]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/blame - block/blk-mq.c
blk-mq: fix race between timeout and freeing request
[mirror_ubuntu-zesty-kernel.git] / block / blk-mq.c
CommitLineData
75bb4625
JA
1/*
2 * Block multiqueue core code
3 *
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
6 */
320ae51f
JA
7#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>
aedcd72f 23#include <linux/crash_dump.h>
320ae51f
JA
24
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
320ae51f
JA
37/*
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
569fd0ce 44 for (i = 0; i < hctx->ctx_map.size; i++)
1429d7c9 45 if (hctx->ctx_map.map[i].word)
320ae51f
JA
46 return true;
47
48 return false;
49}
50
1429d7c9
JA
51static 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
320ae51f
JA
60/*
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{
1429d7c9
JA
66 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);
320ae51f
JA
78}
79
bfd343aa 80static int blk_mq_queue_enter(struct request_queue *q, gfp_t gfp)
320ae51f 81{
add703fd
TH
82 while (true) {
83 int ret;
320ae51f 84
add703fd
TH
85 if (percpu_ref_tryget_live(&q->mq_usage_counter))
86 return 0;
320ae51f 87
bfd343aa
KB
88 if (!(gfp & __GFP_WAIT))
89 return -EBUSY;
90
add703fd 91 ret = wait_event_interruptible(q->mq_freeze_wq,
4ecd4fef
CH
92 !atomic_read(&q->mq_freeze_depth) ||
93 blk_queue_dying(q));
add703fd
TH
94 if (blk_queue_dying(q))
95 return -ENODEV;
96 if (ret)
97 return ret;
98 }
320ae51f
JA
99}
100
101static void blk_mq_queue_exit(struct request_queue *q)
102{
add703fd
TH
103 percpu_ref_put(&q->mq_usage_counter);
104}
105
106static void blk_mq_usage_counter_release(struct percpu_ref *ref)
107{
108 struct request_queue *q =
109 container_of(ref, struct request_queue, mq_usage_counter);
110
111 wake_up_all(&q->mq_freeze_wq);
320ae51f
JA
112}
113
b4c6a028 114void blk_mq_freeze_queue_start(struct request_queue *q)
43a5e4e2 115{
4ecd4fef 116 int freeze_depth;
cddd5d17 117
4ecd4fef
CH
118 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
119 if (freeze_depth == 1) {
9eca8046 120 percpu_ref_kill(&q->mq_usage_counter);
b94ec296 121 blk_mq_run_hw_queues(q, false);
cddd5d17 122 }
f3af020b 123}
b4c6a028 124EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
f3af020b
TH
125
126static void blk_mq_freeze_queue_wait(struct request_queue *q)
127{
add703fd 128 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
43a5e4e2
ML
129}
130
f3af020b
TH
131/*
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}
c761d96b 140EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
f3af020b 141
b4c6a028 142void blk_mq_unfreeze_queue(struct request_queue *q)
320ae51f 143{
4ecd4fef 144 int freeze_depth;
320ae51f 145
4ecd4fef
CH
146 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
147 WARN_ON_ONCE(freeze_depth < 0);
148 if (!freeze_depth) {
add703fd 149 percpu_ref_reinit(&q->mq_usage_counter);
320ae51f 150 wake_up_all(&q->mq_freeze_wq);
add703fd 151 }
320ae51f 152}
b4c6a028 153EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
320ae51f 154
aed3ea94
JA
155void 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);
3fd5940c
KB
163
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);
aed3ea94
JA
170}
171
320ae51f
JA
172bool 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
94eddfbe
JA
178static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
179 struct request *rq, unsigned int rw_flags)
320ae51f 180{
94eddfbe
JA
181 if (blk_queue_io_stat(q))
182 rw_flags |= REQ_IO_STAT;
183
af76e555
CH
184 INIT_LIST_HEAD(&rq->queuelist);
185 /* csd/requeue_work/fifo_time is initialized before use */
186 rq->q = q;
320ae51f 187 rq->mq_ctx = ctx;
0d2602ca 188 rq->cmd_flags |= rw_flags;
af76e555
CH
189 /* do not touch atomic flags, it needs atomic ops against the timer */
190 rq->cpu = -1;
af76e555
CH
191 INIT_HLIST_NODE(&rq->hash);
192 RB_CLEAR_NODE(&rq->rb_node);
af76e555
CH
193 rq->rq_disk = NULL;
194 rq->part = NULL;
3ee32372 195 rq->start_time = jiffies;
af76e555
CH
196#ifdef CONFIG_BLK_CGROUP
197 rq->rl = NULL;
0fec08b4 198 set_start_time_ns(rq);
af76e555
CH
199 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
af76e555
CH
205 rq->special = NULL;
206 /* tag was already set */
207 rq->errors = 0;
af76e555 208
6f4a1626
TB
209 rq->cmd = rq->__cmd;
210
af76e555
CH
211 rq->extra_len = 0;
212 rq->sense_len = 0;
213 rq->resid_len = 0;
214 rq->sense = NULL;
215
af76e555 216 INIT_LIST_HEAD(&rq->timeout_list);
f6be4fb4
JA
217 rq->timeout = 0;
218
af76e555
CH
219 rq->end_io = NULL;
220 rq->end_io_data = NULL;
221 rq->next_rq = NULL;
222
320ae51f
JA
223 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
224}
225
5dee8577 226static struct request *
cb96a42c 227__blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
5dee8577
CH
228{
229 struct request *rq;
230 unsigned int tag;
231
cb96a42c 232 tag = blk_mq_get_tag(data);
5dee8577 233 if (tag != BLK_MQ_TAG_FAIL) {
cb96a42c 234 rq = data->hctx->tags->rqs[tag];
5dee8577 235
cb96a42c 236 if (blk_mq_tag_busy(data->hctx)) {
5dee8577 237 rq->cmd_flags = REQ_MQ_INFLIGHT;
cb96a42c 238 atomic_inc(&data->hctx->nr_active);
5dee8577
CH
239 }
240
241 rq->tag = tag;
cb96a42c 242 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
5dee8577
CH
243 return rq;
244 }
245
246 return NULL;
247}
248
4ce01dd1
CH
249struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
250 bool reserved)
320ae51f 251{
d852564f
CH
252 struct blk_mq_ctx *ctx;
253 struct blk_mq_hw_ctx *hctx;
320ae51f 254 struct request *rq;
cb96a42c 255 struct blk_mq_alloc_data alloc_data;
a492f075 256 int ret;
320ae51f 257
bfd343aa 258 ret = blk_mq_queue_enter(q, gfp);
a492f075
JL
259 if (ret)
260 return ERR_PTR(ret);
320ae51f 261
d852564f
CH
262 ctx = blk_mq_get_ctx(q);
263 hctx = q->mq_ops->map_queue(q, ctx->cpu);
cb96a42c
ML
264 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
265 reserved, ctx, hctx);
d852564f 266
cb96a42c 267 rq = __blk_mq_alloc_request(&alloc_data, rw);
d852564f
CH
268 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);
cb96a42c
ML
274 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;
d852564f
CH
278 }
279 blk_mq_put_ctx(ctx);
c76541a9
KB
280 if (!rq) {
281 blk_mq_queue_exit(q);
a492f075 282 return ERR_PTR(-EWOULDBLOCK);
c76541a9 283 }
320ae51f
JA
284 return rq;
285}
4bb659b1 286EXPORT_SYMBOL(blk_mq_alloc_request);
320ae51f 287
320ae51f
JA
288static 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
0d2602ca
JA
294 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
295 atomic_dec(&hctx->nr_active);
683d0e12 296 rq->cmd_flags = 0;
0d2602ca 297
af76e555 298 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
0d2602ca 299 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
320ae51f
JA
300 blk_mq_queue_exit(q);
301}
302
7c7f2f2b 303void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
320ae51f
JA
304{
305 struct blk_mq_ctx *ctx = rq->mq_ctx;
320ae51f
JA
306
307 ctx->rq_completed[rq_is_sync(rq)]++;
320ae51f 308 __blk_mq_free_request(hctx, ctx, rq);
7c7f2f2b
JA
309
310}
311EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
312
313void blk_mq_free_request(struct request *rq)
314{
315 struct blk_mq_hw_ctx *hctx;
316 struct request_queue *q = rq->q;
317
318 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
319 blk_mq_free_hctx_request(hctx, rq);
320ae51f 320}
1a3b595a 321EXPORT_SYMBOL_GPL(blk_mq_free_request);
320ae51f 322
c8a446ad 323inline void __blk_mq_end_request(struct request *rq, int error)
320ae51f 324{
0d11e6ac
ML
325 blk_account_io_done(rq);
326
91b63639 327 if (rq->end_io) {
320ae51f 328 rq->end_io(rq, error);
91b63639
CH
329 } else {
330 if (unlikely(blk_bidi_rq(rq)))
331 blk_mq_free_request(rq->next_rq);
320ae51f 332 blk_mq_free_request(rq);
91b63639 333 }
320ae51f 334}
c8a446ad 335EXPORT_SYMBOL(__blk_mq_end_request);
63151a44 336
c8a446ad 337void blk_mq_end_request(struct request *rq, int error)
63151a44
CH
338{
339 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
340 BUG();
c8a446ad 341 __blk_mq_end_request(rq, error);
63151a44 342}
c8a446ad 343EXPORT_SYMBOL(blk_mq_end_request);
320ae51f 344
30a91cb4 345static void __blk_mq_complete_request_remote(void *data)
320ae51f 346{
3d6efbf6 347 struct request *rq = data;
320ae51f 348
30a91cb4 349 rq->q->softirq_done_fn(rq);
320ae51f 350}
320ae51f 351
ed851860 352static void blk_mq_ipi_complete_request(struct request *rq)
320ae51f
JA
353{
354 struct blk_mq_ctx *ctx = rq->mq_ctx;
38535201 355 bool shared = false;
320ae51f
JA
356 int cpu;
357
38535201 358 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
30a91cb4
CH
359 rq->q->softirq_done_fn(rq);
360 return;
361 }
320ae51f
JA
362
363 cpu = get_cpu();
38535201
CH
364 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)) {
30a91cb4 368 rq->csd.func = __blk_mq_complete_request_remote;
3d6efbf6
CH
369 rq->csd.info = rq;
370 rq->csd.flags = 0;
c46fff2a 371 smp_call_function_single_async(ctx->cpu, &rq->csd);
3d6efbf6 372 } else {
30a91cb4 373 rq->q->softirq_done_fn(rq);
3d6efbf6 374 }
320ae51f
JA
375 put_cpu();
376}
30a91cb4 377
ed851860
JA
378void __blk_mq_complete_request(struct request *rq)
379{
380 struct request_queue *q = rq->q;
381
382 if (!q->softirq_done_fn)
c8a446ad 383 blk_mq_end_request(rq, rq->errors);
ed851860
JA
384 else
385 blk_mq_ipi_complete_request(rq);
386}
387
30a91cb4
CH
388/**
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{
95f09684
JA
398 struct request_queue *q = rq->q;
399
400 if (unlikely(blk_should_fake_timeout(q)))
30a91cb4 401 return;
ed851860
JA
402 if (!blk_mark_rq_complete(rq))
403 __blk_mq_complete_request(rq);
30a91cb4
CH
404}
405EXPORT_SYMBOL(blk_mq_complete_request);
320ae51f 406
973c0191
KB
407int 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
e2490073 413void blk_mq_start_request(struct request *rq)
320ae51f
JA
414{
415 struct request_queue *q = rq->q;
416
417 trace_block_rq_issue(q, rq);
418
742ee69b 419 rq->resid_len = blk_rq_bytes(rq);
91b63639
CH
420 if (unlikely(blk_bidi_rq(rq)))
421 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
742ee69b 422
2b8393b4 423 blk_add_timer(rq);
87ee7b11 424
538b7534
JA
425 /*
426 * Ensure that ->deadline is visible before set the started
427 * flag and clear the completed flag.
