GNU Linux-libre 4.19.211-gnu1
[releases.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-mq-sched.h"
38 #include "blk-rq-qos.h"
39
40 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie);
41 static void blk_mq_poll_stats_start(struct request_queue *q);
42 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
43
44 static int blk_mq_poll_stats_bkt(const struct request *rq)
45 {
46         int ddir, bytes, bucket;
47
48         ddir = rq_data_dir(rq);
49         bytes = blk_rq_bytes(rq);
50
51         bucket = ddir + 2*(ilog2(bytes) - 9);
52
53         if (bucket < 0)
54                 return -1;
55         else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
56                 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
57
58         return bucket;
59 }
60
61 /*
62  * Check if any of the ctx's have pending work in this hardware queue
63  */
64 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
65 {
66         return !list_empty_careful(&hctx->dispatch) ||
67                 sbitmap_any_bit_set(&hctx->ctx_map) ||
68                         blk_mq_sched_has_work(hctx);
69 }
70
71 /*
72  * Mark this ctx as having pending work in this hardware queue
73  */
74 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
75                                      struct blk_mq_ctx *ctx)
76 {
77         if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
78                 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
79 }
80
81 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
82                                       struct blk_mq_ctx *ctx)
83 {
84         sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
85 }
86
87 struct mq_inflight {
88         struct hd_struct *part;
89         unsigned int *inflight;
90 };
91
92 static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
93                                   struct request *rq, void *priv,
94                                   bool reserved)
95 {
96         struct mq_inflight *mi = priv;
97
98         /*
99          * index[0] counts the specific partition that was asked for. index[1]
100          * counts the ones that are active on the whole device, so increment
101          * that if mi->part is indeed a partition, and not a whole device.
102          */
103         if (rq->part == mi->part)
104                 mi->inflight[0]++;
105         if (mi->part->partno)
106                 mi->inflight[1]++;
107 }
108
109 void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
110                       unsigned int inflight[2])
111 {
112         struct mq_inflight mi = { .part = part, .inflight = inflight, };
113
114         inflight[0] = inflight[1] = 0;
115         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
116 }
117
118 static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
119                                      struct request *rq, void *priv,
120                                      bool reserved)
121 {
122         struct mq_inflight *mi = priv;
123
124         if (rq->part == mi->part)
125                 mi->inflight[rq_data_dir(rq)]++;
126 }
127
128 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
129                          unsigned int inflight[2])
130 {
131         struct mq_inflight mi = { .part = part, .inflight = inflight, };
132
133         inflight[0] = inflight[1] = 0;
134         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
135 }
136
137 void blk_freeze_queue_start(struct request_queue *q)
138 {
139         int freeze_depth;
140
141         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
142         if (freeze_depth == 1) {
143                 percpu_ref_kill(&q->q_usage_counter);
144                 if (q->mq_ops)
145                         blk_mq_run_hw_queues(q, false);
146         }
147 }
148 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
149
150 void blk_mq_freeze_queue_wait(struct request_queue *q)
151 {
152         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
153 }
154 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
155
156 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
157                                      unsigned long timeout)
158 {
159         return wait_event_timeout(q->mq_freeze_wq,
160                                         percpu_ref_is_zero(&q->q_usage_counter),
161                                         timeout);
162 }
163 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
164
165 /*
166  * Guarantee no request is in use, so we can change any data structure of
167  * the queue afterward.
168  */
169 void blk_freeze_queue(struct request_queue *q)
170 {
171         /*
172          * In the !blk_mq case we are only calling this to kill the
173          * q_usage_counter, otherwise this increases the freeze depth
174          * and waits for it to return to zero.  For this reason there is
175          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
176          * exported to drivers as the only user for unfreeze is blk_mq.
177          */
178         blk_freeze_queue_start(q);
179         if (!q->mq_ops)
180                 blk_drain_queue(q);
181         blk_mq_freeze_queue_wait(q);
182 }
183
184 void blk_mq_freeze_queue(struct request_queue *q)
185 {
186         /*
187          * ...just an alias to keep freeze and unfreeze actions balanced
188          * in the blk_mq_* namespace
189          */
190         blk_freeze_queue(q);
191 }
192 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
193
194 void blk_mq_unfreeze_queue(struct request_queue *q)
195 {
196         int freeze_depth;
197
198         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
199         WARN_ON_ONCE(freeze_depth < 0);
200         if (!freeze_depth) {
201                 percpu_ref_reinit(&q->q_usage_counter);
202                 wake_up_all(&q->mq_freeze_wq);
203         }
204 }
205 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
206
207 /*
208  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
209  * mpt3sas driver such that this function can be removed.
210  */
211 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
212 {
213         blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
214 }
215 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
216
217 /**
218  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
219  * @q: request queue.
220  *
221  * Note: this function does not prevent that the struct request end_io()
222  * callback function is invoked. Once this function is returned, we make
223  * sure no dispatch can happen until the queue is unquiesced via
224  * blk_mq_unquiesce_queue().
225  */
226 void blk_mq_quiesce_queue(struct request_queue *q)
227 {
228         struct blk_mq_hw_ctx *hctx;
229         unsigned int i;
230         bool rcu = false;
231
232         blk_mq_quiesce_queue_nowait(q);
233
234         queue_for_each_hw_ctx(q, hctx, i) {
235                 if (hctx->flags & BLK_MQ_F_BLOCKING)
236                         synchronize_srcu(hctx->srcu);
237                 else
238                         rcu = true;
239         }
240         if (rcu)
241                 synchronize_rcu();
242 }
243 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
244
245 /*
246  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
247  * @q: request queue.
248  *
249  * This function recovers queue into the state before quiescing
250  * which is done by blk_mq_quiesce_queue.
251  */
252 void blk_mq_unquiesce_queue(struct request_queue *q)
253 {
254         blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
255
256         /* dispatch requests which are inserted during quiescing */
257         blk_mq_run_hw_queues(q, true);
258 }
259 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
260
261 void blk_mq_wake_waiters(struct request_queue *q)
262 {
263         struct blk_mq_hw_ctx *hctx;
264         unsigned int i;
265
266         queue_for_each_hw_ctx(q, hctx, i)
267                 if (blk_mq_hw_queue_mapped(hctx))
268                         blk_mq_tag_wakeup_all(hctx->tags, true);
269 }
270
271 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
272 {
273         return blk_mq_has_free_tags(hctx->tags);
274 }
275 EXPORT_SYMBOL(blk_mq_can_queue);
276
277 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
278                 unsigned int tag, unsigned int op)
279 {
280         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
281         struct request *rq = tags->static_rqs[tag];
282         req_flags_t rq_flags = 0;
283
284         if (data->flags & BLK_MQ_REQ_INTERNAL) {
285                 rq->tag = -1;
286                 rq->internal_tag = tag;
287         } else {
288                 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
289                         rq_flags = RQF_MQ_INFLIGHT;
290                         atomic_inc(&data->hctx->nr_active);
291                 }
292                 rq->tag = tag;
293                 rq->internal_tag = -1;
294                 data->hctx->tags->rqs[rq->tag] = rq;
295         }
296
297         /* csd/requeue_work/fifo_time is initialized before use */
298         rq->q = data->q;
299         rq->mq_ctx = data->ctx;
300         rq->rq_flags = rq_flags;
301         rq->cpu = -1;
302         rq->cmd_flags = op;
303         if (data->flags & BLK_MQ_REQ_PREEMPT)
304                 rq->rq_flags |= RQF_PREEMPT;
305         if (blk_queue_io_stat(data->q))
306                 rq->rq_flags |= RQF_IO_STAT;
307         INIT_LIST_HEAD(&rq->queuelist);
308         INIT_HLIST_NODE(&rq->hash);
309         RB_CLEAR_NODE(&rq->rb_node);
310         rq->rq_disk = NULL;
311         rq->part = NULL;
312         rq->start_time_ns = ktime_get_ns();
313         rq->io_start_time_ns = 0;
314         rq->nr_phys_segments = 0;
315 #if defined(CONFIG_BLK_DEV_INTEGRITY)
316         rq->nr_integrity_segments = 0;
317 #endif
318         rq->special = NULL;
319         /* tag was already set */
320         rq->extra_len = 0;
321         rq->__deadline = 0;
322
323         INIT_LIST_HEAD(&rq->timeout_list);
324         rq->timeout = 0;
325
326         rq->end_io = NULL;
327         rq->end_io_data = NULL;
328         rq->next_rq = NULL;
329
330 #ifdef CONFIG_BLK_CGROUP
331         rq->rl = NULL;
332 #endif
333
334         data->ctx->rq_dispatched[op_is_sync(op)]++;
335         refcount_set(&rq->ref, 1);
336         return rq;
337 }
338
339 static struct request *blk_mq_get_request(struct request_queue *q,
340                 struct bio *bio, unsigned int op,
341                 struct blk_mq_alloc_data *data)
342 {
343         struct elevator_queue *e = q->elevator;
344         struct request *rq;
345         unsigned int tag;
346         bool put_ctx_on_error = false;
347
348         blk_queue_enter_live(q);
349         data->q = q;
350         if (likely(!data->ctx)) {
351                 data->ctx = blk_mq_get_ctx(q);
352                 put_ctx_on_error = true;
353         }
354         if (likely(!data->hctx))
355                 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
356         if (op & REQ_NOWAIT)
357                 data->flags |= BLK_MQ_REQ_NOWAIT;
358
359         if (e) {
360                 data->flags |= BLK_MQ_REQ_INTERNAL;
361
362                 /*
363                  * Flush requests are special and go directly to the
364                  * dispatch list. Don't include reserved tags in the
365                  * limiting, as it isn't useful.
366                  */
367                 if (!op_is_flush(op) && e->type->ops.mq.limit_depth &&
368                     !(data->flags & BLK_MQ_REQ_RESERVED))
369                         e->type->ops.mq.limit_depth(op, data);
370         } else {
371                 blk_mq_tag_busy(data->hctx);
372         }
373
374         tag = blk_mq_get_tag(data);
375         if (tag == BLK_MQ_TAG_FAIL) {
376                 if (put_ctx_on_error) {
377                         blk_mq_put_ctx(data->ctx);
378                         data->ctx = NULL;
379                 }
380                 blk_queue_exit(q);
381                 return NULL;
382         }
383
384         rq = blk_mq_rq_ctx_init(data, tag, op);
385         if (!op_is_flush(op)) {
386                 rq->elv.icq = NULL;
387                 if (e && e->type->ops.mq.prepare_request) {
388                         if (e->type->icq_cache && rq_ioc(bio))
389                                 blk_mq_sched_assign_ioc(rq, bio);
390
391                         e->type->ops.mq.prepare_request(rq, bio);
392                         rq->rq_flags |= RQF_ELVPRIV;
393                 }
394         }
395         data->hctx->queued++;
396         return rq;
397 }
398
399 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
400                 blk_mq_req_flags_t flags)
401 {
402         struct blk_mq_alloc_data alloc_data = { .flags = flags };
403         struct request *rq;
404         int ret;
405
406         ret = blk_queue_enter(q, flags);
407         if (ret)
408                 return ERR_PTR(ret);
409
410         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
411         blk_queue_exit(q);
412
413         if (!rq)
414                 return ERR_PTR(-EWOULDBLOCK);
415
416         blk_mq_put_ctx(alloc_data.ctx);
417
418         rq->__data_len = 0;
419         rq->__sector = (sector_t) -1;
420         rq->bio = rq->biotail = NULL;
421         return rq;
422 }
423 EXPORT_SYMBOL(blk_mq_alloc_request);
424
425 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
426         unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
427 {
428         struct blk_mq_alloc_data alloc_data = { .flags = flags };
429         struct request *rq;
430         unsigned int cpu;
431         int ret;
432
433         /*
434          * If the tag allocator sleeps we could get an allocation for a
435          * different hardware context.  No need to complicate the low level
436          * allocator for this for the rare use case of a command tied to
437          * a specific queue.
