block/DAC960.c: make some arrays static const, shrinks object size
[sfrench/cifs-2.6.git] / block / blk-mq.c
1 /*
2  * Block multiqueue core code
3  *
4  * Copyright (C) 2013-2014 Jens Axboe
5  * Copyright (C) 2013-2014 Christoph Hellwig
6  */
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
13 #include <linux/mm.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
28
29 #include <trace/events/block.h>
30
31 #include <linux/blk-mq.h>
32 #include "blk.h"
33 #include "blk-mq.h"
34 #include "blk-mq-debugfs.h"
35 #include "blk-mq-tag.h"
36 #include "blk-stat.h"
37 #include "blk-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 (cmpxchg(&rq->state, MQ_RQ_IN_FLIGHT, MQ_RQ_COMPLETE) !=
564                         MQ_RQ_IN_FLIGHT)
565                 return;
566
567         if (rq->internal_tag != -1)
568                 blk_mq_sched_completed_request(rq);
569
570         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
571                 rq->q->softirq_done_fn(rq);
572                 return;
573         }
574
575         cpu = get_cpu();
576         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
577                 shared = cpus_share_cache(cpu, ctx->cpu);
578
579         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
580                 rq->csd.func = __blk_mq_complete_request_remote;
581                 rq->csd.info = rq;
582                 rq->csd.flags = 0;
583                 smp_call_function_single_async(ctx->cpu, &rq->csd);
584         } else {
585                 rq->q->softirq_done_fn(rq);
586         }
587         put_cpu();
588 }
589
590 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
591         __releases(hctx->srcu)
592 {
593         if (!(hctx->flags & BLK_MQ_F_BLOCKING))
594                 rcu_read_unlock();
595         else
596                 srcu_read_unlock(hctx->srcu, srcu_idx);
597 }
598
599 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
600         __acquires(hctx->srcu)
601 {
602         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
603                 /* shut up gcc false positive */
604                 *srcu_idx = 0;
605                 rcu_read_lock();
606         } else
607                 *srcu_idx = srcu_read_lock(hctx->srcu);
608 }
609
610 /**
611  * blk_mq_complete_request - end I/O on a request
612  * @rq:         the request being processed
613  *
614  * Description:
615  *      Ends all I/O on a request. It does not handle partial completions.
616  *      The actual completion happens out-of-order, through a IPI handler.
617  **/
618 void blk_mq_complete_request(struct request *rq)
619 {
620         if (unlikely(blk_should_fake_timeout(rq->q)))
621                 return;
622         __blk_mq_complete_request(rq);
623 }
624 EXPORT_SYMBOL(blk_mq_complete_request);
625
626 int blk_mq_request_started(struct request *rq)
627 {
628         return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
629 }
630 EXPORT_SYMBOL_GPL(blk_mq_request_started);
631
632 void blk_mq_start_request(struct request *rq)
633 {
634         struct request_queue *q = rq->q;
635
636         blk_mq_sched_started_request(rq);
637
638         trace_block_rq_issue(q, rq);
639
640         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
641                 rq->io_start_time_ns = ktime_get_ns();
642 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
643                 rq->throtl_size = blk_rq_sectors(rq);
644 #endif
645                 rq->rq_flags |= RQF_STATS;
646                 rq_qos_issue(q, rq);
647         }
648
649         WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
650
651         blk_add_timer(rq);
652         WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
653
654         if (q->dma_drain_size && blk_rq_bytes(rq)) {
655                 /*
656                  * Make sure space for the drain appears.  We know we can do
657                  * this because max_hw_segments has been adjusted to be one
658                  * fewer than the device can handle.
659                  */
660                 rq->nr_phys_segments++;
661         }
662 }
663 EXPORT_SYMBOL(blk_mq_start_request);
664
665 static void __blk_mq_requeue_request(struct request *rq)
666 {
667         struct request_queue *q = rq->q;
668
669         blk_mq_put_driver_tag(rq);
670
671         trace_block_rq_requeue(q, rq);
672         rq_qos_requeue(q, rq);
673
674         if (blk_mq_request_started(rq)) {
675                 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
676                 rq->rq_flags &= ~RQF_TIMED_OUT;
677                 if (q->dma_drain_size && blk_rq_bytes(rq))
678                         rq->nr_phys_segments--;
679         }
680 }
681
682 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
683 {
684         __blk_mq_requeue_request(rq);
685
686         /* this request will be re-inserted to io scheduler queue */
687         blk_mq_sched_requeue_request(rq);
688
689         BUG_ON(blk_queued_rq(rq));
690         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
691 }
692 EXPORT_SYMBOL(blk_mq_requeue_request);
693
694 static void blk_mq_requeue_work(struct work_struct *work)
695 {
696         struct request_queue *q =
697                 container_of(work, struct request_queue, requeue_work.work);
698         LIST_HEAD(rq_list);
699         struct request *rq, *next;
700
701         spin_lock_irq(&q->requeue_lock);
702         list_splice_init(&q->requeue_list, &rq_list);
703         spin_unlock_irq(&q->requeue_lock);
704
705         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
706                 if (!(rq->rq_flags & RQF_SOFTBARRIER))
707                         continue;
708
709                 rq->rq_flags &= ~RQF_SOFTBARRIER;
710                 list_del_init(&rq->queuelist);
711                 blk_mq_sched_insert_request(rq, true, false, false);
712         }
713
714         while (!list_empty(&rq_list)) {
715                 rq = list_entry(rq_list.next, struct request, queuelist);
716                 list_del_init(&rq->queuelist);
717                 blk_mq_sched_insert_request(rq, false, false, false);
718         }
719
720         blk_mq_run_hw_queues(q, false);
721 }
722
723 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
724                                 bool kick_requeue_list)
725 {
726         struct request_queue *q = rq->q;
727         unsigned long flags;
728
729         /*
730          * We abuse this flag that is otherwise used by the I/O scheduler to
731          * request head insertion from the workqueue.
732          */
733         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
734
735         spin_lock_irqsave(&q->requeue_lock, flags);
736         if (at_head) {
737                 rq->rq_flags |= RQF_SOFTBARRIER;
738                 list_add(&rq->queuelist, &q->requeue_list);
739         } else {
740                 list_add_tail(&rq->queuelist, &q->requeue_list);
741         }
742         spin_unlock_irqrestore(&q->requeue_lock, flags);
743
744         if (kick_requeue_list)
745                 blk_mq_kick_requeue_list(q);
746 }
747 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
748
749 void blk_mq_kick_requeue_list(struct request_queue *q)
750 {
751         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
752 }
753 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
754
755 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
756                                     unsigned long msecs)
757 {
758         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
759                                     msecs_to_jiffies(msecs));
760 }
761 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
762
763 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
764 {
765         if (tag < tags->nr_tags) {
766                 prefetch(tags->rqs[tag]);
767                 return tags->rqs[tag];
768         }
769
770         return NULL;
771 }
772 EXPORT_SYMBOL(blk_mq_tag_to_rq);
773
774 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
775 {
776         req->rq_flags |= RQF_TIMED_OUT;
777         if (req->q->mq_ops->timeout) {
778                 enum blk_eh_timer_return ret;
779
780                 ret = req->q->mq_ops->timeout(req, reserved);
781                 if (ret == BLK_EH_DONE)
782                         return;
783                 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
784         }
785
786         blk_add_timer(req);
787 }
788
789 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
790 {
791         unsigned long deadline;
792
793         if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
794                 return false;
795         if (rq->rq_flags & RQF_TIMED_OUT)
796                 return false;
797
798         deadline = blk_rq_deadline(rq);
799         if (time_after_eq(jiffies, deadline))
800                 return true;
801
802         if (*next == 0)
803                 *next = deadline;
804         else if (time_after(*next, deadline))
805                 *next = deadline;
806         return false;
807 }
808
809 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
810                 struct request *rq, void *priv, bool reserved)
811 {
812         unsigned long *next = priv;
813
814         /*
815          * Just do a quick check if it is expired before locking the request in
816          * so we're not unnecessarilly synchronizing across CPUs.
817          */
818         if (!blk_mq_req_expired(rq, next))
819                 return;
820
821         /*
822          * We have reason to believe the request may be expired. Take a
823          * reference on the request to lock this request lifetime into its
824          * currently allocated context to prevent it from being reallocated in
825          * the event the completion by-passes this timeout handler.
826          *
827          * If the reference was already released, then the driver beat the
828          * timeout handler to posting a natural completion.
829          */
830         if (!refcount_inc_not_zero(&rq->ref))
831                 return;
832
833         /*
834          * The request is now locked and cannot be reallocated underneath the
835          * timeout handler's processing. Re-verify this exact request is truly
836          * expired; if it is not expired, then the request was completed and
837          * reallocated as a new request.
838          */
839         if (blk_mq_req_expired(rq, next))
840                 blk_mq_rq_timed_out(rq, reserved);
841         if (refcount_dec_and_test(&rq->ref))
842                 __blk_mq_free_request(rq);
843 }
844
845 static void blk_mq_timeout_work(struct work_struct *work)
846 {
847         struct request_queue *q =
848                 container_of(work, struct request_queue, timeout_work);
849         unsigned long next = 0;
850         struct blk_mq_hw_ctx *hctx;
851         int i;
852
853         /* A deadlock might occur if a request is stuck requiring a
854          * timeout at the same time a queue freeze is waiting
855          * completion, since the timeout code would not be able to
856          * acquire the queue reference here.
857          *
858          * That's why we don't use blk_queue_enter here; instead, we use
859          * percpu_ref_tryget directly, because we need to be able to
860          * obtain a reference even in the short window between the queue
861          * starting to freeze, by dropping the first reference in
862          * blk_freeze_queue_start, and the moment the last request is
863          * consumed, marked by the instant q_usage_counter reaches
864          * zero.
