211ef367345f270f300cb5cdcb9a63c0a50d2420
[karo-tx-linux.git] / block / blk-mq.c
1 /*
2  * Block multiqueue core code
3  *
4  * Copyright (C) 2013-2014 Jens Axboe
5  * Copyright (C) 2013-2014 Christoph Hellwig
6  */
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/kmemleak.h>
13 #include <linux/mm.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
28
29 #include <trace/events/block.h>
30
31 #include <linux/blk-mq.h>
32 #include "blk.h"
33 #include "blk-mq.h"
34 #include "blk-mq-debugfs.h"
35 #include "blk-mq-tag.h"
36 #include "blk-stat.h"
37 #include "blk-wbt.h"
38 #include "blk-mq-sched.h"
39
40 static void blk_mq_poll_stats_start(struct request_queue *q);
41 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
42
43 static int blk_mq_poll_stats_bkt(const struct request *rq)
44 {
45         int ddir, bytes, bucket;
46
47         ddir = rq_data_dir(rq);
48         bytes = blk_rq_bytes(rq);
49
50         bucket = ddir + 2*(ilog2(bytes) - 9);
51
52         if (bucket < 0)
53                 return -1;
54         else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
55                 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
56
57         return bucket;
58 }
59
60 /*
61  * Check if any of the ctx's have pending work in this hardware queue
62  */
63 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
64 {
65         return sbitmap_any_bit_set(&hctx->ctx_map) ||
66                         !list_empty_careful(&hctx->dispatch) ||
67                         blk_mq_sched_has_work(hctx);
68 }
69
70 /*
71  * Mark this ctx as having pending work in this hardware queue
72  */
73 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
74                                      struct blk_mq_ctx *ctx)
75 {
76         if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
77                 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
78 }
79
80 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
81                                       struct blk_mq_ctx *ctx)
82 {
83         sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
84 }
85
86 void blk_freeze_queue_start(struct request_queue *q)
87 {
88         int freeze_depth;
89
90         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
91         if (freeze_depth == 1) {
92                 percpu_ref_kill(&q->q_usage_counter);
93                 blk_mq_run_hw_queues(q, false);
94         }
95 }
96 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
97
98 void blk_mq_freeze_queue_wait(struct request_queue *q)
99 {
100         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
101 }
102 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
103
104 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
105                                      unsigned long timeout)
106 {
107         return wait_event_timeout(q->mq_freeze_wq,
108                                         percpu_ref_is_zero(&q->q_usage_counter),
109                                         timeout);
110 }
111 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
112
113 /*
114  * Guarantee no request is in use, so we can change any data structure of
115  * the queue afterward.
116  */
117 void blk_freeze_queue(struct request_queue *q)
118 {
119         /*
120          * In the !blk_mq case we are only calling this to kill the
121          * q_usage_counter, otherwise this increases the freeze depth
122          * and waits for it to return to zero.  For this reason there is
123          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
124          * exported to drivers as the only user for unfreeze is blk_mq.
125          */
126         blk_freeze_queue_start(q);
127         blk_mq_freeze_queue_wait(q);
128 }
129
130 void blk_mq_freeze_queue(struct request_queue *q)
131 {
132         /*
133          * ...just an alias to keep freeze and unfreeze actions balanced
134          * in the blk_mq_* namespace
135          */
136         blk_freeze_queue(q);
137 }
138 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
139
140 void blk_mq_unfreeze_queue(struct request_queue *q)
141 {
142         int freeze_depth;
143
144         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
145         WARN_ON_ONCE(freeze_depth < 0);
146         if (!freeze_depth) {
147                 percpu_ref_reinit(&q->q_usage_counter);
148                 wake_up_all(&q->mq_freeze_wq);
149         }
150 }
151 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
152
153 /*
154  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
155  * mpt3sas driver such that this function can be removed.
156  */
157 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
158 {
159         unsigned long flags;
160
161         spin_lock_irqsave(q->queue_lock, flags);
162         queue_flag_set(QUEUE_FLAG_QUIESCED, q);
163         spin_unlock_irqrestore(q->queue_lock, flags);
164 }
165 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
166
167 /**
168  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
169  * @q: request queue.
170  *
171  * Note: this function does not prevent that the struct request end_io()
172  * callback function is invoked. Once this function is returned, we make
173  * sure no dispatch can happen until the queue is unquiesced via
174  * blk_mq_unquiesce_queue().
175  */
176 void blk_mq_quiesce_queue(struct request_queue *q)
177 {
178         struct blk_mq_hw_ctx *hctx;
179         unsigned int i;
180         bool rcu = false;
181
182         blk_mq_quiesce_queue_nowait(q);
183
184         queue_for_each_hw_ctx(q, hctx, i) {
185                 if (hctx->flags & BLK_MQ_F_BLOCKING)
186                         synchronize_srcu(hctx->queue_rq_srcu);
187                 else
188                         rcu = true;
189         }
190         if (rcu)
191                 synchronize_rcu();
192 }
193 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
194
195 /*
196  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
197  * @q: request queue.
198  *
199  * This function recovers queue into the state before quiescing
200  * which is done by blk_mq_quiesce_queue.
201  */
202 void blk_mq_unquiesce_queue(struct request_queue *q)
203 {
204         unsigned long flags;
205
206         spin_lock_irqsave(q->queue_lock, flags);
207         queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
208         spin_unlock_irqrestore(q->queue_lock, flags);
209
210         /* dispatch requests which are inserted during quiescing */
211         blk_mq_run_hw_queues(q, true);
212 }
213 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
214
215 void blk_mq_wake_waiters(struct request_queue *q)
216 {
217         struct blk_mq_hw_ctx *hctx;
218         unsigned int i;
219
220         queue_for_each_hw_ctx(q, hctx, i)
221                 if (blk_mq_hw_queue_mapped(hctx))
222                         blk_mq_tag_wakeup_all(hctx->tags, true);
223
224         /*
225          * If we are called because the queue has now been marked as
226          * dying, we need to ensure that processes currently waiting on
227          * the queue are notified as well.
228          */
229         wake_up_all(&q->mq_freeze_wq);
230 }
231
232 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
233 {
234         return blk_mq_has_free_tags(hctx->tags);
235 }
236 EXPORT_SYMBOL(blk_mq_can_queue);
237
238 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
239                 unsigned int tag, unsigned int op)
240 {
241         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
242         struct request *rq = tags->static_rqs[tag];
243
244         rq->rq_flags = 0;
245
246         if (data->flags & BLK_MQ_REQ_INTERNAL) {
247                 rq->tag = -1;
248                 rq->internal_tag = tag;
249         } else {
250                 if (blk_mq_tag_busy(data->hctx)) {
251                         rq->rq_flags = RQF_MQ_INFLIGHT;
252                         atomic_inc(&data->hctx->nr_active);
253                 }
254                 rq->tag = tag;
255                 rq->internal_tag = -1;
256                 data->hctx->tags->rqs[rq->tag] = rq;
257         }
258
259         INIT_LIST_HEAD(&rq->queuelist);
260         /* csd/requeue_work/fifo_time is initialized before use */
261         rq->q = data->q;
262         rq->mq_ctx = data->ctx;
263         rq->cmd_flags = op;
264         if (blk_queue_io_stat(data->q))
265                 rq->rq_flags |= RQF_IO_STAT;
266         /* do not touch atomic flags, it needs atomic ops against the timer */
267         rq->cpu = -1;
268         INIT_HLIST_NODE(&rq->hash);
269         RB_CLEAR_NODE(&rq->rb_node);
270         rq->rq_disk = NULL;
271         rq->part = NULL;
272         rq->start_time = jiffies;
273 #ifdef CONFIG_BLK_CGROUP
274         rq->rl = NULL;
275         set_start_time_ns(rq);
276         rq->io_start_time_ns = 0;
277 #endif
278         rq->nr_phys_segments = 0;
279 #if defined(CONFIG_BLK_DEV_INTEGRITY)
280         rq->nr_integrity_segments = 0;
281 #endif
282         rq->special = NULL;
283         /* tag was already set */
284         rq->extra_len = 0;
285
286         INIT_LIST_HEAD(&rq->timeout_list);
287         rq->timeout = 0;
288
289         rq->end_io = NULL;
290         rq->end_io_data = NULL;
291         rq->next_rq = NULL;
292
293         data->ctx->rq_dispatched[op_is_sync(op)]++;
294         return rq;
295 }
296
297 static struct request *blk_mq_get_request(struct request_queue *q,
298                 struct bio *bio, unsigned int op,
299                 struct blk_mq_alloc_data *data)
300 {
301         struct elevator_queue *e = q->elevator;
302         struct request *rq;
303         unsigned int tag;
304         struct blk_mq_ctx *local_ctx = NULL;
305
306         blk_queue_enter_live(q);
307         data->q = q;
308         if (likely(!data->ctx))
309                 data->ctx = local_ctx = blk_mq_get_ctx(q);
310         if (likely(!data->hctx))
311                 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
312         if (op & REQ_NOWAIT)
313                 data->flags |= BLK_MQ_REQ_NOWAIT;
314
315         if (e) {
316                 data->flags |= BLK_MQ_REQ_INTERNAL;
317
318                 /*
319                  * Flush requests are special and go directly to the
320                  * dispatch list.
321                  */
322                 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
323                         e->type->ops.mq.limit_depth(op, data);
324         }
325
326         tag = blk_mq_get_tag(data);
327         if (tag == BLK_MQ_TAG_FAIL) {
328                 if (local_ctx) {
329                         blk_mq_put_ctx(local_ctx);
330                         data->ctx = NULL;
331                 }
332                 blk_queue_exit(q);
333                 return NULL;
334         }
335
336         rq = blk_mq_rq_ctx_init(data, tag, op);
337         if (!op_is_flush(op)) {
338                 rq->elv.icq = NULL;
339                 if (e && e->type->ops.mq.prepare_request) {
340                         if (e->type->icq_cache && rq_ioc(bio))
341                                 blk_mq_sched_assign_ioc(rq, bio);
342
343                         e->type->ops.mq.prepare_request(rq, bio);
344                         rq->rq_flags |= RQF_ELVPRIV;
345                 }
346         }
347         data->hctx->queued++;
348         return rq;
349 }
350
351 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
352                 unsigned int flags)
353 {
354         struct blk_mq_alloc_data alloc_data = { .flags = flags };
355         struct request *rq;
356         int ret;
357
358         ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
359         if (ret)
360                 return ERR_PTR(ret);
361
362         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
363
364         if (!rq)
365                 return ERR_PTR(-EWOULDBLOCK);
366
367         blk_mq_put_ctx(alloc_data.ctx);
368         blk_queue_exit(q);
369
370         rq->__data_len = 0;
371         rq->__sector = (sector_t) -1;
372         rq->bio = rq->biotail = NULL;
373         return rq;
374 }
375 EXPORT_SYMBOL(blk_mq_alloc_request);
376
377 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
378                 unsigned int op, unsigned int flags, unsigned int hctx_idx)
379 {
380         struct blk_mq_alloc_data alloc_data = { .flags = flags };
381         struct request *rq;
382         unsigned int cpu;
383         int ret;
384
385         /*
386          * If the tag allocator sleeps we could get an allocation for a
387          * different hardware context.  No need to complicate the low level
388          * allocator for this for the rare use case of a command tied to
389          * a specific queue.
