Loading...
1#include <linux/kernel.h>
2#include <linux/module.h>
3#include <linux/backing-dev.h>
4#include <linux/bio.h>
5#include <linux/blkdev.h>
6#include <linux/mm.h>
7#include <linux/init.h>
8#include <linux/slab.h>
9#include <linux/workqueue.h>
10#include <linux/smp.h>
11#include <linux/llist.h>
12#include <linux/list_sort.h>
13#include <linux/cpu.h>
14#include <linux/cache.h>
15#include <linux/sched/sysctl.h>
16#include <linux/delay.h>
17
18#include <trace/events/block.h>
19
20#include <linux/blk-mq.h>
21#include "blk.h"
22#include "blk-mq.h"
23#include "blk-mq-tag.h"
24
25static DEFINE_MUTEX(all_q_mutex);
26static LIST_HEAD(all_q_list);
27
28static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
29
30static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
31 unsigned int cpu)
32{
33 return per_cpu_ptr(q->queue_ctx, cpu);
34}
35
36/*
37 * This assumes per-cpu software queueing queues. They could be per-node
38 * as well, for instance. For now this is hardcoded as-is. Note that we don't
39 * care about preemption, since we know the ctx's are persistent. This does
40 * mean that we can't rely on ctx always matching the currently running CPU.
41 */
42static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
43{
44 return __blk_mq_get_ctx(q, get_cpu());
45}
46
47static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
48{
49 put_cpu();
50}
51
52/*
53 * Check if any of the ctx's have pending work in this hardware queue
54 */
55static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
56{
57 unsigned int i;
58
59 for (i = 0; i < hctx->nr_ctx_map; i++)
60 if (hctx->ctx_map[i])
61 return true;
62
63 return false;
64}
65
66/*
67 * Mark this ctx as having pending work in this hardware queue
68 */
69static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
70 struct blk_mq_ctx *ctx)
71{
72 if (!test_bit(ctx->index_hw, hctx->ctx_map))
73 set_bit(ctx->index_hw, hctx->ctx_map);
74}
75
76static struct request *__blk_mq_alloc_request(struct blk_mq_hw_ctx *hctx,
77 gfp_t gfp, bool reserved)
78{
79 struct request *rq;
80 unsigned int tag;
81
82 tag = blk_mq_get_tag(hctx->tags, gfp, reserved);
83 if (tag != BLK_MQ_TAG_FAIL) {
84 rq = hctx->rqs[tag];
85 rq->tag = tag;
86
87 return rq;
88 }
89
90 return NULL;
91}
92
93static int blk_mq_queue_enter(struct request_queue *q)
94{
95 int ret;
96
97 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
98 smp_wmb();
99 /* we have problems to freeze the queue if it's initializing */
100 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
101 return 0;
102
103 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
104
105 spin_lock_irq(q->queue_lock);
106 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
107 !blk_queue_bypass(q) || blk_queue_dying(q),
108 *q->queue_lock);
109 /* inc usage with lock hold to avoid freeze_queue runs here */
110 if (!ret && !blk_queue_dying(q))
111 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
112 else if (blk_queue_dying(q))
113 ret = -ENODEV;
114 spin_unlock_irq(q->queue_lock);
115
116 return ret;
117}
118
119static void blk_mq_queue_exit(struct request_queue *q)
120{
121 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
122}
123
124static void __blk_mq_drain_queue(struct request_queue *q)
125{
126 while (true) {
127 s64 count;
128
129 spin_lock_irq(q->queue_lock);
130 count = percpu_counter_sum(&q->mq_usage_counter);
131 spin_unlock_irq(q->queue_lock);
132
133 if (count == 0)
134 break;
135 blk_mq_run_queues(q, false);
136 msleep(10);
137 }
138}
139
140/*
141 * Guarantee no request is in use, so we can change any data structure of
142 * the queue afterward.
143 */
144static void blk_mq_freeze_queue(struct request_queue *q)
145{
146 bool drain;
147
148 spin_lock_irq(q->queue_lock);
149 drain = !q->bypass_depth++;
150 queue_flag_set(QUEUE_FLAG_BYPASS, q);
151 spin_unlock_irq(q->queue_lock);
152
153 if (drain)
154 __blk_mq_drain_queue(q);
155}
156
157void blk_mq_drain_queue(struct request_queue *q)
158{
159 __blk_mq_drain_queue(q);
160}
161
162static void blk_mq_unfreeze_queue(struct request_queue *q)
163{
164 bool wake = false;
165
166 spin_lock_irq(q->queue_lock);
167 if (!--q->bypass_depth) {
168 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
169 wake = true;
170 }
171 WARN_ON_ONCE(q->bypass_depth < 0);
172 spin_unlock_irq(q->queue_lock);
173 if (wake)
174 wake_up_all(&q->mq_freeze_wq);
175}
176
177bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
178{
179 return blk_mq_has_free_tags(hctx->tags);
180}
181EXPORT_SYMBOL(blk_mq_can_queue);
182
183static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
184 struct request *rq, unsigned int rw_flags)
185{
186 if (blk_queue_io_stat(q))
187 rw_flags |= REQ_IO_STAT;
188
189 rq->mq_ctx = ctx;
190 rq->cmd_flags = rw_flags;
191 rq->start_time = jiffies;
192 set_start_time_ns(rq);
193 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
194}
195
196static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
197 int rw, gfp_t gfp,
198 bool reserved)
199{
200 struct request *rq;
201
202 do {
203 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
204 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
205
206 rq = __blk_mq_alloc_request(hctx, gfp & ~__GFP_WAIT, reserved);
