b70a4ad78b635798d0bfa2e25a7a755827079ac9
[sfrench/cifs-2.6.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
305         blk_queue_enter_live(q);
306         data->q = q;
307         if (likely(!data->ctx))
308                 data->ctx = blk_mq_get_ctx(q);
309         if (likely(!data->hctx))
310                 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
311         if (op & REQ_NOWAIT)
312                 data->flags |= BLK_MQ_REQ_NOWAIT;
313
314         if (e) {
315                 data->flags |= BLK_MQ_REQ_INTERNAL;
316
317                 /*
318                  * Flush requests are special and go directly to the
319                  * dispatch list.
320                  */
321                 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
322                         e->type->ops.mq.limit_depth(op, data);
323         }
324
325         tag = blk_mq_get_tag(data);
326         if (tag == BLK_MQ_TAG_FAIL) {
327                 blk_queue_exit(q);
328                 return NULL;
329         }
330
331         rq = blk_mq_rq_ctx_init(data, tag, op);
332         if (!op_is_flush(op)) {
333                 rq->elv.icq = NULL;
334                 if (e && e->type->ops.mq.prepare_request) {
335                         if (e->type->icq_cache && rq_ioc(bio))
336                                 blk_mq_sched_assign_ioc(rq, bio);
337
338                         e->type->ops.mq.prepare_request(rq, bio);
339                         rq->rq_flags |= RQF_ELVPRIV;
340                 }
341         }
342         data->hctx->queued++;
343         return rq;
344 }
345
346 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
347                 unsigned int flags)
348 {
349         struct blk_mq_alloc_data alloc_data = { .flags = flags };
350         struct request *rq;
351         int ret;
352
353         ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
354         if (ret)
355                 return ERR_PTR(ret);
356
357         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
358
359         blk_mq_put_ctx(alloc_data.ctx);
360         blk_queue_exit(q);
361
362         if (!rq)
363                 return ERR_PTR(-EWOULDBLOCK);
364
365         rq->__data_len = 0;
366         rq->__sector = (sector_t) -1;
367         rq->bio = rq->biotail = NULL;
368         return rq;
369 }
370 EXPORT_SYMBOL(blk_mq_alloc_request);
371
372 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
373                 unsigned int op, unsigned int flags, unsigned int hctx_idx)
374 {
375         struct blk_mq_alloc_data alloc_data = { .flags = flags };
376         struct request *rq;
377         unsigned int cpu;
378         int ret;
379
380         /*
381          * If the tag allocator sleeps we could get an allocation for a
382          * different hardware context.  No need to complicate the low level
383          * allocator for this for the rare use case of a command tied to
384          * a specific queue.
385          */
386         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
387                 return ERR_PTR(-EINVAL);
388
389         if (hctx_idx >= q->nr_hw_queues)
390                 return ERR_PTR(-EIO);
391
392         ret = blk_queue_enter(q, true);
393         if (ret)
394                 return ERR_PTR(ret);
395
396         /*
397          * Check if the hardware context is actually mapped to anything.
398          * If not tell the caller that it should skip this queue.
399          */
400         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
401         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
402                 blk_queue_exit(q);
403                 return ERR_PTR(-EXDEV);
404         }
405         cpu = cpumask_first(alloc_data.hctx->cpumask);
406         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
407
408         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
409
410         blk_queue_exit(q);
411
412         if (!rq)
413                 return ERR_PTR(-EWOULDBLOCK);
414
415         return rq;
416 }
417 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
418
419 void blk_mq_free_request(struct request *rq)
420 {
421         struct request_queue *q = rq->q;
422         struct elevator_queue *e = q->elevator;
423         struct blk_mq_ctx *ctx = rq->mq_ctx;
424         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
425         const int sched_tag = rq->internal_tag;
426
427         if (rq->rq_flags & RQF_ELVPRIV) {
428                 if (e && e->type->ops.mq.finish_request)
429                         e->type->ops.mq.finish_request(rq);
430                 if (rq->elv.icq) {
431                         put_io_context(rq->elv.icq->ioc);
432                         rq->elv.icq = NULL;
433                 }
434         }
435
436         ctx->rq_completed[rq_is_sync(rq)]++;
437         if (rq->rq_flags & RQF_MQ_INFLIGHT)
438                 atomic_dec(&hctx->nr_active);
439
440         wbt_done(q->rq_wb, &rq->issue_stat);
441
442         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
443         clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
444         if (rq->tag != -1)
445                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
446         if (sched_tag != -1)
447                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
448         blk_mq_sched_restart(hctx);
449         blk_queue_exit(q);
450 }
451 EXPORT_SYMBOL_GPL(blk_mq_free_request);
452
453 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
454 {
455         blk_account_io_done(rq);
456
457         if (rq->end_io) {
458                 wbt_done(rq->q->rq_wb, &rq->issue_stat);
459                 rq->end_io(rq, error);
460         } else {
461                 if (unlikely(blk_bidi_rq(rq)))
462                         blk_mq_free_request(rq->next_rq);
463                 blk_mq_free_request(rq);
464         }
465 }
466 EXPORT_SYMBOL(__blk_mq_end_request);
467
468 void blk_mq_end_request(struct request *rq, blk_status_t error)
469 {
470         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
471                 BUG();
472         __blk_mq_end_request(rq, error);
473 }
474 EXPORT_SYMBOL(blk_mq_end_request);
475
476 static void __blk_mq_complete_request_remote(void *data)
477 {
478         struct request *rq = data;
479
480         rq->q->softirq_done_fn(rq);
481 }
482
483 static void __blk_mq_complete_request(struct request *rq)
484 {
485         struct blk_mq_ctx *ctx = rq->mq_ctx;
486         bool shared = false;
487         int cpu;
488
489         if (rq->internal_tag != -1)
490                 blk_mq_sched_completed_request(rq);
491         if (rq->rq_flags & RQF_STATS) {
492                 blk_mq_poll_stats_start(rq->q);
493                 blk_stat_add(rq);
494         }
495
496         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
497                 rq->q->softirq_done_fn(rq);
498                 return;
499         }
500
501         cpu = get_cpu();
502         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
503                 shared = cpus_share_cache(cpu, ctx->cpu);
504
505         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
506                 rq->csd.func = __blk_mq_complete_request_remote;
507                 rq->csd.info = rq;
508                 rq->csd.flags = 0;
509                 smp_call_function_single_async(ctx->cpu, &rq->csd);
510         } else {
511                 rq->q->softirq_done_fn(rq);
512         }
513         put_cpu();
514 }
515
516 /**
517  * blk_mq_complete_request - end I/O on a request
518  * @rq:         the request being processed
519  *
520  * Description:
521  *      Ends all I/O on a request. It does not handle partial completions.
522  *      The actual completion happens out-of-order, through a IPI handler.
523  **/
524 void blk_mq_complete_request(struct request *rq)
525 {
526         struct request_queue *q = rq->q;
527
528         if (unlikely(blk_should_fake_timeout(q)))
529                 return;
530         if (!blk_mark_rq_complete(rq))
531                 __blk_mq_complete_request(rq);
532 }
533 EXPORT_SYMBOL(blk_mq_complete_request);
534
535 int blk_mq_request_started(struct request *rq)
536 {
537         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
538 }
539 EXPORT_SYMBOL_GPL(blk_mq_request_started);
540
541 void blk_mq_start_request(struct request *rq)
542 {
543         struct request_queue *q = rq->q;
544
545         blk_mq_sched_started_request(rq);
546
547         trace_block_rq_issue(q, rq);
548
549         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
550                 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
551                 rq->rq_flags |= RQF_STATS;
552                 wbt_issue(q->rq_wb, &rq->issue_stat);
553         }
554
555         blk_add_timer(rq);
556
557         /*
558          * Ensure that ->deadline is visible before set the started
559          * flag and clear the completed flag.
560          */
561         smp_mb__before_atomic();
562
563         /*
564          * Mark us as started and clear complete. Complete might have been
565          * set if requeue raced with timeout, which then marked it as
566          * complete. So be sure to clear complete again when we start
567          * the request, otherwise we'll ignore the completion event.
568          */
569         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
570                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
571         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
572                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
573
574         if (q->dma_drain_size && blk_rq_bytes(rq)) {
575                 /*
576                  * Make sure space for the drain appears.  We know we can do
577                  * this because max_hw_segments has been adjusted to be one
578                  * fewer than the device can handle.
579                  */
580                 rq->nr_phys_segments++;
581         }
582 }
583 EXPORT_SYMBOL(blk_mq_start_request);
584
585 /*
586  * When we reach here because queue is busy, REQ_ATOM_COMPLETE
587  * flag isn't set yet, so there may be race with timeout handler,
588  * but given rq->deadline is just set in .queue_rq() under
589  * this situation, the race won't be possible in reality because
590  * rq->timeout should be set as big enough to cover the window
591  * between blk_mq_start_request() called from .queue_rq() and
592  * clearing REQ_ATOM_STARTED here.