390          */
391         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
392                 return ERR_PTR(-EINVAL);
393
394         if (hctx_idx >= q->nr_hw_queues)
395                 return ERR_PTR(-EIO);
396
397         ret = blk_queue_enter(q, true);
398         if (ret)
399                 return ERR_PTR(ret);
400
401         /*
402          * Check if the hardware context is actually mapped to anything.
403          * If not tell the caller that it should skip this queue.
404          */
405         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
406         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
407                 blk_queue_exit(q);
408                 return ERR_PTR(-EXDEV);
409         }
410         cpu = cpumask_first(alloc_data.hctx->cpumask);
411         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
412
413         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
414
415         if (!rq)
416                 return ERR_PTR(-EWOULDBLOCK);
417
418         blk_queue_exit(q);
419
420         return rq;
421 }
422 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
423
424 void blk_mq_free_request(struct request *rq)
425 {
426         struct request_queue *q = rq->q;
427         struct elevator_queue *e = q->elevator;
428         struct blk_mq_ctx *ctx = rq->mq_ctx;
429         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
430         const int sched_tag = rq->internal_tag;
431
432         if (rq->rq_flags & RQF_ELVPRIV) {
433                 if (e && e->type->ops.mq.finish_request)
434                         e->type->ops.mq.finish_request(rq);
435                 if (rq->elv.icq) {
436                         put_io_context(rq->elv.icq->ioc);
437                         rq->elv.icq = NULL;
438                 }
439         }
440
441         ctx->rq_completed[rq_is_sync(rq)]++;
442         if (rq->rq_flags & RQF_MQ_INFLIGHT)
443                 atomic_dec(&hctx->nr_active);
444
445         wbt_done(q->rq_wb, &rq->issue_stat);
446
447         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
448         clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
449         if (rq->tag != -1)
450                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
451         if (sched_tag != -1)
452                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
453         blk_mq_sched_restart(hctx);
454         blk_queue_exit(q);
455 }
456 EXPORT_SYMBOL_GPL(blk_mq_free_request);
457
458 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
459 {
460         blk_account_io_done(rq);
461
462         if (rq->end_io) {
463                 wbt_done(rq->q->rq_wb, &rq->issue_stat);
464                 rq->end_io(rq, error);
465         } else {
466                 if (unlikely(blk_bidi_rq(rq)))
467                         blk_mq_free_request(rq->next_rq);
468                 blk_mq_free_request(rq);
469         }
470 }
471 EXPORT_SYMBOL(__blk_mq_end_request);
472
473 void blk_mq_end_request(struct request *rq, blk_status_t error)
474 {
475         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
476                 BUG();
477         __blk_mq_end_request(rq, error);
478 }
479 EXPORT_SYMBOL(blk_mq_end_request);
480
481 static void __blk_mq_complete_request_remote(void *data)
482 {
483         struct request *rq = data;
484
485         rq->q->softirq_done_fn(rq);
486 }
487
488 static void __blk_mq_complete_request(struct request *rq)
489 {
490         struct blk_mq_ctx *ctx = rq->mq_ctx;
491         bool shared = false;
492         int cpu;
493
494         if (rq->internal_tag != -1)
495                 blk_mq_sched_completed_request(rq);
496         if (rq->rq_flags & RQF_STATS) {
497                 blk_mq_poll_stats_start(rq->q);
498                 blk_stat_add(rq);
499         }
500
501         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
502                 rq->q->softirq_done_fn(rq);
503                 return;
504         }
505
506         cpu = get_cpu();
507         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
508                 shared = cpus_share_cache(cpu, ctx->cpu);
509
510         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
511                 rq->csd.func = __blk_mq_complete_request_remote;
512                 rq->csd.info = rq;
513                 rq->csd.flags = 0;
514                 smp_call_function_single_async(ctx->cpu, &rq->csd);
515         } else {
516                 rq->q->softirq_done_fn(rq);
517         }
518         put_cpu();
519 }
520
521 /**
522  * blk_mq_complete_request - end I/O on a request
523  * @rq:         the request being processed
524  *
525  * Description:
526  *      Ends all I/O on a request. It does not handle partial completions.
527  *      The actual completion happens out-of-order, through a IPI handler.
528  **/
529 void blk_mq_complete_request(struct request *rq)
530 {
531         struct request_queue *q = rq->q;
532
533         if (unlikely(blk_should_fake_timeout(q)))
534                 return;
535         if (!blk_mark_rq_complete(rq))
536                 __blk_mq_complete_request(rq);
537 }
538 EXPORT_SYMBOL(blk_mq_complete_request);
539
540 int blk_mq_request_started(struct request *rq)
541 {
542         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
543 }
544 EXPORT_SYMBOL_GPL(blk_mq_request_started);
545
546 void blk_mq_start_request(struct request *rq)
547 {
548         struct request_queue *q = rq->q;
549
550         blk_mq_sched_started_request(rq);
551
552         trace_block_rq_issue(q, rq);
553
554         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
555                 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
556                 rq->rq_flags |= RQF_STATS;
557                 wbt_issue(q->rq_wb, &rq->issue_stat);
558         }
559
560         blk_add_timer(rq);
561
562         /*
563          * Ensure that ->deadline is visible before set the started
564          * flag and clear the completed flag.
565          */
566         smp_mb__before_atomic();
567
568         /*
569          * Mark us as started and clear complete. Complete might have been
570          * set if requeue raced with timeout, which then marked it as
571          * complete. So be sure to clear complete again when we start
572          * the request, otherwise we'll ignore the completion event.
573          */
574         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
575                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
576         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
577                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
578
579         if (q->dma_drain_size && blk_rq_bytes(rq)) {
580                 /*
581                  * Make sure space for the drain appears.  We know we can do
582                  * this because max_hw_segments has been adjusted to be one
583                  * fewer than the device can handle.
584                  */
585                 rq->nr_phys_segments++;
586         }
587 }
588 EXPORT_SYMBOL(blk_mq_start_request);
589
590 /*
591  * When we reach here because queue is busy, REQ_ATOM_COMPLETE
592  * flag isn't set yet, so there may be race with timeout handler,
593  * but given rq->deadline is just set in .queue_rq() under
594  * this situation, the race won't be possible in reality because
595  * rq->timeout should be set as big enough to cover the window
596  * between blk_mq_start_request() called from .queue_rq() and
597  * clearing REQ_ATOM_STARTED here.
598  */
599 static void __blk_mq_requeue_request(struct request *rq)
600 {
601         struct request_queue *q = rq->q;
602
603         trace_block_rq_requeue(q, rq);
604         wbt_requeue(q->rq_wb, &rq->issue_stat);
605         blk_mq_sched_requeue_request(rq);
606
607         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
608                 if (q->dma_drain_size && blk_rq_bytes(rq))
609                         rq->nr_phys_segments--;
610         }
611 }
612
613 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
614 {
615         __blk_mq_requeue_request(rq);
616
617         BUG_ON(blk_queued_rq(rq));
618         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
619 }
620 EXPORT_SYMBOL(blk_mq_requeue_request);
621
622 static void blk_mq_requeue_work(struct work_struct *work)
623 {
624         struct request_queue *q =
625                 container_of(work, struct request_queue, requeue_work.work);
626         LIST_HEAD(rq_list);
627         struct request *rq, *next;
628         unsigned long flags;
629
630         spin_lock_irqsave(&q->requeue_lock, flags);
631         list_splice_init(&q->requeue_list, &rq_list);
632         spin_unlock_irqrestore(&q->requeue_lock, flags);
633
634         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
635                 if (!(rq->rq_flags & RQF_SOFTBARRIER))
636                         continue;
637
638                 rq->rq_flags &= ~RQF_SOFTBARRIER;
639                 list_del_init(&rq->queuelist);
640                 blk_mq_sched_insert_request(rq, true, false, false, true);
641         }
642
643         while (!list_empty(&rq_list)) {
644                 rq = list_entry(rq_list.next, struct request, queuelist);
645                 list_del_init(&rq->queuelist);
646                 blk_mq_sched_insert_request(rq, false, false, false, true);
647         }
648
649         blk_mq_run_hw_queues(q, false);
650 }
651
652 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
653                                 bool kick_requeue_list)
654 {
655         struct request_queue *q = rq->q;
656         unsigned long flags;
657
658         /*
659          * We abuse this flag that is otherwise used by the I/O scheduler to
660          * request head insertation from the workqueue.
661          */
662         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
663
664         spin_lock_irqsave(&q->requeue_lock, flags);
665         if (at_head) {
666                 rq->rq_flags |= RQF_SOFTBARRIER;
667                 list_add(&rq->queuelist, &q->requeue_list);
668         } else {
669                 list_add_tail(&rq->queuelist, &q->requeue_list);
670         }
671         spin_unlock_irqrestore(&q->requeue_lock, flags);
672
673         if (kick_requeue_list)
674                 blk_mq_kick_requeue_list(q);
675 }
676 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
677
678 void blk_mq_kick_requeue_list(struct request_queue *q)
679 {
680         kblockd_schedule_delayed_work(&q->requeue_work, 0);
681 }
682 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
683
684 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
685                                     unsigned long msecs)
686 {
687         kblockd_schedule_delayed_work(&q->requeue_work,
688                                       msecs_to_jiffies(msecs));
689 }
690 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
691
692 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
693 {
694         if (tag < tags->nr_tags) {
695                 prefetch(tags->rqs[tag]);
696                 return tags->rqs[tag];
697         }
698
699         return NULL;
700 }
701 EXPORT_SYMBOL(blk_mq_tag_to_rq);
702
703 struct blk_mq_timeout_data {
704         unsigned long next;
705         unsigned int next_set;
706 };
707
708 void blk_mq_rq_timed_out(struct request *req, bool reserved)
709 {
710         const struct blk_mq_ops *ops = req->q->mq_ops;
711         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
712
713         /*
714          * We know that complete is set at this point. If STARTED isn't set
715          * anymore, then the request isn't active and the "timeout" should
716          * just be ignored. This can happen due to the bitflag ordering.