207 if (rq) {
208 blk_mq_rq_ctx_init(q, ctx, rq, rw);
209 break;
210 }
211
212 blk_mq_put_ctx(ctx);
213 if (!(gfp & __GFP_WAIT))
214 break;
215
216 __blk_mq_run_hw_queue(hctx);
217 blk_mq_wait_for_tags(hctx->tags);
218 } while (1);
219
220 return rq;
221}
222
223struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp)
224{
225 struct request *rq;
226
227 if (blk_mq_queue_enter(q))
228 return NULL;
229
230 rq = blk_mq_alloc_request_pinned(q, rw, gfp, false);
231 if (rq)
232 blk_mq_put_ctx(rq->mq_ctx);
233 return rq;
234}
235
236struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
237 gfp_t gfp)
238{
239 struct request *rq;
240
241 if (blk_mq_queue_enter(q))
242 return NULL;
243
244 rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
245 if (rq)
246 blk_mq_put_ctx(rq->mq_ctx);
247 return rq;
248}
249EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
250
251/*
252 * Re-init and set pdu, if we have it
253 */
254void blk_mq_rq_init(struct blk_mq_hw_ctx *hctx, struct request *rq)
255{
256 blk_rq_init(hctx->queue, rq);
257
258 if (hctx->cmd_size)
259 rq->special = blk_mq_rq_to_pdu(rq);
260}
261
262static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
263 struct blk_mq_ctx *ctx, struct request *rq)
264{
265 const int tag = rq->tag;
266 struct request_queue *q = rq->q;
267
268 blk_mq_rq_init(hctx, rq);
269 blk_mq_put_tag(hctx->tags, tag);
270
271 blk_mq_queue_exit(q);
272}
273
274void blk_mq_free_request(struct request *rq)
275{
276 struct blk_mq_ctx *ctx = rq->mq_ctx;
277 struct blk_mq_hw_ctx *hctx;
278 struct request_queue *q = rq->q;
279
280 ctx->rq_completed[rq_is_sync(rq)]++;
281
282 hctx = q->mq_ops->map_queue(q, ctx->cpu);
283 __blk_mq_free_request(hctx, ctx, rq);
284}
285
286bool blk_mq_end_io_partial(struct request *rq, int error, unsigned int nr_bytes)
287{
288 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
289 return true;
290
291 blk_account_io_done(rq);
292
293 if (rq->end_io)
294 rq->end_io(rq, error);
295 else
296 blk_mq_free_request(rq);
297 return false;
298}
299EXPORT_SYMBOL(blk_mq_end_io_partial);
300
301static void __blk_mq_complete_request_remote(void *data)
302{
303 struct request *rq = data;
304
305 rq->q->softirq_done_fn(rq);
306}
307
308void __blk_mq_complete_request(struct request *rq)
309{
310 struct blk_mq_ctx *ctx = rq->mq_ctx;
311 int cpu;
312
313 if (!ctx->ipi_redirect) {
314 rq->q->softirq_done_fn(rq);
315 return;
316 }
317
318 cpu = get_cpu();
319 if (cpu != ctx->cpu && cpu_online(ctx->cpu)) {
320 rq->csd.func = __blk_mq_complete_request_remote;
321 rq->csd.info = rq;
322 rq->csd.flags = 0;
323 smp_call_function_single_async(ctx->cpu, &rq->csd);
324 } else {
325 rq->q->softirq_done_fn(rq);
326 }
327 put_cpu();
328}
329
330/**
331 * blk_mq_complete_request - end I/O on a request
332 * @rq: the request being processed
333 *
334 * Description:
335 * Ends all I/O on a request. It does not handle partial completions.
336 * The actual completion happens out-of-order, through a IPI handler.
337 **/
338void blk_mq_complete_request(struct request *rq)
339{
340 if (unlikely(blk_should_fake_timeout(rq->q)))
341 return;
342 if (!blk_mark_rq_complete(rq))
343 __blk_mq_complete_request(rq);
344}
345EXPORT_SYMBOL(blk_mq_complete_request);
346
347static void blk_mq_start_request(struct request *rq, bool last)
348{
349 struct request_queue *q = rq->q;
350
351 trace_block_rq_issue(q, rq);
352
353 /*
354 * Just mark start time and set the started bit. Due to memory
355 * ordering, we know we'll see the correct deadline as long as
356 * REQ_ATOMIC_STARTED is seen.
357 */
358 rq->deadline = jiffies + q->rq_timeout;
359 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
360
361 if (q->dma_drain_size && blk_rq_bytes(rq)) {
362 /*
363 * Make sure space for the drain appears. We know we can do
364 * this because max_hw_segments has been adjusted to be one
365 * fewer than the device can handle.
366 */
367 rq->nr_phys_segments++;
368 }
369
370 /*
371 * Flag the last request in the series so that drivers know when IO
372 * should be kicked off, if they don't do it on a per-request basis.
373 *
374 * Note: the flag isn't the only condition drivers should do kick off.
375 * If drive is busy, the last request might not have the bit set.
376 */
377 if (last)
378 rq->cmd_flags |= REQ_END;
379}
380
381static void blk_mq_requeue_request(struct request *rq)
382{
383 struct request_queue *q = rq->q;
384
385 trace_block_rq_requeue(q, rq);
386 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
387
388 rq->cmd_flags &= ~REQ_END;
389
390 if (q->dma_drain_size && blk_rq_bytes(rq))
391 rq->nr_phys_segments--;
392}
393
394struct blk_mq_timeout_data {
395 struct blk_mq_hw_ctx *hctx;
396 unsigned long *next;
397 unsigned int *next_set;
398};
399
400static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
401{
402 struct blk_mq_timeout_data *data = __data;
403 struct blk_mq_hw_ctx *hctx = data->hctx;
404 unsigned int tag;
405
406 /* It may not be in flight yet (this is where
407 * the REQ_ATOMIC_STARTED flag comes in). The requests are
408 * statically allocated, so we know it's always safe to access the
409 * memory associated with a bit offset into ->rqs[].