593  */
594 static void __blk_mq_requeue_request(struct request *rq)
595 {
596         struct request_queue *q = rq->q;
597
598         trace_block_rq_requeue(q, rq);
599         wbt_requeue(q->rq_wb, &rq->issue_stat);
600         blk_mq_sched_requeue_request(rq);
601
602         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
603                 if (q->dma_drain_size && blk_rq_bytes(rq))
604                         rq->nr_phys_segments--;
605         }
606 }
607
608 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
609 {
610         __blk_mq_requeue_request(rq);
611
612         BUG_ON(blk_queued_rq(rq));
613         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
614 }
615 EXPORT_SYMBOL(blk_mq_requeue_request);
616
617 static void blk_mq_requeue_work(struct work_struct *work)
618 {
619         struct request_queue *q =
620                 container_of(work, struct request_queue, requeue_work.work);
621         LIST_HEAD(rq_list);
622         struct request *rq, *next;
623
624         spin_lock_irq(&q->requeue_lock);
625         list_splice_init(&q->requeue_list, &rq_list);
626         spin_unlock_irq(&q->requeue_lock);
627
628         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
629                 if (!(rq->rq_flags & RQF_SOFTBARRIER))
630                         continue;
631
632                 rq->rq_flags &= ~RQF_SOFTBARRIER;
633                 list_del_init(&rq->queuelist);
634                 blk_mq_sched_insert_request(rq, true, false, false, true);
635         }
636
637         while (!list_empty(&rq_list)) {
638                 rq = list_entry(rq_list.next, struct request, queuelist);
639                 list_del_init(&rq->queuelist);
640                 blk_mq_sched_insert_request(rq, false, false, false, true);
641         }
642
643         blk_mq_run_hw_queues(q, false);
644 }
645
646 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
647                                 bool kick_requeue_list)
648 {
649         struct request_queue *q = rq->q;
650         unsigned long flags;
651
652         /*
653          * We abuse this flag that is otherwise used by the I/O scheduler to
654          * request head insertation from the workqueue.
655          */
656         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
657
658         spin_lock_irqsave(&q->requeue_lock, flags);
659         if (at_head) {
660                 rq->rq_flags |= RQF_SOFTBARRIER;
661                 list_add(&rq->queuelist, &q->requeue_list);
662         } else {
663                 list_add_tail(&rq->queuelist, &q->requeue_list);
664         }
665         spin_unlock_irqrestore(&q->requeue_lock, flags);
666
667         if (kick_requeue_list)
668                 blk_mq_kick_requeue_list(q);
669 }
670 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
671
672 void blk_mq_kick_requeue_list(struct request_queue *q)
673 {
674         kblockd_schedule_delayed_work(&q->requeue_work, 0);
675 }
676 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
677
678 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
679                                     unsigned long msecs)
680 {
681         kblockd_schedule_delayed_work(&q->requeue_work,
682                                       msecs_to_jiffies(msecs));
683 }
684 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
685
686 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
687 {
688         if (tag < tags->nr_tags) {
689                 prefetch(tags->rqs[tag]);
690                 return tags->rqs[tag];
691         }
692
693         return NULL;
694 }
695 EXPORT_SYMBOL(blk_mq_tag_to_rq);
696
697 struct blk_mq_timeout_data {
698         unsigned long next;
699         unsigned int next_set;
700 };
701
702 void blk_mq_rq_timed_out(struct request *req, bool reserved)
703 {
704         const struct blk_mq_ops *ops = req->q->mq_ops;
705         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
706
707         /*
708          * We know that complete is set at this point. If STARTED isn't set
709          * anymore, then the request isn't active and the "timeout" should
710          * just be ignored. This can happen due to the bitflag ordering.
711          * Timeout first checks if STARTED is set, and if it is, assumes
712          * the request is active. But if we race with completion, then
713          * both flags will get cleared. So check here again, and ignore
714          * a timeout event with a request that isn't active.
715          */
716         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
717                 return;
718
719         if (ops->timeout)
720                 ret = ops->timeout(req, reserved);
721
722         switch (ret) {
723         case BLK_EH_HANDLED:
724                 __blk_mq_complete_request(req);
725                 break;
726         case BLK_EH_RESET_TIMER:
727                 blk_add_timer(req);
728                 blk_clear_rq_complete(req);
729                 break;
730         case BLK_EH_NOT_HANDLED:
731                 break;
732         default:
733                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
734                 break;
735         }
736 }
737
738 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
739                 struct request *rq, void *priv, bool reserved)
740 {
741         struct blk_mq_timeout_data *data = priv;
742
743         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
744                 return;
745
746         /*
747          * The rq being checked may have been freed and reallocated
748          * out already here, we avoid this race by checking rq->deadline
749          * and REQ_ATOM_COMPLETE flag together:
750          *
751          * - if rq->deadline is observed as new value because of
752          *   reusing, the rq won't be timed out because of timing.
753          * - if rq->deadline is observed as previous value,
754          *   REQ_ATOM_COMPLETE flag won't be cleared in reuse path
755          *   because we put a barrier between setting rq->deadline
756          *   and clearing the flag in blk_mq_start_request(), so
757          *   this rq won't be timed out too.
758          */
759         if (time_after_eq(jiffies, rq->deadline)) {
760                 if (!blk_mark_rq_complete(rq))
761                         blk_mq_rq_timed_out(rq, reserved);
762         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
763                 data->next = rq->deadline;
764                 data->next_set = 1;
765         }
766 }
767
768 static void blk_mq_timeout_work(struct work_struct *work)
769 {
770         struct request_queue *q =
771                 container_of(work, struct request_queue, timeout_work);
772         struct blk_mq_timeout_data data = {
773                 .next           = 0,
774                 .next_set       = 0,
775         };
776         int i;
777
778         /* A deadlock might occur if a request is stuck requiring a
779          * timeout at the same time a queue freeze is waiting
780          * completion, since the timeout code would not be able to
781          * acquire the queue reference here.
782          *
783          * That's why we don't use blk_queue_enter here; instead, we use
784          * percpu_ref_tryget directly, because we need to be able to
785          * obtain a reference even in the short window between the queue
786          * starting to freeze, by dropping the first reference in
787          * blk_freeze_queue_start, and the moment the last request is
788          * consumed, marked by the instant q_usage_counter reaches
789          * zero.
790          */
791         if (!percpu_ref_tryget(&q->q_usage_counter))
792                 return;
793
794         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
795
796         if (data.next_set) {
797                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
798                 mod_timer(&q->timeout, data.next);
799         } else {
800                 struct blk_mq_hw_ctx *hctx;
801
802                 queue_for_each_hw_ctx(q, hctx, i) {
803                         /* the hctx may be unmapped, so check it here */
804                         if (blk_mq_hw_queue_mapped(hctx))
805                                 blk_mq_tag_idle(hctx);
806                 }
807         }
808         blk_queue_exit(q);
809 }
810
811 struct flush_busy_ctx_data {
812         struct blk_mq_hw_ctx *hctx;
813         struct list_head *list;
814 };
815
816 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
817 {
818         struct flush_busy_ctx_data *flush_data = data;
819         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
820         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
821
822         sbitmap_clear_bit(sb, bitnr);
823         spin_lock(&ctx->lock);
824         list_splice_tail_init(&ctx->rq_list, flush_data->list);
825         spin_unlock(&ctx->lock);
826         return true;
827 }
828
829 /*
830  * Process software queues that have been marked busy, splicing them
831  * to the for-dispatch
832  */
833 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
834 {
835         struct flush_busy_ctx_data data = {
836                 .hctx = hctx,
837                 .list = list,
838         };
839
840         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
841 }
842 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
843
844 static inline unsigned int queued_to_index(unsigned int queued)
845 {
846         if (!queued)
847                 return 0;
848
849         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
850 }
851
852 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
853                            bool wait)
854 {
855         struct blk_mq_alloc_data data = {
856                 .q = rq->q,
857                 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
858                 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
859         };
860
861         might_sleep_if(wait);
862
863         if (rq->tag != -1)
864                 goto done;
865
866         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
867                 data.flags |= BLK_MQ_REQ_RESERVED;
868
869         rq->tag = blk_mq_get_tag(&data);
870         if (rq->tag >= 0) {
871                 if (blk_mq_tag_busy(data.hctx)) {
872                         rq->rq_flags |= RQF_MQ_INFLIGHT;
873                         atomic_inc(&data.hctx->nr_active);
874                 }
875                 data.hctx->tags->rqs[rq->tag] = rq;
876         }
877
878 done:
879         if (hctx)
880                 *hctx = data.hctx;
881         return rq->tag != -1;
882 }
883
884 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
885                                     struct request *rq)
886 {
887         blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
888         rq->tag = -1;
889
890         if (rq->rq_flags & RQF_MQ_INFLIGHT) {
891                 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
892                 atomic_dec(&hctx->nr_active);
893         }
894 }
895
896 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
897                                        struct request *rq)
898 {
899         if (rq->tag == -1 || rq->internal_tag == -1)
900                 return;
901
902         __blk_mq_put_driver_tag(hctx, rq);
903 }
904
905 static void blk_mq_put_driver_tag(struct request *rq)
906 {
907         struct blk_mq_hw_ctx *hctx;
908
909         if (rq->tag == -1 || rq->internal_tag == -1)
910                 return;
911
912         hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
913         __blk_mq_put_driver_tag(hctx, rq);
914 }
915
916 /*
917  * If we fail getting a driver tag because all the driver tags are already
918  * assigned and on the dispatch list, BUT the first entry does not have a
919  * tag, then we could deadlock. For that case, move entries with assigned
920  * driver tags to the front, leaving the set of tagged requests in the
921  * same order, and the untagged set in the same order.