717          * Timeout first checks if STARTED is set, and if it is, assumes
718          * the request is active. But if we race with completion, then
719          * both flags will get cleared. So check here again, and ignore
720          * a timeout event with a request that isn't active.
721          */
722         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
723                 return;
724
725         if (ops->timeout)
726                 ret = ops->timeout(req, reserved);
727
728         switch (ret) {
729         case BLK_EH_HANDLED:
730                 __blk_mq_complete_request(req);
731                 break;
732         case BLK_EH_RESET_TIMER:
733                 blk_add_timer(req);
734                 blk_clear_rq_complete(req);
735                 break;
736         case BLK_EH_NOT_HANDLED:
737                 break;
738         default:
739                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
740                 break;
741         }
742 }
743
744 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
745                 struct request *rq, void *priv, bool reserved)
746 {
747         struct blk_mq_timeout_data *data = priv;
748
749         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
750                 return;
751
752         /*
753          * The rq being checked may have been freed and reallocated
754          * out already here, we avoid this race by checking rq->deadline
755          * and REQ_ATOM_COMPLETE flag together:
756          *
757          * - if rq->deadline is observed as new value because of
758          *   reusing, the rq won't be timed out because of timing.
759          * - if rq->deadline is observed as previous value,
760          *   REQ_ATOM_COMPLETE flag won't be cleared in reuse path
761          *   because we put a barrier between setting rq->deadline
762          *   and clearing the flag in blk_mq_start_request(), so
763          *   this rq won't be timed out too.
764          */
765         if (time_after_eq(jiffies, rq->deadline)) {
766                 if (!blk_mark_rq_complete(rq))
767                         blk_mq_rq_timed_out(rq, reserved);
768         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
769                 data->next = rq->deadline;
770                 data->next_set = 1;
771         }
772 }
773
774 static void blk_mq_timeout_work(struct work_struct *work)
775 {
776         struct request_queue *q =
777                 container_of(work, struct request_queue, timeout_work);
778         struct blk_mq_timeout_data data = {
779                 .next           = 0,
780                 .next_set       = 0,
781         };
782         int i;
783
784         /* A deadlock might occur if a request is stuck requiring a
785          * timeout at the same time a queue freeze is waiting
786          * completion, since the timeout code would not be able to
787          * acquire the queue reference here.
788          *
789          * That's why we don't use blk_queue_enter here; instead, we use
790          * percpu_ref_tryget directly, because we need to be able to
791          * obtain a reference even in the short window between the queue
792          * starting to freeze, by dropping the first reference in
793          * blk_freeze_queue_start, and the moment the last request is
794          * consumed, marked by the instant q_usage_counter reaches
795          * zero.
796          */
797         if (!percpu_ref_tryget(&q->q_usage_counter))
798                 return;
799
800         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
801
802         if (data.next_set) {
803                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
804                 mod_timer(&q->timeout, data.next);
805         } else {
806                 struct blk_mq_hw_ctx *hctx;
807
808                 queue_for_each_hw_ctx(q, hctx, i) {
809                         /* the hctx may be unmapped, so check it here */
810                         if (blk_mq_hw_queue_mapped(hctx))
811                                 blk_mq_tag_idle(hctx);
812                 }
813         }
814         blk_queue_exit(q);
815 }
816
817 struct flush_busy_ctx_data {
818         struct blk_mq_hw_ctx *hctx;
819         struct list_head *list;
820 };
821
822 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
823 {
824         struct flush_busy_ctx_data *flush_data = data;
825         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
826         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
827
828         sbitmap_clear_bit(sb, bitnr);
829         spin_lock(&ctx->lock);
830         list_splice_tail_init(&ctx->rq_list, flush_data->list);
831         spin_unlock(&ctx->lock);
832         return true;
833 }
834
835 /*
836  * Process software queues that have been marked busy, splicing them
837  * to the for-dispatch
838  */
839 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
840 {
841         struct flush_busy_ctx_data data = {
842                 .hctx = hctx,
843                 .list = list,
844         };
845
846         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
847 }
848 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
849
850 static inline unsigned int queued_to_index(unsigned int queued)
851 {
852         if (!queued)
853                 return 0;
854
855         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
856 }
857
858 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
859                            bool wait)
860 {
861         struct blk_mq_alloc_data data = {
862                 .q = rq->q,
863                 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
864                 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
865         };
866
867         might_sleep_if(wait);
868
869         if (rq->tag != -1)
870                 goto done;
871
872         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
873                 data.flags |= BLK_MQ_REQ_RESERVED;
874
875         rq->tag = blk_mq_get_tag(&data);
876         if (rq->tag >= 0) {
877                 if (blk_mq_tag_busy(data.hctx)) {
878                         rq->rq_flags |= RQF_MQ_INFLIGHT;
879                         atomic_inc(&data.hctx->nr_active);
880                 }
881                 data.hctx->tags->rqs[rq->tag] = rq;
882         }
883
884 done:
885         if (hctx)
886                 *hctx = data.hctx;
887         return rq->tag != -1;
888 }
889
890 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
891                                     struct request *rq)
892 {
893         blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
894         rq->tag = -1;
895
896         if (rq->rq_flags & RQF_MQ_INFLIGHT) {
897                 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
898                 atomic_dec(&hctx->nr_active);
899         }
900 }
901
902 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
903                                        struct request *rq)
904 {
905         if (rq->tag == -1 || rq->internal_tag == -1)
906                 return;
907
908         __blk_mq_put_driver_tag(hctx, rq);
909 }
910
911 static void blk_mq_put_driver_tag(struct request *rq)
912 {
913         struct blk_mq_hw_ctx *hctx;
914
915         if (rq->tag == -1 || rq->internal_tag == -1)
916                 return;
917
918         hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
919         __blk_mq_put_driver_tag(hctx, rq);
920 }
921
922 /*
923  * If we fail getting a driver tag because all the driver tags are already
924  * assigned and on the dispatch list, BUT the first entry does not have a
925  * tag, then we could deadlock. For that case, move entries with assigned
926  * driver tags to the front, leaving the set of tagged requests in the
927  * same order, and the untagged set in the same order.
928  */
929 static bool reorder_tags_to_front(struct list_head *list)
930 {
931         struct request *rq, *tmp, *first = NULL;
932
933         list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
934                 if (rq == first)
935                         break;
936                 if (rq->tag != -1) {
937                         list_move(&rq->queuelist, list);
938                         if (!first)
939                                 first = rq;
940                 }
941         }
942
943         return first != NULL;
944 }
945
946 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
947                                 void *key)
948 {
949         struct blk_mq_hw_ctx *hctx;
950
951         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
952
953         list_del(&wait->entry);
954         clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
955         blk_mq_run_hw_queue(hctx, true);
956         return 1;
957 }
958
959 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
960 {
961         struct sbq_wait_state *ws;
962
963         /*
964          * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
965          * The thread which wins the race to grab this bit adds the hardware
966          * queue to the wait queue.
967          */
968         if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
969             test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
970                 return false;
971
972         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
973         ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
974
975         /*
976          * As soon as this returns, it's no longer safe to fiddle with
977          * hctx->dispatch_wait, since a completion can wake up the wait queue
978          * and unlock the bit.
979          */
980         add_wait_queue(&ws->wait, &hctx->dispatch_wait);
981         return true;
982 }
983
984 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
985 {
986         struct blk_mq_hw_ctx *hctx;
987         struct request *rq;
988         int errors, queued;
989
990         if (list_empty(list))
991                 return false;
992
993         /*
994          * Now process all the entries, sending them to the driver.
995          */
996         errors = queued = 0;
997         do {
998                 struct blk_mq_queue_data bd;
999                 blk_status_t ret;
1000
1001                 rq = list_first_entry(list, struct request, queuelist);
1002                 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1003                         if (!queued && reorder_tags_to_front(list))
1004                                 continue;
1005
1006                         /*
1007                          * The initial allocation attempt failed, so we need to
1008                          * rerun the hardware queue when a tag is freed.
1009                          */
1010                         if (!blk_mq_dispatch_wait_add(hctx))
1011                                 break;
1012
1013                         /*
1014                          * It's possible that a tag was freed in the window
1015                          * between the allocation failure and adding the
1016                          * hardware queue to the wait queue.
1017                          */
1018                         if (!blk_mq_get_driver_tag(rq, &hctx, false))
1019                                 break;
1020                 }
1021
1022                 list_del_init(&rq->queuelist);
1023
1024                 bd.rq = rq;
1025
1026                 /*
1027                  * Flag last if we have no more requests, or if we have more
1028                  * but can't assign a driver tag to it.
1029                  */
1030                 if (list_empty(list))
1031                         bd.last = true;
1032                 else {
1033                         struct request *nxt;
1034
1035                         nxt = list_first_entry(list, struct request, queuelist);
1036                         bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1037                 }
1038
1039                 ret = q->mq_ops->queue_rq(hctx, &bd);
1040                 if (ret == BLK_STS_RESOURCE) {
1041                         blk_mq_put_driver_tag_hctx(hctx, rq);
1042                         list_add(&rq->queuelist, list);
1043                         __blk_mq_requeue_request(rq);
1044                         break;
1045                 }
1046
1047                 if (unlikely(ret != BLK_STS_OK)) {
1048                         errors++;
1049                         blk_mq_end_request(rq, BLK_STS_IOERR);
1050                         continue;
1051                 }
1052
1053                 queued++;
1054         } while (!list_empty(list));
1055
1056         hctx->dispatched[queued_to_index(queued)]++;
1057
1058         /*
1059          * Any items that need requeuing? Stuff them into hctx->dispatch,
1060          * that is where we will continue on next queue run.
1061          */
1062         if (!list_empty(list)) {
1063                 /*
1064                  * If an I/O scheduler has been configured and we got a driver
1065                  * tag for the next request already, free it again.