410 */
411 tag = 0;
412 do {
413 struct request *rq;
414
415 tag = find_next_zero_bit(free_tags, hctx->queue_depth, tag);
416 if (tag >= hctx->queue_depth)
417 break;
418
419 rq = hctx->rqs[tag++];
420
421 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
422 continue;
423
424 blk_rq_check_expired(rq, data->next, data->next_set);
425 } while (1);
426}
427
428static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
429 unsigned long *next,
430 unsigned int *next_set)
431{
432 struct blk_mq_timeout_data data = {
433 .hctx = hctx,
434 .next = next,
435 .next_set = next_set,
436 };
437
438 /*
439 * Ask the tagging code to iterate busy requests, so we can
440 * check them for timeout.
441 */
442 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
443}
444
445static void blk_mq_rq_timer(unsigned long data)
446{
447 struct request_queue *q = (struct request_queue *) data;
448 struct blk_mq_hw_ctx *hctx;
449 unsigned long next = 0;
450 int i, next_set = 0;
451
452 queue_for_each_hw_ctx(q, hctx, i)
453 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
454
455 if (next_set)
456 mod_timer(&q->timeout, round_jiffies_up(next));
457}
458
459/*
460 * Reverse check our software queue for entries that we could potentially
461 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
462 * too much time checking for merges.
463 */
464static bool blk_mq_attempt_merge(struct request_queue *q,
465 struct blk_mq_ctx *ctx, struct bio *bio)
466{
467 struct request *rq;
468 int checked = 8;
469
470 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
471 int el_ret;
472
473 if (!checked--)
474 break;
475
476 if (!blk_rq_merge_ok(rq, bio))
477 continue;
478
479 el_ret = blk_try_merge(rq, bio);
480 if (el_ret == ELEVATOR_BACK_MERGE) {
481 if (bio_attempt_back_merge(q, rq, bio)) {
482 ctx->rq_merged++;
483 return true;
484 }
485 break;
486 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
487 if (bio_attempt_front_merge(q, rq, bio)) {
488 ctx->rq_merged++;
489 return true;
490 }
491 break;
492 }
493 }
494
495 return false;
496}
497
498void blk_mq_add_timer(struct request *rq)
499{
500 __blk_add_timer(rq, NULL);
501}
502
503/*
504 * Run this hardware queue, pulling any software queues mapped to it in.
505 * Note that this function currently has various problems around ordering
506 * of IO. In particular, we'd like FIFO behaviour on handling existing
507 * items on the hctx->dispatch list. Ignore that for now.
508 */
509static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
510{
511 struct request_queue *q = hctx->queue;
512 struct blk_mq_ctx *ctx;
513 struct request *rq;
514 LIST_HEAD(rq_list);
515 int bit, queued;
516
517 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
518 return;
519
520 hctx->run++;
521
522 /*
523 * Touch any software queue that has pending entries.
524 */
525 for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
526 clear_bit(bit, hctx->ctx_map);
527 ctx = hctx->ctxs[bit];
528 BUG_ON(bit != ctx->index_hw);
529
530 spin_lock(&ctx->lock);
531 list_splice_tail_init(&ctx->rq_list, &rq_list);
532 spin_unlock(&ctx->lock);
533 }
534
535 /*
536 * If we have previous entries on our dispatch list, grab them
537 * and stuff them at the front for more fair dispatch.
538 */
539 if (!list_empty_careful(&hctx->dispatch)) {
540 spin_lock(&hctx->lock);
541 if (!list_empty(&hctx->dispatch))
542 list_splice_init(&hctx->dispatch, &rq_list);
543 spin_unlock(&hctx->lock);
544 }
545
546 /*
547 * Delete and return all entries from our dispatch list
548 */
549 queued = 0;
550
551 /*
552 * Now process all the entries, sending them to the driver.
553 */
554 while (!list_empty(&rq_list)) {
555 int ret;
556
557 rq = list_first_entry(&rq_list, struct request, queuelist);
558 list_del_init(&rq->queuelist);
559
560 blk_mq_start_request(rq, list_empty(&rq_list));
561
562 ret = q->mq_ops->queue_rq(hctx, rq);
563 switch (ret) {
564 case BLK_MQ_RQ_QUEUE_OK:
565 queued++;
566 continue;
567 case BLK_MQ_RQ_QUEUE_BUSY:
568 /*
569 * FIXME: we should have a mechanism to stop the queue
570 * like blk_stop_queue, otherwise we will waste cpu
571 * time
572 */
573 list_add(&rq->queuelist, &rq_list);
574 blk_mq_requeue_request(rq);
575 break;
576 default:
577 pr_err("blk-mq: bad return on queue: %d\n", ret);
578 case BLK_MQ_RQ_QUEUE_ERROR:
579 rq->errors = -EIO;
580 blk_mq_end_io(rq, rq->errors);
581 break;
582 }
583
584 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
585 break;
586 }
587
588 if (!queued)
589 hctx->dispatched[0]++;
590 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
591 hctx->dispatched[ilog2(queued) + 1]++;
592
593 /*
594 * Any items that need requeuing? Stuff them into hctx->dispatch,
595 * that is where we will continue on next queue run.