428 */
429 smp_mb__before_atomic();
430
87ee7b11
JA
431 /*
432 * 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 */
4b570521
JA
437 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);
49f5baa5
CH
441
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 }
320ae51f 450}
e2490073 451EXPORT_SYMBOL(blk_mq_start_request);
320ae51f 452
ed0791b2 453static void __blk_mq_requeue_request(struct request *rq)
320ae51f
JA
454{
455 struct request_queue *q = rq->q;
456
457 trace_block_rq_requeue(q, rq);
49f5baa5 458
e2490073
CH
459 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 }
320ae51f
JA
463}
464
ed0791b2
CH
465void blk_mq_requeue_request(struct request *rq)
466{
ed0791b2 467 __blk_mq_requeue_request(rq);
ed0791b2 468
ed0791b2 469 BUG_ON(blk_queued_rq(rq));
6fca6a61 470 blk_mq_add_to_requeue_list(rq, true);
ed0791b2
CH
471}
472EXPORT_SYMBOL(blk_mq_requeue_request);
473
6fca6a61
CH
474static 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
8b957415
JA
501 /*
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);
6fca6a61
CH
506}
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
c68ed59f
KB
530void 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
6fca6a61
CH
536void 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
1885b24d
JA
542void 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
0e62f51f
JA
562struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
563{
0048b483 564 return tags->rqs[tag];
24d2f903
CH
565}
566EXPORT_SYMBOL(blk_mq_tag_to_rq);
567
320ae51f 568struct blk_mq_timeout_data {
46f92d42
CH
569 unsigned long next;
570 unsigned int next_set;
320ae51f
JA
571};
572
90415837 573void blk_mq_rq_timed_out(struct request *req, bool reserved)
320ae51f 574{
46f92d42
CH
575 struct blk_mq_ops *ops = req->q->mq_ops;
576 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
87ee7b11
JA
577
578 /*
579 * We know that complete is set at this point. If STARTED isn't set
580 * anymore, then the request isn't active and the "timeout" should
581 * just be ignored. This can happen due to the bitflag ordering.
582 * Timeout first checks if STARTED is set, and if it is, assumes
583 * the request is active. But if we race with completion, then
584 * we both flags will get cleared. So check here again, and ignore
585 * a timeout event with a request that isn't active.
586 */
46f92d42
CH
587 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
588 return;
87ee7b11 589
46f92d42 590 if (ops->timeout)
0152fb6b 591 ret = ops->timeout(req, reserved);
46f92d42
CH
592
593 switch (ret) {
594 case BLK_EH_HANDLED:
595 __blk_mq_complete_request(req);
596 break;
597 case BLK_EH_RESET_TIMER:
598 blk_add_timer(req);
599 blk_clear_rq_complete(req);
600 break;
601 case BLK_EH_NOT_HANDLED:
602 break;
603 default:
604 printk(KERN_ERR "block: bad eh return: %d\n", ret);
605 break;
606 }
87ee7b11 607}
5b3f25fc 608
81481eb4
CH
609static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
610 struct request *rq, void *priv, bool reserved)
611{
612 struct blk_mq_timeout_data *data = priv;
87ee7b11 613
eb130dbf
KB
614 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
615 /*
616 * If a request wasn't started before the queue was
617 * marked dying, kill it here or it'll go unnoticed.
618 */
619 if (unlikely(blk_queue_dying(rq->q))) {
620 rq->errors = -EIO;
621 blk_mq_complete_request(rq);
622 }
46f92d42 623 return;
eb130dbf 624 }
5b3f25fc
KB
625 if (rq->cmd_flags & REQ_NO_TIMEOUT)
626 return;
87ee7b11 627
46f92d42
CH
628 if (time_after_eq(jiffies, rq->deadline)) {
629 if (!blk_mark_rq_complete(rq))
0152fb6b 630 blk_mq_rq_timed_out(rq, reserved);
46f92d42
CH
631 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
632 data->next = rq->deadline;
633 data->next_set = 1;
634 }
87ee7b11
JA
635}
636
81481eb4 637static void blk_mq_rq_timer(unsigned long priv)
320ae51f 638{
81481eb4
CH
639 struct request_queue *q = (struct request_queue *)priv;
640 struct blk_mq_timeout_data data = {
641 .next = 0,
642 .next_set = 0,
643 };
320ae51f 644 struct blk_mq_hw_ctx *hctx;
81481eb4 645 int i;
320ae51f 646
484b4061
JA
647 queue_for_each_hw_ctx(q, hctx, i) {
648 /*
649 * If not software queues are currently mapped to this
650 * hardware queue, there's nothing to check
651 */
19c66e59 652 if (!blk_mq_hw_queue_mapped(hctx))
484b4061
JA
653 continue;
654
81481eb4 655 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
484b4061 656 }
320ae51f 657
81481eb4
CH
658 if (data.next_set) {
659 data.next = blk_rq_timeout(round_jiffies_up(data.next));
660 mod_timer(&q->timeout, data.next);
0d2602ca 661 } else {
f054b56c
ML
662 queue_for_each_hw_ctx(q, hctx, i) {
663 /* the hctx may be unmapped, so check it here */
664 if (blk_mq_hw_queue_mapped(hctx))
665 blk_mq_tag_idle(hctx);
666 }
0d2602ca 667 }
320ae51f
JA
668}
669
670/*
671 * Reverse check our software queue for entries that we could potentially
672 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
673 * too much time checking for merges.
674 */
675static bool blk_mq_attempt_merge(struct request_queue *q,
676 struct blk_mq_ctx *ctx, struct bio *bio)
677{
678 struct request *rq;
679 int checked = 8;
680
681 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
682 int el_ret;
683
684 if (!checked--)
685 break;
686
687 if (!blk_rq_merge_ok(rq, bio))
688 continue;
689
690 el_ret = blk_try_merge(rq, bio);
691 if (el_ret == ELEVATOR_BACK_MERGE) {
692 if (bio_attempt_back_merge(q, rq, bio)) {
693 ctx->rq_merged++;
694 return true;
695 }
696 break;
697 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
698 if (bio_attempt_front_merge(q, rq, bio)) {
699 ctx->rq_merged++;
700 return true;
701 }
702 break;
703 }
704 }
705
706 return false;
707}
708
1429d7c9
JA
709/*
710 * Process software queues that have been marked busy, splicing them
711 * to the for-dispatch
712 */
713static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
714{
715 struct blk_mq_ctx *ctx;
716 int i;
717
569fd0ce 718 for (i = 0; i < hctx->ctx_map.size; i++) {
1429d7c9
JA
719 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
720 unsigned int off, bit;
721
722 if (!bm->word)
723 continue;
724
725 bit = 0;
726 off = i * hctx->ctx_map.bits_per_word;
727 do {
728 bit = find_next_bit(&bm->word, bm->depth, bit);
729 if (bit >= bm->depth)
730 break;
731
732 ctx = hctx->ctxs[bit + off];
733 clear_bit(bit, &bm->word);
734 spin_lock(&ctx->lock);
735 list_splice_tail_init(&ctx->rq_list, list);
736 spin_unlock(&ctx->lock);
737
738 bit++;
739 } while (1);
740 }
741}
742
320ae51f
JA
743/*
744 * Run this hardware queue, pulling any software queues mapped to it in.