865          */
866         if (!percpu_ref_tryget(&q->q_usage_counter))
867                 return;
868
869         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
870
871         if (next != 0) {
872                 mod_timer(&q->timeout, next);
873         } else {
874                 /*
875                  * Request timeouts are handled as a forward rolling timer. If
876                  * we end up here it means that no requests are pending and
877                  * also that no request has been pending for a while. Mark
878                  * each hctx as idle.
879                  */
880                 queue_for_each_hw_ctx(q, hctx, i) {
881                         /* the hctx may be unmapped, so check it here */
882                         if (blk_mq_hw_queue_mapped(hctx))
883                                 blk_mq_tag_idle(hctx);
884                 }
885         }
886         blk_queue_exit(q);
887 }
888
889 struct flush_busy_ctx_data {
890         struct blk_mq_hw_ctx *hctx;
891         struct list_head *list;
892 };
893
894 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
895 {
896         struct flush_busy_ctx_data *flush_data = data;
897         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
898         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
899
900         spin_lock(&ctx->lock);
901         list_splice_tail_init(&ctx->rq_list, flush_data->list);
902         sbitmap_clear_bit(sb, bitnr);
903         spin_unlock(&ctx->lock);
904         return true;
905 }
906
907 /*
908  * Process software queues that have been marked busy, splicing them
909  * to the for-dispatch
910  */
911 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
912 {
913         struct flush_busy_ctx_data data = {
914                 .hctx = hctx,
915                 .list = list,
916         };
917
918         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
919 }
920 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
921
922 struct dispatch_rq_data {
923         struct blk_mq_hw_ctx *hctx;
924         struct request *rq;
925 };
926
927 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
928                 void *data)
929 {
930         struct dispatch_rq_data *dispatch_data = data;
931         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
932         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
933
934         spin_lock(&ctx->lock);
935         if (!list_empty(&ctx->rq_list)) {
936                 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
937                 list_del_init(&dispatch_data->rq->queuelist);
938                 if (list_empty(&ctx->rq_list))
939                         sbitmap_clear_bit(sb, bitnr);
940         }
941         spin_unlock(&ctx->lock);
942
943         return !dispatch_data->rq;
944 }
945
946 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
947                                         struct blk_mq_ctx *start)
948 {
949         unsigned off = start ? start->index_hw : 0;
950         struct dispatch_rq_data data = {
951                 .hctx = hctx,
952                 .rq   = NULL,
953         };
954
955         __sbitmap_for_each_set(&hctx->ctx_map, off,
956                                dispatch_rq_from_ctx, &data);
957
958         return data.rq;
959 }
960
961 static inline unsigned int queued_to_index(unsigned int queued)
962 {
963         if (!queued)
964                 return 0;
965
966         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
967 }
968
969 bool blk_mq_get_driver_tag(struct request *rq)
970 {
971         struct blk_mq_alloc_data data = {
972                 .q = rq->q,
973                 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
974                 .flags = BLK_MQ_REQ_NOWAIT,
975         };
976         bool shared;
977
978         if (rq->tag != -1)
979                 goto done;
980
981         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
982                 data.flags |= BLK_MQ_REQ_RESERVED;
983
984         shared = blk_mq_tag_busy(data.hctx);
985         rq->tag = blk_mq_get_tag(&data);
986         if (rq->tag >= 0) {
987                 if (shared) {
988                         rq->rq_flags |= RQF_MQ_INFLIGHT;
989                         atomic_inc(&data.hctx->nr_active);
990                 }
991                 data.hctx->tags->rqs[rq->tag] = rq;
992         }
993
994 done:
995         return rq->tag != -1;
996 }
997
998 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
999                                 int flags, void *key)
1000 {
1001         struct blk_mq_hw_ctx *hctx;
1002
1003         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1004
1005         spin_lock(&hctx->dispatch_wait_lock);
1006         list_del_init(&wait->entry);
1007         spin_unlock(&hctx->dispatch_wait_lock);
1008
1009         blk_mq_run_hw_queue(hctx, true);
1010         return 1;
1011 }
1012
1013 /*
1014  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1015  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1016  * restart. For both cases, take care to check the condition again after
1017  * marking us as waiting.
1018  */
1019 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1020                                  struct request *rq)
1021 {
1022         struct wait_queue_head *wq;
1023         wait_queue_entry_t *wait;
1024         bool ret;
1025
1026         if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1027                 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
1028                         set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
1029
1030                 /*
1031                  * It's possible that a tag was freed in the window between the
1032                  * allocation failure and adding the hardware queue to the wait
1033                  * queue.
1034                  *
1035                  * Don't clear RESTART here, someone else could have set it.
1036                  * At most this will cost an extra queue run.
1037                  */
1038                 return blk_mq_get_driver_tag(rq);
1039         }
1040
1041         wait = &hctx->dispatch_wait;
1042         if (!list_empty_careful(&wait->entry))
1043                 return false;
1044
1045         wq = &bt_wait_ptr(&hctx->tags->bitmap_tags, hctx)->wait;
1046
1047         spin_lock_irq(&wq->lock);
1048         spin_lock(&hctx->dispatch_wait_lock);
1049         if (!list_empty(&wait->entry)) {
1050                 spin_unlock(&hctx->dispatch_wait_lock);
1051                 spin_unlock_irq(&wq->lock);
1052                 return false;
1053         }
1054
1055         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1056         __add_wait_queue(wq, wait);
1057
1058         /*
1059          * It's possible that a tag was freed in the window between the
1060          * allocation failure and adding the hardware queue to the wait
1061          * queue.
1062          */
1063         ret = blk_mq_get_driver_tag(rq);
1064         if (!ret) {
1065                 spin_unlock(&hctx->dispatch_wait_lock);
1066                 spin_unlock_irq(&wq->lock);
1067                 return false;
1068         }
1069
1070         /*
1071          * We got a tag, remove ourselves from the wait queue to ensure
1072          * someone else gets the wakeup.
1073          */
1074         list_del_init(&wait->entry);
1075         spin_unlock(&hctx->dispatch_wait_lock);
1076         spin_unlock_irq(&wq->lock);
1077
1078         return true;
1079 }
1080
1081 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  8
1082 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  4
1083 /*
1084  * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1085  * - EWMA is one simple way to compute running average value
1086  * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1087  * - take 4 as factor for avoiding to get too small(0) result, and this
1088  *   factor doesn't matter because EWMA decreases exponentially
1089  */
1090 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1091 {
1092         unsigned int ewma;
1093
1094         if (hctx->queue->elevator)
1095                 return;
1096
1097         ewma = hctx->dispatch_busy;
1098
1099         if (!ewma && !busy)
1100                 return;
1101
1102         ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1103         if (busy)
1104                 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1105         ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1106
1107         hctx->dispatch_busy = ewma;
1108 }
1109
1110 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1111
1112 /*
1113  * Returns true if we did some work AND can potentially do more.
1114  */
1115 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1116                              bool got_budget)
1117 {
1118         struct blk_mq_hw_ctx *hctx;
1119         struct request *rq, *nxt;
1120         bool no_tag = false;
1121         int errors, queued;
1122         blk_status_t ret = BLK_STS_OK;
1123
1124         if (list_empty(list))
1125                 return false;
1126
1127         WARN_ON(!list_is_singular(list) && got_budget);
1128
1129         /*
1130          * Now process all the entries, sending them to the driver.
1131          */
1132         errors = queued = 0;
1133         do {
1134                 struct blk_mq_queue_data bd;
1135
1136                 rq = list_first_entry(list, struct request, queuelist);
1137
1138                 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1139                 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1140                         break;
1141
1142                 if (!blk_mq_get_driver_tag(rq)) {
1143                         /*
1144                          * The initial allocation attempt failed, so we need to
1145                          * rerun the hardware queue when a tag is freed. The
1146                          * waitqueue takes care of that. If the queue is run
1147                          * before we add this entry back on the dispatch list,
1148                          * we'll re-run it below.
1149                          */
1150                         if (!blk_mq_mark_tag_wait(hctx, rq)) {
1151                                 blk_mq_put_dispatch_budget(hctx);
1152                                 /*
1153                                  * For non-shared tags, the RESTART check
1154                                  * will suffice.
1155                                  */
1156                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1157                                         no_tag = true;
1158                                 break;
1159                         }
1160                 }
1161
1162                 list_del_init(&rq->queuelist);
1163
1164                 bd.rq = rq;
1165
1166                 /*
1167                  * Flag last if we have no more requests, or if we have more
1168                  * but can't assign a driver tag to it.
1169                  */
1170                 if (list_empty(list))
1171                         bd.last = true;
1172                 else {
1173                         nxt = list_first_entry(list, struct request, queuelist);
1174                         bd.last = !blk_mq_get_driver_tag(nxt);
1175                 }
1176
1177                 ret = q->mq_ops->queue_rq(hctx, &bd);
1178                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1179                         /*
1180                          * If an I/O scheduler has been configured and we got a
1181                          * driver tag for the next request already, free it
1182                          * again.
1183                          */
1184                         if (!list_empty(list)) {
1185                                 nxt = list_first_entry(list, struct request, queuelist);
1186                                 blk_mq_put_driver_tag(nxt);
1187                         }
1188                         list_add(&rq->queuelist, list);
1189                         __blk_mq_requeue_request(rq);
1190                         break;
1191                 }
1192
1193                 if (unlikely(ret != BLK_STS_OK)) {
1194                         errors++;
1195                         blk_mq_end_request(rq, BLK_STS_IOERR);
1196                         continue;
1197                 }
1198
1199                 queued++;
1200         } while (!list_empty(list));
1201
1202         hctx->dispatched[queued_to_index(queued)]++;
1203
1204         /*
1205          * Any items that need requeuing? Stuff them into hctx->dispatch,
1206          * that is where we will continue on next queue run.
1207          */
1208         if (!list_empty(list)) {
1209                 bool needs_restart;
1210
1211                 spin_lock(&hctx->lock);
1212                 list_splice_init(list, &hctx->dispatch);
1213                 spin_unlock(&hctx->lock);
1214
1215                 /*
1216                  * If SCHED_RESTART was set by the caller of this function and
1217                  * it is no longer set that means that it was cleared by another
1218                  * thread and hence that a queue rerun is needed.