596 */
597 if (!list_empty(&rq_list)) {
598 spin_lock(&hctx->lock);
599 list_splice(&rq_list, &hctx->dispatch);
600 spin_unlock(&hctx->lock);
601 }
602}
603
604void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
605{
606 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
607 return;
608
609 if (!async)
610 __blk_mq_run_hw_queue(hctx);
611 else {
612 struct request_queue *q = hctx->queue;
613
614 kblockd_schedule_delayed_work(q, &hctx->delayed_work, 0);
615 }
616}
617
618void blk_mq_run_queues(struct request_queue *q, bool async)
619{
620 struct blk_mq_hw_ctx *hctx;
621 int i;
622
623 queue_for_each_hw_ctx(q, hctx, i) {
624 if ((!blk_mq_hctx_has_pending(hctx) &&
625 list_empty_careful(&hctx->dispatch)) ||
626 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
627 continue;
628
629 blk_mq_run_hw_queue(hctx, async);
630 }
631}
632EXPORT_SYMBOL(blk_mq_run_queues);
633
634void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
635{
636 cancel_delayed_work(&hctx->delayed_work);
637 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
638}
639EXPORT_SYMBOL(blk_mq_stop_hw_queue);
640
641void blk_mq_stop_hw_queues(struct request_queue *q)
642{
643 struct blk_mq_hw_ctx *hctx;
644 int i;
645
646 queue_for_each_hw_ctx(q, hctx, i)
647 blk_mq_stop_hw_queue(hctx);
648}
649EXPORT_SYMBOL(blk_mq_stop_hw_queues);
650
651void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
652{
653 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
654 __blk_mq_run_hw_queue(hctx);
655}
656EXPORT_SYMBOL(blk_mq_start_hw_queue);
657
658void blk_mq_start_stopped_hw_queues(struct request_queue *q)
659{
660 struct blk_mq_hw_ctx *hctx;
661 int i;
662
663 queue_for_each_hw_ctx(q, hctx, i) {
664 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
665 continue;
666
667 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
668 blk_mq_run_hw_queue(hctx, true);
669 }
670}
671EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
672
673static void blk_mq_work_fn(struct work_struct *work)
674{
675 struct blk_mq_hw_ctx *hctx;
676
677 hctx = container_of(work, struct blk_mq_hw_ctx, delayed_work.work);
678 __blk_mq_run_hw_queue(hctx);
679}
680
681static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
682 struct request *rq, bool at_head)
683{
684 struct blk_mq_ctx *ctx = rq->mq_ctx;
685
686 trace_block_rq_insert(hctx->queue, rq);
687
688 if (at_head)
689 list_add(&rq->queuelist, &ctx->rq_list);
690 else
691 list_add_tail(&rq->queuelist, &ctx->rq_list);
692 blk_mq_hctx_mark_pending(hctx, ctx);
693
694 /*
695 * We do this early, to ensure we are on the right CPU.
696 */
697 blk_mq_add_timer(rq);
698}
699
700void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
701 bool async)
702{
703 struct request_queue *q = rq->q;
704 struct blk_mq_hw_ctx *hctx;
705 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
706
707 current_ctx = blk_mq_get_ctx(q);
708 if (!cpu_online(ctx->cpu))
709 rq->mq_ctx = ctx = current_ctx;
710
711 hctx = q->mq_ops->map_queue(q, ctx->cpu);
712
713 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
714 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
715 blk_insert_flush(rq);
716 } else {
717 spin_lock(&ctx->lock);
718 __blk_mq_insert_request(hctx, rq, at_head);
719 spin_unlock(&ctx->lock);
720 }
721
722 blk_mq_put_ctx(current_ctx);
723
724 if (run_queue)
725 blk_mq_run_hw_queue(hctx, async);
726}
727
728static void blk_mq_insert_requests(struct request_queue *q,
729 struct blk_mq_ctx *ctx,
730 struct list_head *list,
731 int depth,
732 bool from_schedule)
733
734{
735 struct blk_mq_hw_ctx *hctx;
736 struct blk_mq_ctx *current_ctx;
737
738 trace_block_unplug(q, depth, !from_schedule);
739
740 current_ctx = blk_mq_get_ctx(q);
741
742 if (!cpu_online(ctx->cpu))
743 ctx = current_ctx;
744 hctx = q->mq_ops->map_queue(q, ctx->cpu);
745
746 /*
747 * preemption doesn't flush plug list, so it's possible ctx->cpu is
748 * offline now
749 */
750 spin_lock(&ctx->lock);
751 while (!list_empty(list)) {
752 struct request *rq;
753
754 rq = list_first_entry(list, struct request, queuelist);
755 list_del_init(&rq->queuelist);
756 rq->mq_ctx = ctx;
757 __blk_mq_insert_request(hctx, rq, false);
758 }
759 spin_unlock(&ctx->lock);
760
761 blk_mq_put_ctx(current_ctx);
762
763 blk_mq_run_hw_queue(hctx, from_schedule);
764}
765
766static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
767{
768 struct request *rqa = container_of(a, struct request, queuelist);
769 struct request *rqb = container_of(b, struct request, queuelist);
770
771 return !(rqa->mq_ctx < rqb->mq_ctx ||
772 (rqa->mq_ctx == rqb->mq_ctx &&
773 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
774}
775
776void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
777{
778 struct blk_mq_ctx *this_ctx;
779 struct request_queue *this_q;
780 struct request *rq;
781 LIST_HEAD(list);
782 LIST_HEAD(ctx_list);
783 unsigned int depth;
784
785 list_splice_init(&plug->mq_list, &list);
786
787 list_sort(NULL, &list, plug_ctx_cmp);
788
789 this_q = NULL;
790 this_ctx = NULL;
791 depth = 0;
792
793 while (!list_empty(&list)) {
794 rq = list_entry_rq(list.next);
795 list_del_init(&rq->queuelist);
796 BUG_ON(!rq->q);
797 if (rq->mq_ctx != this_ctx) {
798 if (this_ctx) {
799 blk_mq_insert_requests(this_q, this_ctx,
800 &ctx_list, depth,
801 from_schedule);
802 }
803
804 this_ctx = rq->mq_ctx;
805 this_q = rq->q;
806 depth = 0;
807 }
808
809 depth++;
810 list_add_tail(&rq->queuelist, &ctx_list);
811 }
812
813 /*
814 * If 'this_ctx' is set, we know we have entries to complete
815 * on 'ctx_list'. Do those.