745 * Note that this function currently has various problems around ordering
746 * of IO. In particular, we'd like FIFO behaviour on handling existing
747 * items on the hctx->dispatch list. Ignore that for now.
748 */
749static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
750{
751 struct request_queue *q = hctx->queue;
320ae51f
JA
752 struct request *rq;
753 LIST_HEAD(rq_list);
74c45052
JA
754 LIST_HEAD(driver_list);
755 struct list_head *dptr;
1429d7c9 756 int queued;
320ae51f 757
fd1270d5 758 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
e4043dcf 759
5d12f905 760 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
320ae51f
JA
761 return;
762
763 hctx->run++;
764
765 /*
766 * Touch any software queue that has pending entries.
767 */
1429d7c9 768 flush_busy_ctxs(hctx, &rq_list);
320ae51f
JA
769
770 /*
771 * If we have previous entries on our dispatch list, grab them
772 * and stuff them at the front for more fair dispatch.
773 */
774 if (!list_empty_careful(&hctx->dispatch)) {
775 spin_lock(&hctx->lock);
776 if (!list_empty(&hctx->dispatch))
777 list_splice_init(&hctx->dispatch, &rq_list);
778 spin_unlock(&hctx->lock);
779 }
780
74c45052
JA
781 /*
782 * Start off with dptr being NULL, so we start the first request
783 * immediately, even if we have more pending.
784 */
785 dptr = NULL;
786
320ae51f
JA
787 /*
788 * Now process all the entries, sending them to the driver.
789 */
1429d7c9 790 queued = 0;
320ae51f 791 while (!list_empty(&rq_list)) {
74c45052 792 struct blk_mq_queue_data bd;
320ae51f
JA
793 int ret;
794
795 rq = list_first_entry(&rq_list, struct request, queuelist);
796 list_del_init(&rq->queuelist);
320ae51f 797
74c45052
JA
798 bd.rq = rq;
799 bd.list = dptr;
800 bd.last = list_empty(&rq_list);
801
802 ret = q->mq_ops->queue_rq(hctx, &bd);
320ae51f
JA
803 switch (ret) {
804 case BLK_MQ_RQ_QUEUE_OK:
805 queued++;
806 continue;
807 case BLK_MQ_RQ_QUEUE_BUSY:
320ae51f 808 list_add(&rq->queuelist, &rq_list);
ed0791b2 809 __blk_mq_requeue_request(rq);
320ae51f
JA
810 break;
811 default:
812 pr_err("blk-mq: bad return on queue: %d\n", ret);
320ae51f 813 case BLK_MQ_RQ_QUEUE_ERROR:
1e93b8c2 814 rq->errors = -EIO;
c8a446ad 815 blk_mq_end_request(rq, rq->errors);
320ae51f
JA
816 break;
817 }
818
819 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
820 break;
74c45052
JA
821
822 /*
823 * We've done the first request. If we have more than 1
824 * left in the list, set dptr to defer issue.
825 */
826 if (!dptr && rq_list.next != rq_list.prev)
827 dptr = &driver_list;
320ae51f
JA
828 }
829
830 if (!queued)
831 hctx->dispatched[0]++;
832 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
833 hctx->dispatched[ilog2(queued) + 1]++;
834
835 /*
836 * Any items that need requeuing? Stuff them into hctx->dispatch,
837 * that is where we will continue on next queue run.
838 */
839 if (!list_empty(&rq_list)) {
840 spin_lock(&hctx->lock);
841 list_splice(&rq_list, &hctx->dispatch);
842 spin_unlock(&hctx->lock);
9ba52e58
SL
843 /*
844 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
845 * it's possible the queue is stopped and restarted again
846 * before this. Queue restart will dispatch requests. And since
847 * requests in rq_list aren't added into hctx->dispatch yet,
848 * the requests in rq_list might get lost.
849 *
850 * blk_mq_run_hw_queue() already checks the STOPPED bit
851 **/
852 blk_mq_run_hw_queue(hctx, true);
320ae51f
JA
853 }
854}
855
506e931f
JA
856/*
857 * It'd be great if the workqueue API had a way to pass
858 * in a mask and had some smarts for more clever placement.
859 * For now we just round-robin here, switching for every
860 * BLK_MQ_CPU_WORK_BATCH queued items.
861 */
862static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
863{
b657d7e6
CH
864 if (hctx->queue->nr_hw_queues == 1)
865 return WORK_CPU_UNBOUND;
506e931f
JA
866
867 if (--hctx->next_cpu_batch <= 0) {
b657d7e6 868 int cpu = hctx->next_cpu, next_cpu;
506e931f
JA
869
870 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
871 if (next_cpu >= nr_cpu_ids)
872 next_cpu = cpumask_first(hctx->cpumask);
873
874 hctx->next_cpu = next_cpu;
875 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
b657d7e6
CH
876
877 return cpu;
506e931f
JA
878 }
879
b657d7e6 880 return hctx->next_cpu;
506e931f
JA
881}
882
320ae51f
JA
883void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
884{
19c66e59
ML
885 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
886 !blk_mq_hw_queue_mapped(hctx)))
320ae51f
JA
887 return;
888
398205b8 889 if (!async) {
2a90d4aa
PB
890 int cpu = get_cpu();
891 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
398205b8 892 __blk_mq_run_hw_queue(hctx);
2a90d4aa 893 put_cpu();
398205b8
PB
894 return;
895 }
e4043dcf 896
2a90d4aa 897 put_cpu();
e4043dcf 898 }
398205b8 899
b657d7e6
CH
900 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
901 &hctx->run_work, 0);
320ae51f
JA
902}
903
b94ec296 904void blk_mq_run_hw_queues(struct request_queue *q, bool async)
320ae51f
JA
905{
906 struct blk_mq_hw_ctx *hctx;
907 int i;
908
909 queue_for_each_hw_ctx(q, hctx, i) {
910 if ((!blk_mq_hctx_has_pending(hctx) &&
911 list_empty_careful(&hctx->dispatch)) ||
5d12f905 912 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
320ae51f
JA
913 continue;
914
b94ec296 915 blk_mq_run_hw_queue(hctx, async);
320ae51f
JA
916 }
917}
b94ec296 918EXPORT_SYMBOL(blk_mq_run_hw_queues);
320ae51f
JA
919
920void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
921{
70f4db63
CH
922 cancel_delayed_work(&hctx->run_work);
923 cancel_delayed_work(&hctx->delay_work);
320ae51f
JA
924 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
925}
926EXPORT_SYMBOL(blk_mq_stop_hw_queue);
927
280d45f6
CH
928void blk_mq_stop_hw_queues(struct request_queue *q)
929{
930 struct blk_mq_hw_ctx *hctx;
931 int i;
932
933 queue_for_each_hw_ctx(q, hctx, i)
934 blk_mq_stop_hw_queue(hctx);
935}
936EXPORT_SYMBOL(blk_mq_stop_hw_queues);
937
320ae51f
JA
938void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
939{
940 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
e4043dcf 941
0ffbce80 942 blk_mq_run_hw_queue(hctx, false);
320ae51f
JA
943}
944EXPORT_SYMBOL(blk_mq_start_hw_queue);
945
2f268556
CH
946void blk_mq_start_hw_queues(struct request_queue *q)
947{
948 struct blk_mq_hw_ctx *hctx;
949 int i;
950
951 queue_for_each_hw_ctx(q, hctx, i)
952 blk_mq_start_hw_queue(hctx);
953}
954EXPORT_SYMBOL(blk_mq_start_hw_queues);
955
1b4a3258 956void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
320ae51f
JA
957{
958 struct blk_mq_hw_ctx *hctx;
959 int i;
960
961 queue_for_each_hw_ctx(q, hctx, i) {
962 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
963 continue;
964
965 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1b4a3258 966 blk_mq_run_hw_queue(hctx, async);
320ae51f
JA
967 }
968}
969EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
970
70f4db63 971static void blk_mq_run_work_fn(struct work_struct *work)
320ae51f
JA
972{
973 struct blk_mq_hw_ctx *hctx;
974
70f4db63 975 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
e4043dcf 976
320ae51f
JA
977 __blk_mq_run_hw_queue(hctx);
978}
979
70f4db63
CH
980static void blk_mq_delay_work_fn(struct work_struct *work)
981{
982 struct blk_mq_hw_ctx *hctx;
983
984 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
985
986 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
987 __blk_mq_run_hw_queue(hctx);
988}
989
990void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
991{
19c66e59
ML
992 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
993 return;
70f4db63 994
b657d7e6
CH