1066                  */
1067                 rq = list_first_entry(list, struct request, queuelist);
1068                 blk_mq_put_driver_tag(rq);
1069
1070                 spin_lock(&hctx->lock);
1071                 list_splice_init(list, &hctx->dispatch);
1072                 spin_unlock(&hctx->lock);
1073
1074                 /*
1075                  * If SCHED_RESTART was set by the caller of this function and
1076                  * it is no longer set that means that it was cleared by another
1077                  * thread and hence that a queue rerun is needed.
1078                  *
1079                  * If TAG_WAITING is set that means that an I/O scheduler has
1080                  * been configured and another thread is waiting for a driver
1081                  * tag. To guarantee fairness, do not rerun this hardware queue
1082                  * but let the other thread grab the driver tag.
1083                  *
1084                  * If no I/O scheduler has been configured it is possible that
1085                  * the hardware queue got stopped and restarted before requests
1086                  * were pushed back onto the dispatch list. Rerun the queue to
1087                  * avoid starvation. Notes:
1088                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1089                  *   been stopped before rerunning a queue.
1090                  * - Some but not all block drivers stop a queue before
1091                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1092                  *   and dm-rq.
1093                  */
1094                 if (!blk_mq_sched_needs_restart(hctx) &&
1095                     !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1096                         blk_mq_run_hw_queue(hctx, true);
1097         }
1098
1099         return (queued + errors) != 0;
1100 }
1101
1102 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1103 {
1104         int srcu_idx;
1105
1106         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1107                 cpu_online(hctx->next_cpu));
1108
1109         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1110                 rcu_read_lock();
1111                 blk_mq_sched_dispatch_requests(hctx);
1112                 rcu_read_unlock();
1113         } else {
1114                 might_sleep();
1115
1116                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1117                 blk_mq_sched_dispatch_requests(hctx);
1118                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1119         }
1120 }
1121
1122 /*
1123  * It'd be great if the workqueue API had a way to pass
1124  * in a mask and had some smarts for more clever placement.
1125  * For now we just round-robin here, switching for every
1126  * BLK_MQ_CPU_WORK_BATCH queued items.
1127  */
1128 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1129 {
1130         if (hctx->queue->nr_hw_queues == 1)
1131                 return WORK_CPU_UNBOUND;
1132
1133         if (--hctx->next_cpu_batch <= 0) {
1134                 int next_cpu;
1135
1136                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1137                 if (next_cpu >= nr_cpu_ids)
1138                         next_cpu = cpumask_first(hctx->cpumask);
1139
1140                 hctx->next_cpu = next_cpu;
1141                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1142         }
1143
1144         return hctx->next_cpu;
1145 }
1146
1147 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1148                                         unsigned long msecs)
1149 {
1150         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1151                 return;
1152
1153         if (unlikely(blk_mq_hctx_stopped(hctx)))
1154                 return;
1155
1156         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1157                 int cpu = get_cpu();
1158                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1159                         __blk_mq_run_hw_queue(hctx);
1160                         put_cpu();
1161                         return;
1162                 }
1163
1164                 put_cpu();
1165         }
1166
1167         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1168                                          &hctx->run_work,
1169                                          msecs_to_jiffies(msecs));
1170 }
1171
1172 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1173 {
1174         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1175 }
1176 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1177
1178 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1179 {
1180         __blk_mq_delay_run_hw_queue(hctx, async, 0);
1181 }
1182 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1183
1184 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1185 {
1186         struct blk_mq_hw_ctx *hctx;
1187         int i;
1188
1189         queue_for_each_hw_ctx(q, hctx, i) {
1190                 if (!blk_mq_hctx_has_pending(hctx) ||
1191                     blk_mq_hctx_stopped(hctx))
1192                         continue;
1193
1194                 blk_mq_run_hw_queue(hctx, async);
1195         }
1196 }
1197 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1198
1199 /**
1200  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1201  * @q: request queue.
1202  *
1203  * The caller is responsible for serializing this function against
1204  * blk_mq_{start,stop}_hw_queue().
1205  */
1206 bool blk_mq_queue_stopped(struct request_queue *q)
1207 {
1208         struct blk_mq_hw_ctx *hctx;
1209         int i;
1210
1211         queue_for_each_hw_ctx(q, hctx, i)
1212                 if (blk_mq_hctx_stopped(hctx))
1213                         return true;
1214
1215         return false;
1216 }
1217 EXPORT_SYMBOL(blk_mq_queue_stopped);
1218
1219 /*
1220  * This function is often used for pausing .queue_rq() by driver when
1221  * there isn't enough resource or some conditions aren't satisfied, and
1222  * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1223  *
1224  * We do not guarantee that dispatch can be drained or blocked
1225  * after blk_mq_stop_hw_queue() returns. Please use
1226  * blk_mq_quiesce_queue() for that requirement.
1227  */
1228 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1229 {
1230         cancel_delayed_work(&hctx->run_work);
1231
1232         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1233 }
1234 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1235
1236 /*
1237  * This function is often used for pausing .queue_rq() by driver when
1238  * there isn't enough resource or some conditions aren't satisfied, and
1239  * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1240  *
1241  * We do not guarantee that dispatch can be drained or blocked
1242  * after blk_mq_stop_hw_queues() returns. Please use
1243  * blk_mq_quiesce_queue() for that requirement.
1244  */
1245 void blk_mq_stop_hw_queues(struct request_queue *q)
1246 {
1247         struct blk_mq_hw_ctx *hctx;
1248         int i;
1249
1250         queue_for_each_hw_ctx(q, hctx, i)
1251                 blk_mq_stop_hw_queue(hctx);
1252 }
1253 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1254
1255 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1256 {
1257         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1258
1259         blk_mq_run_hw_queue(hctx, false);
1260 }
1261 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1262
1263 void blk_mq_start_hw_queues(struct request_queue *q)
1264 {
1265         struct blk_mq_hw_ctx *hctx;
1266         int i;
1267
1268         queue_for_each_hw_ctx(q, hctx, i)
1269                 blk_mq_start_hw_queue(hctx);
1270 }
1271 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1272
1273 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1274 {
1275         if (!blk_mq_hctx_stopped(hctx))
1276                 return;
1277
1278         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1279         blk_mq_run_hw_queue(hctx, async);
1280 }
1281 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1282
1283 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1284 {
1285         struct blk_mq_hw_ctx *hctx;
1286         int i;
1287
1288         queue_for_each_hw_ctx(q, hctx, i)
1289                 blk_mq_start_stopped_hw_queue(hctx, async);
1290 }
1291 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1292
1293 static void blk_mq_run_work_fn(struct work_struct *work)
1294 {
1295         struct blk_mq_hw_ctx *hctx;
1296
1297         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1298
1299         /*
1300          * If we are stopped, don't run the queue. The exception is if
1301          * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1302          * the STOPPED bit and run it.
1303          */
1304         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1305                 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1306                         return;
1307
1308                 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1309                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1310         }
1311
1312         __blk_mq_run_hw_queue(hctx);
1313 }
1314
1315
1316 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1317 {
1318         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1319                 return;
1320
1321         /*
1322          * Stop the hw queue, then modify currently delayed work.
1323          * This should prevent us from running the queue prematurely.
1324          * Mark the queue as auto-clearing STOPPED when it runs.
1325          */
1326         blk_mq_stop_hw_queue(hctx);
1327         set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1328         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1329                                         &hctx->run_work,
1330                                         msecs_to_jiffies(msecs));
1331 }
1332 EXPORT_SYMBOL(blk_mq_delay_queue);
1333
1334 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1335                                             struct request *rq,
1336                                             bool at_head)
1337 {
1338         struct blk_mq_ctx *ctx = rq->mq_ctx;
1339
1340         lockdep_assert_held(&ctx->lock);
1341
1342         trace_block_rq_insert(hctx->queue, rq);
1343
1344         if (at_head)
1345                 list_add(&rq->queuelist, &ctx->rq_list);
1346         else
1347                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1348 }
1349
1350 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1351                              bool at_head)
1352 {
1353         struct blk_mq_ctx *ctx = rq->mq_ctx;
1354
1355         lockdep_assert_held(&ctx->lock);
1356
1357         __blk_mq_insert_req_list(hctx, rq, at_head);
1358         blk_mq_hctx_mark_pending(hctx, ctx);
1359 }
1360
1361 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1362                             struct list_head *list)
1363
1364 {
1365         /*
1366          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1367          * offline now
1368          */
1369         spin_lock(&ctx->lock);
1370         while (!list_empty(list)) {
1371                 struct request *rq;
1372
1373                 rq = list_first_entry(list, struct request, queuelist);
1374                 BUG_ON(rq->mq_ctx != ctx);
1375                 list_del_init(&rq->queuelist);
1376                 __blk_mq_insert_req_list(hctx, rq, false);
1377         }
1378         blk_mq_hctx_mark_pending(hctx, ctx);
1379         spin_unlock(&ctx->lock);
1380 }
1381
1382 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1383 {
1384         struct request *rqa = container_of(a, struct request, queuelist);
1385         struct request *rqb = container_of(b, struct request, queuelist);
1386
1387         return !(rqa->mq_ctx < rqb->mq_ctx ||
1388                  (rqa->mq_ctx == rqb->mq_ctx &&
1389                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1390 }
1391
1392 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1393 {
1394         struct blk_mq_ctx *this_ctx;
1395         struct request_queue *this_q;
1396         struct request *rq;
1397         LIST_HEAD(list);
1398         LIST_HEAD(ctx_list);
1399         unsigned int depth;
1400
1401         list_splice_init(&plug->mq_list, &list);
1402
1403         list_sort(NULL, &list, plug_ctx_cmp);
1404
1405         this_q = NULL;
1406         this_ctx = NULL;
1407         depth = 0;
1408
1409         while (!list_empty(&list)) {
1410                 rq = list_entry_rq(list.next);
1411                 list_del_init(&rq->queuelist);
1412                 BUG_ON(!rq->q);
1413                 if (rq->mq_ctx != this_ctx) {
1414                         if (this_ctx) {
1415                                 trace_block_unplug(this_q, depth, from_schedule);
1416                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1417                                                                 &ctx_list,
1418                                                                 from_schedule);
1419                         }
1420
1421                         this_ctx = rq->mq_ctx;
1422                         this_q = rq->q;
1423                         depth = 0;
1424                 }
1425
1426                 depth++;
1427                 list_add_tail(&rq->queuelist, &ctx_list);
1428         }
1429
1430         /*
1431          * If 'this_ctx' is set, we know we have entries to complete
1432          * on 'ctx_list'. Do those.