438          */
439         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
440                 return ERR_PTR(-EINVAL);
441
442         if (hctx_idx >= q->nr_hw_queues)
443                 return ERR_PTR(-EIO);
444
445         ret = blk_queue_enter(q, flags);
446         if (ret)
447                 return ERR_PTR(ret);
448
449         /*
450          * Check if the hardware context is actually mapped to anything.
451          * If not tell the caller that it should skip this queue.
452          */
453         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
454         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
455                 blk_queue_exit(q);
456                 return ERR_PTR(-EXDEV);
457         }
458         cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
459         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
460
461         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
462         blk_queue_exit(q);
463
464         if (!rq)
465                 return ERR_PTR(-EWOULDBLOCK);
466
467         return rq;
468 }
469 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
470
471 static void __blk_mq_free_request(struct request *rq)
472 {
473         struct request_queue *q = rq->q;
474         struct blk_mq_ctx *ctx = rq->mq_ctx;
475         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
476         const int sched_tag = rq->internal_tag;
477
478         if (rq->tag != -1)
479                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
480         if (sched_tag != -1)
481                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
482         blk_mq_sched_restart(hctx);
483         blk_queue_exit(q);
484 }
485
486 void blk_mq_free_request(struct request *rq)
487 {
488         struct request_queue *q = rq->q;
489         struct elevator_queue *e = q->elevator;
490         struct blk_mq_ctx *ctx = rq->mq_ctx;
491         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
492
493         if (rq->rq_flags & RQF_ELVPRIV) {
494                 if (e && e->type->ops.mq.finish_request)
495                         e->type->ops.mq.finish_request(rq);
496                 if (rq->elv.icq) {
497                         put_io_context(rq->elv.icq->ioc);
498                         rq->elv.icq = NULL;
499                 }
500         }
501
502         ctx->rq_completed[rq_is_sync(rq)]++;
503         if (rq->rq_flags & RQF_MQ_INFLIGHT)
504                 atomic_dec(&hctx->nr_active);
505
506         if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
507                 laptop_io_completion(q->backing_dev_info);
508
509         rq_qos_done(q, rq);
510
511         if (blk_rq_rl(rq))
512                 blk_put_rl(blk_rq_rl(rq));
513
514         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
515         if (refcount_dec_and_test(&rq->ref))
516                 __blk_mq_free_request(rq);
517 }
518 EXPORT_SYMBOL_GPL(blk_mq_free_request);
519
520 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
521 {
522         u64 now = ktime_get_ns();
523
524         if (rq->rq_flags & RQF_STATS) {
525                 blk_mq_poll_stats_start(rq->q);
526                 blk_stat_add(rq, now);
527         }
528
529         blk_account_io_done(rq, now);
530
531         if (rq->end_io) {
532                 rq_qos_done(rq->q, rq);
533                 rq->end_io(rq, error);
534         } else {
535                 if (unlikely(blk_bidi_rq(rq)))
536                         blk_mq_free_request(rq->next_rq);
537                 blk_mq_free_request(rq);
538         }
539 }
540 EXPORT_SYMBOL(__blk_mq_end_request);
541
542 void blk_mq_end_request(struct request *rq, blk_status_t error)
543 {
544         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
545                 BUG();
546         __blk_mq_end_request(rq, error);
547 }
548 EXPORT_SYMBOL(blk_mq_end_request);
549
550 static void __blk_mq_complete_request_remote(void *data)
551 {
552         struct request *rq = data;
553
554         rq->q->softirq_done_fn(rq);
555 }
556
557 static void __blk_mq_complete_request(struct request *rq)
558 {
559         struct blk_mq_ctx *ctx = rq->mq_ctx;
560         bool shared = false;
561         int cpu;
562
563         if (!blk_mq_mark_complete(rq))
564                 return;
565         if (rq->internal_tag != -1)
566                 blk_mq_sched_completed_request(rq);
567
568         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
569                 rq->q->softirq_done_fn(rq);
570                 return;
571         }
572
573         cpu = get_cpu();
574         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
575                 shared = cpus_share_cache(cpu, ctx->cpu);
576
577         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
578                 rq->csd.func = __blk_mq_complete_request_remote;
579                 rq->csd.info = rq;
580                 rq->csd.flags = 0;
581                 smp_call_function_single_async(ctx->cpu, &rq->csd);
582         } else {
583                 rq->q->softirq_done_fn(rq);
584         }
585         put_cpu();
586 }
587
588 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
589         __releases(hctx->srcu)
590 {
591         if (!(hctx->flags & BLK_MQ_F_BLOCKING))
592                 rcu_read_unlock();
593         else
594                 srcu_read_unlock(hctx->srcu, srcu_idx);
595 }
596
597 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
598         __acquires(hctx->srcu)
599 {
600         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
601                 /* shut up gcc false positive */
602                 *srcu_idx = 0;
603                 rcu_read_lock();
604         } else
605                 *srcu_idx = srcu_read_lock(hctx->srcu);
606 }
607
608 /**
609  * blk_mq_complete_request - end I/O on a request
610  * @rq:         the request being processed
611  *
612  * Description:
613  *      Ends all I/O on a request. It does not handle partial completions.
614  *      The actual completion happens out-of-order, through a IPI handler.
615  **/
616 void blk_mq_complete_request(struct request *rq)
617 {
618         if (unlikely(blk_should_fake_timeout(rq->q)))
619                 return;
620         __blk_mq_complete_request(rq);
621 }
622 EXPORT_SYMBOL(blk_mq_complete_request);
623
624 int blk_mq_request_started(struct request *rq)
625 {
626         return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
627 }
628 EXPORT_SYMBOL_GPL(blk_mq_request_started);
629
630 void blk_mq_start_request(struct request *rq)
631 {
632         struct request_queue *q = rq->q;
633
634         blk_mq_sched_started_request(rq);
635
636         trace_block_rq_issue(q, rq);
637
638         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
639                 rq->io_start_time_ns = ktime_get_ns();
640 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
641                 rq->throtl_size = blk_rq_sectors(rq);
642 #endif
643                 rq->rq_flags |= RQF_STATS;
644                 rq_qos_issue(q, rq);
645         }
646
647         WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
648
649         blk_add_timer(rq);
650         WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
651
652         if (q->dma_drain_size && blk_rq_bytes(rq)) {
653                 /*
654                  * Make sure space for the drain appears.  We know we can do
655                  * this because max_hw_segments has been adjusted to be one
656                  * fewer than the device can handle.
657                  */
658                 rq->nr_phys_segments++;
659         }
660 }
661 EXPORT_SYMBOL(blk_mq_start_request);
662
663 static void __blk_mq_requeue_request(struct request *rq)
664 {
665         struct request_queue *q = rq->q;
666
667         blk_mq_put_driver_tag(rq);
668
669         trace_block_rq_requeue(q, rq);
670         rq_qos_requeue(q, rq);
671
672         if (blk_mq_request_started(rq)) {
673                 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
674                 rq->rq_flags &= ~RQF_TIMED_OUT;
675                 if (q->dma_drain_size && blk_rq_bytes(rq))
676                         rq->nr_phys_segments--;
677         }
678 }
679
680 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
681 {
682         __blk_mq_requeue_request(rq);
683
684         /* this request will be re-inserted to io scheduler queue */
685         blk_mq_sched_requeue_request(rq);
686
687         BUG_ON(blk_queued_rq(rq));
688         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
689 }
690 EXPORT_SYMBOL(blk_mq_requeue_request);
691
692 static void blk_mq_requeue_work(struct work_struct *work)
693 {
694         struct request_queue *q =
695                 container_of(work, struct request_queue, requeue_work.work);
696         LIST_HEAD(rq_list);
697         struct request *rq, *next;
698
699         spin_lock_irq(&q->requeue_lock);
700         list_splice_init(&q->requeue_list, &rq_list);
701         spin_unlock_irq(&q->requeue_lock);
702
703         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
704                 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
705                         continue;
706
707                 rq->rq_flags &= ~RQF_SOFTBARRIER;
708                 list_del_init(&rq->queuelist);
709                 /*
710                  * If RQF_DONTPREP, rq has contained some driver specific
711                  * data, so insert it to hctx dispatch list to avoid any
712                  * merge.
713                  */
714                 if (rq->rq_flags & RQF_DONTPREP)
715                         blk_mq_request_bypass_insert(rq, false);
716                 else
717                         blk_mq_sched_insert_request(rq, true, false, false);
718         }
719
720         while (!list_empty(&rq_list)) {
721                 rq = list_entry(rq_list.next, struct request, queuelist);
722                 list_del_init(&rq->queuelist);
723                 blk_mq_sched_insert_request(rq, false, false, false);
724         }
725
726         blk_mq_run_hw_queues(q, false);
727 }
728
729 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
730                                 bool kick_requeue_list)
731 {
732         struct request_queue *q = rq->q;
733         unsigned long flags;
734
735         /*
736          * We abuse this flag that is otherwise used by the I/O scheduler to
737          * request head insertion from the workqueue.
738          */
739         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
740
741         spin_lock_irqsave(&q->requeue_lock, flags);
742         if (at_head) {
743                 rq->rq_flags |= RQF_SOFTBARRIER;
744                 list_add(&rq->queuelist, &q->requeue_list);
745         } else {
746                 list_add_tail(&rq->queuelist, &q->requeue_list);
747         }
748         spin_unlock_irqrestore(&q->requeue_lock, flags);
749
750         if (kick_requeue_list)
751                 blk_mq_kick_requeue_list(q);
752 }
753 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
754
755 void blk_mq_kick_requeue_list(struct request_queue *q)
756 {
757         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
758 }
759 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
760
761 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
762                                     unsigned long msecs)
763 {
764         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
765                                     msecs_to_jiffies(msecs));
766 }
767 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
768
769 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
770 {
771         if (tag < tags->nr_tags) {
772                 prefetch(tags->rqs[tag]);
773                 return tags->rqs[tag];
774         }
775
776         return NULL;
777 }
778 EXPORT_SYMBOL(blk_mq_tag_to_rq);
779
780 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
781 {
782         req->rq_flags |= RQF_TIMED_OUT;
783         if (req->q->mq_ops->timeout) {
784                 enum blk_eh_timer_return ret;
785
786                 ret = req->q->mq_ops->timeout(req, reserved);
787                 if (ret == BLK_EH_DONE)
788                         return;
789                 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
790         }
791
792         blk_add_timer(req);
793 }
794
795 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
796 {
797         unsigned long deadline;
798
799         if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
800                 return false;
801         if (rq->rq_flags & RQF_TIMED_OUT)
802                 return false;
803
804         deadline = blk_rq_deadline(rq);
805         if (time_after_eq(jiffies, deadline))
806                 return true;
807
808         if (*next == 0)
809                 *next = deadline;
810         else if (time_after(*next, deadline))
811                 *next = deadline;
812         return false;
813 }
814
815 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
816                 struct request *rq, void *priv, bool reserved)
817 {
818         unsigned long *next = priv;
819
820         /*
821          * Just do a quick check if it is expired before locking the request in
822          * so we're not unnecessarilly synchronizing across CPUs.
823          */
824         if (!blk_mq_req_expired(rq, next))
825                 return;
826
827         /*
828          * We have reason to believe the request may be expired. Take a
829          * reference on the request to lock this request lifetime into its
830          * currently allocated context to prevent it from being reallocated in
831          * the event the completion by-passes this timeout handler.
832          *
833          * If the reference was already released, then the driver beat the
834          * timeout handler to posting a natural completion.
835          */
836         if (!refcount_inc_not_zero(&rq->ref))
837                 return;
838
839         /*
840          * The request is now locked and cannot be reallocated underneath the
841          * timeout handler's processing. Re-verify this exact request is truly
842          * expired; if it is not expired, then the request was completed and
843          * reallocated as a new request.