995 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
996 &hctx->delay_work, msecs_to_jiffies(msecs));
70f4db63
CH
997}
998EXPORT_SYMBOL(blk_mq_delay_queue);
999
320ae51f 1000static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
72a0a36e 1001 struct request *rq, bool at_head)
320ae51f
JA
1002{
1003 struct blk_mq_ctx *ctx = rq->mq_ctx;
1004
01b983c9
JA
1005 trace_block_rq_insert(hctx->queue, rq);
1006
72a0a36e
CH
1007 if (at_head)
1008 list_add(&rq->queuelist, &ctx->rq_list);
1009 else
1010 list_add_tail(&rq->queuelist, &ctx->rq_list);
4bb659b1 1011
320ae51f 1012 blk_mq_hctx_mark_pending(hctx, ctx);
320ae51f
JA
1013}
1014
eeabc850
CH
1015void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1016 bool async)
320ae51f 1017{
eeabc850 1018 struct request_queue *q = rq->q;
320ae51f 1019 struct blk_mq_hw_ctx *hctx;
eeabc850
CH
1020 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1021
1022 current_ctx = blk_mq_get_ctx(q);
1023 if (!cpu_online(ctx->cpu))
1024 rq->mq_ctx = ctx = current_ctx;
320ae51f 1025
320ae51f
JA
1026 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1027
a57a178a
CH
1028 spin_lock(&ctx->lock);
1029 __blk_mq_insert_request(hctx, rq, at_head);
1030 spin_unlock(&ctx->lock);
320ae51f 1031
320ae51f
JA
1032 if (run_queue)
1033 blk_mq_run_hw_queue(hctx, async);
e4043dcf
JA
1034
1035 blk_mq_put_ctx(current_ctx);
320ae51f
JA
1036}
1037
1038static void blk_mq_insert_requests(struct request_queue *q,
1039 struct blk_mq_ctx *ctx,
1040 struct list_head *list,
1041 int depth,
1042 bool from_schedule)
1043
1044{
1045 struct blk_mq_hw_ctx *hctx;
1046 struct blk_mq_ctx *current_ctx;
1047
1048 trace_block_unplug(q, depth, !from_schedule);
1049
1050 current_ctx = blk_mq_get_ctx(q);
1051
1052 if (!cpu_online(ctx->cpu))
1053 ctx = current_ctx;
1054 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1055
1056 /*
1057 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1058 * offline now
1059 */
1060 spin_lock(&ctx->lock);
1061 while (!list_empty(list)) {
1062 struct request *rq;
1063
1064 rq = list_first_entry(list, struct request, queuelist);
1065 list_del_init(&rq->queuelist);
1066 rq->mq_ctx = ctx;
72a0a36e 1067 __blk_mq_insert_request(hctx, rq, false);
320ae51f
JA
1068 }
1069 spin_unlock(&ctx->lock);
1070
320ae51f 1071 blk_mq_run_hw_queue(hctx, from_schedule);
e4043dcf 1072 blk_mq_put_ctx(current_ctx);
320ae51f
JA
1073}
1074
1075static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1076{
1077 struct request *rqa = container_of(a, struct request, queuelist);
1078 struct request *rqb = container_of(b, struct request, queuelist);
1079
1080 return !(rqa->mq_ctx < rqb->mq_ctx ||
1081 (rqa->mq_ctx == rqb->mq_ctx &&
1082 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1083}
1084
1085void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1086{
1087 struct blk_mq_ctx *this_ctx;
1088 struct request_queue *this_q;
1089 struct request *rq;
1090 LIST_HEAD(list);
1091 LIST_HEAD(ctx_list);
1092 unsigned int depth;
1093
1094 list_splice_init(&plug->mq_list, &list);
1095
1096 list_sort(NULL, &list, plug_ctx_cmp);
1097
1098 this_q = NULL;
1099 this_ctx = NULL;
1100 depth = 0;
1101
1102 while (!list_empty(&list)) {
1103 rq = list_entry_rq(list.next);
1104 list_del_init(&rq->queuelist);
1105 BUG_ON(!rq->q);
1106 if (rq->mq_ctx != this_ctx) {
1107 if (this_ctx) {
1108 blk_mq_insert_requests(this_q, this_ctx,
1109 &ctx_list, depth,
1110 from_schedule);
1111 }
1112
1113 this_ctx = rq->mq_ctx;
1114 this_q = rq->q;
1115 depth = 0;
1116 }
1117
1118 depth++;
1119 list_add_tail(&rq->queuelist, &ctx_list);
1120 }
1121
1122 /*
1123 * If 'this_ctx' is set, we know we have entries to complete
1124 * on 'ctx_list'. Do those.
1125 */
1126 if (this_ctx) {
1127 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1128 from_schedule);
1129 }
1130}
1131
1132static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1133{
1134 init_request_from_bio(rq, bio);
4b570521 1135
3ee32372 1136 if (blk_do_io_stat(rq))
4b570521 1137 blk_account_io_start(rq, 1);
320ae51f
JA
1138}
1139
274a5843
JA
1140static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1141{
1142 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1143 !blk_queue_nomerges(hctx->queue);
1144}
1145
07068d5b
JA
1146static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1147 struct blk_mq_ctx *ctx,
1148 struct request *rq, struct bio *bio)
320ae51f 1149{
274a5843 1150 if (!hctx_allow_merges(hctx)) {
07068d5b
JA
1151 blk_mq_bio_to_request(rq, bio);
1152 spin_lock(&ctx->lock);
1153insert_rq:
1154 __blk_mq_insert_request(hctx, rq, false);
1155 spin_unlock(&ctx->lock);
1156 return false;
1157 } else {
274a5843
JA
1158 struct request_queue *q = hctx->queue;
1159
07068d5b
JA
1160 spin_lock(&ctx->lock);
1161 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1162 blk_mq_bio_to_request(rq, bio);
1163 goto insert_rq;
1164 }
320ae51f 1165
07068d5b
JA
1166 spin_unlock(&ctx->lock);
1167 __blk_mq_free_request(hctx, ctx, rq);
1168 return true;
14ec77f3 1169 }
07068d5b 1170}
14ec77f3 1171
07068d5b
JA
1172struct blk_map_ctx {
1173 struct blk_mq_hw_ctx *hctx;
1174 struct blk_mq_ctx *ctx;
1175};
1176
1177static struct request *blk_mq_map_request(struct request_queue *q,
1178 struct bio *bio,
1179 struct blk_map_ctx *data)
1180{
1181 struct blk_mq_hw_ctx *hctx;
1182 struct blk_mq_ctx *ctx;
1183 struct request *rq;
1184 int rw = bio_data_dir(bio);
cb96a42c 1185 struct blk_mq_alloc_data alloc_data;
320ae51f 1186
bfd343aa 1187 if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
4246a0b6 1188 bio_io_error(bio);
07068d5b 1189 return NULL;
320ae51f
JA
1190 }
1191
1192 ctx = blk_mq_get_ctx(q);
1193 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1194
07068d5b 1195 if (rw_is_sync(bio->bi_rw))
27fbf4e8 1196 rw |= REQ_SYNC;
07068d5b 1197
320ae51f 1198 trace_block_getrq(q, bio, rw);
cb96a42c
ML
1199 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1200 hctx);
1201 rq = __blk_mq_alloc_request(&alloc_data, rw);
5dee8577 1202 if (unlikely(!rq)) {
793597a6 1203 __blk_mq_run_hw_queue(hctx);
320ae51f
JA
1204 blk_mq_put_ctx(ctx);
1205 trace_block_sleeprq(q, bio, rw);
793597a6
CH
1206
1207 ctx = blk_mq_get_ctx(q);
320ae51f 1208 hctx = q->mq_ops->map_queue(q, ctx->cpu);
cb96a42c
ML
1209 blk_mq_set_alloc_data(&alloc_data, q,
1210 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1211 rq = __blk_mq_alloc_request(&alloc_data, rw);
1212 ctx = alloc_data.ctx;
1213 hctx = alloc_data.hctx;
320ae51f
JA
1214 }
1215
1216 hctx->queued++;
07068d5b
JA
1217 data->hctx = hctx;
1218 data->ctx = ctx;
1219 return rq;
1220}
1221
f984df1f
SL
1222static int blk_mq_direct_issue_request(struct request *rq)
1223{
1224 int ret;
1225 struct request_queue *q = rq->q;
1226 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1227 rq->mq_ctx->cpu);
1228 struct blk_mq_queue_data bd = {
1229 .rq = rq,
1230 .list = NULL,
1231 .last = 1
1232 };
1233
1234 /*
1235 * For OK queue, we are done. For error, kill it. Any other
1236 * error (busy), just add it to our list as we previously
1237 * would have done
1238 */
1239 ret = q->mq_ops->queue_rq(hctx, &bd);
1240 if (ret == BLK_MQ_RQ_QUEUE_OK)
1241 return 0;
1242 else {
1243 __blk_mq_requeue_request(rq);
1244
1245 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1246 rq->errors = -EIO;
1247 blk_mq_end_request(rq, rq->errors);
1248 return 0;
1249 }
1250 return -1;
1251 }
1252}
1253
07068d5b
JA
1254/*
1255 * Multiple hardware queue variant. This will not use per-process plugs,
1256 * but will attempt to bypass the hctx queueing if we can go straight to
1257 * hardware for SYNC IO.