1433          */
1434         if (this_ctx) {
1435                 trace_block_unplug(this_q, depth, from_schedule);
1436                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1437                                                 from_schedule);
1438         }
1439 }
1440
1441 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1442 {
1443         blk_init_request_from_bio(rq, bio);
1444
1445         blk_account_io_start(rq, true);
1446 }
1447
1448 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1449 {
1450         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1451                 !blk_queue_nomerges(hctx->queue);
1452 }
1453
1454 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1455                                    struct blk_mq_ctx *ctx,
1456                                    struct request *rq)
1457 {
1458         spin_lock(&ctx->lock);
1459         __blk_mq_insert_request(hctx, rq, false);
1460         spin_unlock(&ctx->lock);
1461 }
1462
1463 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1464 {
1465         if (rq->tag != -1)
1466                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1467
1468         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1469 }
1470
1471 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1472                                         struct request *rq,
1473                                         blk_qc_t *cookie, bool may_sleep)
1474 {
1475         struct request_queue *q = rq->q;
1476         struct blk_mq_queue_data bd = {
1477                 .rq = rq,
1478                 .last = true,
1479         };
1480         blk_qc_t new_cookie;
1481         blk_status_t ret;
1482         bool run_queue = true;
1483
1484         /* RCU or SRCU read lock is needed before checking quiesced flag */
1485         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1486                 run_queue = false;
1487                 goto insert;
1488         }
1489
1490         if (q->elevator)
1491                 goto insert;
1492
1493         if (!blk_mq_get_driver_tag(rq, NULL, false))
1494                 goto insert;
1495
1496         new_cookie = request_to_qc_t(hctx, rq);
1497
1498         /*
1499          * For OK queue, we are done. For error, kill it. Any other
1500          * error (busy), just add it to our list as we previously
1501          * would have done
1502          */
1503         ret = q->mq_ops->queue_rq(hctx, &bd);
1504         switch (ret) {
1505         case BLK_STS_OK:
1506                 *cookie = new_cookie;
1507                 return;
1508         case BLK_STS_RESOURCE:
1509                 __blk_mq_requeue_request(rq);
1510                 goto insert;
1511         default:
1512                 *cookie = BLK_QC_T_NONE;
1513                 blk_mq_end_request(rq, ret);
1514                 return;
1515         }
1516
1517 insert:
1518         blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1519 }
1520
1521 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1522                 struct request *rq, blk_qc_t *cookie)
1523 {
1524         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1525                 rcu_read_lock();
1526                 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1527                 rcu_read_unlock();
1528         } else {
1529                 unsigned int srcu_idx;
1530
1531                 might_sleep();
1532
1533                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1534                 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1535                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1536         }
1537 }
1538
1539 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1540 {
1541         const int is_sync = op_is_sync(bio->bi_opf);
1542         const int is_flush_fua = op_is_flush(bio->bi_opf);
1543         struct blk_mq_alloc_data data = { .flags = 0 };
1544         struct request *rq;
1545         unsigned int request_count = 0;
1546         struct blk_plug *plug;
1547         struct request *same_queue_rq = NULL;
1548         blk_qc_t cookie;
1549         unsigned int wb_acct;
1550
1551         blk_queue_bounce(q, &bio);
1552
1553         blk_queue_split(q, &bio);
1554
1555         if (!bio_integrity_prep(bio))
1556                 return BLK_QC_T_NONE;
1557
1558         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1559             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1560                 return BLK_QC_T_NONE;
1561
1562         if (blk_mq_sched_bio_merge(q, bio))
1563                 return BLK_QC_T_NONE;
1564
1565         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1566
1567         trace_block_getrq(q, bio, bio->bi_opf);
1568
1569         rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1570         if (unlikely(!rq)) {
1571                 __wbt_done(q->rq_wb, wb_acct);
1572                 if (bio->bi_opf & REQ_NOWAIT)
1573                         bio_wouldblock_error(bio);
1574                 return BLK_QC_T_NONE;
1575         }
1576
1577         wbt_track(&rq->issue_stat, wb_acct);
1578
1579         cookie = request_to_qc_t(data.hctx, rq);
1580
1581         plug = current->plug;
1582         if (unlikely(is_flush_fua)) {
1583                 blk_mq_put_ctx(data.ctx);
1584                 blk_mq_bio_to_request(rq, bio);
1585                 if (q->elevator) {
1586                         blk_mq_sched_insert_request(rq, false, true, true,
1587                                         true);
1588                 } else {
1589                         blk_insert_flush(rq);
1590                         blk_mq_run_hw_queue(data.hctx, true);
1591                 }
1592         } else if (plug && q->nr_hw_queues == 1) {
1593                 struct request *last = NULL;
1594
1595                 blk_mq_put_ctx(data.ctx);
1596                 blk_mq_bio_to_request(rq, bio);
1597
1598                 /*
1599                  * @request_count may become stale because of schedule
1600                  * out, so check the list again.
1601                  */
1602                 if (list_empty(&plug->mq_list))
1603                         request_count = 0;
1604                 else if (blk_queue_nomerges(q))
1605                         request_count = blk_plug_queued_count(q);
1606
1607                 if (!request_count)
1608                         trace_block_plug(q);
1609                 else
1610                         last = list_entry_rq(plug->mq_list.prev);
1611
1612                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1613                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1614                         blk_flush_plug_list(plug, false);
1615                         trace_block_plug(q);
1616                 }
1617
1618                 list_add_tail(&rq->queuelist, &plug->mq_list);
1619         } else if (plug && !blk_queue_nomerges(q)) {
1620                 blk_mq_bio_to_request(rq, bio);
1621
1622                 /*
1623                  * We do limited plugging. If the bio can be merged, do that.
1624                  * Otherwise the existing request in the plug list will be
1625                  * issued. So the plug list will have one request at most
1626                  * The plug list might get flushed before this. If that happens,
1627                  * the plug list is empty, and same_queue_rq is invalid.
1628                  */
1629                 if (list_empty(&plug->mq_list))
1630                         same_queue_rq = NULL;
1631                 if (same_queue_rq)
1632                         list_del_init(&same_queue_rq->queuelist);
1633                 list_add_tail(&rq->queuelist, &plug->mq_list);
1634
1635                 blk_mq_put_ctx(data.ctx);
1636
1637                 if (same_queue_rq) {
1638                         data.hctx = blk_mq_map_queue(q,
1639                                         same_queue_rq->mq_ctx->cpu);
1640                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1641                                         &cookie);
1642                 }
1643         } else if (q->nr_hw_queues > 1 && is_sync) {
1644                 blk_mq_put_ctx(data.ctx);
1645                 blk_mq_bio_to_request(rq, bio);
1646                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1647         } else if (q->elevator) {
1648                 blk_mq_put_ctx(data.ctx);
1649                 blk_mq_bio_to_request(rq, bio);
1650                 blk_mq_sched_insert_request(rq, false, true, true, true);
1651         } else {
1652                 blk_mq_put_ctx(data.ctx);
1653                 blk_mq_bio_to_request(rq, bio);
1654                 blk_mq_queue_io(data.hctx, data.ctx, rq);
1655                 blk_mq_run_hw_queue(data.hctx, true);
1656         }
1657
1658         return cookie;
1659 }
1660
1661 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1662                      unsigned int hctx_idx)
1663 {
1664         struct page *page;
1665
1666         if (tags->rqs && set->ops->exit_request) {
1667                 int i;
1668
1669                 for (i = 0; i < tags->nr_tags; i++) {
1670                         struct request *rq = tags->static_rqs[i];
1671
1672                         if (!rq)
1673                                 continue;
1674                         set->ops->exit_request(set, rq, hctx_idx);
1675                         tags->static_rqs[i] = NULL;
1676                 }
1677         }
1678
1679         while (!list_empty(&tags->page_list)) {
1680                 page = list_first_entry(&tags->page_list, struct page, lru);
1681                 list_del_init(&page->lru);
1682                 /*
1683                  * Remove kmemleak object previously allocated in
1684                  * blk_mq_init_rq_map().
1685                  */
1686                 kmemleak_free(page_address(page));
1687                 __free_pages(page, page->private);
1688         }
1689 }
1690
1691 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1692 {
1693         kfree(tags->rqs);
1694         tags->rqs = NULL;
1695         kfree(tags->static_rqs);
1696         tags->static_rqs = NULL;
1697
1698         blk_mq_free_tags(tags);
1699 }
1700
1701 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1702                                         unsigned int hctx_idx,
1703                                         unsigned int nr_tags,
1704                                         unsigned int reserved_tags)
1705 {
1706         struct blk_mq_tags *tags;
1707         int node;
1708
1709         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1710         if (node == NUMA_NO_NODE)
1711                 node = set->numa_node;
1712
1713         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1714                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1715         if (!tags)
1716                 return NULL;
1717
1718         tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1719                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1720                                  node);
1721         if (!tags->rqs) {
1722                 blk_mq_free_tags(tags);
1723                 return NULL;
1724         }
1725
1726         tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1727                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1728                                  node);
1729         if (!tags->static_rqs) {
1730                 kfree(tags->rqs);
1731                 blk_mq_free_tags(tags);
1732                 return NULL;
1733         }
1734
1735         return tags;
1736 }
1737
1738 static size_t order_to_size(unsigned int order)
1739 {
1740         return (size_t)PAGE_SIZE << order;
1741 }
1742
1743 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1744                      unsigned int hctx_idx, unsigned int depth)
1745 {
1746         unsigned int i, j, entries_per_page, max_order = 4;
1747         size_t rq_size, left;
1748         int node;
1749
1750         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1751         if (node == NUMA_NO_NODE)
1752                 node = set->numa_node;
1753
1754         INIT_LIST_HEAD(&tags->page_list);
1755
1756         /*
1757          * rq_size is the size of the request plus driver payload, rounded
1758          * to the cacheline size
1759          */
1760         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1761                                 cache_line_size());
1762         left = rq_size * depth;
1763
1764         for (i = 0; i < depth; ) {
1765                 int this_order = max_order;
1766                 struct page *page;
1767                 int to_do;
1768                 void *p;
1769
1770                 while (this_order && left < order_to_size(this_order - 1))
1771                         this_order--;
1772
1773                 do {
1774                         page = alloc_pages_node(node,
1775                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1776                                 this_order);
1777                         if (page)
1778                                 break;
1779                         if (!this_order--)
1780                                 break;
1781                         if (order_to_size(this_order) < rq_size)
1782                                 break;
1783                 } while (1);
1784
1785                 if (!page)
1786                         goto fail;
1787
1788                 page->private = this_order;
1789                 list_add_tail(&page->lru, &tags->page_list);
1790
1791                 p = page_address(page);
1792                 /*
1793                  * Allow kmemleak to scan these pages as they contain pointers
1794                  * to additional allocations like via ops->init_request().