922  */
923 static bool reorder_tags_to_front(struct list_head *list)
924 {
925         struct request *rq, *tmp, *first = NULL;
926
927         list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
928                 if (rq == first)
929                         break;
930                 if (rq->tag != -1) {
931                         list_move(&rq->queuelist, list);
932                         if (!first)
933                                 first = rq;
934                 }
935         }
936
937         return first != NULL;
938 }
939
940 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
941                                 void *key)
942 {
943         struct blk_mq_hw_ctx *hctx;
944
945         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
946
947         list_del(&wait->entry);
948         clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
949         blk_mq_run_hw_queue(hctx, true);
950         return 1;
951 }
952
953 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
954 {
955         struct sbq_wait_state *ws;
956
957         /*
958          * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
959          * The thread which wins the race to grab this bit adds the hardware
960          * queue to the wait queue.
961          */
962         if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
963             test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
964                 return false;
965
966         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
967         ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
968
969         /*
970          * As soon as this returns, it's no longer safe to fiddle with
971          * hctx->dispatch_wait, since a completion can wake up the wait queue
972          * and unlock the bit.
973          */
974         add_wait_queue(&ws->wait, &hctx->dispatch_wait);
975         return true;
976 }
977
978 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
979 {
980         struct blk_mq_hw_ctx *hctx;
981         struct request *rq;
982         int errors, queued;
983
984         if (list_empty(list))
985                 return false;
986
987         /*
988          * Now process all the entries, sending them to the driver.
989          */
990         errors = queued = 0;
991         do {
992                 struct blk_mq_queue_data bd;
993                 blk_status_t ret;
994
995                 rq = list_first_entry(list, struct request, queuelist);
996                 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
997                         if (!queued && reorder_tags_to_front(list))
998                                 continue;
999
1000                         /*
1001                          * The initial allocation attempt failed, so we need to
1002                          * rerun the hardware queue when a tag is freed.
1003                          */
1004                         if (!blk_mq_dispatch_wait_add(hctx))
1005                                 break;
1006
1007                         /*
1008                          * It's possible that a tag was freed in the window
1009                          * between the allocation failure and adding the
1010                          * hardware queue to the wait queue.
1011                          */
1012                         if (!blk_mq_get_driver_tag(rq, &hctx, false))
1013                                 break;
1014                 }
1015
1016                 list_del_init(&rq->queuelist);
1017
1018                 bd.rq = rq;
1019
1020                 /*
1021                  * Flag last if we have no more requests, or if we have more
1022                  * but can't assign a driver tag to it.
1023                  */
1024                 if (list_empty(list))
1025                         bd.last = true;
1026                 else {
1027                         struct request *nxt;
1028
1029                         nxt = list_first_entry(list, struct request, queuelist);
1030                         bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1031                 }
1032
1033                 ret = q->mq_ops->queue_rq(hctx, &bd);
1034                 if (ret == BLK_STS_RESOURCE) {
1035                         blk_mq_put_driver_tag_hctx(hctx, rq);
1036                         list_add(&rq->queuelist, list);
1037                         __blk_mq_requeue_request(rq);
1038                         break;
1039                 }
1040
1041                 if (unlikely(ret != BLK_STS_OK)) {
1042                         errors++;
1043                         blk_mq_end_request(rq, BLK_STS_IOERR);
1044                         continue;
1045                 }
1046
1047                 queued++;
1048         } while (!list_empty(list));
1049
1050         hctx->dispatched[queued_to_index(queued)]++;
1051
1052         /*
1053          * Any items that need requeuing? Stuff them into hctx->dispatch,
1054          * that is where we will continue on next queue run.
1055          */
1056         if (!list_empty(list)) {
1057                 /*
1058                  * If an I/O scheduler has been configured and we got a driver
1059                  * tag for the next request already, free it again.
1060                  */
1061                 rq = list_first_entry(list, struct request, queuelist);
1062                 blk_mq_put_driver_tag(rq);
1063
1064                 spin_lock(&hctx->lock);
1065                 list_splice_init(list, &hctx->dispatch);
1066                 spin_unlock(&hctx->lock);
1067
1068                 /*
1069                  * If SCHED_RESTART was set by the caller of this function and
1070                  * it is no longer set that means that it was cleared by another
1071                  * thread and hence that a queue rerun is needed.
1072                  *
1073                  * If TAG_WAITING is set that means that an I/O scheduler has
1074                  * been configured and another thread is waiting for a driver
1075                  * tag. To guarantee fairness, do not rerun this hardware queue
1076                  * but let the other thread grab the driver tag.
1077                  *
1078                  * If no I/O scheduler has been configured it is possible that
1079                  * the hardware queue got stopped and restarted before requests
1080                  * were pushed back onto the dispatch list. Rerun the queue to
1081                  * avoid starvation. Notes:
1082                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1083                  *   been stopped before rerunning a queue.
1084                  * - Some but not all block drivers stop a queue before
1085                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1086                  *   and dm-rq.
1087                  */
1088                 if (!blk_mq_sched_needs_restart(hctx) &&
1089                     !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1090                         blk_mq_run_hw_queue(hctx, true);
1091         }
1092
1093         return (queued + errors) != 0;
1094 }
1095
1096 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1097 {
1098         int srcu_idx;
1099
1100         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1101                 cpu_online(hctx->next_cpu));
1102
1103         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1104                 rcu_read_lock();
1105                 blk_mq_sched_dispatch_requests(hctx);
1106                 rcu_read_unlock();
1107         } else {
1108                 might_sleep();
1109
1110                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1111                 blk_mq_sched_dispatch_requests(hctx);
1112                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1113         }
1114 }
1115
1116 /*
1117  * It'd be great if the workqueue API had a way to pass
1118  * in a mask and had some smarts for more clever placement.
1119  * For now we just round-robin here, switching for every
1120  * BLK_MQ_CPU_WORK_BATCH queued items.
1121  */
1122 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1123 {
1124         if (hctx->queue->nr_hw_queues == 1)
1125                 return WORK_CPU_UNBOUND;
1126
1127         if (--hctx->next_cpu_batch <= 0) {
1128                 int next_cpu;
1129
1130                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1131                 if (next_cpu >= nr_cpu_ids)
1132                         next_cpu = cpumask_first(hctx->cpumask);
1133
1134                 hctx->next_cpu = next_cpu;
1135                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1136         }
1137
1138         return hctx->next_cpu;
1139 }
1140
1141 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1142                                         unsigned long msecs)
1143 {
1144         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1145                 return;
1146
1147         if (unlikely(blk_mq_hctx_stopped(hctx)))
1148                 return;
1149
1150         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1151                 int cpu = get_cpu();
1152                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1153                         __blk_mq_run_hw_queue(hctx);
1154                         put_cpu();
1155                         return;
1156                 }
1157
1158                 put_cpu();
1159         }
1160
1161         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1162                                          &hctx->run_work,
1163                                          msecs_to_jiffies(msecs));
1164 }
1165
1166 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1167 {
1168         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1169 }
1170 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1171
1172 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1173 {
1174         __blk_mq_delay_run_hw_queue(hctx, async, 0);
1175 }
1176 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1177
1178 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1179 {
1180         struct blk_mq_hw_ctx *hctx;
1181         int i;
1182
1183         queue_for_each_hw_ctx(q, hctx, i) {
1184                 if (!blk_mq_hctx_has_pending(hctx) ||
1185                     blk_mq_hctx_stopped(hctx))
1186                         continue;
1187
1188                 blk_mq_run_hw_queue(hctx, async);
1189         }
1190 }
1191 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1192
1193 /**
1194  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1195  * @q: request queue.
1196  *
1197  * The caller is responsible for serializing this function against
1198  * blk_mq_{start,stop}_hw_queue().
1199  */
1200 bool blk_mq_queue_stopped(struct request_queue *q)
1201 {
1202         struct blk_mq_hw_ctx *hctx;
1203         int i;
1204
1205         queue_for_each_hw_ctx(q, hctx, i)
1206                 if (blk_mq_hctx_stopped(hctx))
1207                         return true;
1208
1209         return false;
1210 }
1211 EXPORT_SYMBOL(blk_mq_queue_stopped);
1212
1213 /*
1214  * This function is often used for pausing .queue_rq() by driver when
1215  * there isn't enough resource or some conditions aren't satisfied, and
1216  * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1217  *
1218  * We do not guarantee that dispatch can be drained or blocked
1219  * after blk_mq_stop_hw_queue() returns. Please use
1220  * blk_mq_quiesce_queue() for that requirement.
1221  */
1222 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1223 {
1224         cancel_delayed_work(&hctx->run_work);
1225
1226         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1227 }
1228 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1229
1230 /*
1231  * This function is often used for pausing .queue_rq() by driver when
1232  * there isn't enough resource or some conditions aren't satisfied, and
1233  * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1234  *
1235  * We do not guarantee that dispatch can be drained or blocked
1236  * after blk_mq_stop_hw_queues() returns. Please use
1237  * blk_mq_quiesce_queue() for that requirement.