1258 */
1259static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1260{
1261 const int is_sync = rw_is_sync(bio->bi_rw);
1262 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1263 struct blk_map_ctx data;
1264 struct request *rq;
f984df1f
SL
1265 unsigned int request_count = 0;
1266 struct blk_plug *plug;
5b3f341f 1267 struct request *same_queue_rq = NULL;
07068d5b
JA
1268
1269 blk_queue_bounce(q, &bio);
1270
1271 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
4246a0b6 1272 bio_io_error(bio);
07068d5b
JA
1273 return;
1274 }
1275
54efd50b
KO
1276 blk_queue_split(q, &bio, q->bio_split);
1277
f984df1f 1278 if (!is_flush_fua && !blk_queue_nomerges(q) &&
5b3f341f 1279 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
f984df1f
SL
1280 return;
1281
07068d5b
JA
1282 rq = blk_mq_map_request(q, bio, &data);
1283 if (unlikely(!rq))
1284 return;
1285
1286 if (unlikely(is_flush_fua)) {
1287 blk_mq_bio_to_request(rq, bio);
1288 blk_insert_flush(rq);
1289 goto run_queue;
1290 }
1291
f984df1f 1292 plug = current->plug;
e167dfb5
JA
1293 /*
1294 * If the driver supports defer issued based on 'last', then
1295 * queue it up like normal since we can potentially save some
1296 * CPU this way.
1297 */
f984df1f
SL
1298 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1299 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1300 struct request *old_rq = NULL;
07068d5b
JA
1301
1302 blk_mq_bio_to_request(rq, bio);
07068d5b
JA
1303
1304 /*
f984df1f
SL
1305 * we do limited pluging. If bio can be merged, do merge.
1306 * Otherwise the existing request in the plug list will be
1307 * issued. So the plug list will have one request at most
07068d5b 1308 */
f984df1f 1309 if (plug) {
5b3f341f
SL
1310 /*
1311 * The plug list might get flushed before this. If that
1312 * happens, same_queue_rq is invalid and plug list is empty
1313 **/
1314 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1315 old_rq = same_queue_rq;
f984df1f 1316 list_del_init(&old_rq->queuelist);
07068d5b 1317 }
f984df1f
SL
1318 list_add_tail(&rq->queuelist, &plug->mq_list);
1319 } else /* is_sync */
1320 old_rq = rq;
1321 blk_mq_put_ctx(data.ctx);
1322 if (!old_rq)
239ad215 1323 return;
f984df1f
SL
1324 if (!blk_mq_direct_issue_request(old_rq))
1325 return;
1326 blk_mq_insert_request(old_rq, false, true, true);
1327 return;
07068d5b
JA
1328 }
1329
1330 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1331 /*
1332 * For a SYNC request, send it to the hardware immediately. For
1333 * an ASYNC request, just ensure that we run it later on. The
1334 * latter allows for merging opportunities and more efficient
1335 * dispatching.
1336 */
1337run_queue:
1338 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1339 }
07068d5b
JA
1340 blk_mq_put_ctx(data.ctx);
1341}
1342
1343/*
1344 * Single hardware queue variant. This will attempt to use any per-process
1345 * plug for merging and IO deferral.
1346 */
1347static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1348{
1349 const int is_sync = rw_is_sync(bio->bi_rw);
1350 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
e6c4438b
JM
1351 struct blk_plug *plug;
1352 unsigned int request_count = 0;
07068d5b
JA
1353 struct blk_map_ctx data;
1354 struct request *rq;
1355
07068d5b
JA
1356 blk_queue_bounce(q, &bio);
1357
1358 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
4246a0b6 1359 bio_io_error(bio);
07068d5b
JA
1360 return;
1361 }
1362
54efd50b
KO
1363 blk_queue_split(q, &bio, q->bio_split);
1364
e6c4438b 1365 if (!is_flush_fua && !blk_queue_nomerges(q) &&
5b3f341f 1366 blk_attempt_plug_merge(q, bio, &request_count, NULL))
07068d5b
JA
1367 return;
1368
1369 rq = blk_mq_map_request(q, bio, &data);
ff87bcec
JA
1370 if (unlikely(!rq))
1371 return;
320ae51f
JA
1372
1373 if (unlikely(is_flush_fua)) {
1374 blk_mq_bio_to_request(rq, bio);
320ae51f
JA
1375 blk_insert_flush(rq);
1376 goto run_queue;
1377 }
1378
1379 /*
1380 * A task plug currently exists. Since this is completely lockless,
1381 * utilize that to temporarily store requests until the task is
1382 * either done or scheduled away.
1383 */
e6c4438b
JM
1384 plug = current->plug;
1385 if (plug) {
1386 blk_mq_bio_to_request(rq, bio);
1387 if (list_empty(&plug->mq_list))
1388 trace_block_plug(q);
1389 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1390 blk_flush_plug_list(plug, false);
1391 trace_block_plug(q);
320ae51f 1392 }
e6c4438b
JM
1393 list_add_tail(&rq->queuelist, &plug->mq_list);
1394 blk_mq_put_ctx(data.ctx);
1395 return;
320ae51f
JA
1396 }
1397
07068d5b
JA
1398 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1399 /*
1400 * For a SYNC request, send it to the hardware immediately. For
1401 * an ASYNC request, just ensure that we run it later on. The
1402 * latter allows for merging opportunities and more efficient
1403 * dispatching.
1404 */
1405run_queue:
1406 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
320ae51f
JA
1407 }
1408
07068d5b 1409 blk_mq_put_ctx(data.ctx);
320ae51f
JA
1410}
1411
1412/*
1413 * Default mapping to a software queue, since we use one per CPU.
1414 */
1415struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1416{
1417 return q->queue_hw_ctx[q->mq_map[cpu]];
1418}
1419EXPORT_SYMBOL(blk_mq_map_queue);
1420
24d2f903
CH
1421static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1422 struct blk_mq_tags *tags, unsigned int hctx_idx)
95363efd 1423{
e9b267d9 1424 struct page *page;
320ae51f 1425
24d2f903 1426 if (tags->rqs && set->ops->exit_request) {
e9b267d9 1427 int i;
320ae51f 1428
24d2f903
CH
1429 for (i = 0; i < tags->nr_tags; i++) {
1430 if (!tags->rqs[i])
e9b267d9 1431 continue;
24d2f903
CH
1432 set->ops->exit_request(set->driver_data, tags->rqs[i],
1433 hctx_idx, i);
a5164405 1434 tags->rqs[i] = NULL;
e9b267d9 1435 }
320ae51f 1436 }
320ae51f 1437
24d2f903
CH
1438 while (!list_empty(&tags->page_list)) {
1439 page = list_first_entry(&tags->page_list, struct page, lru);
6753471c 1440 list_del_init(&page->lru);
320ae51f
JA
1441 __free_pages(page, page->private);
1442 }
1443
24d2f903 1444 kfree(tags->rqs);
320ae51f 1445
24d2f903 1446 blk_mq_free_tags(tags);
320ae51f
JA
1447}
1448
1449static size_t order_to_size(unsigned int order)
1450{
4ca08500 1451 return (size_t)PAGE_SIZE << order;
320ae51f
JA
1452}
1453
24d2f903
CH
1454static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1455 unsigned int hctx_idx)
320ae51f 1456{
24d2f903 1457 struct blk_mq_tags *tags;
320ae51f
JA
1458 unsigned int i, j, entries_per_page, max_order = 4;
1459 size_t rq_size, left;
1460
24d2f903 1461 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
24391c0d
SL
1462 set->numa_node,
1463 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
24d2f903
CH
1464 if (!tags)
1465 return NULL;
320ae51f 1466
24d2f903
CH
1467 INIT_LIST_HEAD(&tags->page_list);
1468
a5164405
JA
1469 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1470 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1471 set->numa_node);
24d2f903
CH
1472 if (!tags->rqs) {
1473 blk_mq_free_tags(tags);
1474 return NULL;
1475 }
320ae51f
JA
1476
1477 /*
1478 * rq_size is the size of the request plus driver payload, rounded
1479 * to the cacheline size
1480 */
24d2f903 1481 rq_size = round_up(sizeof(struct request) + set->cmd_size,
320ae51f 1482 cache_line_size());
24d2f903 1483 left = rq_size * set->queue_depth;
320ae51f 1484
24d2f903 1485 for (i = 0; i < set->queue_depth; ) {
320ae51f
JA
1486 int this_order = max_order;
1487 struct page *page;
1488 int to_do;
1489 void *p;
1490
1491 while (left < order_to_size(this_order - 1) && this_order)
1492 this_order--;
1493
1494 do {
a5164405 1495 page = alloc_pages_node(set->numa_node,
ac211175 1496 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
a5164405 1497 this_order);
320ae51f
JA
1498 if (page)
1499 break;
1500 if (!this_order--)
1501 break;
1502 if (order_to_size(this_order) < rq_size)
1503 break;
1504 } while (1);
1505
1506 if (!page)
24d2f903 1507 goto fail;
320ae51f
JA
1508
1509 page->private = this_order;
24d2f903 1510 list_add_tail(&page->lru, &tags->page_list);
320ae51f
JA
1511
1512 p = page_address(page);
1513 entries_per_page = order_to_size(this_order) / rq_size;
24d2f903 1514 to_do = min(entries_per_page, set->queue_depth - i);
320ae51f
JA
1515 left -= to_do * rq_size;
1516 for (j = 0; j < to_do; j++) {
24d2f903
CH
1517 tags->rqs[i] = p;
1518 if (set->ops->init_request) {
1519 if (set->ops->init_request(set->driver_data,
1520 tags->rqs[i], hctx_idx, i,
a5164405
JA
1521 set->numa_node)) {
1522 tags->rqs[i] = NULL;
24d2f903 1523 goto fail;
a5164405 1524 }
e9b267d9
CH
1525 }
1526
320ae51f
JA
1527 p += rq_size;
1528 i++;
1529 }
1530 }
24d2f903 1531 return tags;
320ae51f 1532
24d2f903 1533fail:
24d2f903
CH
1534 blk_mq_free_rq_map(set, tags, hctx_idx);
1535 return NULL;
320ae51f
JA
1536}
1537
1429d7c9
JA
1538static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1539{
1540 kfree(bitmap->map);
1541}
1542
1543static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1544{
1545 unsigned int bpw = 8, total, num_maps, i;
1546
1547 bitmap->bits_per_word = bpw;
1548
1549 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1550 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1551 GFP_KERNEL, node);
1552 if (!bitmap->map)
1553 return -ENOMEM;
1554
1429d7c9
JA
1555 total = nr_cpu_ids;
1556 for (i = 0; i < num_maps; i++) {
1557 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1558 total -= bitmap->map[i].depth;
1559 }
1560
1561 return 0;
1562}
1563
484b4061
JA
1564static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1565{
1566 struct request_queue *q = hctx->queue;
1567 struct blk_mq_ctx *ctx;
1568 LIST_HEAD(tmp);
1569
1570 /*
1571 * Move ctx entries to new CPU, if this one is going away.