816 */
817 if (this_ctx) {
818 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
819 from_schedule);
820 }
821}
822
823static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
824{
825 init_request_from_bio(rq, bio);
826 blk_account_io_start(rq, 1);
827}
828
829static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
830{
831 struct blk_mq_hw_ctx *hctx;
832 struct blk_mq_ctx *ctx;
833 const int is_sync = rw_is_sync(bio->bi_rw);
834 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
835 int rw = bio_data_dir(bio);
836 struct request *rq;
837 unsigned int use_plug, request_count = 0;
838
839 /*
840 * If we have multiple hardware queues, just go directly to
841 * one of those for sync IO.
842 */
843 use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
844
845 blk_queue_bounce(q, &bio);
846
847 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
848 bio_endio(bio, -EIO);
849 return;
850 }
851
852 if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
853 return;
854
855 if (blk_mq_queue_enter(q)) {
856 bio_endio(bio, -EIO);
857 return;
858 }
859
860 ctx = blk_mq_get_ctx(q);
861 hctx = q->mq_ops->map_queue(q, ctx->cpu);
862
863 if (is_sync)
864 rw |= REQ_SYNC;
865 trace_block_getrq(q, bio, rw);
866 rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
867 if (likely(rq))
868 blk_mq_rq_ctx_init(q, ctx, rq, rw);
869 else {
870 blk_mq_put_ctx(ctx);
871 trace_block_sleeprq(q, bio, rw);
872 rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
873 false);
874 ctx = rq->mq_ctx;
875 hctx = q->mq_ops->map_queue(q, ctx->cpu);
876 }
877
878 hctx->queued++;
879
880 if (unlikely(is_flush_fua)) {
881 blk_mq_bio_to_request(rq, bio);
882 blk_mq_put_ctx(ctx);
883 blk_insert_flush(rq);
884 goto run_queue;
885 }
886
887 /*
888 * A task plug currently exists. Since this is completely lockless,
889 * utilize that to temporarily store requests until the task is
890 * either done or scheduled away.
891 */
892 if (use_plug) {
893 struct blk_plug *plug = current->plug;
894
895 if (plug) {
896 blk_mq_bio_to_request(rq, bio);
897 if (list_empty(&plug->mq_list))
898 trace_block_plug(q);
899 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
900 blk_flush_plug_list(plug, false);
901 trace_block_plug(q);
902 }
903 list_add_tail(&rq->queuelist, &plug->mq_list);
904 blk_mq_put_ctx(ctx);
905 return;
906 }
907 }
908
909 spin_lock(&ctx->lock);
910
911 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
912 blk_mq_attempt_merge(q, ctx, bio))
913 __blk_mq_free_request(hctx, ctx, rq);
914 else {
915 blk_mq_bio_to_request(rq, bio);
916 __blk_mq_insert_request(hctx, rq, false);
917 }
918
919 spin_unlock(&ctx->lock);
920 blk_mq_put_ctx(ctx);
921
922 /*
923 * For a SYNC request, send it to the hardware immediately. For an
924 * ASYNC request, just ensure that we run it later on. The latter
925 * allows for merging opportunities and more efficient dispatching.
926 */
927run_queue:
928 blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
929}
930
931/*
932 * Default mapping to a software queue, since we use one per CPU.
933 */
934struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
935{
936 return q->queue_hw_ctx[q->mq_map[cpu]];
937}
938EXPORT_SYMBOL(blk_mq_map_queue);
939
940struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_reg *reg,
941 unsigned int hctx_index)
942{
943 return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
944 GFP_KERNEL | __GFP_ZERO, reg->numa_node);
945}
946EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
947
948void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
949 unsigned int hctx_index)
950{
951 kfree(hctx);
952}
953EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
954
955static void blk_mq_hctx_notify(void *data, unsigned long action,
956 unsigned int cpu)
957{
958 struct blk_mq_hw_ctx *hctx = data;
959 struct request_queue *q = hctx->queue;
960 struct blk_mq_ctx *ctx;
961 LIST_HEAD(tmp);
962
963 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
964 return;
965
966 /*
967 * Move ctx entries to new CPU, if this one is going away.