844          */
845         if (blk_mq_req_expired(rq, next))
846                 blk_mq_rq_timed_out(rq, reserved);
847
848         if (is_flush_rq(rq, hctx))
849                 rq->end_io(rq, 0);
850         else if (refcount_dec_and_test(&rq->ref))
851                 __blk_mq_free_request(rq);
852 }
853
854 static void blk_mq_timeout_work(struct work_struct *work)
855 {
856         struct request_queue *q =
857                 container_of(work, struct request_queue, timeout_work);
858         unsigned long next = 0;
859         struct blk_mq_hw_ctx *hctx;
860         int i;
861
862         /* A deadlock might occur if a request is stuck requiring a
863          * timeout at the same time a queue freeze is waiting
864          * completion, since the timeout code would not be able to
865          * acquire the queue reference here.
866          *
867          * That's why we don't use blk_queue_enter here; instead, we use
868          * percpu_ref_tryget directly, because we need to be able to
869          * obtain a reference even in the short window between the queue
870          * starting to freeze, by dropping the first reference in
871          * blk_freeze_queue_start, and the moment the last request is
872          * consumed, marked by the instant q_usage_counter reaches
873          * zero.
874          */
875         if (!percpu_ref_tryget(&q->q_usage_counter))
876                 return;
877
878         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
879
880         if (next != 0) {
881                 mod_timer(&q->timeout, next);
882         } else {
883                 /*
884                  * Request timeouts are handled as a forward rolling timer. If
885                  * we end up here it means that no requests are pending and
886                  * also that no request has been pending for a while. Mark
887                  * each hctx as idle.
888                  */
889                 queue_for_each_hw_ctx(q, hctx, i) {
890                         /* the hctx may be unmapped, so check it here */
891                         if (blk_mq_hw_queue_mapped(hctx))
892                                 blk_mq_tag_idle(hctx);
893                 }
894         }
895         blk_queue_exit(q);
896 }
897
898 struct flush_busy_ctx_data {
899         struct blk_mq_hw_ctx *hctx;
900         struct list_head *list;
901 };
902
903 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
904 {
905         struct flush_busy_ctx_data *flush_data = data;
906         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
907         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
908
909         spin_lock(&ctx->lock);
910         list_splice_tail_init(&ctx->rq_list, flush_data->list);
911         sbitmap_clear_bit(sb, bitnr);
912         spin_unlock(&ctx->lock);
913         return true;
914 }
915
916 /*
917  * Process software queues that have been marked busy, splicing them
918  * to the for-dispatch
919  */
920 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
921 {
922         struct flush_busy_ctx_data data = {
923                 .hctx = hctx,
924                 .list = list,
925         };
926
927         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
928 }
929 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
930
931 struct dispatch_rq_data {
932         struct blk_mq_hw_ctx *hctx;
933         struct request *rq;
934 };
935
936 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
937                 void *data)
938 {
939         struct dispatch_rq_data *dispatch_data = data;
940         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
941         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
942
943         spin_lock(&ctx->lock);
944         if (!list_empty(&ctx->rq_list)) {
945                 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
946                 list_del_init(&dispatch_data->rq->queuelist);
947                 if (list_empty(&ctx->rq_list))
948                         sbitmap_clear_bit(sb, bitnr);
949         }
950         spin_unlock(&ctx->lock);
951
952         return !dispatch_data->rq;
953 }
954
955 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
956                                         struct blk_mq_ctx *start)
957 {
958         unsigned off = start ? start->index_hw : 0;
959         struct dispatch_rq_data data = {
960                 .hctx = hctx,
961                 .rq   = NULL,
962         };
963
964         __sbitmap_for_each_set(&hctx->ctx_map, off,
965                                dispatch_rq_from_ctx, &data);
966
967         return data.rq;
968 }
969
970 static inline unsigned int queued_to_index(unsigned int queued)
971 {
972         if (!queued)
973                 return 0;
974
975         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
976 }
977
978 bool blk_mq_get_driver_tag(struct request *rq)
979 {
980         struct blk_mq_alloc_data data = {
981                 .q = rq->q,
982                 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
983                 .flags = BLK_MQ_REQ_NOWAIT,
984         };
985         bool shared;
986
987         if (rq->tag != -1)
988                 goto done;
989
990         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
991                 data.flags |= BLK_MQ_REQ_RESERVED;
992
993         shared = blk_mq_tag_busy(data.hctx);
994         rq->tag = blk_mq_get_tag(&data);
995         if (rq->tag >= 0) {
996                 if (shared) {
997                         rq->rq_flags |= RQF_MQ_INFLIGHT;
998                         atomic_inc(&data.hctx->nr_active);
999                 }
1000                 data.hctx->tags->rqs[rq->tag] = rq;
1001         }
1002
1003 done:
1004         return rq->tag != -1;
1005 }
1006
1007 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1008                                 int flags, void *key)
1009 {
1010         struct blk_mq_hw_ctx *hctx;
1011
1012         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1013
1014         spin_lock(&hctx->dispatch_wait_lock);
1015         list_del_init(&wait->entry);
1016         spin_unlock(&hctx->dispatch_wait_lock);
1017
1018         blk_mq_run_hw_queue(hctx, true);
1019         return 1;
1020 }
1021
1022 /*
1023  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1024  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1025  * restart. For both cases, take care to check the condition again after
1026  * marking us as waiting.
1027  */
1028 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1029                                  struct request *rq)
1030 {
1031         struct wait_queue_head *wq;
1032         wait_queue_entry_t *wait;
1033         bool ret;
1034
1035         if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1036                 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
1037                         set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
1038
1039                 /*
1040                  * It's possible that a tag was freed in the window between the
1041                  * allocation failure and adding the hardware queue to the wait
1042                  * queue.
1043                  *
1044                  * Don't clear RESTART here, someone else could have set it.
1045                  * At most this will cost an extra queue run.
1046                  */
1047                 return blk_mq_get_driver_tag(rq);
1048         }
1049
1050         wait = &hctx->dispatch_wait;
1051         if (!list_empty_careful(&wait->entry))
1052                 return false;
1053
1054         wq = &bt_wait_ptr(&hctx->tags->bitmap_tags, hctx)->wait;
1055
1056         spin_lock_irq(&wq->lock);
1057         spin_lock(&hctx->dispatch_wait_lock);
1058         if (!list_empty(&wait->entry)) {
1059                 spin_unlock(&hctx->dispatch_wait_lock);
1060                 spin_unlock_irq(&wq->lock);
1061                 return false;
1062         }
1063
1064         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1065         __add_wait_queue(wq, wait);
1066
1067         /*
1068          * It's possible that a tag was freed in the window between the
1069          * allocation failure and adding the hardware queue to the wait
1070          * queue.
1071          */
1072         ret = blk_mq_get_driver_tag(rq);
1073         if (!ret) {
1074                 spin_unlock(&hctx->dispatch_wait_lock);
1075                 spin_unlock_irq(&wq->lock);
1076                 return false;
1077         }
1078
1079         /*
1080          * We got a tag, remove ourselves from the wait queue to ensure
1081          * someone else gets the wakeup.
1082          */
1083         list_del_init(&wait->entry);
1084         spin_unlock(&hctx->dispatch_wait_lock);
1085         spin_unlock_irq(&wq->lock);
1086
1087         return true;
1088 }
1089
1090 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  8
1091 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  4
1092 /*
1093  * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1094  * - EWMA is one simple way to compute running average value
1095  * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1096  * - take 4 as factor for avoiding to get too small(0) result, and this
1097  *   factor doesn't matter because EWMA decreases exponentially
1098  */
1099 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1100 {
1101         unsigned int ewma;
1102
1103         if (hctx->queue->elevator)
1104                 return;
1105
1106         ewma = hctx->dispatch_busy;
1107
1108         if (!ewma && !busy)
1109                 return;
1110
1111         ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1112         if (busy)
1113                 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1114         ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1115
1116         hctx->dispatch_busy = ewma;
1117 }
1118
1119 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1120
1121 static void blk_mq_handle_dev_resource(struct request *rq,
1122                                        struct list_head *list)
1123 {
1124         struct request *next =
1125                 list_first_entry_or_null(list, struct request, queuelist);
1126
1127         /*
1128          * If an I/O scheduler has been configured and we got a driver tag for
1129          * the next request already, free it.
1130          */
1131         if (next)
1132                 blk_mq_put_driver_tag(next);
1133
1134         list_add(&rq->queuelist, list);
1135         __blk_mq_requeue_request(rq);
1136 }
1137
1138 /*
1139  * Returns true if we did some work AND can potentially do more.
1140  */
1141 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1142                              bool got_budget)
1143 {
1144         struct blk_mq_hw_ctx *hctx;
1145         struct request *rq, *nxt;
1146         bool no_tag = false;
1147         int errors, queued;
1148         blk_status_t ret = BLK_STS_OK;
1149
1150         if (list_empty(list))
1151                 return false;
1152
1153         WARN_ON(!list_is_singular(list) && got_budget);
1154
1155         /*
1156          * Now process all the entries, sending them to the driver.
1157          */
1158         errors = queued = 0;
1159         do {
1160                 struct blk_mq_queue_data bd;
1161
1162                 rq = list_first_entry(list, struct request, queuelist);
1163
1164                 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1165                 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1166                         break;
1167
1168                 if (!blk_mq_get_driver_tag(rq)) {
1169                         /*
1170                          * The initial allocation attempt failed, so we need to
1171                          * rerun the hardware queue when a tag is freed. The
1172                          * waitqueue takes care of that. If the queue is run
1173                          * before we add this entry back on the dispatch list,
1174                          * we'll re-run it below.
1175                          */
1176                         if (!blk_mq_mark_tag_wait(hctx, rq)) {
1177                                 blk_mq_put_dispatch_budget(hctx);
1178                                 /*
1179                                  * For non-shared tags, the RESTART check
1180                                  * will suffice.
1181                                  */
1182                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1183                                         no_tag = true;
1184                                 break;
1185                         }
1186                 }
1187
1188                 list_del_init(&rq->queuelist);
1189
1190                 bd.rq = rq;
1191
1192                 /*
1193                  * Flag last if we have no more requests, or if we have more
1194                  * but can't assign a driver tag to it.
1195                  */
1196                 if (list_empty(list))
1197                         bd.last = true;
1198                 else {
1199                         nxt = list_first_entry(list, struct request, queuelist);
1200                         bd.last = !blk_mq_get_driver_tag(nxt);
1201                 }
1202
1203                 ret = q->mq_ops->queue_rq(hctx, &bd);
1204                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1205                         blk_mq_handle_dev_resource(rq, list);
1206                         break;
1207                 }
1208
1209                 if (unlikely(ret != BLK_STS_OK)) {
1210                         errors++;
1211                         blk_mq_end_request(rq, BLK_STS_IOERR);
1212                         continue;
1213                 }
1214
1215                 queued++;
1216         } while (!list_empty(list));
1217
1218         hctx->dispatched[queued_to_index(queued)]++;
1219
1220         /*
1221          * Any items that need requeuing? Stuff them into hctx->dispatch,
1222          * that is where we will continue on next queue run.
1223          */
1224         if (!list_empty(list)) {
1225                 bool needs_restart;
1226
1227                 spin_lock(&hctx->lock);
1228                 list_splice_init(list, &hctx->dispatch);
1229                 spin_unlock(&hctx->lock);
1230
1231                 /*
1232                  * Order adding requests to hctx->dispatch and checking
1233                  * SCHED_RESTART flag. The pair of this smp_mb() is the one
1234                  * in blk_mq_sched_restart(). Avoid restart code path to
1235                  * miss the new added requests to hctx->dispatch, meantime
1236                  * SCHED_RESTART is observed here.
1237                  */
1238                 smp_mb();
1239
1240                 /*
1241                  * If SCHED_RESTART was set by the caller of this function and
1242                  * it is no longer set that means that it was cleared by another
1243                  * thread and hence that a queue rerun is needed.