1572 */
1573 ctx = __blk_mq_get_ctx(q, cpu);
1574
1575 spin_lock(&ctx->lock);
1576 if (!list_empty(&ctx->rq_list)) {
1577 list_splice_init(&ctx->rq_list, &tmp);
1578 blk_mq_hctx_clear_pending(hctx, ctx);
1579 }
1580 spin_unlock(&ctx->lock);
1581
1582 if (list_empty(&tmp))
1583 return NOTIFY_OK;
1584
1585 ctx = blk_mq_get_ctx(q);
1586 spin_lock(&ctx->lock);
1587
1588 while (!list_empty(&tmp)) {
1589 struct request *rq;
1590
1591 rq = list_first_entry(&tmp, struct request, queuelist);
1592 rq->mq_ctx = ctx;
1593 list_move_tail(&rq->queuelist, &ctx->rq_list);
1594 }
1595
1596 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1597 blk_mq_hctx_mark_pending(hctx, ctx);
1598
1599 spin_unlock(&ctx->lock);
1600
1601 blk_mq_run_hw_queue(hctx, true);
1602 blk_mq_put_ctx(ctx);
1603 return NOTIFY_OK;
1604}
1605
484b4061
JA
1606static int blk_mq_hctx_notify(void *data, unsigned long action,
1607 unsigned int cpu)
1608{
1609 struct blk_mq_hw_ctx *hctx = data;
1610
1611 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1612 return blk_mq_hctx_cpu_offline(hctx, cpu);
2a34c087
ML
1613
1614 /*
1615 * In case of CPU online, tags may be reallocated
1616 * in blk_mq_map_swqueue() after mapping is updated.
1617 */
484b4061
JA
1618
1619 return NOTIFY_OK;
1620}
1621
c3b4afca 1622/* hctx->ctxs will be freed in queue's release handler */
08e98fc6
ML
1623static void blk_mq_exit_hctx(struct request_queue *q,
1624 struct blk_mq_tag_set *set,
1625 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1626{
f70ced09
ML
1627 unsigned flush_start_tag = set->queue_depth;
1628
08e98fc6
ML
1629 blk_mq_tag_idle(hctx);
1630
f70ced09
ML
1631 if (set->ops->exit_request)
1632 set->ops->exit_request(set->driver_data,
1633 hctx->fq->flush_rq, hctx_idx,
1634 flush_start_tag + hctx_idx);
1635
08e98fc6
ML
1636 if (set->ops->exit_hctx)
1637 set->ops->exit_hctx(hctx, hctx_idx);
1638
1639 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
f70ced09 1640 blk_free_flush_queue(hctx->fq);
08e98fc6
ML
1641 blk_mq_free_bitmap(&hctx->ctx_map);
1642}
1643
624dbe47
ML
1644static void blk_mq_exit_hw_queues(struct request_queue *q,
1645 struct blk_mq_tag_set *set, int nr_queue)
1646{
1647 struct blk_mq_hw_ctx *hctx;
1648 unsigned int i;
1649
1650 queue_for_each_hw_ctx(q, hctx, i) {
1651 if (i == nr_queue)
1652 break;
08e98fc6 1653 blk_mq_exit_hctx(q, set, hctx, i);
624dbe47 1654 }
624dbe47
ML
1655}
1656
1657static void blk_mq_free_hw_queues(struct request_queue *q,
1658 struct blk_mq_tag_set *set)
1659{
1660 struct blk_mq_hw_ctx *hctx;
1661 unsigned int i;
1662
e09aae7e 1663 queue_for_each_hw_ctx(q, hctx, i)
624dbe47 1664 free_cpumask_var(hctx->cpumask);
624dbe47
ML
1665}
1666
08e98fc6
ML
1667static int blk_mq_init_hctx(struct request_queue *q,
1668 struct blk_mq_tag_set *set,
1669 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
320ae51f 1670{
08e98fc6 1671 int node;
f70ced09 1672 unsigned flush_start_tag = set->queue_depth;
08e98fc6
ML
1673
1674 node = hctx->numa_node;
1675 if (node == NUMA_NO_NODE)
1676 node = hctx->numa_node = set->numa_node;
1677
1678 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1679 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1680 spin_lock_init(&hctx->lock);
1681 INIT_LIST_HEAD(&hctx->dispatch);
1682 hctx->queue = q;
1683 hctx->queue_num = hctx_idx;
1684 hctx->flags = set->flags;
08e98fc6
ML
1685
1686 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1687 blk_mq_hctx_notify, hctx);
1688 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1689
1690 hctx->tags = set->tags[hctx_idx];
320ae51f
JA
1691
1692 /*
08e98fc6
ML
1693 * Allocate space for all possible cpus to avoid allocation at
1694 * runtime
320ae51f 1695 */
08e98fc6
ML
1696 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1697 GFP_KERNEL, node);
1698 if (!hctx->ctxs)
1699 goto unregister_cpu_notifier;
320ae51f 1700
08e98fc6
ML
1701 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1702 goto free_ctxs;
320ae51f 1703
08e98fc6 1704 hctx->nr_ctx = 0;
320ae51f 1705
08e98fc6
ML
1706 if (set->ops->init_hctx &&
1707 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1708 goto free_bitmap;
320ae51f 1709
f70ced09
ML
1710 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1711 if (!hctx->fq)
1712 goto exit_hctx;
320ae51f 1713
f70ced09
ML
1714 if (set->ops->init_request &&
1715 set->ops->init_request(set->driver_data,
1716 hctx->fq->flush_rq, hctx_idx,
1717 flush_start_tag + hctx_idx, node))
1718 goto free_fq;
320ae51f 1719
08e98fc6 1720 return 0;
320ae51f 1721
f70ced09
ML
1722 free_fq:
1723 kfree(hctx->fq);
1724 exit_hctx:
1725 if (set->ops->exit_hctx)
1726 set->ops->exit_hctx(hctx, hctx_idx);
08e98fc6
ML
1727 free_bitmap:
1728 blk_mq_free_bitmap(&hctx->ctx_map);
1729 free_ctxs:
1730 kfree(hctx->ctxs);
1731 unregister_cpu_notifier:
1732 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
320ae51f 1733
08e98fc6
ML
1734 return -1;
1735}
320ae51f 1736
08e98fc6
ML
1737static int blk_mq_init_hw_queues(struct request_queue *q,
1738 struct blk_mq_tag_set *set)
1739{
1740 struct blk_mq_hw_ctx *hctx;
1741 unsigned int i;
320ae51f 1742
08e98fc6
ML
1743 /*
1744 * Initialize hardware queues
1745 */
1746 queue_for_each_hw_ctx(q, hctx, i) {
1747 if (blk_mq_init_hctx(q, set, hctx, i))
320ae51f
JA
1748 break;
1749 }
1750
1751 if (i == q->nr_hw_queues)
1752 return 0;
1753
1754 /*
1755 * Init failed
1756 */
624dbe47 1757 blk_mq_exit_hw_queues(q, set, i);
320ae51f
JA
1758
1759 return 1;
1760}
1761
1762static void blk_mq_init_cpu_queues(struct request_queue *q,
1763 unsigned int nr_hw_queues)
1764{
1765 unsigned int i;
1766
1767 for_each_possible_cpu(i) {
1768 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1769 struct blk_mq_hw_ctx *hctx;
1770
1771 memset(__ctx, 0, sizeof(*__ctx));
1772 __ctx->cpu = i;
1773 spin_lock_init(&__ctx->lock);
1774 INIT_LIST_HEAD(&__ctx->rq_list);
1775 __ctx->queue = q;
1776
1777 /* If the cpu isn't online, the cpu is mapped to first hctx */
320ae51f
JA
1778 if (!cpu_online(i))
1779 continue;
1780
e4043dcf 1781 hctx = q->mq_ops->map_queue(q, i);
e4043dcf 1782
320ae51f
JA
1783 /*
1784 * Set local node, IFF we have more than one hw queue. If
1785 * not, we remain on the home node of the device
1786 */
1787 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1788 hctx->numa_node = cpu_to_node(i);
1789 }
1790}
1791
1792static void blk_mq_map_swqueue(struct request_queue *q)
1793{
1794 unsigned int i;
1795 struct blk_mq_hw_ctx *hctx;
1796 struct blk_mq_ctx *ctx;
2a34c087 1797 struct blk_mq_tag_set *set = q->tag_set;
320ae51f
JA
1798
1799 queue_for_each_hw_ctx(q, hctx, i) {
e4043dcf 1800 cpumask_clear(hctx->cpumask);
320ae51f
JA
1801 hctx->nr_ctx = 0;
1802 }
1803
1804 /*
1805 * Map software to hardware queues
1806 */
1807 queue_for_each_ctx(q, ctx, i) {
1808 /* If the cpu isn't online, the cpu is mapped to first hctx */
e4043dcf
JA
1809 if (!cpu_online(i))
1810 continue;
1811
320ae51f 1812 hctx = q->mq_ops->map_queue(q, i);
e4043dcf 1813 cpumask_set_cpu(i, hctx->cpumask);
f26cdc85 1814 cpumask_set_cpu(i, hctx->tags->cpumask);
320ae51f
JA
1815 ctx->index_hw = hctx->nr_ctx;
1816 hctx->ctxs[hctx->nr_ctx++] = ctx;
1817 }
506e931f
JA
1818
1819 queue_for_each_hw_ctx(q, hctx, i) {
889fa31f
CY
1820 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1821
484b4061 1822 /*
a68aafa5
JA
1823 * If no software queues are mapped to this hardware queue,
1824 * disable it and free the request entries.