1795                  */
1796                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1797                 entries_per_page = order_to_size(this_order) / rq_size;
1798                 to_do = min(entries_per_page, depth - i);
1799                 left -= to_do * rq_size;
1800                 for (j = 0; j < to_do; j++) {
1801                         struct request *rq = p;
1802
1803                         tags->static_rqs[i] = rq;
1804                         if (set->ops->init_request) {
1805                                 if (set->ops->init_request(set, rq, hctx_idx,
1806                                                 node)) {
1807                                         tags->static_rqs[i] = NULL;
1808                                         goto fail;
1809                                 }
1810                         }
1811
1812                         p += rq_size;
1813                         i++;
1814                 }
1815         }
1816         return 0;
1817
1818 fail:
1819         blk_mq_free_rqs(set, tags, hctx_idx);
1820         return -ENOMEM;
1821 }
1822
1823 /*
1824  * 'cpu' is going away. splice any existing rq_list entries from this
1825  * software queue to the hw queue dispatch list, and ensure that it
1826  * gets run.
1827  */
1828 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1829 {
1830         struct blk_mq_hw_ctx *hctx;
1831         struct blk_mq_ctx *ctx;
1832         LIST_HEAD(tmp);
1833
1834         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1835         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1836
1837         spin_lock(&ctx->lock);
1838         if (!list_empty(&ctx->rq_list)) {
1839                 list_splice_init(&ctx->rq_list, &tmp);
1840                 blk_mq_hctx_clear_pending(hctx, ctx);
1841         }
1842         spin_unlock(&ctx->lock);
1843
1844         if (list_empty(&tmp))
1845                 return 0;
1846
1847         spin_lock(&hctx->lock);
1848         list_splice_tail_init(&tmp, &hctx->dispatch);
1849         spin_unlock(&hctx->lock);
1850
1851         blk_mq_run_hw_queue(hctx, true);
1852         return 0;
1853 }
1854
1855 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1856 {
1857         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1858                                             &hctx->cpuhp_dead);
1859 }
1860
1861 /* hctx->ctxs will be freed in queue's release handler */
1862 static void blk_mq_exit_hctx(struct request_queue *q,
1863                 struct blk_mq_tag_set *set,
1864                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1865 {
1866         blk_mq_debugfs_unregister_hctx(hctx);
1867
1868         blk_mq_tag_idle(hctx);
1869
1870         if (set->ops->exit_request)
1871                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1872
1873         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1874
1875         if (set->ops->exit_hctx)
1876                 set->ops->exit_hctx(hctx, hctx_idx);
1877
1878         if (hctx->flags & BLK_MQ_F_BLOCKING)
1879                 cleanup_srcu_struct(hctx->queue_rq_srcu);
1880
1881         blk_mq_remove_cpuhp(hctx);
1882         blk_free_flush_queue(hctx->fq);
1883         sbitmap_free(&hctx->ctx_map);
1884 }
1885
1886 static void blk_mq_exit_hw_queues(struct request_queue *q,
1887                 struct blk_mq_tag_set *set, int nr_queue)
1888 {
1889         struct blk_mq_hw_ctx *hctx;
1890         unsigned int i;
1891
1892         queue_for_each_hw_ctx(q, hctx, i) {
1893                 if (i == nr_queue)
1894                         break;
1895                 blk_mq_exit_hctx(q, set, hctx, i);
1896         }
1897 }
1898
1899 static int blk_mq_init_hctx(struct request_queue *q,
1900                 struct blk_mq_tag_set *set,
1901                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1902 {
1903         int node;
1904
1905         node = hctx->numa_node;
1906         if (node == NUMA_NO_NODE)
1907                 node = hctx->numa_node = set->numa_node;
1908
1909         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1910         spin_lock_init(&hctx->lock);
1911         INIT_LIST_HEAD(&hctx->dispatch);
1912         hctx->queue = q;
1913         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1914
1915         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1916
1917         hctx->tags = set->tags[hctx_idx];
1918
1919         /*
1920          * Allocate space for all possible cpus to avoid allocation at
1921          * runtime
1922          */
1923         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1924                                         GFP_KERNEL, node);
1925         if (!hctx->ctxs)
1926                 goto unregister_cpu_notifier;
1927
1928         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1929                               node))
1930                 goto free_ctxs;
1931
1932         hctx->nr_ctx = 0;
1933
1934         if (set->ops->init_hctx &&
1935             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1936                 goto free_bitmap;
1937
1938         if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1939                 goto exit_hctx;
1940
1941         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1942         if (!hctx->fq)
1943                 goto sched_exit_hctx;
1944
1945         if (set->ops->init_request &&
1946             set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
1947                                    node))
1948                 goto free_fq;
1949
1950         if (hctx->flags & BLK_MQ_F_BLOCKING)
1951                 init_srcu_struct(hctx->queue_rq_srcu);
1952
1953         blk_mq_debugfs_register_hctx(q, hctx);
1954
1955         return 0;
1956
1957  free_fq:
1958         kfree(hctx->fq);
1959  sched_exit_hctx:
1960         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1961  exit_hctx:
1962         if (set->ops->exit_hctx)
1963                 set->ops->exit_hctx(hctx, hctx_idx);
1964  free_bitmap:
1965         sbitmap_free(&hctx->ctx_map);
1966  free_ctxs:
1967         kfree(hctx->ctxs);
1968  unregister_cpu_notifier:
1969         blk_mq_remove_cpuhp(hctx);
1970         return -1;
1971 }
1972
1973 static void blk_mq_init_cpu_queues(struct request_queue *q,
1974                                    unsigned int nr_hw_queues)
1975 {
1976         unsigned int i;
1977
1978         for_each_possible_cpu(i) {
1979                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1980                 struct blk_mq_hw_ctx *hctx;
1981
1982                 __ctx->cpu = i;
1983                 spin_lock_init(&__ctx->lock);
1984                 INIT_LIST_HEAD(&__ctx->rq_list);
1985                 __ctx->queue = q;
1986
1987                 /* If the cpu isn't present, the cpu is mapped to first hctx */
1988                 if (!cpu_present(i))
1989                         continue;
1990
1991                 hctx = blk_mq_map_queue(q, i);
1992
1993                 /*
1994                  * Set local node, IFF we have more than one hw queue. If
1995                  * not, we remain on the home node of the device
1996                  */
1997                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1998                         hctx->numa_node = local_memory_node(cpu_to_node(i));
1999         }
2000 }
2001
2002 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2003 {
2004         int ret = 0;
2005
2006         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2007                                         set->queue_depth, set->reserved_tags);
2008         if (!set->tags[hctx_idx])
2009                 return false;
2010
2011         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2012                                 set->queue_depth);
2013         if (!ret)
2014                 return true;
2015
2016         blk_mq_free_rq_map(set->tags[hctx_idx]);
2017         set->tags[hctx_idx] = NULL;
2018         return false;
2019 }
2020
2021 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2022                                          unsigned int hctx_idx)
2023 {
2024         if (set->tags[hctx_idx]) {
2025                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2026                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2027                 set->tags[hctx_idx] = NULL;
2028         }
2029 }
2030
2031 static void blk_mq_map_swqueue(struct request_queue *q)
2032 {
2033         unsigned int i, hctx_idx;
2034         struct blk_mq_hw_ctx *hctx;
2035         struct blk_mq_ctx *ctx;
2036         struct blk_mq_tag_set *set = q->tag_set;
2037
2038         /*
2039          * Avoid others reading imcomplete hctx->cpumask through sysfs
2040          */
2041         mutex_lock(&q->sysfs_lock);
2042
2043         queue_for_each_hw_ctx(q, hctx, i) {
2044                 cpumask_clear(hctx->cpumask);
2045                 hctx->nr_ctx = 0;
2046         }
2047
2048         /*
2049          * Map software to hardware queues.
2050          *
2051          * If the cpu isn't present, the cpu is mapped to first hctx.
2052          */
2053         for_each_present_cpu(i) {
2054                 hctx_idx = q->mq_map[i];
2055                 /* unmapped hw queue can be remapped after CPU topo changed */
2056                 if (!set->tags[hctx_idx] &&
2057                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2058                         /*
2059                          * If tags initialization fail for some hctx,
2060                          * that hctx won't be brought online.  In this
2061                          * case, remap the current ctx to hctx[0] which
2062                          * is guaranteed to always have tags allocated
2063                          */
2064                         q->mq_map[i] = 0;
2065                 }
2066
2067                 ctx = per_cpu_ptr(q->queue_ctx, i);
2068                 hctx = blk_mq_map_queue(q, i);
2069
2070                 cpumask_set_cpu(i, hctx->cpumask);
2071                 ctx->index_hw = hctx->nr_ctx;
2072                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2073         }
2074
2075         mutex_unlock(&q->sysfs_lock);
2076
2077         queue_for_each_hw_ctx(q, hctx, i) {
2078                 /*
2079                  * If no software queues are mapped to this hardware queue,
2080                  * disable it and free the request entries.
2081                  */
2082                 if (!hctx->nr_ctx) {
2083                         /* Never unmap queue 0.  We need it as a
2084                          * fallback in case of a new remap fails
2085                          * allocation
2086                          */
2087                         if (i && set->tags[i])
2088                                 blk_mq_free_map_and_requests(set, i);
2089
2090                         hctx->tags = NULL;
2091                         continue;
2092                 }
2093
2094                 hctx->tags = set->tags[i];
2095                 WARN_ON(!hctx->tags);
2096
2097                 /*
2098                  * Set the map size to the number of mapped software queues.
2099                  * This is more accurate and more efficient than looping
2100                  * over all possibly mapped software queues.
2101                  */
2102                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2103
2104                 /*
2105                  * Initialize batch roundrobin counts
2106                  */
2107                 hctx->next_cpu = cpumask_first(hctx->cpumask);
2108                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2109         }
2110 }
2111
2112 /*
2113  * Caller needs to ensure that we're either frozen/quiesced, or that
2114  * the queue isn't live yet.