1238  */
1239 void blk_mq_stop_hw_queues(struct request_queue *q)
1240 {
1241         struct blk_mq_hw_ctx *hctx;
1242         int i;
1243
1244         queue_for_each_hw_ctx(q, hctx, i)
1245                 blk_mq_stop_hw_queue(hctx);
1246 }
1247 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1248
1249 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1250 {
1251         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1252
1253         blk_mq_run_hw_queue(hctx, false);
1254 }
1255 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1256
1257 void blk_mq_start_hw_queues(struct request_queue *q)
1258 {
1259         struct blk_mq_hw_ctx *hctx;
1260         int i;
1261
1262         queue_for_each_hw_ctx(q, hctx, i)
1263                 blk_mq_start_hw_queue(hctx);
1264 }
1265 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1266
1267 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1268 {
1269         if (!blk_mq_hctx_stopped(hctx))
1270                 return;
1271
1272         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1273         blk_mq_run_hw_queue(hctx, async);
1274 }
1275 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1276
1277 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1278 {
1279         struct blk_mq_hw_ctx *hctx;
1280         int i;
1281
1282         queue_for_each_hw_ctx(q, hctx, i)
1283                 blk_mq_start_stopped_hw_queue(hctx, async);
1284 }
1285 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1286
1287 static void blk_mq_run_work_fn(struct work_struct *work)
1288 {
1289         struct blk_mq_hw_ctx *hctx;
1290
1291         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1292
1293         /*
1294          * If we are stopped, don't run the queue. The exception is if
1295          * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1296          * the STOPPED bit and run it.
1297          */
1298         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1299                 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1300                         return;
1301
1302                 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1303                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1304         }
1305
1306         __blk_mq_run_hw_queue(hctx);
1307 }
1308
1309
1310 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1311 {
1312         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1313                 return;
1314
1315         /*
1316          * Stop the hw queue, then modify currently delayed work.
1317          * This should prevent us from running the queue prematurely.
1318          * Mark the queue as auto-clearing STOPPED when it runs.
1319          */
1320         blk_mq_stop_hw_queue(hctx);
1321         set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1322         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1323                                         &hctx->run_work,
1324                                         msecs_to_jiffies(msecs));
1325 }
1326 EXPORT_SYMBOL(blk_mq_delay_queue);
1327
1328 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1329                                             struct request *rq,
1330                                             bool at_head)
1331 {
1332         struct blk_mq_ctx *ctx = rq->mq_ctx;
1333
1334         lockdep_assert_held(&ctx->lock);
1335
1336         trace_block_rq_insert(hctx->queue, rq);
1337
1338         if (at_head)
1339                 list_add(&rq->queuelist, &ctx->rq_list);
1340         else
1341                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1342 }
1343
1344 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1345                              bool at_head)
1346 {
1347         struct blk_mq_ctx *ctx = rq->mq_ctx;
1348
1349         lockdep_assert_held(&ctx->lock);
1350
1351         __blk_mq_insert_req_list(hctx, rq, at_head);
1352         blk_mq_hctx_mark_pending(hctx, ctx);
1353 }
1354
1355 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1356                             struct list_head *list)
1357
1358 {
1359         /*
1360          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1361          * offline now
1362          */
1363         spin_lock(&ctx->lock);
1364         while (!list_empty(list)) {
1365                 struct request *rq;
1366
1367                 rq = list_first_entry(list, struct request, queuelist);
1368                 BUG_ON(rq->mq_ctx != ctx);
1369                 list_del_init(&rq->queuelist);
1370                 __blk_mq_insert_req_list(hctx, rq, false);
1371         }
1372         blk_mq_hctx_mark_pending(hctx, ctx);
1373         spin_unlock(&ctx->lock);
1374 }
1375
1376 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1377 {
1378         struct request *rqa = container_of(a, struct request, queuelist);
1379         struct request *rqb = container_of(b, struct request, queuelist);
1380
1381         return !(rqa->mq_ctx < rqb->mq_ctx ||
1382                  (rqa->mq_ctx == rqb->mq_ctx &&
1383                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1384 }
1385
1386 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1387 {
1388         struct blk_mq_ctx *this_ctx;
1389         struct request_queue *this_q;
1390         struct request *rq;
1391         LIST_HEAD(list);
1392         LIST_HEAD(ctx_list);
1393         unsigned int depth;
1394
1395         list_splice_init(&plug->mq_list, &list);
1396
1397         list_sort(NULL, &list, plug_ctx_cmp);
1398
1399         this_q = NULL;
1400         this_ctx = NULL;
1401         depth = 0;
1402
1403         while (!list_empty(&list)) {
1404                 rq = list_entry_rq(list.next);
1405                 list_del_init(&rq->queuelist);
1406                 BUG_ON(!rq->q);
1407                 if (rq->mq_ctx != this_ctx) {
1408                         if (this_ctx) {
1409                                 trace_block_unplug(this_q, depth, from_schedule);
1410                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1411                                                                 &ctx_list,
1412                                                                 from_schedule);
1413                         }
1414
1415                         this_ctx = rq->mq_ctx;
1416                         this_q = rq->q;
1417                         depth = 0;
1418                 }
1419
1420                 depth++;
1421                 list_add_tail(&rq->queuelist, &ctx_list);
1422         }
1423
1424         /*
1425          * If 'this_ctx' is set, we know we have entries to complete
1426          * on 'ctx_list'. Do those.
1427          */
1428         if (this_ctx) {
1429                 trace_block_unplug(this_q, depth, from_schedule);
1430                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1431                                                 from_schedule);
1432         }
1433 }
1434
1435 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1436 {
1437         blk_init_request_from_bio(rq, bio);
1438
1439         blk_account_io_start(rq, true);
1440 }
1441
1442 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1443 {
1444         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1445                 !blk_queue_nomerges(hctx->queue);
1446 }
1447
1448 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1449                                    struct blk_mq_ctx *ctx,
1450                                    struct request *rq)
1451 {
1452         spin_lock(&ctx->lock);
1453         __blk_mq_insert_request(hctx, rq, false);
1454         spin_unlock(&ctx->lock);
1455 }
1456
1457 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1458 {
1459         if (rq->tag != -1)
1460                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1461
1462         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1463 }
1464
1465 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1466                                         struct request *rq,
1467                                         blk_qc_t *cookie, bool may_sleep)
1468 {
1469         struct request_queue *q = rq->q;
1470         struct blk_mq_queue_data bd = {
1471                 .rq = rq,
1472                 .last = true,
1473         };
1474         blk_qc_t new_cookie;
1475         blk_status_t ret;
1476         bool run_queue = true;
1477
1478         /* RCU or SRCU read lock is needed before checking quiesced flag */
1479         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1480                 run_queue = false;
1481                 goto insert;
1482         }
1483
1484         if (q->elevator)
1485                 goto insert;
1486
1487         if (!blk_mq_get_driver_tag(rq, NULL, false))
1488                 goto insert;
1489
1490         new_cookie = request_to_qc_t(hctx, rq);
1491
1492         /*
1493          * For OK queue, we are done. For error, kill it. Any other
1494          * error (busy), just add it to our list as we previously
1495          * would have done
1496          */
1497         ret = q->mq_ops->queue_rq(hctx, &bd);
1498         switch (ret) {
1499         case BLK_STS_OK:
1500                 *cookie = new_cookie;
1501                 return;
1502         case BLK_STS_RESOURCE:
1503                 __blk_mq_requeue_request(rq);
1504                 goto insert;
1505         default:
1506                 *cookie = BLK_QC_T_NONE;
1507                 blk_mq_end_request(rq, ret);
1508                 return;
1509         }
1510
1511 insert:
1512         blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1513 }
1514
1515 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1516                 struct request *rq, blk_qc_t *cookie)
1517 {
1518         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1519                 rcu_read_lock();
1520                 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1521                 rcu_read_unlock();
1522         } else {
1523                 unsigned int srcu_idx;
1524
1525                 might_sleep();
1526
1527                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1528                 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1529                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1530         }
1531 }
1532
1533 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1534 {
1535         const int is_sync = op_is_sync(bio->bi_opf);
1536         const int is_flush_fua = op_is_flush(bio->bi_opf);
1537         struct blk_mq_alloc_data data = { .flags = 0 };
1538         struct request *rq;
1539         unsigned int request_count = 0;
1540         struct blk_plug *plug;
1541         struct request *same_queue_rq = NULL;
1542         blk_qc_t cookie;
1543         unsigned int wb_acct;
1544
1545         blk_queue_bounce(q, &bio);
1546
1547         blk_queue_split(q, &bio);
1548
1549         if (!bio_integrity_prep(bio))
1550                 return BLK_QC_T_NONE;
1551
1552         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1553             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1554                 return BLK_QC_T_NONE;
1555
1556         if (blk_mq_sched_bio_merge(q, bio))
1557                 return BLK_QC_T_NONE;
1558
1559         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1560
1561         trace_block_getrq(q, bio, bio->bi_opf);
1562
1563         rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1564         if (unlikely(!rq)) {
1565                 __wbt_done(q->rq_wb, wb_acct);
1566                 if (bio->bi_opf & REQ_NOWAIT)
1567                         bio_wouldblock_error(bio);
1568                 return BLK_QC_T_NONE;
1569         }
1570
1571         wbt_track(&rq->issue_stat, wb_acct);
1572
1573         cookie = request_to_qc_t(data.hctx, rq);
1574
1575         plug = current->plug;
1576         if (unlikely(is_flush_fua)) {
1577                 blk_mq_put_ctx(data.ctx);
1578                 blk_mq_bio_to_request(rq, bio);
1579                 if (q->elevator) {
1580                         blk_mq_sched_insert_request(rq, false, true, true,
1581                                         true);
1582                 } else {
1583                         blk_insert_flush(rq);
1584                         blk_mq_run_hw_queue(data.hctx, true);
1585                 }
1586         } else if (plug && q->nr_hw_queues == 1) {
1587                 struct request *last = NULL;
1588
1589                 blk_mq_put_ctx(data.ctx);
1590                 blk_mq_bio_to_request(rq, bio);
1591
1592                 /*
1593                  * @request_count may become stale because of schedule
1594                  * out, so check the list again.