1219                  *
1220                  * If 'no_tag' is set, that means that we failed getting
1221                  * a driver tag with an I/O scheduler attached. If our dispatch
1222                  * waitqueue is no longer active, ensure that we run the queue
1223                  * AFTER adding our entries back to the list.
1224                  *
1225                  * If no I/O scheduler has been configured it is possible that
1226                  * the hardware queue got stopped and restarted before requests
1227                  * were pushed back onto the dispatch list. Rerun the queue to
1228                  * avoid starvation. Notes:
1229                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1230                  *   been stopped before rerunning a queue.
1231                  * - Some but not all block drivers stop a queue before
1232                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1233                  *   and dm-rq.
1234                  *
1235                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1236                  * bit is set, run queue after a delay to avoid IO stalls
1237                  * that could otherwise occur if the queue is idle.
1238                  */
1239                 needs_restart = blk_mq_sched_needs_restart(hctx);
1240                 if (!needs_restart ||
1241                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1242                         blk_mq_run_hw_queue(hctx, true);
1243                 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1244                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1245
1246                 blk_mq_update_dispatch_busy(hctx, true);
1247                 return false;
1248         } else
1249                 blk_mq_update_dispatch_busy(hctx, false);
1250
1251         /*
1252          * If the host/device is unable to accept more work, inform the
1253          * caller of that.
1254          */
1255         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1256                 return false;
1257
1258         return (queued + errors) != 0;
1259 }
1260
1261 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1262 {
1263         int srcu_idx;
1264
1265         /*
1266          * We should be running this queue from one of the CPUs that
1267          * are mapped to it.
1268          *
1269          * There are at least two related races now between setting
1270          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1271          * __blk_mq_run_hw_queue():
1272          *
1273          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1274          *   but later it becomes online, then this warning is harmless
1275          *   at all
1276          *
1277          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1278          *   but later it becomes offline, then the warning can't be
1279          *   triggered, and we depend on blk-mq timeout handler to
1280          *   handle dispatched requests to this hctx
1281          */
1282         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1283                 cpu_online(hctx->next_cpu)) {
1284                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1285                         raw_smp_processor_id(),
1286                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1287                 dump_stack();
1288         }
1289
1290         /*
1291          * We can't run the queue inline with ints disabled. Ensure that
1292          * we catch bad users of this early.
1293          */
1294         WARN_ON_ONCE(in_interrupt());
1295
1296         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1297
1298         hctx_lock(hctx, &srcu_idx);
1299         blk_mq_sched_dispatch_requests(hctx);
1300         hctx_unlock(hctx, srcu_idx);
1301 }
1302
1303 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1304 {
1305         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1306
1307         if (cpu >= nr_cpu_ids)
1308                 cpu = cpumask_first(hctx->cpumask);
1309         return cpu;
1310 }
1311
1312 /*
1313  * It'd be great if the workqueue API had a way to pass
1314  * in a mask and had some smarts for more clever placement.
1315  * For now we just round-robin here, switching for every
1316  * BLK_MQ_CPU_WORK_BATCH queued items.
1317  */
1318 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1319 {
1320         bool tried = false;
1321         int next_cpu = hctx->next_cpu;
1322
1323         if (hctx->queue->nr_hw_queues == 1)
1324                 return WORK_CPU_UNBOUND;
1325
1326         if (--hctx->next_cpu_batch <= 0) {
1327 select_cpu:
1328                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1329                                 cpu_online_mask);
1330                 if (next_cpu >= nr_cpu_ids)
1331                         next_cpu = blk_mq_first_mapped_cpu(hctx);
1332                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1333         }
1334
1335         /*
1336          * Do unbound schedule if we can't find a online CPU for this hctx,
1337          * and it should only happen in the path of handling CPU DEAD.
1338          */
1339         if (!cpu_online(next_cpu)) {
1340                 if (!tried) {
1341                         tried = true;
1342                         goto select_cpu;
1343                 }
1344
1345                 /*
1346                  * Make sure to re-select CPU next time once after CPUs
1347                  * in hctx->cpumask become online again.
1348                  */
1349                 hctx->next_cpu = next_cpu;
1350                 hctx->next_cpu_batch = 1;
1351                 return WORK_CPU_UNBOUND;
1352         }
1353
1354         hctx->next_cpu = next_cpu;
1355         return next_cpu;
1356 }
1357
1358 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1359                                         unsigned long msecs)
1360 {
1361         if (unlikely(blk_mq_hctx_stopped(hctx)))
1362                 return;
1363
1364         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1365                 int cpu = get_cpu();
1366                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1367                         __blk_mq_run_hw_queue(hctx);
1368                         put_cpu();
1369                         return;
1370                 }
1371
1372                 put_cpu();
1373         }
1374
1375         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1376                                     msecs_to_jiffies(msecs));
1377 }
1378
1379 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1380 {
1381         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1382 }
1383 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1384
1385 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1386 {
1387         int srcu_idx;
1388         bool need_run;
1389
1390         /*
1391          * When queue is quiesced, we may be switching io scheduler, or
1392          * updating nr_hw_queues, or other things, and we can't run queue
1393          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1394          *
1395          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1396          * quiesced.
1397          */
1398         hctx_lock(hctx, &srcu_idx);
1399         need_run = !blk_queue_quiesced(hctx->queue) &&
1400                 blk_mq_hctx_has_pending(hctx);
1401         hctx_unlock(hctx, srcu_idx);
1402
1403         if (need_run) {
1404                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1405                 return true;
1406         }
1407
1408         return false;
1409 }
1410 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1411
1412 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1413 {
1414         struct blk_mq_hw_ctx *hctx;
1415         int i;
1416
1417         queue_for_each_hw_ctx(q, hctx, i) {
1418                 if (blk_mq_hctx_stopped(hctx))
1419                         continue;
1420
1421                 blk_mq_run_hw_queue(hctx, async);
1422         }
1423 }
1424 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1425
1426 /**
1427  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1428  * @q: request queue.
1429  *
1430  * The caller is responsible for serializing this function against
1431  * blk_mq_{start,stop}_hw_queue().
1432  */
1433 bool blk_mq_queue_stopped(struct request_queue *q)
1434 {
1435         struct blk_mq_hw_ctx *hctx;
1436         int i;
1437
1438         queue_for_each_hw_ctx(q, hctx, i)
1439                 if (blk_mq_hctx_stopped(hctx))
1440                         return true;
1441
1442         return false;
1443 }
1444 EXPORT_SYMBOL(blk_mq_queue_stopped);
1445
1446 /*
1447  * This function is often used for pausing .queue_rq() by driver when
1448  * there isn't enough resource or some conditions aren't satisfied, and
1449  * BLK_STS_RESOURCE is usually returned.
1450  *
1451  * We do not guarantee that dispatch can be drained or blocked
1452  * after blk_mq_stop_hw_queue() returns. Please use
1453  * blk_mq_quiesce_queue() for that requirement.
1454  */
1455 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1456 {
1457         cancel_delayed_work(&hctx->run_work);
1458
1459         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1460 }
1461 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1462
1463 /*
1464  * This function is often used for pausing .queue_rq() by driver when
1465  * there isn't enough resource or some conditions aren't satisfied, and
1466  * BLK_STS_RESOURCE is usually returned.
1467  *
1468  * We do not guarantee that dispatch can be drained or blocked
1469  * after blk_mq_stop_hw_queues() returns. Please use
1470  * blk_mq_quiesce_queue() for that requirement.
1471  */
1472 void blk_mq_stop_hw_queues(struct request_queue *q)
1473 {
1474         struct blk_mq_hw_ctx *hctx;
1475         int i;
1476
1477         queue_for_each_hw_ctx(q, hctx, i)
1478                 blk_mq_stop_hw_queue(hctx);
1479 }
1480 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1481
1482 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1483 {
1484         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1485
1486         blk_mq_run_hw_queue(hctx, false);
1487 }
1488 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1489
1490 void blk_mq_start_hw_queues(struct request_queue *q)
1491 {
1492         struct blk_mq_hw_ctx *hctx;
1493         int i;
1494
1495         queue_for_each_hw_ctx(q, hctx, i)
1496                 blk_mq_start_hw_queue(hctx);
1497 }
1498 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1499
1500 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1501 {
1502         if (!blk_mq_hctx_stopped(hctx))
1503                 return;
1504
1505         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1506         blk_mq_run_hw_queue(hctx, async);
1507 }
1508 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1509
1510 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1511 {
1512         struct blk_mq_hw_ctx *hctx;
1513         int i;
1514
1515         queue_for_each_hw_ctx(q, hctx, i)
1516                 blk_mq_start_stopped_hw_queue(hctx, async);
1517 }
1518 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1519
1520 static void blk_mq_run_work_fn(struct work_struct *work)
1521 {
1522         struct blk_mq_hw_ctx *hctx;
1523
1524         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1525
1526         /*
1527          * If we are stopped, don't run the queue.
1528          */
1529         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1530                 return;
1531
1532         __blk_mq_run_hw_queue(hctx);
1533 }
1534
1535 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1536                                             struct request *rq,
1537                                             bool at_head)
1538 {
1539         struct blk_mq_ctx *ctx = rq->mq_ctx;
1540
1541         lockdep_assert_held(&ctx->lock);
1542
1543         trace_block_rq_insert(hctx->queue, rq);
1544
1545         if (at_head)
1546                 list_add(&rq->queuelist, &ctx->rq_list);
1547         else
1548                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1549 }
1550
1551 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1552                              bool at_head)
1553 {
1554         struct blk_mq_ctx *ctx = rq->mq_ctx;
1555
1556         lockdep_assert_held(&ctx->lock);
1557
1558         __blk_mq_insert_req_list(hctx, rq, at_head);
1559         blk_mq_hctx_mark_pending(hctx, ctx);
1560 }
1561
1562 /*
1563  * Should only be used carefully, when the caller knows we want to
1564  * bypass a potential IO scheduler on the target device.