2115  */
2116 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2117 {
2118         struct blk_mq_hw_ctx *hctx;
2119         int i;
2120
2121         queue_for_each_hw_ctx(q, hctx, i) {
2122                 if (shared) {
2123                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2124                                 atomic_inc(&q->shared_hctx_restart);
2125                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2126                 } else {
2127                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2128                                 atomic_dec(&q->shared_hctx_restart);
2129                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2130                 }
2131         }
2132 }
2133
2134 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2135                                         bool shared)
2136 {
2137         struct request_queue *q;
2138
2139         lockdep_assert_held(&set->tag_list_lock);
2140
2141         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2142                 blk_mq_freeze_queue(q);
2143                 queue_set_hctx_shared(q, shared);
2144                 blk_mq_unfreeze_queue(q);
2145         }
2146 }
2147
2148 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2149 {
2150         struct blk_mq_tag_set *set = q->tag_set;
2151
2152         mutex_lock(&set->tag_list_lock);
2153         list_del_rcu(&q->tag_set_list);
2154         INIT_LIST_HEAD(&q->tag_set_list);
2155         if (list_is_singular(&set->tag_list)) {
2156                 /* just transitioned to unshared */
2157                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2158                 /* update existing queue */
2159                 blk_mq_update_tag_set_depth(set, false);
2160         }
2161         mutex_unlock(&set->tag_list_lock);
2162
2163         synchronize_rcu();
2164 }
2165
2166 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2167                                      struct request_queue *q)
2168 {
2169         q->tag_set = set;
2170
2171         mutex_lock(&set->tag_list_lock);
2172
2173         /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2174         if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2175                 set->flags |= BLK_MQ_F_TAG_SHARED;
2176                 /* update existing queue */
2177                 blk_mq_update_tag_set_depth(set, true);
2178         }
2179         if (set->flags & BLK_MQ_F_TAG_SHARED)
2180                 queue_set_hctx_shared(q, true);
2181         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2182
2183         mutex_unlock(&set->tag_list_lock);
2184 }
2185
2186 /*
2187  * It is the actual release handler for mq, but we do it from
2188  * request queue's release handler for avoiding use-after-free
2189  * and headache because q->mq_kobj shouldn't have been introduced,
2190  * but we can't group ctx/kctx kobj without it.
2191  */
2192 void blk_mq_release(struct request_queue *q)
2193 {
2194         struct blk_mq_hw_ctx *hctx;
2195         unsigned int i;
2196
2197         /* hctx kobj stays in hctx */
2198         queue_for_each_hw_ctx(q, hctx, i) {
2199                 if (!hctx)
2200                         continue;
2201                 kobject_put(&hctx->kobj);
2202         }
2203
2204         q->mq_map = NULL;
2205
2206         kfree(q->queue_hw_ctx);
2207
2208         /*
2209          * release .mq_kobj and sw queue's kobject now because
2210          * both share lifetime with request queue.
2211          */
2212         blk_mq_sysfs_deinit(q);
2213
2214         free_percpu(q->queue_ctx);
2215 }
2216
2217 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2218 {
2219         struct request_queue *uninit_q, *q;
2220
2221         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2222         if (!uninit_q)
2223                 return ERR_PTR(-ENOMEM);
2224
2225         q = blk_mq_init_allocated_queue(set, uninit_q);
2226         if (IS_ERR(q))
2227                 blk_cleanup_queue(uninit_q);
2228
2229         return q;
2230 }
2231 EXPORT_SYMBOL(blk_mq_init_queue);
2232
2233 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2234 {
2235         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2236
2237         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2238                            __alignof__(struct blk_mq_hw_ctx)) !=
2239                      sizeof(struct blk_mq_hw_ctx));
2240
2241         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2242                 hw_ctx_size += sizeof(struct srcu_struct);
2243
2244         return hw_ctx_size;
2245 }
2246
2247 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2248                                                 struct request_queue *q)
2249 {
2250         int i, j;
2251         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2252
2253         blk_mq_sysfs_unregister(q);
2254         for (i = 0; i < set->nr_hw_queues; i++) {
2255                 int node;
2256
2257                 if (hctxs[i])
2258                         continue;
2259
2260                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2261                 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2262                                         GFP_KERNEL, node);
2263                 if (!hctxs[i])
2264                         break;
2265
2266                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2267                                                 node)) {
2268                         kfree(hctxs[i]);
2269                         hctxs[i] = NULL;
2270                         break;
2271                 }
2272
2273                 atomic_set(&hctxs[i]->nr_active, 0);
2274                 hctxs[i]->numa_node = node;
2275                 hctxs[i]->queue_num = i;
2276
2277                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2278                         free_cpumask_var(hctxs[i]->cpumask);
2279                         kfree(hctxs[i]);
2280                         hctxs[i] = NULL;
2281                         break;
2282                 }
2283                 blk_mq_hctx_kobj_init(hctxs[i]);
2284         }
2285         for (j = i; j < q->nr_hw_queues; j++) {
2286                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2287
2288                 if (hctx) {
2289                         if (hctx->tags)
2290                                 blk_mq_free_map_and_requests(set, j);
2291                         blk_mq_exit_hctx(q, set, hctx, j);
2292                         kobject_put(&hctx->kobj);
2293                         hctxs[j] = NULL;
2294
2295                 }
2296         }
2297         q->nr_hw_queues = i;
2298         blk_mq_sysfs_register(q);
2299 }
2300
2301 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2302                                                   struct request_queue *q)
2303 {
2304         /* mark the queue as mq asap */
2305         q->mq_ops = set->ops;
2306
2307         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2308                                              blk_mq_poll_stats_bkt,
2309                                              BLK_MQ_POLL_STATS_BKTS, q);
2310         if (!q->poll_cb)
2311                 goto err_exit;
2312
2313         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2314         if (!q->queue_ctx)
2315                 goto err_exit;
2316
2317         /* init q->mq_kobj and sw queues' kobjects */
2318         blk_mq_sysfs_init(q);
2319
2320         q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2321                                                 GFP_KERNEL, set->numa_node);
2322         if (!q->queue_hw_ctx)
2323                 goto err_percpu;
2324
2325         q->mq_map = set->mq_map;
2326
2327         blk_mq_realloc_hw_ctxs(set, q);
2328         if (!q->nr_hw_queues)
2329                 goto err_hctxs;
2330
2331         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2332         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2333
2334         q->nr_queues = nr_cpu_ids;
2335
2336         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2337
2338         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2339                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2340
2341         q->sg_reserved_size = INT_MAX;
2342
2343         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2344         INIT_LIST_HEAD(&q->requeue_list);
2345         spin_lock_init(&q->requeue_lock);
2346
2347         blk_queue_make_request(q, blk_mq_make_request);
2348
2349         /*
2350          * Do this after blk_queue_make_request() overrides it...
2351          */
2352         q->nr_requests = set->queue_depth;
2353
2354         /*
2355          * Default to classic polling
2356          */
2357         q->poll_nsec = -1;
2358
2359         if (set->ops->complete)
2360                 blk_queue_softirq_done(q, set->ops->complete);
2361
2362         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2363         blk_mq_add_queue_tag_set(set, q);
2364         blk_mq_map_swqueue(q);
2365
2366         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2367                 int ret;
2368
2369                 ret = blk_mq_sched_init(q);
2370                 if (ret)
2371                         return ERR_PTR(ret);
2372         }
2373
2374         return q;
2375
2376 err_hctxs:
2377         kfree(q->queue_hw_ctx);
2378 err_percpu:
2379         free_percpu(q->queue_ctx);
2380 err_exit:
2381         q->mq_ops = NULL;
2382         return ERR_PTR(-ENOMEM);
2383 }
2384 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2385
2386 void blk_mq_free_queue(struct request_queue *q)
2387 {
2388         struct blk_mq_tag_set   *set = q->tag_set;
2389
2390         blk_mq_del_queue_tag_set(q);
2391         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2392 }
2393
2394 /* Basically redo blk_mq_init_queue with queue frozen */
2395 static void blk_mq_queue_reinit(struct request_queue *q)
2396 {
2397         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2398
2399         blk_mq_debugfs_unregister_hctxs(q);
2400         blk_mq_sysfs_unregister(q);
2401
2402         /*
2403          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2404          * we should change hctx numa_node according to new topology (this
2405          * involves free and re-allocate memory, worthy doing?)
2406          */
2407
2408         blk_mq_map_swqueue(q);
2409
2410         blk_mq_sysfs_register(q);
2411         blk_mq_debugfs_register_hctxs(q);
2412 }
2413
2414 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2415 {
2416         int i;
2417
2418         for (i = 0; i < set->nr_hw_queues; i++)
2419                 if (!__blk_mq_alloc_rq_map(set, i))
2420                         goto out_unwind;
2421
2422         return 0;
2423
2424 out_unwind:
2425         while (--i >= 0)
2426                 blk_mq_free_rq_map(set->tags[i]);
2427
2428         return -ENOMEM;
2429 }
2430
2431 /*
2432  * Allocate the request maps associated with this tag_set. Note that this
2433  * may reduce the depth asked for, if memory is tight. set->queue_depth
2434  * will be updated to reflect the allocated depth.
2435  */
2436 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2437 {
2438         unsigned int depth;
2439         int err;
2440
2441         depth = set->queue_depth;
2442         do {
2443                 err = __blk_mq_alloc_rq_maps(set);
2444                 if (!err)
2445                         break;
2446
2447                 set->queue_depth >>= 1;
2448                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2449                         err = -ENOMEM;
2450                         break;
2451                 }
2452         } while (set->queue_depth);
2453
2454         if (!set->queue_depth || err) {
2455                 pr_err("blk-mq: failed to allocate request map\n");
2456                 return -ENOMEM;
2457         }
2458
2459         if (depth != set->queue_depth)
2460                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2461                                                 depth, set->queue_depth);
2462
2463         return 0;
2464 }
2465
2466 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2467 {
2468         if (set->ops->map_queues)
2469                 return set->ops->map_queues(set);
2470         else
2471                 return blk_mq_map_queues(set);
2472 }
2473
2474 /*
2475  * Alloc a tag set to be associated with one or more request queues.
2476  * May fail with EINVAL for various error conditions. May adjust the
2477  * requested depth down, if if it too large. In that case, the set
2478  * value will be stored in set->queue_depth.