1595                  */
1596                 if (list_empty(&plug->mq_list))
1597                         request_count = 0;
1598                 else if (blk_queue_nomerges(q))
1599                         request_count = blk_plug_queued_count(q);
1600
1601                 if (!request_count)
1602                         trace_block_plug(q);
1603                 else
1604                         last = list_entry_rq(plug->mq_list.prev);
1605
1606                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1607                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1608                         blk_flush_plug_list(plug, false);
1609                         trace_block_plug(q);
1610                 }
1611
1612                 list_add_tail(&rq->queuelist, &plug->mq_list);
1613         } else if (plug && !blk_queue_nomerges(q)) {
1614                 blk_mq_bio_to_request(rq, bio);
1615
1616                 /*
1617                  * We do limited plugging. If the bio can be merged, do that.
1618                  * Otherwise the existing request in the plug list will be
1619                  * issued. So the plug list will have one request at most
1620                  * The plug list might get flushed before this. If that happens,
1621                  * the plug list is empty, and same_queue_rq is invalid.
1622                  */
1623                 if (list_empty(&plug->mq_list))
1624                         same_queue_rq = NULL;
1625                 if (same_queue_rq)
1626                         list_del_init(&same_queue_rq->queuelist);
1627                 list_add_tail(&rq->queuelist, &plug->mq_list);
1628
1629                 blk_mq_put_ctx(data.ctx);
1630
1631                 if (same_queue_rq) {
1632                         data.hctx = blk_mq_map_queue(q,
1633                                         same_queue_rq->mq_ctx->cpu);
1634                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1635                                         &cookie);
1636                 }
1637         } else if (q->nr_hw_queues > 1 && is_sync) {
1638                 blk_mq_put_ctx(data.ctx);
1639                 blk_mq_bio_to_request(rq, bio);
1640                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1641         } else if (q->elevator) {
1642                 blk_mq_put_ctx(data.ctx);
1643                 blk_mq_bio_to_request(rq, bio);
1644                 blk_mq_sched_insert_request(rq, false, true, true, true);
1645         } else {
1646                 blk_mq_put_ctx(data.ctx);
1647                 blk_mq_bio_to_request(rq, bio);
1648                 blk_mq_queue_io(data.hctx, data.ctx, rq);
1649                 blk_mq_run_hw_queue(data.hctx, true);
1650         }
1651
1652         return cookie;
1653 }
1654
1655 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1656                      unsigned int hctx_idx)
1657 {
1658         struct page *page;
1659
1660         if (tags->rqs && set->ops->exit_request) {
1661                 int i;
1662
1663                 for (i = 0; i < tags->nr_tags; i++) {
1664                         struct request *rq = tags->static_rqs[i];
1665
1666                         if (!rq)
1667                                 continue;
1668                         set->ops->exit_request(set, rq, hctx_idx);
1669                         tags->static_rqs[i] = NULL;
1670                 }
1671         }
1672
1673         while (!list_empty(&tags->page_list)) {
1674                 page = list_first_entry(&tags->page_list, struct page, lru);
1675                 list_del_init(&page->lru);
1676                 /*
1677                  * Remove kmemleak object previously allocated in
1678                  * blk_mq_init_rq_map().
1679                  */
1680                 kmemleak_free(page_address(page));
1681                 __free_pages(page, page->private);
1682         }
1683 }
1684
1685 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1686 {
1687         kfree(tags->rqs);
1688         tags->rqs = NULL;
1689         kfree(tags->static_rqs);
1690         tags->static_rqs = NULL;
1691
1692         blk_mq_free_tags(tags);
1693 }
1694
1695 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1696                                         unsigned int hctx_idx,
1697                                         unsigned int nr_tags,
1698                                         unsigned int reserved_tags)
1699 {
1700         struct blk_mq_tags *tags;
1701         int node;
1702
1703         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1704         if (node == NUMA_NO_NODE)
1705                 node = set->numa_node;
1706
1707         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1708                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1709         if (!tags)
1710                 return NULL;
1711
1712         tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1713                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1714                                  node);
1715         if (!tags->rqs) {
1716                 blk_mq_free_tags(tags);
1717                 return NULL;
1718         }
1719
1720         tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1721                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1722                                  node);
1723         if (!tags->static_rqs) {
1724                 kfree(tags->rqs);
1725                 blk_mq_free_tags(tags);
1726                 return NULL;
1727         }
1728
1729         return tags;
1730 }
1731
1732 static size_t order_to_size(unsigned int order)
1733 {
1734         return (size_t)PAGE_SIZE << order;
1735 }
1736
1737 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1738                      unsigned int hctx_idx, unsigned int depth)
1739 {
1740         unsigned int i, j, entries_per_page, max_order = 4;
1741         size_t rq_size, left;
1742         int node;
1743
1744         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1745         if (node == NUMA_NO_NODE)
1746                 node = set->numa_node;
1747
1748         INIT_LIST_HEAD(&tags->page_list);
1749
1750         /*
1751          * rq_size is the size of the request plus driver payload, rounded
1752          * to the cacheline size
1753          */
1754         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1755                                 cache_line_size());
1756         left = rq_size * depth;
1757
1758         for (i = 0; i < depth; ) {
1759                 int this_order = max_order;
1760                 struct page *page;
1761                 int to_do;
1762                 void *p;
1763
1764                 while (this_order && left < order_to_size(this_order - 1))
1765                         this_order--;
1766
1767                 do {
1768                         page = alloc_pages_node(node,
1769                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1770                                 this_order);
1771                         if (page)
1772                                 break;
1773                         if (!this_order--)
1774                                 break;
1775                         if (order_to_size(this_order) < rq_size)
1776                                 break;
1777                 } while (1);
1778
1779                 if (!page)
1780                         goto fail;
1781
1782                 page->private = this_order;
1783                 list_add_tail(&page->lru, &tags->page_list);
1784
1785                 p = page_address(page);
1786                 /*
1787                  * Allow kmemleak to scan these pages as they contain pointers
1788                  * to additional allocations like via ops->init_request().
1789                  */
1790                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1791                 entries_per_page = order_to_size(this_order) / rq_size;
1792                 to_do = min(entries_per_page, depth - i);
1793                 left -= to_do * rq_size;
1794                 for (j = 0; j < to_do; j++) {
1795                         struct request *rq = p;
1796
1797                         tags->static_rqs[i] = rq;
1798                         if (set->ops->init_request) {
1799                                 if (set->ops->init_request(set, rq, hctx_idx,
1800                                                 node)) {
1801                                         tags->static_rqs[i] = NULL;
1802                                         goto fail;
1803                                 }
1804                         }
1805
1806                         p += rq_size;
1807                         i++;
1808                 }
1809         }
1810         return 0;
1811
1812 fail:
1813         blk_mq_free_rqs(set, tags, hctx_idx);
1814         return -ENOMEM;
1815 }
1816
1817 /*
1818  * 'cpu' is going away. splice any existing rq_list entries from this
1819  * software queue to the hw queue dispatch list, and ensure that it
1820  * gets run.