1244                  *
1245                  * If 'no_tag' is set, that means that we failed getting
1246                  * a driver tag with an I/O scheduler attached. If our dispatch
1247                  * waitqueue is no longer active, ensure that we run the queue
1248                  * AFTER adding our entries back to the list.
1249                  *
1250                  * If no I/O scheduler has been configured it is possible that
1251                  * the hardware queue got stopped and restarted before requests
1252                  * were pushed back onto the dispatch list. Rerun the queue to
1253                  * avoid starvation. Notes:
1254                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1255                  *   been stopped before rerunning a queue.
1256                  * - Some but not all block drivers stop a queue before
1257                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1258                  *   and dm-rq.
1259                  *
1260                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1261                  * bit is set, run queue after a delay to avoid IO stalls
1262                  * that could otherwise occur if the queue is idle.
1263                  */
1264                 needs_restart = blk_mq_sched_needs_restart(hctx);
1265                 if (!needs_restart ||
1266                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1267                         blk_mq_run_hw_queue(hctx, true);
1268                 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1269                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1270
1271                 blk_mq_update_dispatch_busy(hctx, true);
1272                 return false;
1273         } else
1274                 blk_mq_update_dispatch_busy(hctx, false);
1275
1276         /*
1277          * If the host/device is unable to accept more work, inform the
1278          * caller of that.
1279          */
1280         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1281                 return false;
1282
1283         return (queued + errors) != 0;
1284 }
1285
1286 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1287 {
1288         int srcu_idx;
1289
1290         /*
1291          * We should be running this queue from one of the CPUs that
1292          * are mapped to it.
1293          *
1294          * There are at least two related races now between setting
1295          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1296          * __blk_mq_run_hw_queue():
1297          *
1298          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1299          *   but later it becomes online, then this warning is harmless
1300          *   at all
1301          *
1302          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1303          *   but later it becomes offline, then the warning can't be
1304          *   triggered, and we depend on blk-mq timeout handler to
1305          *   handle dispatched requests to this hctx
1306          */
1307         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1308                 cpu_online(hctx->next_cpu)) {
1309                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1310                         raw_smp_processor_id(),
1311                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1312                 dump_stack();
1313         }
1314
1315         /*
1316          * We can't run the queue inline with ints disabled. Ensure that
1317          * we catch bad users of this early.
1318          */
1319         WARN_ON_ONCE(in_interrupt());
1320
1321         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1322
1323         hctx_lock(hctx, &srcu_idx);
1324         blk_mq_sched_dispatch_requests(hctx);
1325         hctx_unlock(hctx, srcu_idx);
1326 }
1327
1328 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1329 {
1330         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1331
1332         if (cpu >= nr_cpu_ids)
1333                 cpu = cpumask_first(hctx->cpumask);
1334         return cpu;
1335 }
1336
1337 /*
1338  * It'd be great if the workqueue API had a way to pass
1339  * in a mask and had some smarts for more clever placement.
1340  * For now we just round-robin here, switching for every
1341  * BLK_MQ_CPU_WORK_BATCH queued items.
1342  */
1343 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1344 {
1345         bool tried = false;
1346         int next_cpu = hctx->next_cpu;
1347
1348         if (hctx->queue->nr_hw_queues == 1)
1349                 return WORK_CPU_UNBOUND;
1350
1351         if (--hctx->next_cpu_batch <= 0) {
1352 select_cpu:
1353                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1354                                 cpu_online_mask);
1355                 if (next_cpu >= nr_cpu_ids)
1356                         next_cpu = blk_mq_first_mapped_cpu(hctx);
1357                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1358         }
1359
1360         /*
1361          * Do unbound schedule if we can't find a online CPU for this hctx,
1362          * and it should only happen in the path of handling CPU DEAD.
1363          */
1364         if (!cpu_online(next_cpu)) {
1365                 if (!tried) {
1366                         tried = true;
1367                         goto select_cpu;
1368                 }
1369
1370                 /*
1371                  * Make sure to re-select CPU next time once after CPUs
1372                  * in hctx->cpumask become online again.
1373                  */
1374                 hctx->next_cpu = next_cpu;
1375                 hctx->next_cpu_batch = 1;
1376                 return WORK_CPU_UNBOUND;
1377         }
1378
1379         hctx->next_cpu = next_cpu;
1380         return next_cpu;
1381 }
1382
1383 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1384                                         unsigned long msecs)
1385 {
1386         if (unlikely(blk_mq_hctx_stopped(hctx)))
1387                 return;
1388
1389         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1390                 int cpu = get_cpu();
1391                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1392                         __blk_mq_run_hw_queue(hctx);
1393                         put_cpu();
1394                         return;
1395                 }
1396
1397                 put_cpu();
1398         }
1399
1400         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1401                                     msecs_to_jiffies(msecs));
1402 }
1403
1404 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1405 {
1406         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1407 }
1408 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1409
1410 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1411 {
1412         int srcu_idx;
1413         bool need_run;
1414
1415         /*
1416          * When queue is quiesced, we may be switching io scheduler, or
1417          * updating nr_hw_queues, or other things, and we can't run queue
1418          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1419          *
1420          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1421          * quiesced.
1422          */
1423         hctx_lock(hctx, &srcu_idx);
1424         need_run = !blk_queue_quiesced(hctx->queue) &&
1425                 blk_mq_hctx_has_pending(hctx);
1426         hctx_unlock(hctx, srcu_idx);
1427
1428         if (need_run) {
1429                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1430                 return true;
1431         }
1432
1433         return false;
1434 }
1435 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1436
1437 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1438 {
1439         struct blk_mq_hw_ctx *hctx;
1440         int i;
1441
1442         queue_for_each_hw_ctx(q, hctx, i) {
1443                 if (blk_mq_hctx_stopped(hctx))
1444                         continue;
1445
1446                 blk_mq_run_hw_queue(hctx, async);
1447         }
1448 }
1449 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1450
1451 /**
1452  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1453  * @q: request queue.
1454  *
1455  * The caller is responsible for serializing this function against
1456  * blk_mq_{start,stop}_hw_queue().
1457  */
1458 bool blk_mq_queue_stopped(struct request_queue *q)
1459 {
1460         struct blk_mq_hw_ctx *hctx;
1461         int i;
1462
1463         queue_for_each_hw_ctx(q, hctx, i)
1464                 if (blk_mq_hctx_stopped(hctx))
1465                         return true;
1466
1467         return false;
1468 }
1469 EXPORT_SYMBOL(blk_mq_queue_stopped);
1470
1471 /*
1472  * This function is often used for pausing .queue_rq() by driver when
1473  * there isn't enough resource or some conditions aren't satisfied, and
1474  * BLK_STS_RESOURCE is usually returned.
1475  *
1476  * We do not guarantee that dispatch can be drained or blocked
1477  * after blk_mq_stop_hw_queue() returns. Please use
1478  * blk_mq_quiesce_queue() for that requirement.
1479  */
1480 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1481 {
1482         cancel_delayed_work(&hctx->run_work);
1483
1484         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1485 }
1486 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1487
1488 /*
1489  * This function is often used for pausing .queue_rq() by driver when
1490  * there isn't enough resource or some conditions aren't satisfied, and
1491  * BLK_STS_RESOURCE is usually returned.
1492  *
1493  * We do not guarantee that dispatch can be drained or blocked
1494  * after blk_mq_stop_hw_queues() returns. Please use
1495  * blk_mq_quiesce_queue() for that requirement.
1496  */
1497 void blk_mq_stop_hw_queues(struct request_queue *q)
1498 {
1499         struct blk_mq_hw_ctx *hctx;
1500         int i;
1501
1502         queue_for_each_hw_ctx(q, hctx, i)
1503                 blk_mq_stop_hw_queue(hctx);
1504 }
1505 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1506
1507 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1508 {
1509         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1510
1511         blk_mq_run_hw_queue(hctx, false);
1512 }
1513 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1514
1515 void blk_mq_start_hw_queues(struct request_queue *q)
1516 {
1517         struct blk_mq_hw_ctx *hctx;
1518         int i;
1519
1520         queue_for_each_hw_ctx(q, hctx, i)
1521                 blk_mq_start_hw_queue(hctx);
1522 }
1523 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1524
1525 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1526 {
1527         if (!blk_mq_hctx_stopped(hctx))
1528                 return;
1529
1530         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1531         blk_mq_run_hw_queue(hctx, async);
1532 }
1533 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1534
1535 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1536 {
1537         struct blk_mq_hw_ctx *hctx;
1538         int i;
1539
1540         queue_for_each_hw_ctx(q, hctx, i)
1541                 blk_mq_start_stopped_hw_queue(hctx, async);
1542 }
1543 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1544
1545 static void blk_mq_run_work_fn(struct work_struct *work)
1546 {
1547         struct blk_mq_hw_ctx *hctx;
1548
1549         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1550
1551         /*
1552          * If we are stopped, don't run the queue.
1553          */
1554         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1555                 return;
1556
1557         __blk_mq_run_hw_queue(hctx);
1558 }
1559
1560 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1561                                             struct request *rq,
1562                                             bool at_head)
1563 {
1564         struct blk_mq_ctx *ctx = rq->mq_ctx;
1565
1566         lockdep_assert_held(&ctx->lock);
1567
1568         trace_block_rq_insert(hctx->queue, rq);
1569
1570         if (at_head)
1571                 list_add(&rq->queuelist, &ctx->rq_list);
1572         else
1573                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1574 }
1575
1576 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1577                              bool at_head)
1578 {
1579         struct blk_mq_ctx *ctx = rq->mq_ctx;
1580
1581         lockdep_assert_held(&ctx->lock);
1582
1583         __blk_mq_insert_req_list(hctx, rq, at_head);
1584         blk_mq_hctx_mark_pending(hctx, ctx);
1585 }
1586
1587 /*
1588  * Should only be used carefully, when the caller knows we want to
1589  * bypass a potential IO scheduler on the target device.
1590  */
1591 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1592 {
1593         struct blk_mq_ctx *ctx = rq->mq_ctx;
1594         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1595
1596         spin_lock(&hctx->lock);
1597         list_add_tail(&rq->queuelist, &hctx->dispatch);
1598         spin_unlock(&hctx->lock);
1599
1600         if (run_queue)
1601                 blk_mq_run_hw_queue(hctx, false);
1602 }
1603
1604 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1605                             struct list_head *list)
1606
1607 {
1608         struct request *rq;
1609
1610         /*
1611          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1612          * offline now
1613          */
1614         list_for_each_entry(rq, list, queuelist) {
1615                 BUG_ON(rq->mq_ctx != ctx);
1616                 trace_block_rq_insert(hctx->queue, rq);
1617         }
1618
1619         spin_lock(&ctx->lock);
1620         list_splice_tail_init(list, &ctx->rq_list);
1621         blk_mq_hctx_mark_pending(hctx, ctx);
1622         spin_unlock(&ctx->lock);
1623 }
1624
1625 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1626 {
1627         struct request *rqa = container_of(a, struct request, queuelist);
1628         struct request *rqb = container_of(b, struct request, queuelist);
1629
1630         return !(rqa->mq_ctx < rqb->mq_ctx ||
1631                  (rqa->mq_ctx == rqb->mq_ctx &&
1632                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1633 }
1634
1635 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1636 {
1637         struct blk_mq_ctx *this_ctx;
1638         struct request_queue *this_q;
1639         struct request *rq;
1640         LIST_HEAD(list);
1641         LIST_HEAD(ctx_list);
1642         unsigned int depth;
1643
1644         list_splice_init(&plug->mq_list, &list);
1645
1646         list_sort(NULL, &list, plug_ctx_cmp);
1647
1648         this_q = NULL;
1649         this_ctx = NULL;
1650         depth = 0;
1651
1652         while (!list_empty(&list)) {
1653                 rq = list_entry_rq(list.next);
1654                 list_del_init(&rq->queuelist);
1655                 BUG_ON(!rq->q);
1656                 if (rq->mq_ctx != this_ctx) {
1657                         if (this_ctx) {
1658                                 trace_block_unplug(this_q, depth, !from_schedule);
1659                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1660                                                                 &ctx_list,
1661                                                                 from_schedule);
1662                         }
1663
1664                         this_ctx = rq->mq_ctx;
1665                         this_q = rq->q;
1666                         depth = 0;
1667                 }
1668
1669                 depth++;
1670                 list_add_tail(&rq->queuelist, &ctx_list);
1671         }
1672
1673         /*
1674          * If 'this_ctx' is set, we know we have entries to complete
1675          * on 'ctx_list'. Do those.
