_hw_ctx *hctx;
2219 int i, ret;
2220
2221 if (!set || nr > set->queue_depth)
2222 return -EINVAL;
2223
2224 ret = 0;
2225 queue_for_each_hw_ctx(q, hctx, i) {
2226 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2227 if (ret)
2228 break;
2229 }
2230
2231 if (!ret)
2232 q->nr_requests = nr;
2233
2234 return ret;
2235}
2236
Jens Axboe676141e2014-03-20 13:29:18 -06002237void blk_mq_disable_hotplug(void)
2238{
2239 mutex_lock(&all_q_mutex);
2240}
2241
2242void blk_mq_enable_hotplug(void)
2243{
2244 mutex_unlock(&all_q_mutex);
2245}
2246
Jens Axboe320ae512013-10-24 09:20:05 +01002247static int __init blk_mq_init(void)
2248{
Jens Axboe320ae512013-10-24 09:20:05 +01002249 blk_mq_cpu_init();
2250
Tejun Heoadd703f2014-07-01 10:34:38 -06002251 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
Jens Axboe320ae512013-10-24 09:20:05 +01002252
2253 return 0;
2254}
2255subsys_initcall(blk_mq_init);