1565  */
1566 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1567 {
1568         struct blk_mq_ctx *ctx = rq->mq_ctx;
1569         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1570
1571         spin_lock(&hctx->lock);
1572         list_add_tail(&rq->queuelist, &hctx->dispatch);
1573         spin_unlock(&hctx->lock);
1574
1575         if (run_queue)
1576                 blk_mq_run_hw_queue(hctx, false);
1577 }
1578
1579 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1580                             struct list_head *list)
1581
1582 {
1583         struct request *rq;
1584
1585         /*
1586          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1587          * offline now
1588          */
1589         list_for_each_entry(rq, list, queuelist) {
1590                 BUG_ON(rq->mq_ctx != ctx);
1591                 trace_block_rq_insert(hctx->queue, rq);
1592         }
1593
1594         spin_lock(&ctx->lock);
1595         list_splice_tail_init(list, &ctx->rq_list);
1596         blk_mq_hctx_mark_pending(hctx, ctx);
1597         spin_unlock(&ctx->lock);
1598 }
1599
1600 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1601 {
1602         struct request *rqa = container_of(a, struct request, queuelist);
1603         struct request *rqb = container_of(b, struct request, queuelist);
1604
1605         return !(rqa->mq_ctx < rqb->mq_ctx ||
1606                  (rqa->mq_ctx == rqb->mq_ctx &&
1607                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1608 }
1609
1610 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1611 {
1612         struct blk_mq_ctx *this_ctx;
1613         struct request_queue *this_q;
1614         struct request *rq;
1615         LIST_HEAD(list);
1616         LIST_HEAD(ctx_list);
1617         unsigned int depth;
1618
1619         list_splice_init(&plug->mq_list, &list);
1620
1621         list_sort(NULL, &list, plug_ctx_cmp);
1622
1623         this_q = NULL;
1624         this_ctx = NULL;
1625         depth = 0;
1626
1627         while (!list_empty(&list)) {
1628                 rq = list_entry_rq(list.next);
1629                 list_del_init(&rq->queuelist);
1630                 BUG_ON(!rq->q);
1631                 if (rq->mq_ctx != this_ctx) {
1632                         if (this_ctx) {
1633                                 trace_block_unplug(this_q, depth, from_schedule);
1634                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1635                                                                 &ctx_list,
1636                                                                 from_schedule);
1637                         }
1638
1639                         this_ctx = rq->mq_ctx;
1640                         this_q = rq->q;
1641                         depth = 0;
1642                 }
1643
1644                 depth++;
1645                 list_add_tail(&rq->queuelist, &ctx_list);
1646         }
1647
1648         /*
1649          * If 'this_ctx' is set, we know we have entries to complete
1650          * on 'ctx_list'. Do those.
1651          */
1652         if (this_ctx) {
1653                 trace_block_unplug(this_q, depth, from_schedule);
1654                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1655                                                 from_schedule);
1656         }
1657 }
1658
1659 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1660 {
1661         blk_init_request_from_bio(rq, bio);
1662
1663         blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1664
1665         blk_account_io_start(rq, true);
1666 }
1667
1668 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1669 {
1670         if (rq->tag != -1)
1671                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1672
1673         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1674 }
1675
1676 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1677                                             struct request *rq,
1678                                             blk_qc_t *cookie)
1679 {
1680         struct request_queue *q = rq->q;
1681         struct blk_mq_queue_data bd = {
1682                 .rq = rq,
1683                 .last = true,
1684         };
1685         blk_qc_t new_cookie;
1686         blk_status_t ret;
1687
1688         new_cookie = request_to_qc_t(hctx, rq);
1689
1690         /*
1691          * For OK queue, we are done. For error, caller may kill it.
1692          * Any other error (busy), just add it to our list as we
1693          * previously would have done.
1694          */
1695         ret = q->mq_ops->queue_rq(hctx, &bd);
1696         switch (ret) {
1697         case BLK_STS_OK:
1698                 blk_mq_update_dispatch_busy(hctx, false);
1699                 *cookie = new_cookie;
1700                 break;
1701         case BLK_STS_RESOURCE:
1702         case BLK_STS_DEV_RESOURCE:
1703                 blk_mq_update_dispatch_busy(hctx, true);
1704                 __blk_mq_requeue_request(rq);
1705                 break;
1706         default:
1707                 blk_mq_update_dispatch_busy(hctx, false);
1708                 *cookie = BLK_QC_T_NONE;
1709                 break;
1710         }
1711
1712         return ret;
1713 }
1714
1715 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1716                                                 struct request *rq,
1717                                                 blk_qc_t *cookie,
1718                                                 bool bypass_insert)
1719 {
1720         struct request_queue *q = rq->q;
1721         bool run_queue = true;
1722
1723         /*
1724          * RCU or SRCU read lock is needed before checking quiesced flag.
1725          *
1726          * When queue is stopped or quiesced, ignore 'bypass_insert' from
1727          * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1728          * and avoid driver to try to dispatch again.
1729          */
1730         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1731                 run_queue = false;
1732                 bypass_insert = false;
1733                 goto insert;
1734         }
1735
1736         if (q->elevator && !bypass_insert)
1737                 goto insert;
1738
1739         if (!blk_mq_get_dispatch_budget(hctx))
1740                 goto insert;
1741
1742         if (!blk_mq_get_driver_tag(rq)) {
1743                 blk_mq_put_dispatch_budget(hctx);
1744                 goto insert;
1745         }
1746
1747         return __blk_mq_issue_directly(hctx, rq, cookie);
1748 insert:
1749         if (bypass_insert)
1750                 return BLK_STS_RESOURCE;
1751
1752         blk_mq_sched_insert_request(rq, false, run_queue, false);
1753         return BLK_STS_OK;
1754 }
1755
1756 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1757                 struct request *rq, blk_qc_t *cookie)
1758 {
1759         blk_status_t ret;
1760         int srcu_idx;
1761
1762         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1763
1764         hctx_lock(hctx, &srcu_idx);
1765
1766         ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1767         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1768                 blk_mq_sched_insert_request(rq, false, true, false);
1769         else if (ret != BLK_STS_OK)
1770                 blk_mq_end_request(rq, ret);
1771
1772         hctx_unlock(hctx, srcu_idx);
1773 }
1774
1775 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1776 {
1777         blk_status_t ret;
1778         int srcu_idx;
1779         blk_qc_t unused_cookie;
1780         struct blk_mq_ctx *ctx = rq->mq_ctx;
1781         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1782
1783         hctx_lock(hctx, &srcu_idx);
1784         ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1785         hctx_unlock(hctx, srcu_idx);
1786
1787         return ret;
1788 }
1789
1790 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1791                 struct list_head *list)
1792 {
1793         while (!list_empty(list)) {
1794                 blk_status_t ret;
1795                 struct request *rq = list_first_entry(list, struct request,
1796                                 queuelist);
1797
1798                 list_del_init(&rq->queuelist);
1799                 ret = blk_mq_request_issue_directly(rq);
1800                 if (ret != BLK_STS_OK) {
1801                         if (ret == BLK_STS_RESOURCE ||
1802                                         ret == BLK_STS_DEV_RESOURCE) {
1803                                 list_add(&rq->queuelist, list);
1804                                 break;
1805                         }
1806                         blk_mq_end_request(rq, ret);
1807                 }
1808         }
1809 }
1810
1811 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1812 {
1813         const int is_sync = op_is_sync(bio->bi_opf);
1814         const int is_flush_fua = op_is_flush(bio->bi_opf);
1815         struct blk_mq_alloc_data data = { .flags = 0 };
1816         struct request *rq;
1817         unsigned int request_count = 0;
1818         struct blk_plug *plug;
1819         struct request *same_queue_rq = NULL;
1820         blk_qc_t cookie;
1821
1822         blk_queue_bounce(q, &bio);
1823
1824         blk_queue_split(q, &bio);
1825
1826         if (!bio_integrity_prep(bio))
1827                 return BLK_QC_T_NONE;
1828
1829         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1830             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1831                 return BLK_QC_T_NONE;
1832
1833         if (blk_mq_sched_bio_merge(q, bio))
1834                 return BLK_QC_T_NONE;
1835
1836         rq_qos_throttle(q, bio, NULL);
1837
1838         trace_block_getrq(q, bio, bio->bi_opf);
1839
1840         rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1841         if (unlikely(!rq)) {
1842                 rq_qos_cleanup(q, bio);
1843                 if (bio->bi_opf & REQ_NOWAIT)
1844                         bio_wouldblock_error(bio);
1845                 return BLK_QC_T_NONE;
1846         }
1847
1848         rq_qos_track(q, rq, bio);
1849
1850         cookie = request_to_qc_t(data.hctx, rq);
1851
1852         plug = current->plug;
1853         if (unlikely(is_flush_fua)) {
1854                 blk_mq_put_ctx(data.ctx);
1855                 blk_mq_bio_to_request(rq, bio);
1856
1857                 /* bypass scheduler for flush rq */
1858                 blk_insert_flush(rq);
1859                 blk_mq_run_hw_queue(data.hctx, true);
1860         } else if (plug && q->nr_hw_queues == 1) {
1861                 struct request *last = NULL;
1862
1863                 blk_mq_put_ctx(data.ctx);
1864                 blk_mq_bio_to_request(rq, bio);
1865
1866                 /*
1867                  * @request_count may become stale because of schedule
1868                  * out, so check the list again.
1869                  */
1870                 if (list_empty(&plug->mq_list))
1871                         request_count = 0;
1872                 else if (blk_queue_nomerges(q))
1873                         request_count = blk_plug_queued_count(q);
1874
1875                 if (!request_count)
1876                         trace_block_plug(q);
1877                 else
1878                         last = list_entry_rq(plug->mq_list.prev);
1879
1880                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1881                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1882                         blk_flush_plug_list(plug, false);
1883                         trace_block_plug(q);
1884                 }
1885
1886                 list_add_tail(&rq->queuelist, &plug->mq_list);
1887         } else if (plug && !blk_queue_nomerges(q)) {
1888                 blk_mq_bio_to_request(rq, bio);
1889
1890                 /*
1891                  * We do limited plugging. If the bio can be merged, do that.