484b4061
JA
1825 */
1826 if (!hctx->nr_ctx) {
484b4061
JA
1827 if (set->tags[i]) {
1828 blk_mq_free_rq_map(set, set->tags[i], i);
1829 set->tags[i] = NULL;
484b4061 1830 }
2a34c087 1831 hctx->tags = NULL;
484b4061
JA
1832 continue;
1833 }
1834
2a34c087
ML
1835 /* unmapped hw queue can be remapped after CPU topo changed */
1836 if (!set->tags[i])
1837 set->tags[i] = blk_mq_init_rq_map(set, i);
1838 hctx->tags = set->tags[i];
1839 WARN_ON(!hctx->tags);
1840
889fa31f
CY
1841 /*
1842 * Set the map size to the number of mapped software queues.
1843 * This is more accurate and more efficient than looping
1844 * over all possibly mapped software queues.
1845 */
569fd0ce 1846 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
889fa31f 1847
484b4061
JA
1848 /*
1849 * Initialize batch roundrobin counts
1850 */
506e931f
JA
1851 hctx->next_cpu = cpumask_first(hctx->cpumask);
1852 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1853 }
320ae51f
JA
1854}
1855
0d2602ca
JA
1856static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1857{
1858 struct blk_mq_hw_ctx *hctx;
1859 struct request_queue *q;
1860 bool shared;
1861 int i;
1862
1863 if (set->tag_list.next == set->tag_list.prev)
1864 shared = false;
1865 else
1866 shared = true;
1867
1868 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1869 blk_mq_freeze_queue(q);
1870
1871 queue_for_each_hw_ctx(q, hctx, i) {
1872 if (shared)
1873 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1874 else
1875 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1876 }
1877 blk_mq_unfreeze_queue(q);
1878 }
1879}
1880
1881static void blk_mq_del_queue_tag_set(struct request_queue *q)
1882{
1883 struct blk_mq_tag_set *set = q->tag_set;
1884
0d2602ca
JA
1885 mutex_lock(&set->tag_list_lock);
1886 list_del_init(&q->tag_set_list);
1887 blk_mq_update_tag_set_depth(set);
1888 mutex_unlock(&set->tag_list_lock);
0d2602ca
JA
1889}
1890
1891static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1892 struct request_queue *q)
1893{
1894 q->tag_set = set;
1895
1896 mutex_lock(&set->tag_list_lock);
1897 list_add_tail(&q->tag_set_list, &set->tag_list);
1898 blk_mq_update_tag_set_depth(set);
1899 mutex_unlock(&set->tag_list_lock);
1900}
1901
e09aae7e
ML
1902/*
1903 * It is the actual release handler for mq, but we do it from
1904 * request queue's release handler for avoiding use-after-free
1905 * and headache because q->mq_kobj shouldn't have been introduced,
1906 * but we can't group ctx/kctx kobj without it.
1907 */
1908void blk_mq_release(struct request_queue *q)
1909{
1910 struct blk_mq_hw_ctx *hctx;
1911 unsigned int i;
1912
1913 /* hctx kobj stays in hctx */
c3b4afca
ML
1914 queue_for_each_hw_ctx(q, hctx, i) {
1915 if (!hctx)
1916 continue;
1917 kfree(hctx->ctxs);
e09aae7e 1918 kfree(hctx);
c3b4afca 1919 }
e09aae7e
ML
1920
1921 kfree(q->queue_hw_ctx);
1922
1923 /* ctx kobj stays in queue_ctx */
1924 free_percpu(q->queue_ctx);
1925}
1926
24d2f903 1927struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
b62c21b7
MS
1928{
1929 struct request_queue *uninit_q, *q;
1930
1931 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1932 if (!uninit_q)
1933 return ERR_PTR(-ENOMEM);
1934
1935 q = blk_mq_init_allocated_queue(set, uninit_q);
1936 if (IS_ERR(q))
1937 blk_cleanup_queue(uninit_q);
1938
1939 return q;
1940}
1941EXPORT_SYMBOL(blk_mq_init_queue);
1942
1943struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1944 struct request_queue *q)
320ae51f
JA
1945{
1946 struct blk_mq_hw_ctx **hctxs;
e6cdb092 1947 struct blk_mq_ctx __percpu *ctx;
f14bbe77 1948 unsigned int *map;
320ae51f
JA
1949 int i;
1950
320ae51f
JA
1951 ctx = alloc_percpu(struct blk_mq_ctx);
1952 if (!ctx)
1953 return ERR_PTR(-ENOMEM);
1954
24d2f903
CH
1955 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1956 set->numa_node);
320ae51f
JA
1957
1958 if (!hctxs)
1959 goto err_percpu;
1960
f14bbe77
JA
1961 map = blk_mq_make_queue_map(set);
1962 if (!map)
1963 goto err_map;
1964
24d2f903 1965 for (i = 0; i < set->nr_hw_queues; i++) {
f14bbe77
JA
1966 int node = blk_mq_hw_queue_to_node(map, i);
1967
cdef54dd
CH
1968 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1969 GFP_KERNEL, node);
320ae51f
JA
1970 if (!hctxs[i])
1971 goto err_hctxs;
1972
a86073e4
JA
1973 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1974 node))
e4043dcf
JA
1975 goto err_hctxs;
1976
0d2602ca 1977 atomic_set(&hctxs[i]->nr_active, 0);
f14bbe77 1978 hctxs[i]->numa_node = node;
320ae51f
JA
1979 hctxs[i]->queue_num = i;
1980 }
1981
17497acb
TH
1982 /*
1983 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1984 * See blk_register_queue() for details.
1985 */
a34375ef 1986 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
17497acb 1987 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
b62c21b7 1988 goto err_hctxs;
3d2936f4 1989
320ae51f 1990 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
e56f698b 1991 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
320ae51f
JA
1992
1993 q->nr_queues = nr_cpu_ids;
24d2f903 1994 q->nr_hw_queues = set->nr_hw_queues;
f14bbe77 1995 q->mq_map = map;
320ae51f
JA
1996
1997 q->queue_ctx = ctx;
1998 q->queue_hw_ctx = hctxs;
1999
24d2f903 2000 q->mq_ops = set->ops;
94eddfbe 2001 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
320ae51f 2002
05f1dd53
JA
2003 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2004 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2005
1be036e9
CH
2006 q->sg_reserved_size = INT_MAX;
2007
6fca6a61
CH
2008 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2009 INIT_LIST_HEAD(&q->requeue_list);
2010 spin_lock_init(&q->requeue_lock);
2011
07068d5b
JA
2012 if (q->nr_hw_queues > 1)
2013 blk_queue_make_request(q, blk_mq_make_request);
2014 else
2015 blk_queue_make_request(q, blk_sq_make_request);
2016
eba71768
JA
2017 /*
2018 * Do this after blk_queue_make_request() overrides it...