1676          */
1677         if (this_ctx) {
1678                 trace_block_unplug(this_q, depth, !from_schedule);
1679                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1680                                                 from_schedule);
1681         }
1682 }
1683
1684 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1685 {
1686         blk_init_request_from_bio(rq, bio);
1687
1688         blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1689
1690         blk_account_io_start(rq, true);
1691 }
1692
1693 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1694 {
1695         if (rq->tag != -1)
1696                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1697
1698         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1699 }
1700
1701 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1702                                             struct request *rq,
1703                                             blk_qc_t *cookie)
1704 {
1705         struct request_queue *q = rq->q;
1706         struct blk_mq_queue_data bd = {
1707                 .rq = rq,
1708                 .last = true,
1709         };
1710         blk_qc_t new_cookie;
1711         blk_status_t ret;
1712
1713         new_cookie = request_to_qc_t(hctx, rq);
1714
1715         /*
1716          * For OK queue, we are done. For error, caller may kill it.
1717          * Any other error (busy), just add it to our list as we
1718          * previously would have done.
1719          */
1720         ret = q->mq_ops->queue_rq(hctx, &bd);
1721         switch (ret) {
1722         case BLK_STS_OK:
1723                 blk_mq_update_dispatch_busy(hctx, false);
1724                 *cookie = new_cookie;
1725                 break;
1726         case BLK_STS_RESOURCE:
1727         case BLK_STS_DEV_RESOURCE:
1728                 blk_mq_update_dispatch_busy(hctx, true);
1729                 __blk_mq_requeue_request(rq);
1730                 break;
1731         default:
1732                 blk_mq_update_dispatch_busy(hctx, false);
1733                 *cookie = BLK_QC_T_NONE;
1734                 break;
1735         }
1736
1737         return ret;
1738 }
1739
1740 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1741                                                 struct request *rq,
1742                                                 blk_qc_t *cookie,
1743                                                 bool bypass_insert)
1744 {
1745         struct request_queue *q = rq->q;
1746         bool run_queue = true;
1747
1748         /*
1749          * RCU or SRCU read lock is needed before checking quiesced flag.
1750          *
1751          * When queue is stopped or quiesced, ignore 'bypass_insert' from
1752          * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1753          * and avoid driver to try to dispatch again.
1754          */
1755         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1756                 run_queue = false;
1757                 bypass_insert = false;
1758                 goto insert;
1759         }
1760
1761         if (q->elevator && !bypass_insert)
1762                 goto insert;
1763
1764         if (!blk_mq_get_dispatch_budget(hctx))
1765                 goto insert;
1766
1767         if (!blk_mq_get_driver_tag(rq)) {
1768                 blk_mq_put_dispatch_budget(hctx);
1769                 goto insert;
1770         }
1771
1772         return __blk_mq_issue_directly(hctx, rq, cookie);
1773 insert:
1774         if (bypass_insert)
1775                 return BLK_STS_RESOURCE;
1776
1777         blk_mq_request_bypass_insert(rq, run_queue);
1778         return BLK_STS_OK;
1779 }
1780
1781 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1782                 struct request *rq, blk_qc_t *cookie)
1783 {
1784         blk_status_t ret;
1785         int srcu_idx;
1786
1787         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1788
1789         hctx_lock(hctx, &srcu_idx);
1790
1791         ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1792         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1793                 blk_mq_request_bypass_insert(rq, true);
1794         else if (ret != BLK_STS_OK)
1795                 blk_mq_end_request(rq, ret);
1796
1797         hctx_unlock(hctx, srcu_idx);
1798 }
1799
1800 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1801 {
1802         blk_status_t ret;
1803         int srcu_idx;
1804         blk_qc_t unused_cookie;
1805         struct blk_mq_ctx *ctx = rq->mq_ctx;
1806         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1807
1808         hctx_lock(hctx, &srcu_idx);
1809         ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1810         hctx_unlock(hctx, srcu_idx);
1811
1812         return ret;
1813 }
1814
1815 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1816                 struct list_head *list)
1817 {
1818         while (!list_empty(list)) {
1819                 blk_status_t ret;
1820                 struct request *rq = list_first_entry(list, struct request,
1821                                 queuelist);
1822
1823                 list_del_init(&rq->queuelist);
1824                 ret = blk_mq_request_issue_directly(rq);
1825                 if (ret != BLK_STS_OK) {
1826                         if (ret == BLK_STS_RESOURCE ||
1827                                         ret == BLK_STS_DEV_RESOURCE) {
1828                                 blk_mq_request_bypass_insert(rq,
1829                                                         list_empty(list));
1830                                 break;
1831                         }
1832                         blk_mq_end_request(rq, ret);
1833                 }
1834         }
1835 }
1836
1837 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1838 {
1839         const int is_sync = op_is_sync(bio->bi_opf);
1840         const int is_flush_fua = op_is_flush(bio->bi_opf);
1841         struct blk_mq_alloc_data data = { .flags = 0 };
1842         struct request *rq;
1843         unsigned int request_count = 0;
1844         struct blk_plug *plug;
1845         struct request *same_queue_rq = NULL;
1846         blk_qc_t cookie;
1847
1848         blk_queue_bounce(q, &bio);
1849
1850         blk_queue_split(q, &bio);
1851
1852         if (!bio_integrity_prep(bio))
1853                 return BLK_QC_T_NONE;
1854
1855         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1856             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1857                 return BLK_QC_T_NONE;
1858
1859         if (blk_mq_sched_bio_merge(q, bio))
1860                 return BLK_QC_T_NONE;
1861
1862         rq_qos_throttle(q, bio, NULL);
1863
1864         trace_block_getrq(q, bio, bio->bi_opf);
1865
1866         rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1867         if (unlikely(!rq)) {
1868                 rq_qos_cleanup(q, bio);
1869                 if (bio->bi_opf & REQ_NOWAIT)
1870                         bio_wouldblock_error(bio);
1871                 return BLK_QC_T_NONE;
1872         }
1873
1874         rq_qos_track(q, rq, bio);
1875
1876         cookie = request_to_qc_t(data.hctx, rq);
1877
1878         plug = current->plug;
1879         if (unlikely(is_flush_fua)) {
1880                 blk_mq_put_ctx(data.ctx);
1881                 blk_mq_bio_to_request(rq, bio);
1882
1883                 /* bypass scheduler for flush rq */
1884                 blk_insert_flush(rq);
1885                 blk_mq_run_hw_queue(data.hctx, true);
1886         } else if (plug && q->nr_hw_queues == 1) {
1887                 struct request *last = NULL;
1888
1889                 blk_mq_put_ctx(data.ctx);
1890                 blk_mq_bio_to_request(rq, bio);
1891
1892                 /*
1893                  * @request_count may become stale because of schedule
1894                  * out, so check the list again.
1895                  */
1896                 if (list_empty(&plug->mq_list))
1897                         request_count = 0;
1898                 else if (blk_queue_nomerges(q))
1899                         request_count = blk_plug_queued_count(q);
1900
1901                 if (!request_count)
1902                         trace_block_plug(q);
1903                 else
1904                         last = list_entry_rq(plug->mq_list.prev);
1905
1906                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1907                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1908                         blk_flush_plug_list(plug, false);
1909                         trace_block_plug(q);
1910                 }
1911
1912                 list_add_tail(&rq->queuelist, &plug->mq_list);
1913         } else if (plug && !blk_queue_nomerges(q)) {
1914                 blk_mq_bio_to_request(rq, bio);
1915
1916                 /*
1917                  * We do limited plugging. If the bio can be merged, do that.
1918                  * Otherwise the existing request in the plug list will be
1919                  * issued. So the plug list will have one request at most
1920                  * The plug list might get flushed before this. If that happens,
1921                  * the plug list is empty, and same_queue_rq is invalid.
1922                  */
1923                 if (list_empty(&plug->mq_list))
1924                         same_queue_rq = NULL;
1925                 if (same_queue_rq)
1926                         list_del_init(&same_queue_rq->queuelist);
1927                 list_add_tail(&rq->queuelist, &plug->mq_list);
1928
1929                 blk_mq_put_ctx(data.ctx);
1930
1931                 if (same_queue_rq) {
1932                         data.hctx = blk_mq_map_queue(q,
1933                                         same_queue_rq->mq_ctx->cpu);
1934                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1935                                         &cookie);
1936                 }
1937         } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
1938                         !data.hctx->dispatch_busy)) {
1939                 blk_mq_put_ctx(data.ctx);
1940                 blk_mq_bio_to_request(rq, bio);
1941                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1942         } else {
1943                 blk_mq_put_ctx(data.ctx);
1944                 blk_mq_bio_to_request(rq, bio);
1945                 blk_mq_sched_insert_request(rq, false, true, true);
1946         }
1947
1948         return cookie;
1949 }
1950
1951 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1952                      unsigned int hctx_idx)
1953 {
1954         struct page *page;
1955
1956         if (tags->rqs && set->ops->exit_request) {
1957                 int i;
1958
1959                 for (i = 0; i < tags->nr_tags; i++) {
1960                         struct request *rq = tags->static_rqs[i];
1961
1962                         if (!rq)
1963                                 continue;
1964                         set->ops->exit_request(set, rq, hctx_idx);
1965                         tags->static_rqs[i] = NULL;
1966                 }
1967         }
1968
1969         while (!list_empty(&tags->page_list)) {
1970                 page = list_first_entry(&tags->page_list, struct page, lru);
1971                 list_del_init(&page->lru);
1972                 /*
1973                  * Remove kmemleak object previously allocated in
1974                  * blk_mq_init_rq_map().