1892                  * Otherwise the existing request in the plug list will be
1893                  * issued. So the plug list will have one request at most
1894                  * The plug list might get flushed before this. If that happens,
1895                  * the plug list is empty, and same_queue_rq is invalid.
1896                  */
1897                 if (list_empty(&plug->mq_list))
1898                         same_queue_rq = NULL;
1899                 if (same_queue_rq)
1900                         list_del_init(&same_queue_rq->queuelist);
1901                 list_add_tail(&rq->queuelist, &plug->mq_list);
1902
1903                 blk_mq_put_ctx(data.ctx);
1904
1905                 if (same_queue_rq) {
1906                         data.hctx = blk_mq_map_queue(q,
1907                                         same_queue_rq->mq_ctx->cpu);
1908                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1909                                         &cookie);
1910                 }
1911         } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
1912                         !data.hctx->dispatch_busy)) {
1913                 blk_mq_put_ctx(data.ctx);
1914                 blk_mq_bio_to_request(rq, bio);
1915                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1916         } else {
1917                 blk_mq_put_ctx(data.ctx);
1918                 blk_mq_bio_to_request(rq, bio);
1919                 blk_mq_sched_insert_request(rq, false, true, true);
1920         }
1921
1922         return cookie;
1923 }
1924
1925 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1926                      unsigned int hctx_idx)
1927 {
1928         struct page *page;
1929
1930         if (tags->rqs && set->ops->exit_request) {
1931                 int i;
1932
1933                 for (i = 0; i < tags->nr_tags; i++) {
1934                         struct request *rq = tags->static_rqs[i];
1935
1936                         if (!rq)
1937                                 continue;
1938                         set->ops->exit_request(set, rq, hctx_idx);
1939                         tags->static_rqs[i] = NULL;
1940                 }
1941         }
1942
1943         while (!list_empty(&tags->page_list)) {
1944                 page = list_first_entry(&tags->page_list, struct page, lru);
1945                 list_del_init(&page->lru);
1946                 /*
1947                  * Remove kmemleak object previously allocated in
1948                  * blk_mq_init_rq_map().
1949                  */
1950                 kmemleak_free(page_address(page));
1951                 __free_pages(page, page->private);
1952         }
1953 }
1954
1955 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1956 {
1957         kfree(tags->rqs);
1958         tags->rqs = NULL;
1959         kfree(tags->static_rqs);
1960         tags->static_rqs = NULL;
1961
1962         blk_mq_free_tags(tags);
1963 }
1964
1965 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1966                                         unsigned int hctx_idx,
1967                                         unsigned int nr_tags,
1968                                         unsigned int reserved_tags)
1969 {
1970         struct blk_mq_tags *tags;
1971         int node;
1972
1973         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1974         if (node == NUMA_NO_NODE)
1975                 node = set->numa_node;
1976
1977         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1978                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1979         if (!tags)
1980                 return NULL;
1981
1982         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
1983                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1984                                  node);
1985         if (!tags->rqs) {
1986                 blk_mq_free_tags(tags);
1987                 return NULL;
1988         }
1989
1990         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
1991                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1992                                         node);
1993         if (!tags->static_rqs) {
1994                 kfree(tags->rqs);
1995                 blk_mq_free_tags(tags);
1996                 return NULL;
1997         }
1998
1999         return tags;
2000 }
2001
2002 static size_t order_to_size(unsigned int order)
2003 {
2004         return (size_t)PAGE_SIZE << order;
2005 }
2006
2007 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2008                                unsigned int hctx_idx, int node)
2009 {
2010         int ret;
2011
2012         if (set->ops->init_request) {
2013                 ret = set->ops->init_request(set, rq, hctx_idx, node);
2014                 if (ret)
2015                         return ret;
2016         }
2017
2018         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2019         return 0;
2020 }
2021
2022 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2023                      unsigned int hctx_idx, unsigned int depth)
2024 {
2025         unsigned int i, j, entries_per_page, max_order = 4;
2026         size_t rq_size, left;
2027         int node;
2028
2029         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2030         if (node == NUMA_NO_NODE)
2031                 node = set->numa_node;
2032
2033         INIT_LIST_HEAD(&tags->page_list);
2034
2035         /*
2036          * rq_size is the size of the request plus driver payload, rounded
2037          * to the cacheline size
2038          */
2039         rq_size = round_up(sizeof(struct request) + set->cmd_size,
2040                                 cache_line_size());
2041         left = rq_size * depth;
2042
2043         for (i = 0; i < depth; ) {
2044                 int this_order = max_order;
2045                 struct page *page;
2046                 int to_do;
2047                 void *p;
2048
2049                 while (this_order && left < order_to_size(this_order - 1))
2050                         this_order--;
2051
2052                 do {
2053                         page = alloc_pages_node(node,
2054                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2055                                 this_order);
2056                         if (page)
2057                                 break;
2058                         if (!this_order--)
2059                                 break;
2060                         if (order_to_size(this_order) < rq_size)
2061                                 break;
2062                 } while (1);
2063
2064                 if (!page)
2065                         goto fail;
2066
2067                 page->private = this_order;
2068                 list_add_tail(&page->lru, &tags->page_list);
2069
2070                 p = page_address(page);
2071                 /*
2072                  * Allow kmemleak to scan these pages as they contain pointers
2073                  * to additional allocations like via ops->init_request().
2074                  */
2075                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2076                 entries_per_page = order_to_size(this_order) / rq_size;
2077                 to_do = min(entries_per_page, depth - i);
2078                 left -= to_do * rq_size;
2079                 for (j = 0; j < to_do; j++) {
2080                         struct request *rq = p;
2081
2082                         tags->static_rqs[i] = rq;
2083                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2084                                 tags->static_rqs[i] = NULL;
2085                                 goto fail;
2086                         }
2087
2088                         p += rq_size;
2089                         i++;
2090                 }
2091         }
2092         return 0;
2093
2094 fail:
2095         blk_mq_free_rqs(set, tags, hctx_idx);
2096         return -ENOMEM;
2097 }
2098
2099 /*
2100  * 'cpu' is going away. splice any existing rq_list entries from this
2101  * software queue to the hw queue dispatch list, and ensure that it
2102  * gets run.
2103  */
2104 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2105 {
2106         struct blk_mq_hw_ctx *hctx;
2107         struct blk_mq_ctx *ctx;
2108         LIST_HEAD(tmp);
2109
2110         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2111         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2112
2113         spin_lock(&ctx->lock);
2114         if (!list_empty(&ctx->rq_list)) {
2115                 list_splice_init(&ctx->rq_list, &tmp);
2116                 blk_mq_hctx_clear_pending(hctx, ctx);
2117         }
2118         spin_unlock(&ctx->lock);
2119
2120         if (list_empty(&tmp))
2121                 return 0;
2122
2123         spin_lock(&hctx->lock);
2124         list_splice_tail_init(&tmp, &hctx->dispatch);
2125         spin_unlock(&hctx->lock);
2126
2127         blk_mq_run_hw_queue(hctx, true);
2128         return 0;
2129 }
2130
2131 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2132 {
2133         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2134                                             &hctx->cpuhp_dead);
2135 }
2136
2137 /* hctx->ctxs will be freed in queue's release handler */
2138 static void blk_mq_exit_hctx(struct request_queue *q,
2139                 struct blk_mq_tag_set *set,
2140                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2141 {
2142         blk_mq_debugfs_unregister_hctx(hctx);
2143
2144         if (blk_mq_hw_queue_mapped(hctx))
2145                 blk_mq_tag_idle(hctx);
2146
2147         if (set->ops->exit_request)
2148                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2149
2150         if (set->ops->exit_hctx)
2151                 set->ops->exit_hctx(hctx, hctx_idx);
2152
2153         if (hctx->flags & BLK_MQ_F_BLOCKING)
2154                 cleanup_srcu_struct(hctx->srcu);
2155
2156         blk_mq_remove_cpuhp(hctx);
2157         blk_free_flush_queue(hctx->fq);
2158         sbitmap_free(&hctx->ctx_map);
2159 }
2160
2161 static void blk_mq_exit_hw_queues(struct request_queue *q,
2162                 struct blk_mq_tag_set *set, int nr_queue)
2163 {
2164         struct blk_mq_hw_ctx *hctx;
2165         unsigned int i;
2166
2167         queue_for_each_hw_ctx(q, hctx, i) {
2168                 if (i == nr_queue)
2169                         break;
2170                 blk_mq_exit_hctx(q, set, hctx, i);
2171         }
2172 }
2173
2174 static int blk_mq_init_hctx(struct request_queue *q,
2175                 struct blk_mq_tag_set *set,
2176                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2177 {
2178         int node;
2179
2180         node = hctx->numa_node;
2181         if (node == NUMA_NO_NODE)