1821  */
1822 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1823 {
1824         struct blk_mq_hw_ctx *hctx;
1825         struct blk_mq_ctx *ctx;
1826         LIST_HEAD(tmp);
1827
1828         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1829         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1830
1831         spin_lock(&ctx->lock);
1832         if (!list_empty(&ctx->rq_list)) {
1833                 list_splice_init(&ctx->rq_list, &tmp);
1834                 blk_mq_hctx_clear_pending(hctx, ctx);
1835         }
1836         spin_unlock(&ctx->lock);
1837
1838         if (list_empty(&tmp))
1839                 return 0;
1840
1841         spin_lock(&hctx->lock);
1842         list_splice_tail_init(&tmp, &hctx->dispatch);
1843         spin_unlock(&hctx->lock);
1844
1845         blk_mq_run_hw_queue(hctx, true);
1846         return 0;
1847 }
1848
1849 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1850 {
1851         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1852                                             &hctx->cpuhp_dead);
1853 }
1854
1855 /* hctx->ctxs will be freed in queue's release handler */
1856 static void blk_mq_exit_hctx(struct request_queue *q,
1857                 struct blk_mq_tag_set *set,
1858                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1859 {
1860         blk_mq_debugfs_unregister_hctx(hctx);
1861
1862         blk_mq_tag_idle(hctx);
1863
1864         if (set->ops->exit_request)
1865                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1866
1867         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1868
1869         if (set->ops->exit_hctx)
1870                 set->ops->exit_hctx(hctx, hctx_idx);
1871
1872         if (hctx->flags & BLK_MQ_F_BLOCKING)
1873                 cleanup_srcu_struct(hctx->queue_rq_srcu);
1874
1875         blk_mq_remove_cpuhp(hctx);
1876         blk_free_flush_queue(hctx->fq);
1877         sbitmap_free(&hctx->ctx_map);
1878 }
1879
1880 static void blk_mq_exit_hw_queues(struct request_queue *q,
1881                 struct blk_mq_tag_set *set, int nr_queue)
1882 {
1883         struct blk_mq_hw_ctx *hctx;
1884         unsigned int i;
1885
1886         queue_for_each_hw_ctx(q, hctx, i) {
1887                 if (i == nr_queue)
1888                         break;
1889                 blk_mq_exit_hctx(q, set, hctx, i);
1890         }
1891 }
1892
1893 static int blk_mq_init_hctx(struct request_queue *q,
1894                 struct blk_mq_tag_set *set,
1895                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1896 {
1897         int node;
1898
1899         node = hctx->numa_node;
1900         if (node == NUMA_NO_NODE)
1901                 node = hctx->numa_node = set->numa_node;
1902
1903         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1904         spin_lock_init(&hctx->lock);
1905         INIT_LIST_HEAD(&hctx->dispatch);
1906         hctx->queue = q;
1907         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1908
1909         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1910
1911         hctx->tags = set->tags[hctx_idx];
1912
1913         /*
1914          * Allocate space for all possible cpus to avoid allocation at
1915          * runtime
1916          */
1917         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1918                                         GFP_KERNEL, node);
1919         if (!hctx->ctxs)
1920                 goto unregister_cpu_notifier;
1921
1922         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1923                               node))
1924                 goto free_ctxs;
1925
1926         hctx->nr_ctx = 0;
1927
1928         if (set->ops->init_hctx &&
1929             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1930                 goto free_bitmap;
1931
1932         if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1933                 goto exit_hctx;
1934
1935         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1936         if (!hctx->fq)
1937                 goto sched_exit_hctx;
1938
1939         if (set->ops->init_request &&
1940             set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
1941                                    node))
1942                 goto free_fq;
1943
1944         if (hctx->flags & BLK_MQ_F_BLOCKING)
1945                 init_srcu_struct(hctx->queue_rq_srcu);
1946
1947         blk_mq_debugfs_register_hctx(q, hctx);
1948
1949         return 0;
1950
1951  free_fq:
1952         kfree(hctx->fq);
1953  sched_exit_hctx:
1954         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1955  exit_hctx:
1956         if (set->ops->exit_hctx)
1957                 set->ops->exit_hctx(hctx, hctx_idx);
1958  free_bitmap:
1959         sbitmap_free(&hctx->ctx_map);
1960  free_ctxs:
1961         kfree(hctx->ctxs);
1962  unregister_cpu_notifier:
1963         blk_mq_remove_cpuhp(hctx);
1964         return -1;
1965 }
1966
1967 static void blk_mq_init_cpu_queues(struct request_queue *q,
1968                                    unsigned int nr_hw_queues)
1969 {
1970         unsigned int i;
1971
1972         for_each_possible_cpu(i) {
1973                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1974                 struct blk_mq_hw_ctx *hctx;
1975
1976                 __ctx->cpu = i;
1977                 spin_lock_init(&__ctx->lock);
1978                 INIT_LIST_HEAD(&__ctx->rq_list);
1979                 __ctx->queue = q;
1980
1981                 /* If the cpu isn't present, the cpu is mapped to first hctx */
1982                 if (!cpu_present(i))
1983                         continue;
1984
1985                 hctx = blk_mq_map_queue(q, i);
1986
1987                 /*
1988                  * Set local node, IFF we have more than one hw queue. If
1989                  * not, we remain on the home node of the device
1990                  */
1991                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1992                         hctx->numa_node = local_memory_node(cpu_to_node(i));
1993         }
1994 }
1995
1996 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1997 {
1998         int ret = 0;
1999
2000         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2001                                         set->queue_depth, set->reserved_tags);
2002         if (!set->tags[hctx_idx])
2003                 return false;
2004
2005         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2006                                 set->queue_depth);
2007         if (!ret)
2008                 return true;
2009
2010         blk_mq_free_rq_map(set->tags[hctx_idx]);
2011         set->tags[hctx_idx] = NULL;
2012         return false;
2013 }
2014
2015 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2016                                          unsigned int hctx_idx)
2017 {
2018         if (set->tags[hctx_idx]) {
2019                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2020                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2021                 set->tags[hctx_idx] = NULL;
2022         }
2023 }
2024
2025 static void blk_mq_map_swqueue(struct request_queue *q)
2026 {
2027         unsigned int i, hctx_idx;
2028         struct blk_mq_hw_ctx *hctx;
2029         struct blk_mq_ctx *ctx;
2030         struct blk_mq_tag_set *set = q->tag_set;
2031
2032         /*
2033          * Avoid others reading imcomplete hctx->cpumask through sysfs
2034          */
2035         mutex_lock(&q->sysfs_lock);
2036
2037         queue_for_each_hw_ctx(q, hctx, i) {
2038                 cpumask_clear(hctx->cpumask);
2039                 hctx->nr_ctx = 0;
2040         }
2041
2042         /*
2043          * Map software to hardware queues.
2044          *
2045          * If the cpu isn't present, the cpu is mapped to first hctx.
2046          */
2047         for_each_present_cpu(i) {
2048                 hctx_idx = q->mq_map[i];
2049                 /* unmapped hw queue can be remapped after CPU topo changed */
2050                 if (!set->tags[hctx_idx] &&
2051                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2052                         /*
2053                          * If tags initialization fail for some hctx,
2054                          * that hctx won't be brought online.  In this
2055                          * case, remap the current ctx to hctx[0] which
2056                          * is guaranteed to always have tags allocated
2057                          */
2058                         q->mq_map[i] = 0;
2059                 }
2060
2061                 ctx = per_cpu_ptr(q->queue_ctx, i);
2062                 hctx = blk_mq_map_queue(q, i);
2063
2064                 cpumask_set_cpu(i, hctx->cpumask);
2065                 ctx->index_hw = hctx->nr_ctx;
2066                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2067         }
2068
2069         mutex_unlock(&q->sysfs_lock);
2070
2071         queue_for_each_hw_ctx(q, hctx, i) {
2072                 /*
2073                  * If no software queues are mapped to this hardware queue,
2074                  * disable it and free the request entries.
2075                  */
2076                 if (!hctx->nr_ctx) {
2077                         /* Never unmap queue 0.  We need it as a
2078                          * fallback in case of a new remap fails
2079                          * allocation
2080                          */
2081                         if (i && set->tags[i])
2082                                 blk_mq_free_map_and_requests(set, i);
2083
2084                         hctx->tags = NULL;
2085                         continue;
2086                 }
2087
2088                 hctx->tags = set->tags[i];
2089                 WARN_ON(!hctx->tags);
2090
2091                 /*
2092                  * Set the map size to the number of mapped software queues.
2093                  * This is more accurate and more efficient than looping
2094                  * over all possibly mapped software queues.
2095                  */
2096                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2097
2098                 /*
2099                  * Initialize batch roundrobin counts
2100                  */
2101                 hctx->next_cpu = cpumask_first(hctx->cpumask);
2102                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2103         }
2104 }
2105
2106 /*
2107  * Caller needs to ensure that we're either frozen/quiesced, or that
2108  * the queue isn't live yet.
2109  */
2110 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2111 {
2112         struct blk_mq_hw_ctx *hctx;
2113         int i;
2114
2115         queue_for_each_hw_ctx(q, hctx, i) {
2116                 if (shared) {
2117                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2118                                 atomic_inc(&q->shared_hctx_restart);
2119                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2120                 } else {
2121                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2122                                 atomic_dec(&q->shared_hctx_restart);
2123                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2124                 }
2125         }
2126 }
2127
2128 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2129                                         bool shared)
2130 {
2131         struct request_queue *q;
2132
2133         lockdep_assert_held(&set->tag_list_lock);
2134
2135         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2136                 blk_mq_freeze_queue(q);
2137                 queue_set_hctx_shared(q, shared);
2138                 blk_mq_unfreeze_queue(q);
2139         }
2140 }
2141
2142 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2143 {
2144         struct blk_mq_tag_set *set = q->tag_set;
2145
2146         mutex_lock(&set->tag_list_lock);
2147         list_del_rcu(&q->tag_set_list);
2148         INIT_LIST_HEAD(&q->tag_set_list);
2149         if (list_is_singular(&set->tag_list)) {
2150                 /* just transitioned to unshared */
2151                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2152                 /* update existing queue */
2153                 blk_mq_update_tag_set_depth(set, false);
2154         }
2155         mutex_unlock(&set->tag_list_lock);
2156
2157         synchronize_rcu();
2158 }
2159
2160 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2161                                      struct request_queue *q)
2162 {
2163         q->tag_set = set;
2164
2165         mutex_lock(&set->tag_list_lock);
2166
2167         /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2168         if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2169                 set->flags |= BLK_MQ_F_TAG_SHARED;
2170                 /* update existing queue */
2171                 blk_mq_update_tag_set_depth(set, true);
2172         }
2173         if (set->flags & BLK_MQ_F_TAG_SHARED)
2174                 queue_set_hctx_shared(q, true);
2175         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2176
2177         mutex_unlock(&set->tag_list_lock);
2178 }
2179
2180 /*
2181  * It is the actual release handler for mq, but we do it from
2182  * request queue's release handler for avoiding use-after-free
2183  * and headache because q->mq_kobj shouldn't have been introduced,
2184  * but we can't group ctx/kctx kobj without it.