968 */
969 ctx = __blk_mq_get_ctx(q, cpu);
970
971 spin_lock(&ctx->lock);
972 if (!list_empty(&ctx->rq_list)) {
973 list_splice_init(&ctx->rq_list, &tmp);
974 clear_bit(ctx->index_hw, hctx->ctx_map);
975 }
976 spin_unlock(&ctx->lock);
977
978 if (list_empty(&tmp))
979 return;
980
981 ctx = blk_mq_get_ctx(q);
982 spin_lock(&ctx->lock);
983
984 while (!list_empty(&tmp)) {
985 struct request *rq;
986
987 rq = list_first_entry(&tmp, struct request, queuelist);
988 rq->mq_ctx = ctx;
989 list_move_tail(&rq->queuelist, &ctx->rq_list);
990 }
991
992 hctx = q->mq_ops->map_queue(q, ctx->cpu);
993 blk_mq_hctx_mark_pending(hctx, ctx);
994
995 spin_unlock(&ctx->lock);
996 blk_mq_put_ctx(ctx);
997
998 blk_mq_run_hw_queue(hctx, true);
999}
1000
1001static int blk_mq_init_hw_commands(struct blk_mq_hw_ctx *hctx,
1002 int (*init)(void *, struct blk_mq_hw_ctx *,
1003 struct request *, unsigned int),
1004 void *data)
1005{
1006 unsigned int i;
1007 int ret = 0;
1008
1009 for (i = 0; i < hctx->queue_depth; i++) {
1010 struct request *rq = hctx->rqs[i];
1011
1012 ret = init(data, hctx, rq, i);
1013 if (ret)
1014 break;
1015 }
1016
1017 return ret;
1018}
1019
1020int blk_mq_init_commands(struct request_queue *q,
1021 int (*init)(void *, struct blk_mq_hw_ctx *,
1022 struct request *, unsigned int),
1023 void *data)
1024{
1025 struct blk_mq_hw_ctx *hctx;
1026 unsigned int i;
1027 int ret = 0;
1028
1029 queue_for_each_hw_ctx(q, hctx, i) {
1030 ret = blk_mq_init_hw_commands(hctx, init, data);
1031 if (ret)
1032 break;
1033 }
1034
1035 return ret;
1036}
1037EXPORT_SYMBOL(blk_mq_init_commands);
1038
1039static void blk_mq_free_hw_commands(struct blk_mq_hw_ctx *hctx,
1040 void (*free)(void *, struct blk_mq_hw_ctx *,
1041 struct request *, unsigned int),
1042 void *data)
1043{
1044 unsigned int i;
1045
1046 for (i = 0; i < hctx->queue_depth; i++) {
1047 struct request *rq = hctx->rqs[i];
1048
1049 free(data, hctx, rq, i);
1050 }
1051}
1052
1053void blk_mq_free_commands(struct request_queue *q,
1054 void (*free)(void *, struct blk_mq_hw_ctx *,
1055 struct request *, unsigned int),
1056 void *data)
1057{
1058 struct blk_mq_hw_ctx *hctx;
1059 unsigned int i;
1060
1061 queue_for_each_hw_ctx(q, hctx, i)
1062 blk_mq_free_hw_commands(hctx, free, data);
1063}
1064EXPORT_SYMBOL(blk_mq_free_commands);
1065
1066static void blk_mq_free_rq_map(struct blk_mq_hw_ctx *hctx)
1067{
1068 struct page *page;
1069
1070 while (!list_empty(&hctx->page_list)) {
1071 page = list_first_entry(&hctx->page_list, struct page, lru);
1072 list_del_init(&page->lru);
1073 __free_pages(page, page->private);
1074 }
1075
1076 kfree(hctx->rqs);
1077
1078 if (hctx->tags)
1079 blk_mq_free_tags(hctx->tags);
1080}
1081
1082static size_t order_to_size(unsigned int order)
1083{
1084 size_t ret = PAGE_SIZE;
1085
1086 while (order--)
1087 ret *= 2;
1088
1089 return ret;
1090}
1091
1092static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
1093 unsigned int reserved_tags, int node)
1094{
1095 unsigned int i, j, entries_per_page, max_order = 4;
1096 size_t rq_size, left;
1097
1098 INIT_LIST_HEAD(&hctx->page_list);
1099
1100 hctx->rqs = kmalloc_node(hctx->queue_depth * sizeof(struct request *),
1101 GFP_KERNEL, node);
1102 if (!hctx->rqs)
1103 return -ENOMEM;
1104
1105 /*
1106 * rq_size is the size of the request plus driver payload, rounded
1107 * to the cacheline size
1108 */
1109 rq_size = round_up(sizeof(struct request) + hctx->cmd_size,
1110 cache_line_size());
1111 left = rq_size * hctx->queue_depth;
1112
1113 for (i = 0; i < hctx->queue_depth;) {
1114 int this_order = max_order;
1115 struct page *page;
1116 int to_do;
1117 void *p;
1118
1119 while (left < order_to_size(this_order - 1) && this_order)
1120 this_order--;
1121
1122 do {
1123 page = alloc_pages_node(node, GFP_KERNEL, this_order);
1124 if (page)
1125 break;
1126 if (!this_order--)
1127 break;
1128 if (order_to_size(this_order) < rq_size)
1129 break;
1130 } while (1);
1131
1132 if (!page)
1133 break;
1134
1135 page->private = this_order;
1136 list_add_tail(&page->lru, &hctx->page_list);
1137
1138 p = page_address(page);
1139 entries_per_page = order_to_size(this_order) / rq_size;
1140 to_do = min(entries_per_page, hctx->queue_depth - i);
1141 left -= to_do * rq_size;
1142 for (j = 0; j < to_do; j++) {
1143 hctx->rqs[i] = p;
1144 blk_mq_rq_init(hctx, hctx->rqs[i]);
1145 p += rq_size;
1146 i++;
1147 }
1148 }
1149
1150 if (i < (reserved_tags + BLK_MQ_TAG_MIN))
1151 goto err_rq_map;
1152 else if (i != hctx->queue_depth) {
1153 hctx->queue_depth = i;