1975                  */
1976                 kmemleak_free(page_address(page));
1977                 __free_pages(page, page->private);
1978         }
1979 }
1980
1981 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1982 {
1983         kfree(tags->rqs);
1984         tags->rqs = NULL;
1985         kfree(tags->static_rqs);
1986         tags->static_rqs = NULL;
1987
1988         blk_mq_free_tags(tags);
1989 }
1990
1991 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1992                                         unsigned int hctx_idx,
1993                                         unsigned int nr_tags,
1994                                         unsigned int reserved_tags)
1995 {
1996         struct blk_mq_tags *tags;
1997         int node;
1998
1999         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2000         if (node == NUMA_NO_NODE)
2001                 node = set->numa_node;
2002
2003         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2004                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2005         if (!tags)
2006                 return NULL;
2007
2008         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2009                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2010                                  node);
2011         if (!tags->rqs) {
2012                 blk_mq_free_tags(tags);
2013                 return NULL;
2014         }
2015
2016         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2017                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2018                                         node);
2019         if (!tags->static_rqs) {
2020                 kfree(tags->rqs);
2021                 blk_mq_free_tags(tags);
2022                 return NULL;
2023         }
2024
2025         return tags;
2026 }
2027
2028 static size_t order_to_size(unsigned int order)
2029 {
2030         return (size_t)PAGE_SIZE << order;
2031 }
2032
2033 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2034                                unsigned int hctx_idx, int node)
2035 {
2036         int ret;
2037
2038         if (set->ops->init_request) {
2039                 ret = set->ops->init_request(set, rq, hctx_idx, node);
2040                 if (ret)
2041                         return ret;
2042         }
2043
2044         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2045         return 0;
2046 }
2047
2048 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2049                      unsigned int hctx_idx, unsigned int depth)
2050 {
2051         unsigned int i, j, entries_per_page, max_order = 4;
2052         size_t rq_size, left;
2053         int node;
2054
2055         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2056         if (node == NUMA_NO_NODE)
2057                 node = set->numa_node;
2058
2059         INIT_LIST_HEAD(&tags->page_list);
2060
2061         /*
2062          * rq_size is the size of the request plus driver payload, rounded
2063          * to the cacheline size
2064          */
2065         rq_size = round_up(sizeof(struct request) + set->cmd_size,
2066                                 cache_line_size());
2067         left = rq_size * depth;
2068
2069         for (i = 0; i < depth; ) {
2070                 int this_order = max_order;
2071                 struct page *page;
2072                 int to_do;
2073                 void *p;
2074
2075                 while (this_order && left < order_to_size(this_order - 1))
2076                         this_order--;
2077
2078                 do {
2079                         page = alloc_pages_node(node,
2080                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2081                                 this_order);
2082                         if (page)
2083                                 break;
2084                         if (!this_order--)
2085                                 break;
2086                         if (order_to_size(this_order) < rq_size)
2087                                 break;
2088                 } while (1);
2089
2090                 if (!page)
2091                         goto fail;
2092
2093                 page->private = this_order;
2094                 list_add_tail(&page->lru, &tags->page_list);
2095
2096                 p = page_address(page);
2097                 /*
2098                  * Allow kmemleak to scan these pages as they contain pointers
2099                  * to additional allocations like via ops->init_request().
2100                  */
2101                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2102                 entries_per_page = order_to_size(this_order) / rq_size;
2103                 to_do = min(entries_per_page, depth - i);
2104                 left -= to_do * rq_size;
2105                 for (j = 0; j < to_do; j++) {
2106                         struct request *rq = p;
2107
2108                         tags->static_rqs[i] = rq;
2109                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2110                                 tags->static_rqs[i] = NULL;
2111                                 goto fail;
2112                         }
2113
2114                         p += rq_size;
2115                         i++;
2116                 }
2117         }
2118         return 0;
2119
2120 fail:
2121         blk_mq_free_rqs(set, tags, hctx_idx);
2122         return -ENOMEM;
2123 }
2124
2125 /*
2126  * 'cpu' is going away. splice any existing rq_list entries from this
2127  * software queue to the hw queue dispatch list, and ensure that it
2128  * gets run.
2129  */
2130 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2131 {
2132         struct blk_mq_hw_ctx *hctx;
2133         struct blk_mq_ctx *ctx;
2134         LIST_HEAD(tmp);
2135
2136         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2137         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2138
2139         spin_lock(&ctx->lock);
2140         if (!list_empty(&ctx->rq_list)) {
2141                 list_splice_init(&ctx->rq_list, &tmp);
2142                 blk_mq_hctx_clear_pending(hctx, ctx);
2143         }
2144         spin_unlock(&ctx->lock);
2145
2146         if (list_empty(&tmp))
2147                 return 0;
2148
2149         spin_lock(&hctx->lock);
2150         list_splice_tail_init(&tmp, &hctx->dispatch);
2151         spin_unlock(&hctx->lock);
2152
2153         blk_mq_run_hw_queue(hctx, true);
2154         return 0;
2155 }
2156
2157 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2158 {
2159         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2160                                             &hctx->cpuhp_dead);
2161 }
2162
2163 /* hctx->ctxs will be freed in queue's release handler */
2164 static void blk_mq_exit_hctx(struct request_queue *q,
2165                 struct blk_mq_tag_set *set,
2166                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2167 {
2168         blk_mq_debugfs_unregister_hctx(hctx);
2169
2170         if (blk_mq_hw_queue_mapped(hctx))
2171                 blk_mq_tag_idle(hctx);
2172
2173         if (set->ops->exit_request)
2174                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2175
2176         if (set->ops->exit_hctx)
2177                 set->ops->exit_hctx(hctx, hctx_idx);
2178
2179         blk_mq_remove_cpuhp(hctx);
2180 }
2181
2182 static void blk_mq_exit_hw_queues(struct request_queue *q,
2183                 struct blk_mq_tag_set *set, int nr_queue)
2184 {
2185         struct blk_mq_hw_ctx *hctx;
2186         unsigned int i;
2187
2188         queue_for_each_hw_ctx(q, hctx, i) {
2189                 if (i == nr_queue)
2190                         break;
2191                 blk_mq_exit_hctx(q, set, hctx, i);
2192         }
2193 }
2194
2195 static int blk_mq_init_hctx(struct request_queue *q,
2196                 struct blk_mq_tag_set *set,
2197                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2198 {
2199         int node;
2200
2201         node = hctx->numa_node;
2202         if (node == NUMA_NO_NODE)
2203                 node = hctx->numa_node = set->numa_node;
2204
2205         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2206         spin_lock_init(&hctx->lock);
2207         INIT_LIST_HEAD(&hctx->dispatch);
2208         hctx->queue = q;
2209         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2210
2211         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2212
2213         hctx->tags = set->tags[hctx_idx];
2214
2215         /*
2216          * Allocate space for all possible cpus to avoid allocation at
2217          * runtime
2218          */
2219         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2220                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node);
2221         if (!hctx->ctxs)
2222                 goto unregister_cpu_notifier;
2223
2224         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2225                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node))
2226                 goto free_ctxs;
2227
2228         hctx->nr_ctx = 0;
2229
2230         spin_lock_init(&hctx->dispatch_wait_lock);
2231         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2232         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2233
2234         if (set->ops->init_hctx &&
2235             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2236                 goto free_bitmap;
2237
2238         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2239                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
2240         if (!hctx->fq)
2241                 goto exit_hctx;
2242
2243         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2244                 goto free_fq;
2245
2246         if (hctx->flags & BLK_MQ_F_BLOCKING)
2247                 init_srcu_struct(hctx->srcu);
2248
2249         blk_mq_debugfs_register_hctx(q, hctx);
2250
2251         return 0;
2252
2253  free_fq:
2254         blk_free_flush_queue(hctx->fq);
2255  exit_hctx:
2256         if (set->ops->exit_hctx)
2257                 set->ops->exit_hctx(hctx, hctx_idx);
2258  free_bitmap:
2259         sbitmap_free(&hctx->ctx_map);
2260  free_ctxs:
2261         kfree(hctx->ctxs);
2262  unregister_cpu_notifier:
2263         blk_mq_remove_cpuhp(hctx);
2264         return -1;
2265 }
2266
2267 static void blk_mq_init_cpu_queues(struct request_queue *q,
2268                                    unsigned int nr_hw_queues)
2269 {
2270         unsigned int i;
2271
2272         for_each_possible_cpu(i) {
2273                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2274                 struct blk_mq_hw_ctx *hctx;
2275
2276                 __ctx->cpu = i;
2277                 spin_lock_init(&__ctx->lock);
2278                 INIT_LIST_HEAD(&__ctx->rq_list);
2279                 __ctx->queue = q;
2280
2281                 /*
2282                  * Set local node, IFF we have more than one hw queue. If
2283                  * not, we remain on the home node of the device
2284                  */
2285                 hctx = blk_mq_map_queue(q, i);
2286                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2287                         hctx->numa_node = local_memory_node(cpu_to_node(i));
2288         }
2289 }
2290
2291 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2292 {
2293         int ret = 0;
2294
2295         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2296                                         set->queue_depth, set->reserved_tags);
2297         if (!set->tags[hctx_idx])
2298                 return false;
2299
2300         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2301                                 set->queue_depth);
2302         if (!ret)
2303                 return true;
2304
2305         blk_mq_free_rq_map(set->tags[hctx_idx]);
2306         set->tags[hctx_idx] = NULL;
2307         return false;
2308 }
2309
2310 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2311                                          unsigned int hctx_idx)
2312 {
2313         if (set->tags[hctx_idx]) {
2314                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2315                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2316                 set->tags[hctx_idx] = NULL;
2317         }
2318 }
2319
2320 static void blk_mq_map_swqueue(struct request_queue *q)
2321 {
2322         unsigned int i, hctx_idx;
2323         struct blk_mq_hw_ctx *hctx;
2324         struct blk_mq_ctx *ctx;
2325         struct blk_mq_tag_set *set = q->tag_set;
2326
2327         queue_for_each_hw_ctx(q, hctx, i) {
2328                 cpumask_clear(hctx->cpumask);
2329                 hctx->nr_ctx = 0;
2330                 hctx->dispatch_from = NULL;
2331         }
2332
2333         /*
2334          * Map software to hardware queues.
2335          *
2336          * If the cpu isn't present, the cpu is mapped to first hctx.
2337          */
2338         for_each_possible_cpu(i) {
2339                 hctx_idx = q->mq_map[i];
2340                 /* unmapped hw queue can be remapped after CPU topo changed */
2341                 if (!set->tags[hctx_idx] &&
2342                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2343                         /*
2344                          * If tags initialization fail for some hctx,
2345                          * that hctx won't be brought online.  In this
2346                          * case, remap the current ctx to hctx[0] which
2347                          * is guaranteed to always have tags allocated
2348                          */
2349                         q->mq_map[i] = 0;
2350                 }
2351
2352                 ctx = per_cpu_ptr(q->queue_ctx, i);
2353                 hctx = blk_mq_map_queue(q, i);
2354
2355                 cpumask_set_cpu(i, hctx->cpumask);
2356                 ctx->index_hw = hctx->nr_ctx;
2357                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2358         }
2359
2360         queue_for_each_hw_ctx(q, hctx, i) {
2361                 /*
2362                  * If no software queues are mapped to this hardware queue,
2363                  * disable it and free the request entries.
2364                  */
2365                 if (!hctx->nr_ctx) {
2366                         /* Never unmap queue 0.  We need it as a
2367                          * fallback in case of a new remap fails
2368                          * allocation
2369                          */
2370                         if (i && set->tags[i])
2371                                 blk_mq_free_map_and_requests(set, i);
2372
2373                         hctx->tags = NULL;
2374                         continue;
2375                 }
2376
2377                 hctx->tags = set->tags[i];
2378                 WARN_ON(!hctx->tags);
2379
2380                 /*
2381                  * Set the map size to the number of mapped software queues.
2382                  * This is more accurate and more efficient than looping
2383                  * over all possibly mapped software queues.
2384                  */
2385                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2386
2387                 /*
2388                  * Initialize batch roundrobin counts
2389                  */
2390                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2391                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2392         }
2393 }
2394
2395 /*
2396  * Caller needs to ensure that we're either frozen/quiesced, or that
2397  * the queue isn't live yet.
2398  */
2399 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2400 {
2401         struct blk_mq_hw_ctx *hctx;
2402         int i;
2403
2404         queue_for_each_hw_ctx(q, hctx, i) {
2405                 if (shared)
2406                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2407                 else
2408                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2409         }
2410 }
2411
2412 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2413                                         bool shared)
2414 {
2415         struct request_queue *q;
2416
2417         lockdep_assert_held(&set->tag_list_lock);
2418
2419         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2420                 blk_mq_freeze_queue(q);
2421                 queue_set_hctx_shared(q, shared);
2422                 blk_mq_unfreeze_queue(q);
2423         }
2424 }
2425
2426 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2427 {
2428         struct blk_mq_tag_set *set = q->tag_set;
2429
2430         mutex_lock(&set->tag_list_lock);
2431         list_del_rcu(&q->tag_set_list);
2432         if (list_is_singular(&set->tag_list)) {
2433                 /* just transitioned to unshared */
2434                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2435                 /* update existing queue */
2436                 blk_mq_update_tag_set_depth(set, false);
2437         }
2438         mutex_unlock(&set->tag_list_lock);
2439         INIT_LIST_HEAD(&q->tag_set_list);
2440 }
2441
2442 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2443                                      struct request_queue *q)
2444 {
2445         q->tag_set = set;
2446
2447         mutex_lock(&set->tag_list_lock);
2448
2449         /*
2450          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2451          */
2452         if (!list_empty(&set->tag_list) &&
2453             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2454                 set->flags |= BLK_MQ_F_TAG_SHARED;
2455                 /* update existing queue */
2456                 blk_mq_update_tag_set_depth(set, true);
2457         }
2458         if (set->flags & BLK_MQ_F_TAG_SHARED)
2459                 queue_set_hctx_shared(q, true);
2460         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2461
2462         mutex_unlock(&set->tag_list_lock);
2463 }
2464
2465 /*
2466  * It is the actual release handler for mq, but we do it from
2467  * request queue's release handler for avoiding use-after-free
2468  * and headache because q->mq_kobj shouldn't have been introduced,
2469  * but we can't group ctx/kctx kobj without it.