2479  */
2480 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2481 {
2482         int ret;
2483
2484         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2485
2486         if (!set->nr_hw_queues)
2487                 return -EINVAL;
2488         if (!set->queue_depth)
2489                 return -EINVAL;
2490         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2491                 return -EINVAL;
2492
2493         if (!set->ops->queue_rq)
2494                 return -EINVAL;
2495
2496         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2497                 pr_info("blk-mq: reduced tag depth to %u\n",
2498                         BLK_MQ_MAX_DEPTH);
2499                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2500         }
2501
2502         /*
2503          * If a crashdump is active, then we are potentially in a very
2504          * memory constrained environment. Limit us to 1 queue and
2505          * 64 tags to prevent using too much memory.
2506          */
2507         if (is_kdump_kernel()) {
2508                 set->nr_hw_queues = 1;
2509                 set->queue_depth = min(64U, set->queue_depth);
2510         }
2511         /*
2512          * There is no use for more h/w queues than cpus.
2513          */
2514         if (set->nr_hw_queues > nr_cpu_ids)
2515                 set->nr_hw_queues = nr_cpu_ids;
2516
2517         set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2518                                  GFP_KERNEL, set->numa_node);
2519         if (!set->tags)
2520                 return -ENOMEM;
2521
2522         ret = -ENOMEM;
2523         set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2524                         GFP_KERNEL, set->numa_node);
2525         if (!set->mq_map)
2526                 goto out_free_tags;
2527
2528         ret = blk_mq_update_queue_map(set);
2529         if (ret)
2530                 goto out_free_mq_map;
2531
2532         ret = blk_mq_alloc_rq_maps(set);
2533         if (ret)
2534                 goto out_free_mq_map;
2535
2536         mutex_init(&set->tag_list_lock);
2537         INIT_LIST_HEAD(&set->tag_list);
2538
2539         return 0;
2540
2541 out_free_mq_map:
2542         kfree(set->mq_map);
2543         set->mq_map = NULL;
2544 out_free_tags:
2545         kfree(set->tags);
2546         set->tags = NULL;
2547         return ret;
2548 }
2549 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2550
2551 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2552 {
2553         int i;
2554
2555         for (i = 0; i < nr_cpu_ids; i++)
2556                 blk_mq_free_map_and_requests(set, i);
2557
2558         kfree(set->mq_map);
2559         set->mq_map = NULL;
2560
2561         kfree(set->tags);
2562         set->tags = NULL;
2563 }
2564 EXPORT_SYMBOL(blk_mq_free_tag_set);
2565
2566 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2567 {
2568         struct blk_mq_tag_set *set = q->tag_set;
2569         struct blk_mq_hw_ctx *hctx;
2570         int i, ret;
2571
2572         if (!set)
2573                 return -EINVAL;
2574
2575         blk_mq_freeze_queue(q);
2576
2577         ret = 0;
2578         queue_for_each_hw_ctx(q, hctx, i) {
2579                 if (!hctx->tags)
2580                         continue;
2581                 /*
2582                  * If we're using an MQ scheduler, just update the scheduler
2583                  * queue depth. This is similar to what the old code would do.
2584                  */
2585                 if (!hctx->sched_tags) {
2586                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2587                                                         min(nr, set->queue_depth),
2588                                                         false);
2589                 } else {
2590                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2591                                                         nr, true);
2592                 }
2593                 if (ret)
2594                         break;
2595         }
2596
2597         if (!ret)
2598                 q->nr_requests = nr;
2599
2600         blk_mq_unfreeze_queue(q);
2601
2602         return ret;
2603 }
2604
2605 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2606                                                         int nr_hw_queues)
2607 {
2608         struct request_queue *q;
2609
2610         lockdep_assert_held(&set->tag_list_lock);
2611
2612         if (nr_hw_queues > nr_cpu_ids)
2613                 nr_hw_queues = nr_cpu_ids;
2614         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2615                 return;
2616
2617         list_for_each_entry(q, &set->tag_list, tag_set_list)
2618                 blk_mq_freeze_queue(q);
2619
2620         set->nr_hw_queues = nr_hw_queues;
2621         blk_mq_update_queue_map(set);
2622         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2623                 blk_mq_realloc_hw_ctxs(set, q);
2624                 blk_mq_queue_reinit(q);
2625         }
2626
2627         list_for_each_entry(q, &set->tag_list, tag_set_list)
2628                 blk_mq_unfreeze_queue(q);
2629 }
2630
2631 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2632 {
2633         mutex_lock(&set->tag_list_lock);
2634         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2635         mutex_unlock(&set->tag_list_lock);
2636 }
2637 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2638
2639 /* Enable polling stats and return whether they were already enabled. */
2640 static bool blk_poll_stats_enable(struct request_queue *q)
2641 {
2642         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2643             test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2644                 return true;
2645         blk_stat_add_callback(q, q->poll_cb);
2646         return false;
2647 }
2648
2649 static void blk_mq_poll_stats_start(struct request_queue *q)
2650 {
2651         /*
2652          * We don't arm the callback if polling stats are not enabled or the
2653          * callback is already active.
2654          */
2655         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2656             blk_stat_is_active(q->poll_cb))
2657                 return;
2658
2659         blk_stat_activate_msecs(q->poll_cb, 100);
2660 }
2661
2662 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2663 {
2664         struct request_queue *q = cb->data;
2665         int bucket;
2666
2667         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2668                 if (cb->stat[bucket].nr_samples)
2669                         q->poll_stat[bucket] = cb->stat[bucket];
2670         }
2671 }
2672
2673 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2674                                        struct blk_mq_hw_ctx *hctx,
2675                                        struct request *rq)
2676 {
2677         unsigned long ret = 0;
2678         int bucket;
2679
2680         /*
2681          * If stats collection isn't on, don't sleep but turn it on for
2682          * future users
2683          */
2684         if (!blk_poll_stats_enable(q))
2685                 return 0;
2686
2687         /*
2688          * As an optimistic guess, use half of the mean service time
2689          * for this type of request. We can (and should) make this smarter.
2690          * For instance, if the completion latencies are tight, we can
2691          * get closer than just half the mean. This is especially
2692          * important on devices where the completion latencies are longer
2693          * than ~10 usec. We do use the stats for the relevant IO size
2694          * if available which does lead to better estimates.
2695          */
2696         bucket = blk_mq_poll_stats_bkt(rq);
2697         if (bucket < 0)
2698                 return ret;
2699
2700         if (q->poll_stat[bucket].nr_samples)
2701                 ret = (q->poll_stat[bucket].mean + 1) / 2;
2702
2703         return ret;
2704 }
2705
2706 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2707                                      struct blk_mq_hw_ctx *hctx,
2708                                      struct request *rq)
2709 {
2710         struct hrtimer_sleeper hs;
2711         enum hrtimer_mode mode;
2712         unsigned int nsecs;
2713         ktime_t kt;
2714
2715         if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2716                 return false;
2717
2718         /*
2719          * poll_nsec can be:
2720          *
2721          * -1:  don't ever hybrid sleep
2722          *  0:  use half of prev avg
2723          * >0:  use this specific value
2724          */
2725         if (q->poll_nsec == -1)
2726                 return false;
2727         else if (q->poll_nsec > 0)
2728                 nsecs = q->poll_nsec;
2729         else
2730                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2731
2732         if (!nsecs)
2733                 return false;
2734
2735         set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2736
2737         /*
2738          * This will be replaced with the stats tracking code, using
2739          * 'avg_completion_time / 2' as the pre-sleep target.
2740          */
2741         kt = nsecs;
2742
2743         mode = HRTIMER_MODE_REL;
2744         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2745         hrtimer_set_expires(&hs.timer, kt);
2746
2747         hrtimer_init_sleeper(&hs, current);
2748         do {
2749                 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2750                         break;
2751                 set_current_state(TASK_UNINTERRUPTIBLE);
2752                 hrtimer_start_expires(&hs.timer, mode);
2753                 if (hs.task)
2754                         io_schedule();
2755                 hrtimer_cancel(&hs.timer);
2756                 mode = HRTIMER_MODE_ABS;
2757         } while (hs.task && !signal_pending(current));
2758
2759         __set_current_state(TASK_RUNNING);
2760         destroy_hrtimer_on_stack(&hs.timer);
2761         return true;
2762 }
2763
2764 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2765 {
2766         struct request_queue *q = hctx->queue;
2767         long state;
2768
2769         /*
2770          * If we sleep, have the caller restart the poll loop to reset
2771          * the state. Like for the other success return cases, the
2772          * caller is responsible for checking if the IO completed. If
2773          * the IO isn't complete, we'll get called again and will go
2774          * straight to the busy poll loop.
2775          */
2776         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2777                 return true;
2778
2779         hctx->poll_considered++;
2780
2781         state = current->state;
2782         while (!need_resched()) {
2783                 int ret;
2784
2785                 hctx->poll_invoked++;
2786
2787                 ret = q->mq_ops->poll(hctx, rq->tag);
2788                 if (ret > 0) {
2789                         hctx->poll_success++;
2790                         set_current_state(TASK_RUNNING);
2791                         return true;
2792                 }
2793
2794                 if (signal_pending_state(state, current))
2795                         set_current_state(TASK_RUNNING);
2796
2797                 if (current->state == TASK_RUNNING)
2798                         return true;
2799                 if (ret < 0)
2800                         break;
2801                 cpu_relax();
2802         }
2803
2804         return false;
2805 }
2806
2807 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2808 {
2809         struct blk_mq_hw_ctx *hctx;
2810         struct blk_plug *plug;
2811         struct request *rq;
2812
2813         if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2814             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2815                 return false;
2816
2817         plug = current->plug;
2818         if (plug)
2819                 blk_flush_plug_list(plug, false);
2820
2821         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2822         if (!blk_qc_t_is_internal(cookie))
2823                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2824         else {
2825                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2826                 /*
2827                  * With scheduling, if the request has completed, we'll
2828                  * get a NULL return here, as we clear the sched tag when
2829                  * that happens. The request still remains valid, like always,
2830                  * so we should be safe with just the NULL check.
2831                  */
2832                 if (!rq)
2833                         return false;
2834         }
2835
2836         return __blk_mq_poll(hctx, rq);
2837 }
2838 EXPORT_SYMBOL_GPL(blk_mq_poll);
2839
2840 static int __init blk_mq_init(void)
2841 {
2842         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2843                                 blk_mq_hctx_notify_dead);
2844         return 0;
2845 }
2846 subsys_initcall(blk_mq_init);