1154 pr_warn("%s: queue depth set to %u because of low memory\n",
1155 __func__, i);
1156 }
1157
1158 hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node);
1159 if (!hctx->tags) {
1160err_rq_map:
1161 blk_mq_free_rq_map(hctx);
1162 return -ENOMEM;
1163 }
1164
1165 return 0;
1166}
1167
1168static int blk_mq_init_hw_queues(struct request_queue *q,
1169 struct blk_mq_reg *reg, void *driver_data)
1170{
1171 struct blk_mq_hw_ctx *hctx;
1172 unsigned int i, j;
1173
1174 /*
1175 * Initialize hardware queues
1176 */
1177 queue_for_each_hw_ctx(q, hctx, i) {
1178 unsigned int num_maps;
1179 int node;
1180
1181 node = hctx->numa_node;
1182 if (node == NUMA_NO_NODE)
1183 node = hctx->numa_node = reg->numa_node;
1184
1185 INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn);
1186 spin_lock_init(&hctx->lock);
1187 INIT_LIST_HEAD(&hctx->dispatch);
1188 hctx->queue = q;
1189 hctx->queue_num = i;
1190 hctx->flags = reg->flags;
1191 hctx->queue_depth = reg->queue_depth;
1192 hctx->cmd_size = reg->cmd_size;
1193
1194 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1195 blk_mq_hctx_notify, hctx);
1196 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1197
1198 if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node))
1199 break;
1200
1201 /*
1202 * Allocate space for all possible cpus to avoid allocation in
1203 * runtime
1204 */
1205 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1206 GFP_KERNEL, node);
1207 if (!hctx->ctxs)
1208 break;
1209
1210 num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1211 hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1212 GFP_KERNEL, node);
1213 if (!hctx->ctx_map)
1214 break;
1215
1216 hctx->nr_ctx_map = num_maps;
1217 hctx->nr_ctx = 0;
1218
1219 if (reg->ops->init_hctx &&
1220 reg->ops->init_hctx(hctx, driver_data, i))
1221 break;
1222 }
1223
1224 if (i == q->nr_hw_queues)
1225 return 0;
1226
1227 /*
1228 * Init failed
1229 */
1230 queue_for_each_hw_ctx(q, hctx, j) {
1231 if (i == j)
1232 break;
1233
1234 if (reg->ops->exit_hctx)
1235 reg->ops->exit_hctx(hctx, j);
1236
1237 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1238 blk_mq_free_rq_map(hctx);
1239 kfree(hctx->ctxs);
1240 }
1241
1242 return 1;
1243}
1244
1245static void blk_mq_init_cpu_queues(struct request_queue *q,
1246 unsigned int nr_hw_queues)
1247{
1248 unsigned int i;
1249
1250 for_each_possible_cpu(i) {
1251 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1252 struct blk_mq_hw_ctx *hctx;
1253
1254 memset(__ctx, 0, sizeof(*__ctx));
1255 __ctx->cpu = i;
1256 spin_lock_init(&__ctx->lock);
1257 INIT_LIST_HEAD(&__ctx->rq_list);
1258 __ctx->queue = q;
1259
1260 /* If the cpu isn't online, the cpu is mapped to first hctx */
1261 hctx = q->mq_ops->map_queue(q, i);
1262 hctx->nr_ctx++;
1263
1264 if (!cpu_online(i))
1265 continue;
1266
1267 /*
1268 * Set local node, IFF we have more than one hw queue. If
1269 * not, we remain on the home node of the device
1270 */
1271 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1272 hctx->numa_node = cpu_to_node(i);
1273 }
1274}
1275
1276static void blk_mq_map_swqueue(struct request_queue *q)
1277{
1278 unsigned int i;
1279 struct blk_mq_hw_ctx *hctx;
1280 struct blk_mq_ctx *ctx;
1281
1282 queue_for_each_hw_ctx(q, hctx, i) {
1283 hctx->nr_ctx = 0;
1284 }
1285
1286 /*
1287 * Map software to hardware queues
1288 */
1289 queue_for_each_ctx(q, ctx, i) {
1290 /* If the cpu isn't online, the cpu is mapped to first hctx */
1291 hctx = q->mq_ops->map_queue(q, i);
1292 ctx->index_hw = hctx->nr_ctx;
1293 hctx->ctxs[hctx->nr_ctx++] = ctx;
1294 }
1295}
1296
1297struct request_queue *blk_mq_init_queue(struct blk_mq_reg *reg,
1298 void *driver_data)
1299{
1300 struct blk_mq_hw_ctx **hctxs;
1301 struct blk_mq_ctx *ctx;
1302 struct request_queue *q;
1303 int i;
1304
1305 if (!reg->nr_hw_queues ||
1306 !reg->ops->queue_rq || !reg->ops->map_queue ||
1307 !reg->ops->alloc_hctx || !reg->ops->free_hctx)
1308 return ERR_PTR(-EINVAL);
1309
1310 if (!reg->queue_depth)
1311 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1312 else if (reg->queue_depth > BLK_MQ_MAX_DEPTH) {
1313 pr_err("blk-mq: queuedepth too large (%u)\n", reg->queue_depth);
1314 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1315 }
1316
1317 if (reg->queue_depth < (reg->reserved_tags + BLK_MQ_TAG_MIN))
1318 return ERR_PTR(-EINVAL);
1319
1320 ctx = alloc_percpu(struct blk_mq_ctx);
1321 if (!ctx)
1322 return ERR_PTR(-ENOMEM);
1323
1324 hctxs = kmalloc_node(reg->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1325 reg->numa_node);