2182                 node = hctx->numa_node = set->numa_node;
2183
2184         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2185         spin_lock_init(&hctx->lock);
2186         INIT_LIST_HEAD(&hctx->dispatch);
2187         hctx->queue = q;
2188         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2189
2190         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2191
2192         hctx->tags = set->tags[hctx_idx];
2193
2194         /*
2195          * Allocate space for all possible cpus to avoid allocation at
2196          * runtime
2197          */
2198         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2199                                         GFP_KERNEL, node);
2200         if (!hctx->ctxs)
2201                 goto unregister_cpu_notifier;
2202
2203         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2204                               node))
2205                 goto free_ctxs;
2206
2207         hctx->nr_ctx = 0;
2208
2209         spin_lock_init(&hctx->dispatch_wait_lock);
2210         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2211         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2212
2213         if (set->ops->init_hctx &&
2214             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2215                 goto free_bitmap;
2216
2217         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2218         if (!hctx->fq)
2219                 goto exit_hctx;
2220
2221         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2222                 goto free_fq;
2223
2224         if (hctx->flags & BLK_MQ_F_BLOCKING)
2225                 init_srcu_struct(hctx->srcu);
2226
2227         blk_mq_debugfs_register_hctx(q, hctx);
2228
2229         return 0;
2230
2231  free_fq:
2232         kfree(hctx->fq);
2233  exit_hctx:
2234         if (set->ops->exit_hctx)
2235                 set->ops->exit_hctx(hctx, hctx_idx);
2236  free_bitmap:
2237         sbitmap_free(&hctx->ctx_map);
2238  free_ctxs:
2239         kfree(hctx->ctxs);
2240  unregister_cpu_notifier:
2241         blk_mq_remove_cpuhp(hctx);
2242         return -1;
2243 }
2244
2245 static void blk_mq_init_cpu_queues(struct request_queue *q,
2246                                    unsigned int nr_hw_queues)
2247 {
2248         unsigned int i;
2249
2250         for_each_possible_cpu(i) {
2251                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2252                 struct blk_mq_hw_ctx *hctx;
2253
2254                 __ctx->cpu = i;
2255                 spin_lock_init(&__ctx->lock);
2256                 INIT_LIST_HEAD(&__ctx->rq_list);
2257                 __ctx->queue = q;
2258
2259                 /*
2260                  * Set local node, IFF we have more than one hw queue. If
2261                  * not, we remain on the home node of the device
2262                  */
2263                 hctx = blk_mq_map_queue(q, i);
2264                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2265                         hctx->numa_node = local_memory_node(cpu_to_node(i));
2266         }
2267 }
2268
2269 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2270 {
2271         int ret = 0;
2272
2273         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2274                                         set->queue_depth, set->reserved_tags);
2275         if (!set->tags[hctx_idx])
2276                 return false;
2277
2278         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2279                                 set->queue_depth);
2280         if (!ret)
2281                 return true;
2282
2283         blk_mq_free_rq_map(set->tags[hctx_idx]);
2284         set->tags[hctx_idx] = NULL;
2285         return false;
2286 }
2287
2288 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2289                                          unsigned int hctx_idx)
2290 {
2291         if (set->tags[hctx_idx]) {
2292                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2293                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2294                 set->tags[hctx_idx] = NULL;
2295         }
2296 }
2297
2298 static void blk_mq_map_swqueue(struct request_queue *q)
2299 {
2300         unsigned int i, hctx_idx;
2301         struct blk_mq_hw_ctx *hctx;
2302         struct blk_mq_ctx *ctx;
2303         struct blk_mq_tag_set *set = q->tag_set;
2304
2305         /*
2306          * Avoid others reading imcomplete hctx->cpumask through sysfs
2307          */
2308         mutex_lock(&q->sysfs_lock);
2309
2310         queue_for_each_hw_ctx(q, hctx, i) {
2311                 cpumask_clear(hctx->cpumask);
2312                 hctx->nr_ctx = 0;
2313                 hctx->dispatch_from = NULL;
2314         }
2315
2316         /*
2317          * Map software to hardware queues.
2318          *
2319          * If the cpu isn't present, the cpu is mapped to first hctx.
2320          */
2321         for_each_possible_cpu(i) {
2322                 hctx_idx = q->mq_map[i];
2323                 /* unmapped hw queue can be remapped after CPU topo changed */
2324                 if (!set->tags[hctx_idx] &&
2325                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2326                         /*
2327                          * If tags initialization fail for some hctx,
2328                          * that hctx won't be brought online.  In this
2329                          * case, remap the current ctx to hctx[0] which
2330                          * is guaranteed to always have tags allocated
2331                          */
2332                         q->mq_map[i] = 0;
2333                 }
2334
2335                 ctx = per_cpu_ptr(q->queue_ctx, i);
2336                 hctx = blk_mq_map_queue(q, i);
2337
2338                 cpumask_set_cpu(i, hctx->cpumask);
2339                 ctx->index_hw = hctx->nr_ctx;
2340                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2341         }
2342
2343         mutex_unlock(&q->sysfs_lock);
2344
2345         queue_for_each_hw_ctx(q, hctx, i) {
2346                 /*
2347                  * If no software queues are mapped to this hardware queue,
2348                  * disable it and free the request entries.
2349                  */
2350                 if (!hctx->nr_ctx) {
2351                         /* Never unmap queue 0.  We need it as a
2352                          * fallback in case of a new remap fails
2353                          * allocation
2354                          */
2355                         if (i && set->tags[i])
2356                                 blk_mq_free_map_and_requests(set, i);
2357
2358                         hctx->tags = NULL;
2359                         continue;
2360                 }
2361
2362                 hctx->tags = set->tags[i];
2363                 WARN_ON(!hctx->tags);
2364
2365                 /*
2366                  * Set the map size to the number of mapped software queues.
2367                  * This is more accurate and more efficient than looping
2368                  * over all possibly mapped software queues.
2369                  */
2370                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2371
2372                 /*
2373                  * Initialize batch roundrobin counts
2374                  */
2375                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2376                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2377         }
2378 }
2379
2380 /*
2381  * Caller needs to ensure that we're either frozen/quiesced, or that
2382  * the queue isn't live yet.
2383  */
2384 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2385 {
2386         struct blk_mq_hw_ctx *hctx;
2387         int i;
2388
2389         queue_for_each_hw_ctx(q, hctx, i) {
2390                 if (shared)
2391                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2392                 else
2393                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2394         }
2395 }
2396
2397 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2398                                         bool shared)
2399 {
2400         struct request_queue *q;
2401
2402         lockdep_assert_held(&set->tag_list_lock);
2403
2404         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2405                 blk_mq_freeze_queue(q);
2406                 queue_set_hctx_shared(q, shared);
2407                 blk_mq_unfreeze_queue(q);
2408         }
2409 }
2410
2411 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2412 {
2413         struct blk_mq_tag_set *set = q->tag_set;
2414
2415         mutex_lock(&set->tag_list_lock);
2416         list_del_rcu(&q->tag_set_list);
2417         if (list_is_singular(&set->tag_list)) {
2418                 /* just transitioned to unshared */
2419                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2420                 /* update existing queue */
2421                 blk_mq_update_tag_set_depth(set, false);
2422         }
2423         mutex_unlock(&set->tag_list_lock);
2424         INIT_LIST_HEAD(&q->tag_set_list);
2425 }
2426
2427 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2428                                      struct request_queue *q)
2429 {
2430         q->tag_set = set;
2431
2432         mutex_lock(&set->tag_list_lock);
2433
2434         /*
2435          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2436          */
2437         if (!list_empty(&set->tag_list) &&
2438             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2439                 set->flags |= BLK_MQ_F_TAG_SHARED;
2440                 /* update existing queue */
2441                 blk_mq_update_tag_set_depth(set, true);
2442         }
2443         if (set->flags & BLK_MQ_F_TAG_SHARED)
2444                 queue_set_hctx_shared(q, true);
2445         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2446
2447         mutex_unlock(&set->tag_list_lock);
2448 }
2449
2450 /*
2451  * It is the actual release handler for mq, but we do it from
2452  * request queue's release handler for avoiding use-after-free
2453  * and headache because q->mq_kobj shouldn't have been introduced,
2454  * but we can't group ctx/kctx kobj without it.
2455  */
2456 void blk_mq_release(struct request_queue *q)
2457 {
2458         struct blk_mq_hw_ctx *hctx;
2459         unsigned int i;
2460
2461         /* hctx kobj stays in hctx */
2462         queue_for_each_hw_ctx(q, hctx, i) {
2463                 if (!hctx)
2464                         continue;
2465                 kobject_put(&hctx->kobj);
2466         }
2467
2468         q->mq_map = NULL;
2469
2470         kfree(q->queue_hw_ctx);
2471
2472         /*
2473          * release .mq_kobj and sw queue's kobject now because
2474          * both share lifetime with request queue.