2185  */
2186 void blk_mq_release(struct request_queue *q)
2187 {
2188         struct blk_mq_hw_ctx *hctx;
2189         unsigned int i;
2190
2191         /* hctx kobj stays in hctx */
2192         queue_for_each_hw_ctx(q, hctx, i) {
2193                 if (!hctx)
2194                         continue;
2195                 kobject_put(&hctx->kobj);
2196         }
2197
2198         q->mq_map = NULL;
2199
2200         kfree(q->queue_hw_ctx);
2201
2202         /*
2203          * release .mq_kobj and sw queue's kobject now because
2204          * both share lifetime with request queue.
2205          */
2206         blk_mq_sysfs_deinit(q);
2207
2208         free_percpu(q->queue_ctx);
2209 }
2210
2211 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2212 {
2213         struct request_queue *uninit_q, *q;
2214
2215         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2216         if (!uninit_q)
2217                 return ERR_PTR(-ENOMEM);
2218
2219         q = blk_mq_init_allocated_queue(set, uninit_q);
2220         if (IS_ERR(q))
2221                 blk_cleanup_queue(uninit_q);
2222
2223         return q;
2224 }
2225 EXPORT_SYMBOL(blk_mq_init_queue);
2226
2227 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2228 {
2229         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2230
2231         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2232                            __alignof__(struct blk_mq_hw_ctx)) !=
2233                      sizeof(struct blk_mq_hw_ctx));
2234
2235         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2236                 hw_ctx_size += sizeof(struct srcu_struct);
2237
2238         return hw_ctx_size;
2239 }
2240
2241 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2242                                                 struct request_queue *q)
2243 {
2244         int i, j;
2245         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2246
2247         blk_mq_sysfs_unregister(q);
2248         for (i = 0; i < set->nr_hw_queues; i++) {
2249                 int node;
2250
2251                 if (hctxs[i])
2252                         continue;
2253
2254                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2255                 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2256                                         GFP_KERNEL, node);
2257                 if (!hctxs[i])
2258                         break;
2259
2260                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2261                                                 node)) {
2262                         kfree(hctxs[i]);
2263                         hctxs[i] = NULL;
2264                         break;
2265                 }
2266
2267                 atomic_set(&hctxs[i]->nr_active, 0);
2268                 hctxs[i]->numa_node = node;
2269                 hctxs[i]->queue_num = i;
2270
2271                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2272                         free_cpumask_var(hctxs[i]->cpumask);
2273                         kfree(hctxs[i]);
2274                         hctxs[i] = NULL;
2275                         break;
2276                 }
2277                 blk_mq_hctx_kobj_init(hctxs[i]);
2278         }
2279         for (j = i; j < q->nr_hw_queues; j++) {
2280                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2281
2282                 if (hctx) {
2283                         if (hctx->tags)
2284                                 blk_mq_free_map_and_requests(set, j);
2285                         blk_mq_exit_hctx(q, set, hctx, j);
2286                         kobject_put(&hctx->kobj);
2287                         hctxs[j] = NULL;
2288
2289                 }
2290         }
2291         q->nr_hw_queues = i;
2292         blk_mq_sysfs_register(q);
2293 }
2294
2295 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2296                                                   struct request_queue *q)
2297 {
2298         /* mark the queue as mq asap */
2299         q->mq_ops = set->ops;
2300
2301         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2302                                              blk_mq_poll_stats_bkt,
2303                                              BLK_MQ_POLL_STATS_BKTS, q);
2304         if (!q->poll_cb)
2305                 goto err_exit;
2306
2307         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2308         if (!q->queue_ctx)
2309                 goto err_exit;
2310
2311         /* init q->mq_kobj and sw queues' kobjects */
2312         blk_mq_sysfs_init(q);
2313
2314         q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2315                                                 GFP_KERNEL, set->numa_node);
2316         if (!q->queue_hw_ctx)
2317                 goto err_percpu;
2318
2319         q->mq_map = set->mq_map;
2320
2321         blk_mq_realloc_hw_ctxs(set, q);
2322         if (!q->nr_hw_queues)
2323                 goto err_hctxs;
2324
2325         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2326         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2327
2328         q->nr_queues = nr_cpu_ids;
2329
2330         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2331
2332         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2333                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2334
2335         q->sg_reserved_size = INT_MAX;
2336
2337         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2338         INIT_LIST_HEAD(&q->requeue_list);
2339         spin_lock_init(&q->requeue_lock);
2340
2341         blk_queue_make_request(q, blk_mq_make_request);
2342
2343         /*
2344          * Do this after blk_queue_make_request() overrides it...
2345          */
2346         q->nr_requests = set->queue_depth;
2347
2348         /*
2349          * Default to classic polling
2350          */
2351         q->poll_nsec = -1;
2352
2353         if (set->ops->complete)
2354                 blk_queue_softirq_done(q, set->ops->complete);
2355
2356         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2357         blk_mq_add_queue_tag_set(set, q);
2358         blk_mq_map_swqueue(q);
2359
2360         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2361                 int ret;
2362
2363                 ret = blk_mq_sched_init(q);
2364                 if (ret)
2365                         return ERR_PTR(ret);
2366         }
2367
2368         return q;
2369
2370 err_hctxs:
2371         kfree(q->queue_hw_ctx);
2372 err_percpu:
2373         free_percpu(q->queue_ctx);
2374 err_exit:
2375         q->mq_ops = NULL;
2376         return ERR_PTR(-ENOMEM);
2377 }
2378 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2379
2380 void blk_mq_free_queue(struct request_queue *q)
2381 {
2382         struct blk_mq_tag_set   *set = q->tag_set;
2383
2384         blk_mq_del_queue_tag_set(q);
2385         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2386 }
2387
2388 /* Basically redo blk_mq_init_queue with queue frozen */
2389 static void blk_mq_queue_reinit(struct request_queue *q)
2390 {
2391         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2392
2393         blk_mq_debugfs_unregister_hctxs(q);
2394         blk_mq_sysfs_unregister(q);
2395
2396         /*
2397          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2398          * we should change hctx numa_node according to new topology (this
2399          * involves free and re-allocate memory, worthy doing?)
2400          */
2401
2402         blk_mq_map_swqueue(q);
2403
2404         blk_mq_sysfs_register(q);
2405         blk_mq_debugfs_register_hctxs(q);
2406 }
2407
2408 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2409 {
2410         int i;
2411
2412         for (i = 0; i < set->nr_hw_queues; i++)
2413                 if (!__blk_mq_alloc_rq_map(set, i))
2414                         goto out_unwind;
2415
2416         return 0;
2417
2418 out_unwind:
2419         while (--i >= 0)
2420                 blk_mq_free_rq_map(set->tags[i]);
2421
2422         return -ENOMEM;
2423 }
2424
2425 /*
2426  * Allocate the request maps associated with this tag_set. Note that this
2427  * may reduce the depth asked for, if memory is tight. set->queue_depth
2428  * will be updated to reflect the allocated depth.
2429  */
2430 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2431 {
2432         unsigned int depth;
2433         int err;
2434
2435         depth = set->queue_depth;
2436         do {
2437                 err = __blk_mq_alloc_rq_maps(set);
2438                 if (!err)
2439                         break;
2440
2441                 set->queue_depth >>= 1;
2442                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2443                         err = -ENOMEM;
2444                         break;
2445                 }
2446         } while (set->queue_depth);
2447
2448         if (!set->queue_depth || err) {
2449                 pr_err("blk-mq: failed to allocate request map\n");
2450                 return -ENOMEM;
2451         }
2452
2453         if (depth != set->queue_depth)
2454                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2455                                                 depth, set->queue_depth);
2456
2457         return 0;
2458 }
2459
2460 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2461 {
2462         if (set->ops->map_queues)
2463                 return set->ops->map_queues(set);
2464         else
2465                 return blk_mq_map_queues(set);
2466 }
2467
2468 /*
2469  * Alloc a tag set to be associated with one or more request queues.
2470  * May fail with EINVAL for various error conditions. May adjust the
2471  * requested depth down, if if it too large. In that case, the set
2472  * value will be stored in set->queue_depth.
2473  */
2474 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2475 {
2476         int ret;
2477
2478         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2479
2480         if (!set->nr_hw_queues)
2481                 return -EINVAL;
2482         if (!set->queue_depth)
2483                 return -EINVAL;
2484         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2485                 return -EINVAL;
2486
2487         if (!set->ops->queue_rq)
2488                 return -EINVAL;
2489
2490         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2491                 pr_info("blk-mq: reduced tag depth to %u\n",
2492                         BLK_MQ_MAX_DEPTH);
2493                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2494         }
2495
2496         /*
2497          * If a crashdump is active, then we are potentially in a very
2498          * memory constrained environment. Limit us to 1 queue and
2499          * 64 tags to prevent using too much memory.
2500          */
2501         if (is_kdump_kernel()) {
2502                 set->nr_hw_queues = 1;
2503                 set->queue_depth = min(64U, set->queue_depth);
2504         }
2505         /*
2506          * There is no use for more h/w queues than cpus.