2470  */
2471 void blk_mq_release(struct request_queue *q)
2472 {
2473         struct blk_mq_hw_ctx *hctx;
2474         unsigned int i;
2475
2476         /* hctx kobj stays in hctx */
2477         queue_for_each_hw_ctx(q, hctx, i) {
2478                 if (!hctx)
2479                         continue;
2480                 kobject_put(&hctx->kobj);
2481         }
2482
2483         q->mq_map = NULL;
2484
2485         kfree(q->queue_hw_ctx);
2486
2487         /*
2488          * release .mq_kobj and sw queue's kobject now because
2489          * both share lifetime with request queue.
2490          */
2491         blk_mq_sysfs_deinit(q);
2492
2493         free_percpu(q->queue_ctx);
2494 }
2495
2496 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2497 {
2498         struct request_queue *uninit_q, *q;
2499
2500         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2501         if (!uninit_q)
2502                 return ERR_PTR(-ENOMEM);
2503
2504         q = blk_mq_init_allocated_queue(set, uninit_q);
2505         if (IS_ERR(q))
2506                 blk_cleanup_queue(uninit_q);
2507
2508         return q;
2509 }
2510 EXPORT_SYMBOL(blk_mq_init_queue);
2511
2512 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2513 {
2514         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2515
2516         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2517                            __alignof__(struct blk_mq_hw_ctx)) !=
2518                      sizeof(struct blk_mq_hw_ctx));
2519
2520         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2521                 hw_ctx_size += sizeof(struct srcu_struct);
2522
2523         return hw_ctx_size;
2524 }
2525
2526 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2527                                                 struct request_queue *q)
2528 {
2529         int i, j;
2530         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2531
2532         blk_mq_sysfs_unregister(q);
2533
2534         /* protect against switching io scheduler  */
2535         mutex_lock(&q->sysfs_lock);
2536         for (i = 0; i < set->nr_hw_queues; i++) {
2537                 int node;
2538
2539                 if (hctxs[i])
2540                         continue;
2541
2542                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2543                 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2544                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2545                                 node);
2546                 if (!hctxs[i])
2547                         break;
2548
2549                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask,
2550                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2551                                         node)) {
2552                         kfree(hctxs[i]);
2553                         hctxs[i] = NULL;
2554                         break;
2555                 }
2556
2557                 atomic_set(&hctxs[i]->nr_active, 0);
2558                 hctxs[i]->numa_node = node;
2559                 hctxs[i]->queue_num = i;
2560
2561                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2562                         free_cpumask_var(hctxs[i]->cpumask);
2563                         kfree(hctxs[i]);
2564                         hctxs[i] = NULL;
2565                         break;
2566                 }
2567                 blk_mq_hctx_kobj_init(hctxs[i]);
2568         }
2569         for (j = i; j < q->nr_hw_queues; j++) {
2570                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2571
2572                 if (hctx) {
2573                         if (hctx->tags)
2574                                 blk_mq_free_map_and_requests(set, j);
2575                         blk_mq_exit_hctx(q, set, hctx, j);
2576                         kobject_put(&hctx->kobj);
2577                         hctxs[j] = NULL;
2578
2579                 }
2580         }
2581         q->nr_hw_queues = i;
2582         mutex_unlock(&q->sysfs_lock);
2583         blk_mq_sysfs_register(q);
2584 }
2585
2586 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2587                                                   struct request_queue *q)
2588 {
2589         /* mark the queue as mq asap */
2590         q->mq_ops = set->ops;
2591
2592         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2593                                              blk_mq_poll_stats_bkt,
2594                                              BLK_MQ_POLL_STATS_BKTS, q);
2595         if (!q->poll_cb)
2596                 goto err_exit;
2597
2598         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2599         if (!q->queue_ctx)
2600                 goto err_exit;
2601
2602         /* init q->mq_kobj and sw queues' kobjects */
2603         blk_mq_sysfs_init(q);
2604
2605         q->queue_hw_ctx = kcalloc_node(nr_cpu_ids, sizeof(*(q->queue_hw_ctx)),
2606                                                 GFP_KERNEL, set->numa_node);
2607         if (!q->queue_hw_ctx)
2608                 goto err_percpu;
2609
2610         q->mq_map = set->mq_map;
2611
2612         blk_mq_realloc_hw_ctxs(set, q);
2613         if (!q->nr_hw_queues)
2614                 goto err_hctxs;
2615
2616         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2617         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2618
2619         q->nr_queues = nr_cpu_ids;
2620
2621         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2622
2623         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2624                 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2625
2626         q->sg_reserved_size = INT_MAX;
2627
2628         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2629         INIT_LIST_HEAD(&q->requeue_list);
2630         spin_lock_init(&q->requeue_lock);
2631
2632         blk_queue_make_request(q, blk_mq_make_request);
2633         if (q->mq_ops->poll)
2634                 q->poll_fn = blk_mq_poll;
2635
2636         /*
2637          * Do this after blk_queue_make_request() overrides it...
2638          */
2639         q->nr_requests = set->queue_depth;
2640
2641         /*
2642          * Default to classic polling
2643          */
2644         q->poll_nsec = -1;
2645
2646         if (set->ops->complete)
2647                 blk_queue_softirq_done(q, set->ops->complete);
2648
2649         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2650         blk_mq_add_queue_tag_set(set, q);
2651         blk_mq_map_swqueue(q);
2652
2653         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2654                 int ret;
2655
2656                 ret = elevator_init_mq(q);
2657                 if (ret)
2658                         return ERR_PTR(ret);
2659         }
2660
2661         return q;
2662
2663 err_hctxs:
2664         kfree(q->queue_hw_ctx);
2665 err_percpu:
2666         free_percpu(q->queue_ctx);
2667 err_exit:
2668         q->mq_ops = NULL;
2669         return ERR_PTR(-ENOMEM);
2670 }
2671 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2672
2673 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2674 void blk_mq_exit_queue(struct request_queue *q)
2675 {
2676         struct blk_mq_tag_set *set = q->tag_set;
2677
2678         /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
2679         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2680         /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
2681         blk_mq_del_queue_tag_set(q);
2682 }
2683
2684 /* Basically redo blk_mq_init_queue with queue frozen */
2685 static void blk_mq_queue_reinit(struct request_queue *q)
2686 {
2687         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2688
2689         blk_mq_debugfs_unregister_hctxs(q);
2690         blk_mq_sysfs_unregister(q);
2691
2692         /*
2693          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2694          * we should change hctx numa_node according to the new topology (this
2695          * involves freeing and re-allocating memory, worth doing?)
2696          */
2697         blk_mq_map_swqueue(q);
2698
2699         blk_mq_sysfs_register(q);
2700         blk_mq_debugfs_register_hctxs(q);
2701 }
2702
2703 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2704 {
2705         int i;
2706
2707         for (i = 0; i < set->nr_hw_queues; i++)
2708                 if (!__blk_mq_alloc_rq_map(set, i))
2709                         goto out_unwind;
2710
2711         return 0;
2712
2713 out_unwind:
2714         while (--i >= 0)
2715                 blk_mq_free_rq_map(set->tags[i]);
2716
2717         return -ENOMEM;
2718 }
2719
2720 /*
2721  * Allocate the request maps associated with this tag_set. Note that this
2722  * may reduce the depth asked for, if memory is tight. set->queue_depth
2723  * will be updated to reflect the allocated depth.
2724  */
2725 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2726 {
2727         unsigned int depth;
2728         int err;
2729
2730         depth = set->queue_depth;
2731         do {
2732                 err = __blk_mq_alloc_rq_maps(set);
2733                 if (!err)
2734                         break;
2735
2736                 set->queue_depth >>= 1;
2737                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2738                         err = -ENOMEM;
2739                         break;
2740                 }
2741         } while (set->queue_depth);
2742
2743         if (!set->queue_depth || err) {
2744                 pr_err("blk-mq: failed to allocate request map\n");
2745                 return -ENOMEM;
2746         }
2747
2748         if (depth != set->queue_depth)
2749                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2750                                                 depth, set->queue_depth);
2751
2752         return 0;
2753 }
2754
2755 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2756 {
2757         if (set->ops->map_queues) {
2758                 /*
2759                  * transport .map_queues is usually done in the following
2760                  * way:
2761                  *
2762                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2763                  *      mask = get_cpu_mask(queue)
2764                  *      for_each_cpu(cpu, mask)
2765                  *              set->mq_map[cpu] = queue;
2766                  * }
2767                  *
2768                  * When we need to remap, the table has to be cleared for
2769                  * killing stale mapping since one CPU may not be mapped
2770                  * to any hw queue.
2771                  */
2772                 blk_mq_clear_mq_map(set);
2773
2774                 return set->ops->map_queues(set);
2775         } else
2776                 return blk_mq_map_queues(set);
2777 }
2778
2779 /*
2780  * Alloc a tag set to be associated with one or more request queues.
2781  * May fail with EINVAL for various error conditions. May adjust the
2782  * requested depth down, if it's too large. In that case, the set
2783  * value will be stored in set->queue_depth.
2784  */
2785 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2786 {
2787         int ret;
2788
2789         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2790
2791         if (!set->nr_hw_queues)
2792                 return -EINVAL;
2793         if (!set->queue_depth)
2794                 return -EINVAL;
2795         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2796                 return -EINVAL;
2797
2798         if (!set->ops->queue_rq)
2799                 return -EINVAL;
2800
2801         if (!set->ops->get_budget ^ !set->ops->put_budget)
2802                 return -EINVAL;
2803
2804         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2805                 pr_info("blk-mq: reduced tag depth to %u\n",
2806                         BLK_MQ_MAX_DEPTH);
2807                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2808         }
2809
2810         /*
2811          * If a crashdump is active, then we are potentially in a very
2812          * memory constrained environment. Limit us to 1 queue and
2813          * 64 tags to prevent using too much memory.
2814          */
2815         if (is_kdump_kernel()) {
2816                 set->nr_hw_queues = 1;
2817                 set->queue_depth = min(64U, set->queue_depth);
2818         }
2819         /*
2820          * There is no use for more h/w queues than cpus.