2475          */
2476         blk_mq_sysfs_deinit(q);
2477
2478         free_percpu(q->queue_ctx);
2479 }
2480
2481 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2482 {
2483         struct request_queue *uninit_q, *q;
2484
2485         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2486         if (!uninit_q)
2487                 return ERR_PTR(-ENOMEM);
2488
2489         q = blk_mq_init_allocated_queue(set, uninit_q);
2490         if (IS_ERR(q))
2491                 blk_cleanup_queue(uninit_q);
2492
2493         return q;
2494 }
2495 EXPORT_SYMBOL(blk_mq_init_queue);
2496
2497 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2498 {
2499         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2500
2501         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2502                            __alignof__(struct blk_mq_hw_ctx)) !=
2503                      sizeof(struct blk_mq_hw_ctx));
2504
2505         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2506                 hw_ctx_size += sizeof(struct srcu_struct);
2507
2508         return hw_ctx_size;
2509 }
2510
2511 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2512                                                 struct request_queue *q)
2513 {
2514         int i, j;
2515         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2516
2517         blk_mq_sysfs_unregister(q);
2518
2519         /* protect against switching io scheduler  */
2520         mutex_lock(&q->sysfs_lock);
2521         for (i = 0; i < set->nr_hw_queues; i++) {
2522                 int node;
2523
2524                 if (hctxs[i])
2525                         continue;
2526
2527                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2528                 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2529                                         GFP_KERNEL, node);
2530                 if (!hctxs[i])
2531                         break;
2532
2533                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2534                                                 node)) {
2535                         kfree(hctxs[i]);
2536                         hctxs[i] = NULL;
2537                         break;
2538                 }
2539
2540                 atomic_set(&hctxs[i]->nr_active, 0);
2541                 hctxs[i]->numa_node = node;
2542                 hctxs[i]->queue_num = i;
2543
2544                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2545                         free_cpumask_var(hctxs[i]->cpumask);
2546                         kfree(hctxs[i]);
2547                         hctxs[i] = NULL;
2548                         break;
2549                 }
2550                 blk_mq_hctx_kobj_init(hctxs[i]);
2551         }
2552         for (j = i; j < q->nr_hw_queues; j++) {
2553                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2554
2555                 if (hctx) {
2556                         if (hctx->tags)
2557                                 blk_mq_free_map_and_requests(set, j);
2558                         blk_mq_exit_hctx(q, set, hctx, j);
2559                         kobject_put(&hctx->kobj);
2560                         hctxs[j] = NULL;
2561
2562                 }
2563         }
2564         q->nr_hw_queues = i;
2565         mutex_unlock(&q->sysfs_lock);
2566         blk_mq_sysfs_register(q);
2567 }
2568
2569 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2570                                                   struct request_queue *q)
2571 {
2572         /* mark the queue as mq asap */
2573         q->mq_ops = set->ops;
2574
2575         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2576                                              blk_mq_poll_stats_bkt,
2577                                              BLK_MQ_POLL_STATS_BKTS, q);
2578         if (!q->poll_cb)
2579                 goto err_exit;
2580
2581         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2582         if (!q->queue_ctx)
2583                 goto err_exit;
2584
2585         /* init q->mq_kobj and sw queues' kobjects */
2586         blk_mq_sysfs_init(q);
2587
2588         q->queue_hw_ctx = kcalloc_node(nr_cpu_ids, sizeof(*(q->queue_hw_ctx)),
2589                                                 GFP_KERNEL, set->numa_node);
2590         if (!q->queue_hw_ctx)
2591                 goto err_percpu;
2592
2593         q->mq_map = set->mq_map;
2594
2595         blk_mq_realloc_hw_ctxs(set, q);
2596         if (!q->nr_hw_queues)
2597                 goto err_hctxs;
2598
2599         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2600         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2601
2602         q->nr_queues = nr_cpu_ids;
2603
2604         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2605
2606         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2607                 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2608
2609         q->sg_reserved_size = INT_MAX;
2610
2611         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2612         INIT_LIST_HEAD(&q->requeue_list);
2613         spin_lock_init(&q->requeue_lock);
2614
2615         blk_queue_make_request(q, blk_mq_make_request);
2616         if (q->mq_ops->poll)
2617                 q->poll_fn = blk_mq_poll;
2618
2619         /*
2620          * Do this after blk_queue_make_request() overrides it...
2621          */
2622         q->nr_requests = set->queue_depth;
2623
2624         /*
2625          * Default to classic polling
2626          */
2627         q->poll_nsec = -1;
2628
2629         if (set->ops->complete)
2630                 blk_queue_softirq_done(q, set->ops->complete);
2631
2632         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2633         blk_mq_add_queue_tag_set(set, q);
2634         blk_mq_map_swqueue(q);
2635
2636         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2637                 int ret;
2638
2639                 ret = elevator_init_mq(q);
2640                 if (ret)
2641                         return ERR_PTR(ret);
2642         }
2643
2644         return q;
2645
2646 err_hctxs:
2647         kfree(q->queue_hw_ctx);
2648 err_percpu:
2649         free_percpu(q->queue_ctx);
2650 err_exit:
2651         q->mq_ops = NULL;
2652         return ERR_PTR(-ENOMEM);
2653 }
2654 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2655
2656 void blk_mq_free_queue(struct request_queue *q)
2657 {
2658         struct blk_mq_tag_set   *set = q->tag_set;
2659
2660         blk_mq_del_queue_tag_set(q);
2661         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2662 }
2663
2664 /* Basically redo blk_mq_init_queue with queue frozen */
2665 static void blk_mq_queue_reinit(struct request_queue *q)
2666 {
2667         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2668
2669         blk_mq_debugfs_unregister_hctxs(q);
2670         blk_mq_sysfs_unregister(q);
2671
2672         /*
2673          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2674          * we should change hctx numa_node according to the new topology (this
2675          * involves freeing and re-allocating memory, worth doing?)
2676          */
2677         blk_mq_map_swqueue(q);
2678
2679         blk_mq_sysfs_register(q);
2680         blk_mq_debugfs_register_hctxs(q);
2681 }
2682
2683 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2684 {
2685         int i;
2686
2687         for (i = 0; i < set->nr_hw_queues; i++)
2688                 if (!__blk_mq_alloc_rq_map(set, i))
2689                         goto out_unwind;
2690
2691         return 0;
2692
2693 out_unwind:
2694         while (--i >= 0)
2695                 blk_mq_free_rq_map(set->tags[i]);
2696
2697         return -ENOMEM;
2698 }
2699
2700 /*
2701  * Allocate the request maps associated with this tag_set. Note that this
2702  * may reduce the depth asked for, if memory is tight. set->queue_depth
2703  * will be updated to reflect the allocated depth.
2704  */
2705 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2706 {
2707         unsigned int depth;
2708         int err;
2709
2710         depth = set->queue_depth;
2711         do {
2712                 err = __blk_mq_alloc_rq_maps(set);
2713                 if (!err)
2714                         break;
2715
2716                 set->queue_depth >>= 1;
2717                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2718                         err = -ENOMEM;
2719                         break;
2720                 }
2721         } while (set->queue_depth);
2722
2723         if (!set->queue_depth || err) {
2724                 pr_err("blk-mq: failed to allocate request map\n");
2725                 return -ENOMEM;
2726         }
2727
2728         if (depth != set->queue_depth)
2729                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2730                                                 depth, set->queue_depth);
2731
2732         return 0;
2733 }
2734
2735 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2736 {
2737         if (set->ops->map_queues) {
2738                 /*
2739                  * transport .map_queues is usually done in the following
2740                  * way:
2741                  *
2742                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2743                  *      mask = get_cpu_mask(queue)
2744                  *      for_each_cpu(cpu, mask)
2745                  *              set->mq_map[cpu] = queue;
2746                  * }
2747                  *
2748                  * When we need to remap, the table has to be cleared for
2749                  * killing stale mapping since one CPU may not be mapped
2750                  * to any hw queue.
2751                  */
2752                 blk_mq_clear_mq_map(set);
2753
2754                 return set->ops->map_queues(set);
2755         } else
2756                 return blk_mq_map_queues(set);
2757 }
2758
2759 /*
2760  * Alloc a tag set to be associated with one or more request queues.
2761  * May fail with EINVAL for various error conditions. May adjust the
2762  * requested depth down, if it's too large. In that case, the set
2763  * value will be stored in set->queue_depth.
2764  */
2765 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2766 {
2767         int ret;
2768
2769         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2770
2771         if (!set->nr_hw_queues)
2772                 return -EINVAL;
2773         if (!set->queue_depth)
2774                 return -EINVAL;
2775         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2776                 return -EINVAL;
2777
2778         if (!set->ops->queue_rq)
2779                 return -EINVAL;
2780
2781         if (!set->ops->get_budget ^ !set->ops->put_budget)
2782                 return -EINVAL;
2783
2784         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2785                 pr_info("blk-mq: reduced tag depth to %u\n",
2786                         BLK_MQ_MAX_DEPTH);
2787                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2788         }
2789
2790         /*
2791          * If a crashdump is active, then we are potentially in a very
2792          * memory constrained environment. Limit us to 1 queue and
2793          * 64 tags to prevent using too much memory.
2794          */
2795         if (is_kdump_kernel()) {
2796                 set->nr_hw_queues = 1;
2797                 set->queue_depth = min(64U, set->queue_depth);
2798         }
2799         /*
2800          * There is no use for more h/w queues than cpus.
2801          */
2802         if (set->nr_hw_queues > nr_cpu_ids)
2803                 set->nr_hw_queues = nr_cpu_ids;
2804
2805         set->tags = kcalloc_node(nr_cpu_ids, sizeof(struct blk_mq_tags *),
2806                                  GFP_KERNEL, set->numa_node);
2807         if (!set->tags)
2808                 return -ENOMEM;
2809
2810         ret = -ENOMEM;
2811         set->mq_map = kcalloc_node(nr_cpu_ids, sizeof(*set->mq_map),
2812                                    GFP_KERNEL, set->numa_node);
2813         if (!set->mq_map)
2814                 goto out_free_tags;
2815
2816         ret = blk_mq_update_queue_map(set);
2817         if (ret)
2818                 goto out_free_mq_map;
2819
2820         ret = blk_mq_alloc_rq_maps(set);
2821         if (ret)
2822                 goto out_free_mq_map;
2823
2824         mutex_init(&set->tag_list_lock);
2825         INIT_LIST_HEAD(&set->tag_list);
2826
2827         return 0;
2828
2829 out_free_mq_map:
2830         kfree(set->mq_map);
2831         set->mq_map = NULL;
2832 out_free_tags:
2833         kfree(set->tags);
2834         set->tags = NULL;
2835         return ret;
2836 }
2837 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2838
2839 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2840 {
2841         int i;
2842
2843         for (i = 0; i < nr_cpu_ids; i++)
2844                 blk_mq_free_map_and_requests(set, i);
2845
2846         kfree(set->mq_map);
2847         set->mq_map = NULL;
2848
2849         kfree(set->tags);
2850         set->tags = NULL;
2851 }
2852 EXPORT_SYMBOL(blk_mq_free_tag_set);
2853
2854 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2855 {
2856         struct blk_mq_tag_set *set = q->tag_set;
2857         struct blk_mq_hw_ctx *hctx;
2858         int i, ret;
2859
2860         if (!set)
2861                 return -EINVAL;
2862
2863         blk_mq_freeze_queue(q);
2864         blk_mq_quiesce_queue(q);
2865
2866         ret = 0;
2867         queue_for_each_hw_ctx(q, hctx, i) {
2868                 if (!hctx->tags)
2869                         continue;
2870                 /*
2871                  * If we're using an MQ scheduler, just update the scheduler
2872                  * queue depth. This is similar to what the old code would do.