2507          */
2508         if (set->nr_hw_queues > nr_cpu_ids)
2509                 set->nr_hw_queues = nr_cpu_ids;
2510
2511         set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2512                                  GFP_KERNEL, set->numa_node);
2513         if (!set->tags)
2514                 return -ENOMEM;
2515
2516         ret = -ENOMEM;
2517         set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2518                         GFP_KERNEL, set->numa_node);
2519         if (!set->mq_map)
2520                 goto out_free_tags;
2521
2522         ret = blk_mq_update_queue_map(set);
2523         if (ret)
2524                 goto out_free_mq_map;
2525
2526         ret = blk_mq_alloc_rq_maps(set);
2527         if (ret)
2528                 goto out_free_mq_map;
2529
2530         mutex_init(&set->tag_list_lock);
2531         INIT_LIST_HEAD(&set->tag_list);
2532
2533         return 0;
2534
2535 out_free_mq_map:
2536         kfree(set->mq_map);
2537         set->mq_map = NULL;
2538 out_free_tags:
2539         kfree(set->tags);
2540         set->tags = NULL;
2541         return ret;
2542 }
2543 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2544
2545 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2546 {
2547         int i;
2548
2549         for (i = 0; i < nr_cpu_ids; i++)
2550                 blk_mq_free_map_and_requests(set, i);
2551
2552         kfree(set->mq_map);
2553         set->mq_map = NULL;
2554
2555         kfree(set->tags);
2556         set->tags = NULL;
2557 }
2558 EXPORT_SYMBOL(blk_mq_free_tag_set);
2559
2560 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2561 {
2562         struct blk_mq_tag_set *set = q->tag_set;
2563         struct blk_mq_hw_ctx *hctx;
2564         int i, ret;
2565
2566         if (!set)
2567                 return -EINVAL;
2568
2569         blk_mq_freeze_queue(q);
2570
2571         ret = 0;
2572         queue_for_each_hw_ctx(q, hctx, i) {
2573                 if (!hctx->tags)
2574                         continue;
2575                 /*
2576                  * If we're using an MQ scheduler, just update the scheduler
2577                  * queue depth. This is similar to what the old code would do.
2578                  */
2579                 if (!hctx->sched_tags) {
2580                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2581                                                         min(nr, set->queue_depth),
2582                                                         false);
2583                 } else {
2584                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2585                                                         nr, true);
2586                 }
2587                 if (ret)
2588                         break;
2589         }
2590
2591         if (!ret)
2592                 q->nr_requests = nr;
2593
2594         blk_mq_unfreeze_queue(q);
2595
2596         return ret;
2597 }
2598
2599 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2600                                                         int nr_hw_queues)
2601 {
2602         struct request_queue *q;
2603
2604         lockdep_assert_held(&set->tag_list_lock);
2605
2606         if (nr_hw_queues > nr_cpu_ids)
2607                 nr_hw_queues = nr_cpu_ids;
2608         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2609                 return;
2610
2611         list_for_each_entry(q, &set->tag_list, tag_set_list)
2612                 blk_mq_freeze_queue(q);
2613
2614         set->nr_hw_queues = nr_hw_queues;
2615         blk_mq_update_queue_map(set);
2616         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2617                 blk_mq_realloc_hw_ctxs(set, q);
2618                 blk_mq_queue_reinit(q);
2619         }
2620
2621         list_for_each_entry(q, &set->tag_list, tag_set_list)
2622                 blk_mq_unfreeze_queue(q);
2623 }
2624
2625 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2626 {
2627         mutex_lock(&set->tag_list_lock);
2628         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2629         mutex_unlock(&set->tag_list_lock);
2630 }
2631 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2632
2633 /* Enable polling stats and return whether they were already enabled. */
2634 static bool blk_poll_stats_enable(struct request_queue *q)
2635 {
2636         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2637             test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2638                 return true;
2639         blk_stat_add_callback(q, q->poll_cb);
2640         return false;
2641 }
2642
2643 static void blk_mq_poll_stats_start(struct request_queue *q)
2644 {
2645         /*
2646          * We don't arm the callback if polling stats are not enabled or the
2647          * callback is already active.
2648          */
2649         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2650             blk_stat_is_active(q->poll_cb))
2651                 return;
2652
2653         blk_stat_activate_msecs(q->poll_cb, 100);
2654 }
2655
2656 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2657 {
2658         struct request_queue *q = cb->data;
2659         int bucket;
2660
2661         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2662                 if (cb->stat[bucket].nr_samples)
2663                         q->poll_stat[bucket] = cb->stat[bucket];
2664         }
2665 }
2666
2667 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2668                                        struct blk_mq_hw_ctx *hctx,
2669                                        struct request *rq)
2670 {
2671         unsigned long ret = 0;
2672         int bucket;
2673
2674         /*
2675          * If stats collection isn't on, don't sleep but turn it on for
2676          * future users
2677          */
2678         if (!blk_poll_stats_enable(q))
2679                 return 0;
2680
2681         /*
2682          * As an optimistic guess, use half of the mean service time
2683          * for this type of request. We can (and should) make this smarter.
2684          * For instance, if the completion latencies are tight, we can
2685          * get closer than just half the mean. This is especially
2686          * important on devices where the completion latencies are longer
2687          * than ~10 usec. We do use the stats for the relevant IO size
2688          * if available which does lead to better estimates.
2689          */
2690         bucket = blk_mq_poll_stats_bkt(rq);
2691         if (bucket < 0)
2692                 return ret;
2693
2694         if (q->poll_stat[bucket].nr_samples)
2695                 ret = (q->poll_stat[bucket].mean + 1) / 2;
2696
2697         return ret;
2698 }
2699
2700 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2701                                      struct blk_mq_hw_ctx *hctx,
2702                                      struct request *rq)
2703 {
2704         struct hrtimer_sleeper hs;
2705         enum hrtimer_mode mode;
2706         unsigned int nsecs;
2707         ktime_t kt;
2708
2709         if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2710                 return false;
2711
2712         /*
2713          * poll_nsec can be:
2714          *
2715          * -1:  don't ever hybrid sleep
2716          *  0:  use half of prev avg
2717          * >0:  use this specific value
2718          */
2719         if (q->poll_nsec == -1)
2720                 return false;
2721         else if (q->poll_nsec > 0)
2722                 nsecs = q->poll_nsec;
2723         else
2724                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2725
2726         if (!nsecs)
2727                 return false;
2728
2729         set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2730
2731         /*
2732          * This will be replaced with the stats tracking code, using
2733          * 'avg_completion_time / 2' as the pre-sleep target.
2734          */
2735         kt = nsecs;
2736
2737         mode = HRTIMER_MODE_REL;
2738         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2739         hrtimer_set_expires(&hs.timer, kt);
2740
2741         hrtimer_init_sleeper(&hs, current);
2742         do {
2743                 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2744                         break;
2745                 set_current_state(TASK_UNINTERRUPTIBLE);
2746                 hrtimer_start_expires(&hs.timer, mode);
2747                 if (hs.task)
2748                         io_schedule();
2749                 hrtimer_cancel(&hs.timer);
2750                 mode = HRTIMER_MODE_ABS;
2751         } while (hs.task && !signal_pending(current));
2752
2753         __set_current_state(TASK_RUNNING);
2754         destroy_hrtimer_on_stack(&hs.timer);
2755         return true;
2756 }
2757
2758 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2759 {
2760         struct request_queue *q = hctx->queue;
2761         long state;
2762
2763         /*
2764          * If we sleep, have the caller restart the poll loop to reset
2765          * the state. Like for the other success return cases, the
2766          * caller is responsible for checking if the IO completed. If
2767          * the IO isn't complete, we'll get called again and will go
2768          * straight to the busy poll loop.
2769          */
2770         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2771                 return true;
2772
2773         hctx->poll_considered++;
2774
2775         state = current->state;
2776         while (!need_resched()) {
2777                 int ret;
2778
2779                 hctx->poll_invoked++;
2780
2781                 ret = q->mq_ops->poll(hctx, rq->tag);
2782                 if (ret > 0) {
2783                         hctx->poll_success++;
2784                         set_current_state(TASK_RUNNING);
2785                         return true;
2786                 }
2787
2788                 if (signal_pending_state(state, current))
2789                         set_current_state(TASK_RUNNING);
2790
2791                 if (current->state == TASK_RUNNING)
2792                         return true;
2793                 if (ret < 0)
2794                         break;
2795                 cpu_relax();
2796         }
2797
2798         return false;
2799 }
2800
2801 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2802 {
2803         struct blk_mq_hw_ctx *hctx;
2804         struct blk_plug *plug;
2805         struct request *rq;
2806
2807         if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2808             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2809                 return false;
2810
2811         plug = current->plug;
2812         if (plug)
2813                 blk_flush_plug_list(plug, false);
2814
2815         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2816         if (!blk_qc_t_is_internal(cookie))
2817                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2818         else {
2819                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2820                 /*
2821                  * With scheduling, if the request has completed, we'll
2822                  * get a NULL return here, as we clear the sched tag when
2823                  * that happens. The request still remains valid, like always,
2824                  * so we should be safe with just the NULL check.
2825                  */
2826                 if (!rq)
2827                         return false;
2828         }
2829
2830         return __blk_mq_poll(hctx, rq);
2831 }
2832 EXPORT_SYMBOL_GPL(blk_mq_poll);
2833
2834 static int __init blk_mq_init(void)
2835 {
2836         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2837                                 blk_mq_hctx_notify_dead);
2838         return 0;
2839 }
2840 subsys_initcall(blk_mq_init);