2019 */
2020 q->nr_requests = set->queue_depth;
2021
24d2f903
CH
2022 if (set->ops->complete)
2023 blk_queue_softirq_done(q, set->ops->complete);
30a91cb4 2024
24d2f903 2025 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
320ae51f 2026
24d2f903 2027 if (blk_mq_init_hw_queues(q, set))
b62c21b7 2028 goto err_hctxs;
18741986 2029
320ae51f
JA
2030 mutex_lock(&all_q_mutex);
2031 list_add_tail(&q->all_q_node, &all_q_list);
2032 mutex_unlock(&all_q_mutex);
2033
0d2602ca
JA
2034 blk_mq_add_queue_tag_set(set, q);
2035
484b4061
JA
2036 blk_mq_map_swqueue(q);
2037
320ae51f 2038 return q;
18741986 2039
320ae51f 2040err_hctxs:
f14bbe77 2041 kfree(map);
24d2f903 2042 for (i = 0; i < set->nr_hw_queues; i++) {
320ae51f
JA
2043 if (!hctxs[i])
2044 break;
e4043dcf 2045 free_cpumask_var(hctxs[i]->cpumask);
cdef54dd 2046 kfree(hctxs[i]);
320ae51f 2047 }
f14bbe77 2048err_map:
320ae51f
JA
2049 kfree(hctxs);
2050err_percpu:
2051 free_percpu(ctx);
2052 return ERR_PTR(-ENOMEM);
2053}
b62c21b7 2054EXPORT_SYMBOL(blk_mq_init_allocated_queue);
320ae51f
JA
2055
2056void blk_mq_free_queue(struct request_queue *q)
2057{
624dbe47 2058 struct blk_mq_tag_set *set = q->tag_set;
320ae51f 2059
0d2602ca
JA
2060 blk_mq_del_queue_tag_set(q);
2061
624dbe47
ML
2062 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2063 blk_mq_free_hw_queues(q, set);
320ae51f 2064
add703fd 2065 percpu_ref_exit(&q->mq_usage_counter);
3d2936f4 2066
320ae51f
JA
2067 kfree(q->mq_map);
2068
320ae51f
JA
2069 q->mq_map = NULL;
2070
2071 mutex_lock(&all_q_mutex);
2072 list_del_init(&q->all_q_node);
2073 mutex_unlock(&all_q_mutex);
2074}
320ae51f
JA
2075
2076/* Basically redo blk_mq_init_queue with queue frozen */
f618ef7c 2077static void blk_mq_queue_reinit(struct request_queue *q)
320ae51f 2078{
4ecd4fef 2079 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
320ae51f 2080
67aec14c
JA
2081 blk_mq_sysfs_unregister(q);
2082
320ae51f
JA
2083 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2084
2085 /*
2086 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2087 * we should change hctx numa_node according to new topology (this
2088 * involves free and re-allocate memory, worthy doing?)
2089 */
2090
2091 blk_mq_map_swqueue(q);
2092
67aec14c 2093 blk_mq_sysfs_register(q);
320ae51f
JA
2094}
2095
f618ef7c
PG
2096static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2097 unsigned long action, void *hcpu)
320ae51f
JA
2098{
2099 struct request_queue *q;
2100
2101 /*
9fccfed8
JA
2102 * Before new mappings are established, hotadded cpu might already
2103 * start handling requests. This doesn't break anything as we map
2104 * offline CPUs to first hardware queue. We will re-init the queue
2105 * below to get optimal settings.
320ae51f
JA
2106 */
2107 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2108 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2109 return NOTIFY_OK;
2110
2111 mutex_lock(&all_q_mutex);
f3af020b
TH
2112
2113 /*
2114 * We need to freeze and reinit all existing queues. Freezing
2115 * involves synchronous wait for an RCU grace period and doing it
2116 * one by one may take a long time. Start freezing all queues in
2117 * one swoop and then wait for the completions so that freezing can
2118 * take place in parallel.
2119 */
2120 list_for_each_entry(q, &all_q_list, all_q_node)
2121 blk_mq_freeze_queue_start(q);
f054b56c 2122 list_for_each_entry(q, &all_q_list, all_q_node) {
f3af020b
TH
2123 blk_mq_freeze_queue_wait(q);
2124
f054b56c
ML
2125 /*
2126 * timeout handler can't touch hw queue during the
2127 * reinitialization
2128 */
2129 del_timer_sync(&q->timeout);
2130 }
2131
320ae51f
JA
2132 list_for_each_entry(q, &all_q_list, all_q_node)
2133 blk_mq_queue_reinit(q);
f3af020b
TH
2134
2135 list_for_each_entry(q, &all_q_list, all_q_node)
2136 blk_mq_unfreeze_queue(q);
2137
320ae51f
JA
2138 mutex_unlock(&all_q_mutex);
2139 return NOTIFY_OK;
2140}
2141
a5164405
JA
2142static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2143{
2144 int i;
2145
2146 for (i = 0; i < set->nr_hw_queues; i++) {
2147 set->tags[i] = blk_mq_init_rq_map(set, i);
2148 if (!set->tags[i])
2149 goto out_unwind;
2150 }
2151
2152 return 0;
2153
2154out_unwind:
2155 while (--i >= 0)
2156 blk_mq_free_rq_map(set, set->tags[i], i);
2157
a5164405
JA
2158 return -ENOMEM;
2159}
2160
2161/*
2162 * Allocate the request maps associated with this tag_set. Note that this
2163 * may reduce the depth asked for, if memory is tight. set->queue_depth
2164 * will be updated to reflect the allocated depth.
2165 */
2166static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2167{
2168 unsigned int depth;
2169 int err;
2170
2171 depth = set->queue_depth;
2172 do {
2173 err = __blk_mq_alloc_rq_maps(set);
2174 if (!err)
2175 break;
2176
2177 set->queue_depth >>= 1;
2178 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2179 err = -ENOMEM;
2180 break;
2181 }
2182 } while (set->queue_depth);
2183
2184 if (!set->queue_depth || err) {
2185 pr_err("blk-mq: failed to allocate request map\n");
2186 return -ENOMEM;
2187 }
2188
2189 if (depth != set->queue_depth)
2190 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2191 depth, set->queue_depth);
2192
2193 return 0;
2194}
2195
f26cdc85
KB
2196struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2197{
2198 return tags->cpumask;
2199}
2200EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2201
a4391c64
JA
2202/*
2203 * Alloc a tag set to be associated with one or more request queues.
2204 * May fail with EINVAL for various error conditions. May adjust the
2205 * requested depth down, if if it too large. In that case, the set
2206 * value will be stored in set->queue_depth.
2207 */
24d2f903
CH
2208int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2209{
205fb5f5
BVA
2210 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2211
24d2f903
CH
2212 if (!set->nr_hw_queues)
2213 return -EINVAL;
a4391c64 2214 if (!set->queue_depth)
24d2f903
CH
2215 return -EINVAL;
2216 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2217 return -EINVAL;
2218
f9018ac9 2219 if (!set->ops->queue_rq || !set->ops->map_queue)
24d2f903
CH
2220 return -EINVAL;
2221
a4391c64
JA
2222 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2223 pr_info("blk-mq: reduced tag depth to %u\n",
2224 BLK_MQ_MAX_DEPTH);
2225 set->queue_depth = BLK_MQ_MAX_DEPTH;
2226 }
24d2f903 2227
6637fadf
SL
2228 /*
2229 * If a crashdump is active, then we are potentially in a very
2230 * memory constrained environment. Limit us to 1 queue and
2231 * 64 tags to prevent using too much memory.
2232 */
2233 if (is_kdump_kernel()) {
2234 set->nr_hw_queues = 1;
2235 set->queue_depth = min(64U, set->queue_depth);
2236 }
2237
48479005
ML
2238 set->tags = kmalloc_node(set->nr_hw_queues *
2239 sizeof(struct blk_mq_tags *),
24d2f903
CH
2240 GFP_KERNEL, set->numa_node);
2241 if (!set->tags)
a5164405 2242 return -ENOMEM;
24d2f903 2243
a5164405
JA
2244 if (blk_mq_alloc_rq_maps(set))
2245 goto enomem;
24d2f903 2246
0d2602ca
JA
2247 mutex_init(&set->tag_list_lock);
2248 INIT_LIST_HEAD(&set->tag_list);
2249
24d2f903 2250 return 0;
a5164405 2251enomem:
5676e7b6
RE
2252 kfree(set->tags);
2253 set->tags = NULL;
24d2f903
CH
2254 return -ENOMEM;
2255}
2256EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2257
2258void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2259{
2260 int i;
2261
484b4061 2262 for (i = 0; i < set->nr_hw_queues; i++) {
f26cdc85 2263 if (set->tags[i]) {
484b4061 2264 blk_mq_free_rq_map(set, set->tags[i], i);
f26cdc85
KB
2265 free_cpumask_var(set->tags[i]->cpumask);
2266 }
484b4061
JA
2267 }
2268
981bd189 2269 kfree(set->tags);
5676e7b6 2270 set->tags = NULL;
24d2f903
CH
2271}
2272EXPORT_SYMBOL(blk_mq_free_tag_set);
2273
e3a2b3f9
JA
2274int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2275{
2276 struct blk_mq_tag_set *set = q->tag_set;
2277 struct blk_mq_hw_ctx *hctx;
2278 int i, ret;
2279
2280 if (!set || nr > set->queue_depth)
2281 return -EINVAL;
2282
2283 ret = 0;
2284 queue_for_each_hw_ctx(q, hctx, i) {
2285 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2286 if (ret)
2287 break;
2288 }
2289
2290 if (!ret)
2291 q->nr_requests = nr;
2292
2293 return ret;
2294}
2295
676141e4
JA
2296void blk_mq_disable_hotplug(void)
2297{
2298 mutex_lock(&all_q_mutex);
2299}
2300
2301void blk_mq_enable_hotplug(void)
2302{
2303 mutex_unlock(&all_q_mutex);
2304}
2305
320ae51f
JA
2306static int __init blk_mq_init(void)
2307{
320ae51f
JA
2308 blk_mq_cpu_init();
2309
add703fd 2310 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
320ae51f
JA
2311
2312 return 0;
2313}
2314subsys_initcall(blk_mq_init);