2873                  */
2874                 if (!hctx->sched_tags) {
2875                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2876                                                         false);
2877                 } else {
2878                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2879                                                         nr, true);
2880                 }
2881                 if (ret)
2882                         break;
2883         }
2884
2885         if (!ret)
2886                 q->nr_requests = nr;
2887
2888         blk_mq_unquiesce_queue(q);
2889         blk_mq_unfreeze_queue(q);
2890
2891         return ret;
2892 }
2893
2894 /*
2895  * request_queue and elevator_type pair.
2896  * It is just used by __blk_mq_update_nr_hw_queues to cache
2897  * the elevator_type associated with a request_queue.
2898  */
2899 struct blk_mq_qe_pair {
2900         struct list_head node;
2901         struct request_queue *q;
2902         struct elevator_type *type;
2903 };
2904
2905 /*
2906  * Cache the elevator_type in qe pair list and switch the
2907  * io scheduler to 'none'
2908  */
2909 static bool blk_mq_elv_switch_none(struct list_head *head,
2910                 struct request_queue *q)
2911 {
2912         struct blk_mq_qe_pair *qe;
2913
2914         if (!q->elevator)
2915                 return true;
2916
2917         qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
2918         if (!qe)
2919                 return false;
2920
2921         INIT_LIST_HEAD(&qe->node);
2922         qe->q = q;
2923         qe->type = q->elevator->type;
2924         list_add(&qe->node, head);
2925
2926         mutex_lock(&q->sysfs_lock);
2927         /*
2928          * After elevator_switch_mq, the previous elevator_queue will be
2929          * released by elevator_release. The reference of the io scheduler
2930          * module get by elevator_get will also be put. So we need to get
2931          * a reference of the io scheduler module here to prevent it to be
2932          * removed.
2933          */
2934         __module_get(qe->type->elevator_owner);
2935         elevator_switch_mq(q, NULL);
2936         mutex_unlock(&q->sysfs_lock);
2937
2938         return true;
2939 }
2940
2941 static void blk_mq_elv_switch_back(struct list_head *head,
2942                 struct request_queue *q)
2943 {
2944         struct blk_mq_qe_pair *qe;
2945         struct elevator_type *t = NULL;
2946
2947         list_for_each_entry(qe, head, node)
2948                 if (qe->q == q) {
2949                         t = qe->type;
2950                         break;
2951                 }
2952
2953         if (!t)
2954                 return;
2955
2956         list_del(&qe->node);
2957         kfree(qe);
2958
2959         mutex_lock(&q->sysfs_lock);
2960         elevator_switch_mq(q, t);
2961         mutex_unlock(&q->sysfs_lock);
2962 }
2963
2964 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2965                                                         int nr_hw_queues)
2966 {
2967         struct request_queue *q;
2968         LIST_HEAD(head);
2969
2970         lockdep_assert_held(&set->tag_list_lock);
2971
2972         if (nr_hw_queues > nr_cpu_ids)
2973                 nr_hw_queues = nr_cpu_ids;
2974         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2975                 return;
2976
2977         list_for_each_entry(q, &set->tag_list, tag_set_list)
2978                 blk_mq_freeze_queue(q);
2979         /*
2980          * Sync with blk_mq_queue_tag_busy_iter.
2981          */
2982         synchronize_rcu();
2983         /*
2984          * Switch IO scheduler to 'none', cleaning up the data associated
2985          * with the previous scheduler. We will switch back once we are done
2986          * updating the new sw to hw queue mappings.
2987          */
2988         list_for_each_entry(q, &set->tag_list, tag_set_list)
2989                 if (!blk_mq_elv_switch_none(&head, q))
2990                         goto switch_back;
2991
2992         set->nr_hw_queues = nr_hw_queues;
2993         blk_mq_update_queue_map(set);
2994         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2995                 blk_mq_realloc_hw_ctxs(set, q);
2996                 blk_mq_queue_reinit(q);
2997         }
2998
2999 switch_back:
3000         list_for_each_entry(q, &set->tag_list, tag_set_list)
3001                 blk_mq_elv_switch_back(&head, q);
3002
3003         list_for_each_entry(q, &set->tag_list, tag_set_list)
3004                 blk_mq_unfreeze_queue(q);
3005 }
3006
3007 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3008 {
3009         mutex_lock(&set->tag_list_lock);
3010         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3011         mutex_unlock(&set->tag_list_lock);
3012 }
3013 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3014
3015 /* Enable polling stats and return whether they were already enabled. */
3016 static bool blk_poll_stats_enable(struct request_queue *q)
3017 {
3018         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3019             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3020                 return true;
3021         blk_stat_add_callback(q, q->poll_cb);
3022         return false;
3023 }
3024
3025 static void blk_mq_poll_stats_start(struct request_queue *q)
3026 {
3027         /*
3028          * We don't arm the callback if polling stats are not enabled or the
3029          * callback is already active.
3030          */
3031         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3032             blk_stat_is_active(q->poll_cb))
3033                 return;
3034
3035         blk_stat_activate_msecs(q->poll_cb, 100);
3036 }
3037
3038 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3039 {
3040         struct request_queue *q = cb->data;
3041         int bucket;
3042
3043         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3044                 if (cb->stat[bucket].nr_samples)
3045                         q->poll_stat[bucket] = cb->stat[bucket];
3046         }
3047 }
3048
3049 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3050                                        struct blk_mq_hw_ctx *hctx,
3051                                        struct request *rq)
3052 {
3053         unsigned long ret = 0;
3054         int bucket;
3055
3056         /*
3057          * If stats collection isn't on, don't sleep but turn it on for
3058          * future users
3059          */
3060         if (!blk_poll_stats_enable(q))
3061                 return 0;
3062
3063         /*
3064          * As an optimistic guess, use half of the mean service time
3065          * for this type of request. We can (and should) make this smarter.
3066          * For instance, if the completion latencies are tight, we can
3067          * get closer than just half the mean. This is especially
3068          * important on devices where the completion latencies are longer
3069          * than ~10 usec. We do use the stats for the relevant IO size
3070          * if available which does lead to better estimates.
3071          */
3072         bucket = blk_mq_poll_stats_bkt(rq);
3073         if (bucket < 0)
3074                 return ret;
3075
3076         if (q->poll_stat[bucket].nr_samples)
3077                 ret = (q->poll_stat[bucket].mean + 1) / 2;
3078
3079         return ret;
3080 }
3081
3082 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3083                                      struct blk_mq_hw_ctx *hctx,
3084                                      struct request *rq)
3085 {
3086         struct hrtimer_sleeper hs;
3087         enum hrtimer_mode mode;
3088         unsigned int nsecs;
3089         ktime_t kt;
3090
3091         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3092                 return false;
3093
3094         /*
3095          * poll_nsec can be:
3096          *
3097          * -1:  don't ever hybrid sleep
3098          *  0:  use half of prev avg
3099          * >0:  use this specific value
3100          */
3101         if (q->poll_nsec == -1)
3102                 return false;
3103         else if (q->poll_nsec > 0)
3104                 nsecs = q->poll_nsec;
3105         else
3106                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3107
3108         if (!nsecs)
3109                 return false;
3110
3111         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3112
3113         /*
3114          * This will be replaced with the stats tracking code, using
3115          * 'avg_completion_time / 2' as the pre-sleep target.
3116          */
3117         kt = nsecs;
3118
3119         mode = HRTIMER_MODE_REL;
3120         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3121         hrtimer_set_expires(&hs.timer, kt);
3122
3123         hrtimer_init_sleeper(&hs, current);
3124         do {
3125                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3126                         break;
3127                 set_current_state(TASK_UNINTERRUPTIBLE);
3128                 hrtimer_start_expires(&hs.timer, mode);
3129                 if (hs.task)
3130                         io_schedule();
3131                 hrtimer_cancel(&hs.timer);
3132                 mode = HRTIMER_MODE_ABS;
3133         } while (hs.task && !signal_pending(current));
3134
3135         __set_current_state(TASK_RUNNING);
3136         destroy_hrtimer_on_stack(&hs.timer);
3137         return true;
3138 }
3139
3140 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3141 {
3142         struct request_queue *q = hctx->queue;
3143         long state;
3144
3145         /*
3146          * If we sleep, have the caller restart the poll loop to reset
3147          * the state. Like for the other success return cases, the
3148          * caller is responsible for checking if the IO completed. If
3149          * the IO isn't complete, we'll get called again and will go
3150          * straight to the busy poll loop.
3151          */
3152         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3153                 return true;
3154
3155         hctx->poll_considered++;
3156
3157         state = current->state;
3158         while (!need_resched()) {
3159                 int ret;
3160
3161                 hctx->poll_invoked++;
3162
3163                 ret = q->mq_ops->poll(hctx, rq->tag);
3164                 if (ret > 0) {
3165                         hctx->poll_success++;
3166                         set_current_state(TASK_RUNNING);
3167                         return true;
3168                 }
3169
3170                 if (signal_pending_state(state, current))
3171                         set_current_state(TASK_RUNNING);
3172
3173                 if (current->state == TASK_RUNNING)
3174                         return true;
3175                 if (ret < 0)
3176                         break;
3177                 cpu_relax();
3178         }
3179
3180         __set_current_state(TASK_RUNNING);
3181         return false;
3182 }
3183
3184 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3185 {
3186         struct blk_mq_hw_ctx *hctx;
3187         struct request *rq;
3188
3189         if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3190                 return false;
3191
3192         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3193         if (!blk_qc_t_is_internal(cookie))
3194                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3195         else {
3196                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3197                 /*
3198                  * With scheduling, if the request has completed, we'll
3199                  * get a NULL return here, as we clear the sched tag when
3200                  * that happens. The request still remains valid, like always,
3201                  * so we should be safe with just the NULL check.
3202                  */
3203                 if (!rq)
3204                         return false;
3205         }
3206
3207         return __blk_mq_poll(hctx, rq);
3208 }
3209
3210 static int __init blk_mq_init(void)
3211 {
3212         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3213                                 blk_mq_hctx_notify_dead);
3214         return 0;
3215 }
3216 subsys_initcall(blk_mq_init);