2821          */
2822         if (set->nr_hw_queues > nr_cpu_ids)
2823                 set->nr_hw_queues = nr_cpu_ids;
2824
2825         set->tags = kcalloc_node(nr_cpu_ids, sizeof(struct blk_mq_tags *),
2826                                  GFP_KERNEL, set->numa_node);
2827         if (!set->tags)
2828                 return -ENOMEM;
2829
2830         ret = -ENOMEM;
2831         set->mq_map = kcalloc_node(nr_cpu_ids, sizeof(*set->mq_map),
2832                                    GFP_KERNEL, set->numa_node);
2833         if (!set->mq_map)
2834                 goto out_free_tags;
2835
2836         ret = blk_mq_update_queue_map(set);
2837         if (ret)
2838                 goto out_free_mq_map;
2839
2840         ret = blk_mq_alloc_rq_maps(set);
2841         if (ret)
2842                 goto out_free_mq_map;
2843
2844         mutex_init(&set->tag_list_lock);
2845         INIT_LIST_HEAD(&set->tag_list);
2846
2847         return 0;
2848
2849 out_free_mq_map:
2850         kfree(set->mq_map);
2851         set->mq_map = NULL;
2852 out_free_tags:
2853         kfree(set->tags);
2854         set->tags = NULL;
2855         return ret;
2856 }
2857 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2858
2859 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2860 {
2861         int i;
2862
2863         for (i = 0; i < nr_cpu_ids; i++)
2864                 blk_mq_free_map_and_requests(set, i);
2865
2866         kfree(set->mq_map);
2867         set->mq_map = NULL;
2868
2869         kfree(set->tags);
2870         set->tags = NULL;
2871 }
2872 EXPORT_SYMBOL(blk_mq_free_tag_set);
2873
2874 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2875 {
2876         struct blk_mq_tag_set *set = q->tag_set;
2877         struct blk_mq_hw_ctx *hctx;
2878         int i, ret;
2879
2880         if (!set)
2881                 return -EINVAL;
2882
2883         blk_mq_freeze_queue(q);
2884         blk_mq_quiesce_queue(q);
2885
2886         ret = 0;
2887         queue_for_each_hw_ctx(q, hctx, i) {
2888                 if (!hctx->tags)
2889                         continue;
2890                 /*
2891                  * If we're using an MQ scheduler, just update the scheduler
2892                  * queue depth. This is similar to what the old code would do.
2893                  */
2894                 if (!hctx->sched_tags) {
2895                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2896                                                         false);
2897                 } else {
2898                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2899                                                         nr, true);
2900                 }
2901                 if (ret)
2902                         break;
2903                 if (q->elevator && q->elevator->type->ops.mq.depth_updated)
2904                         q->elevator->type->ops.mq.depth_updated(hctx);
2905         }
2906
2907         if (!ret)
2908                 q->nr_requests = nr;
2909
2910         blk_mq_unquiesce_queue(q);
2911         blk_mq_unfreeze_queue(q);
2912
2913         return ret;
2914 }
2915
2916 /*
2917  * request_queue and elevator_type pair.
2918  * It is just used by __blk_mq_update_nr_hw_queues to cache
2919  * the elevator_type associated with a request_queue.
2920  */
2921 struct blk_mq_qe_pair {
2922         struct list_head node;
2923         struct request_queue *q;
2924         struct elevator_type *type;
2925 };
2926
2927 /*
2928  * Cache the elevator_type in qe pair list and switch the
2929  * io scheduler to 'none'
2930  */
2931 static bool blk_mq_elv_switch_none(struct list_head *head,
2932                 struct request_queue *q)
2933 {
2934         struct blk_mq_qe_pair *qe;
2935
2936         if (!q->elevator)
2937                 return true;
2938
2939         qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
2940         if (!qe)
2941                 return false;
2942
2943         INIT_LIST_HEAD(&qe->node);
2944         qe->q = q;
2945         qe->type = q->elevator->type;
2946         list_add(&qe->node, head);
2947
2948         mutex_lock(&q->sysfs_lock);
2949         /*
2950          * After elevator_switch_mq, the previous elevator_queue will be
2951          * released by elevator_release. The reference of the io scheduler
2952          * module get by elevator_get will also be put. So we need to get
2953          * a reference of the io scheduler module here to prevent it to be
2954          * removed.
2955          */
2956         __module_get(qe->type->elevator_owner);
2957         elevator_switch_mq(q, NULL);
2958         mutex_unlock(&q->sysfs_lock);
2959
2960         return true;
2961 }
2962
2963 static void blk_mq_elv_switch_back(struct list_head *head,
2964                 struct request_queue *q)
2965 {
2966         struct blk_mq_qe_pair *qe;
2967         struct elevator_type *t = NULL;
2968
2969         list_for_each_entry(qe, head, node)
2970                 if (qe->q == q) {
2971                         t = qe->type;
2972                         break;
2973                 }
2974
2975         if (!t)
2976                 return;
2977
2978         list_del(&qe->node);
2979         kfree(qe);
2980
2981         mutex_lock(&q->sysfs_lock);
2982         elevator_switch_mq(q, t);
2983         mutex_unlock(&q->sysfs_lock);
2984 }
2985
2986 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2987                                                         int nr_hw_queues)
2988 {
2989         struct request_queue *q;
2990         LIST_HEAD(head);
2991
2992         lockdep_assert_held(&set->tag_list_lock);
2993
2994         if (nr_hw_queues > nr_cpu_ids)
2995                 nr_hw_queues = nr_cpu_ids;
2996         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2997                 return;
2998
2999         list_for_each_entry(q, &set->tag_list, tag_set_list)
3000                 blk_mq_freeze_queue(q);
3001         /*
3002          * Sync with blk_mq_queue_tag_busy_iter.
3003          */
3004         synchronize_rcu();
3005         /*
3006          * Switch IO scheduler to 'none', cleaning up the data associated
3007          * with the previous scheduler. We will switch back once we are done
3008          * updating the new sw to hw queue mappings.
3009          */
3010         list_for_each_entry(q, &set->tag_list, tag_set_list)
3011                 if (!blk_mq_elv_switch_none(&head, q))
3012                         goto switch_back;
3013
3014         set->nr_hw_queues = nr_hw_queues;
3015         blk_mq_update_queue_map(set);
3016         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3017                 blk_mq_realloc_hw_ctxs(set, q);
3018                 blk_mq_queue_reinit(q);
3019         }
3020
3021 switch_back:
3022         list_for_each_entry(q, &set->tag_list, tag_set_list)
3023                 blk_mq_elv_switch_back(&head, q);
3024
3025         list_for_each_entry(q, &set->tag_list, tag_set_list)
3026                 blk_mq_unfreeze_queue(q);
3027 }
3028
3029 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3030 {
3031         mutex_lock(&set->tag_list_lock);
3032         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3033         mutex_unlock(&set->tag_list_lock);
3034 }
3035 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3036
3037 /* Enable polling stats and return whether they were already enabled. */
3038 static bool blk_poll_stats_enable(struct request_queue *q)
3039 {
3040         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3041             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3042                 return true;
3043         blk_stat_add_callback(q, q->poll_cb);
3044         return false;
3045 }
3046
3047 static void blk_mq_poll_stats_start(struct request_queue *q)
3048 {
3049         /*
3050          * We don't arm the callback if polling stats are not enabled or the
3051          * callback is already active.
3052          */
3053         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3054             blk_stat_is_active(q->poll_cb))
3055                 return;
3056
3057         blk_stat_activate_msecs(q->poll_cb, 100);
3058 }
3059
3060 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3061 {
3062         struct request_queue *q = cb->data;
3063         int bucket;
3064
3065         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3066                 if (cb->stat[bucket].nr_samples)
3067                         q->poll_stat[bucket] = cb->stat[bucket];
3068         }
3069 }
3070
3071 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3072                                        struct blk_mq_hw_ctx *hctx,
3073                                        struct request *rq)
3074 {
3075         unsigned long ret = 0;
3076         int bucket;
3077
3078         /*
3079          * If stats collection isn't on, don't sleep but turn it on for
3080          * future users
3081          */
3082         if (!blk_poll_stats_enable(q))
3083                 return 0;
3084
3085         /*
3086          * As an optimistic guess, use half of the mean service time
3087          * for this type of request. We can (and should) make this smarter.
3088          * For instance, if the completion latencies are tight, we can
3089          * get closer than just half the mean. This is especially
3090          * important on devices where the completion latencies are longer
3091          * than ~10 usec. We do use the stats for the relevant IO size
3092          * if available which does lead to better estimates.
3093          */
3094         bucket = blk_mq_poll_stats_bkt(rq);
3095         if (bucket < 0)
3096                 return ret;
3097
3098         if (q->poll_stat[bucket].nr_samples)
3099                 ret = (q->poll_stat[bucket].mean + 1) / 2;
3100
3101         return ret;
3102 }
3103
3104 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3105                                      struct blk_mq_hw_ctx *hctx,
3106                                      struct request *rq)
3107 {
3108         struct hrtimer_sleeper hs;
3109         enum hrtimer_mode mode;
3110         unsigned int nsecs;
3111         ktime_t kt;
3112
3113         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3114                 return false;
3115
3116         /*
3117          * poll_nsec can be:
3118          *
3119          * -1:  don't ever hybrid sleep
3120          *  0:  use half of prev avg
3121          * >0:  use this specific value
3122          */
3123         if (q->poll_nsec == -1)
3124                 return false;
3125         else if (q->poll_nsec > 0)
3126                 nsecs = q->poll_nsec;
3127         else
3128                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3129
3130         if (!nsecs)
3131                 return false;
3132
3133         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3134
3135         /*
3136          * This will be replaced with the stats tracking code, using
3137          * 'avg_completion_time / 2' as the pre-sleep target.
3138          */
3139         kt = nsecs;
3140
3141         mode = HRTIMER_MODE_REL;
3142         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3143         hrtimer_set_expires(&hs.timer, kt);
3144
3145         hrtimer_init_sleeper(&hs, current);
3146         do {
3147                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3148                         break;
3149                 set_current_state(TASK_UNINTERRUPTIBLE);
3150                 hrtimer_start_expires(&hs.timer, mode);
3151                 if (hs.task)
3152                         io_schedule();
3153                 hrtimer_cancel(&hs.timer);
3154                 mode = HRTIMER_MODE_ABS;
3155         } while (hs.task && !signal_pending(current));
3156
3157         __set_current_state(TASK_RUNNING);
3158         destroy_hrtimer_on_stack(&hs.timer);
3159         return true;
3160 }
3161
3162 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3163 {
3164         struct request_queue *q = hctx->queue;
3165         long state;
3166
3167         /*
3168          * If we sleep, have the caller restart the poll loop to reset
3169          * the state. Like for the other success return cases, the
3170          * caller is responsible for checking if the IO completed. If
3171          * the IO isn't complete, we'll get called again and will go
3172          * straight to the busy poll loop.
3173          */
3174         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3175                 return true;
3176
3177         hctx->poll_considered++;
3178
3179         state = current->state;
3180         while (!need_resched()) {
3181                 int ret;
3182
3183                 hctx->poll_invoked++;
3184
3185                 ret = q->mq_ops->poll(hctx, rq->tag);
3186                 if (ret > 0) {
3187                         hctx->poll_success++;
3188                         set_current_state(TASK_RUNNING);
3189                         return true;
3190                 }
3191
3192                 if (signal_pending_state(state, current))
3193                         set_current_state(TASK_RUNNING);
3194
3195                 if (current->state == TASK_RUNNING)
3196                         return true;
3197                 if (ret < 0)
3198                         break;
3199                 cpu_relax();
3200         }
3201
3202         __set_current_state(TASK_RUNNING);
3203         return false;
3204 }
3205
3206 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3207 {
3208         struct blk_mq_hw_ctx *hctx;
3209         struct request *rq;
3210
3211         if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3212                 return false;
3213
3214         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3215         if (!blk_qc_t_is_internal(cookie))
3216                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3217         else {
3218                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3219                 /*
3220                  * With scheduling, if the request has completed, we'll
3221                  * get a NULL return here, as we clear the sched tag when
3222                  * that happens. The request still remains valid, like always,
3223                  * so we should be safe with just the NULL check.
3224                  */
3225                 if (!rq)
3226                         return false;
3227         }
3228
3229         return __blk_mq_poll(hctx, rq);
3230 }
3231
3232 static int __init blk_mq_init(void)
3233 {
3234         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3235                                 blk_mq_hctx_notify_dead);
3236         return 0;
3237 }
3238 subsys_initcall(blk_mq_init);