1326
1327 if (!hctxs)
1328 goto err_percpu;
1329
1330 for (i = 0; i < reg->nr_hw_queues; i++) {
1331 hctxs[i] = reg->ops->alloc_hctx(reg, i);
1332 if (!hctxs[i])
1333 goto err_hctxs;
1334
1335 hctxs[i]->numa_node = NUMA_NO_NODE;
1336 hctxs[i]->queue_num = i;
1337 }
1338
1339 q = blk_alloc_queue_node(GFP_KERNEL, reg->numa_node);
1340 if (!q)
1341 goto err_hctxs;
1342
1343 q->mq_map = blk_mq_make_queue_map(reg);
1344 if (!q->mq_map)
1345 goto err_map;
1346
1347 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1348 blk_queue_rq_timeout(q, 30000);
1349
1350 q->nr_queues = nr_cpu_ids;
1351 q->nr_hw_queues = reg->nr_hw_queues;
1352
1353 q->queue_ctx = ctx;
1354 q->queue_hw_ctx = hctxs;
1355
1356 q->mq_ops = reg->ops;
1357 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1358
1359 q->sg_reserved_size = INT_MAX;
1360
1361 blk_queue_make_request(q, blk_mq_make_request);
1362 blk_queue_rq_timed_out(q, reg->ops->timeout);
1363 if (reg->timeout)
1364 blk_queue_rq_timeout(q, reg->timeout);
1365
1366 if (reg->ops->complete)
1367 blk_queue_softirq_done(q, reg->ops->complete);
1368
1369 blk_mq_init_flush(q);
1370 blk_mq_init_cpu_queues(q, reg->nr_hw_queues);
1371
1372 q->flush_rq = kzalloc(round_up(sizeof(struct request) + reg->cmd_size,
1373 cache_line_size()), GFP_KERNEL);
1374 if (!q->flush_rq)
1375 goto err_hw;
1376
1377 if (blk_mq_init_hw_queues(q, reg, driver_data))
1378 goto err_flush_rq;
1379
1380 blk_mq_map_swqueue(q);
1381
1382 mutex_lock(&all_q_mutex);
1383 list_add_tail(&q->all_q_node, &all_q_list);
1384 mutex_unlock(&all_q_mutex);
1385
1386 return q;
1387
1388err_flush_rq:
1389 kfree(q->flush_rq);
1390err_hw:
1391 kfree(q->mq_map);
1392err_map:
1393 blk_cleanup_queue(q);
1394err_hctxs:
1395 for (i = 0; i < reg->nr_hw_queues; i++) {
1396 if (!hctxs[i])
1397 break;
1398 reg->ops->free_hctx(hctxs[i], i);
1399 }
1400 kfree(hctxs);
1401err_percpu:
1402 free_percpu(ctx);
1403 return ERR_PTR(-ENOMEM);
1404}
1405EXPORT_SYMBOL(blk_mq_init_queue);
1406
1407void blk_mq_free_queue(struct request_queue *q)
1408{
1409 struct blk_mq_hw_ctx *hctx;
1410 int i;
1411
1412 queue_for_each_hw_ctx(q, hctx, i) {
1413 kfree(hctx->ctx_map);
1414 kfree(hctx->ctxs);
1415 blk_mq_free_rq_map(hctx);
1416 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1417 if (q->mq_ops->exit_hctx)
1418 q->mq_ops->exit_hctx(hctx, i);
1419 q->mq_ops->free_hctx(hctx, i);
1420 }
1421
1422 free_percpu(q->queue_ctx);
1423 kfree(q->queue_hw_ctx);
1424 kfree(q->mq_map);
1425
1426 q->queue_ctx = NULL;
1427 q->queue_hw_ctx = NULL;
1428 q->mq_map = NULL;
1429
1430 mutex_lock(&all_q_mutex);
1431 list_del_init(&q->all_q_node);
1432 mutex_unlock(&all_q_mutex);
1433}
1434
1435/* Basically redo blk_mq_init_queue with queue frozen */
1436static void blk_mq_queue_reinit(struct request_queue *q)
1437{
1438 blk_mq_freeze_queue(q);
1439
1440 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1441
1442 /*
1443 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1444 * we should change hctx numa_node according to new topology (this
1445 * involves free and re-allocate memory, worthy doing?)
1446 */
1447
1448 blk_mq_map_swqueue(q);
1449
1450 blk_mq_unfreeze_queue(q);
1451}
1452
1453static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1454 unsigned long action, void *hcpu)
1455{
1456 struct request_queue *q;
1457
1458 /*
1459 * Before new mapping is established, hotadded cpu might already start
1460 * handling requests. This doesn't break anything as we map offline
1461 * CPUs to first hardware queue. We will re-init queue below to get
1462 * optimal settings.
1463 */
1464 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1465 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1466 return NOTIFY_OK;
1467
1468 mutex_lock(&all_q_mutex);
1469 list_for_each_entry(q, &all_q_list, all_q_node)
1470 blk_mq_queue_reinit(q);
1471 mutex_unlock(&all_q_mutex);
1472 return NOTIFY_OK;
1473}
1474
1475void blk_mq_disable_hotplug(void)
1476{
1477 mutex_lock(&all_q_mutex);
1478}
1479
1480void blk_mq_enable_hotplug(void)
1481{
1482 mutex_unlock(&all_q_mutex);
1483}
1484
1485static int __init blk_mq_init(void)
1486{
1487 blk_mq_cpu_init();
1488
1489 /* Must be called after percpu_counter_hotcpu_callback() */
1490 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1491
1492 return 0;
1493}
1494subsys_initcall(blk_mq_init);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Block multiqueue core code
4 *
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
7 */
8#include <linux/kernel.h>
9#include <linux/module.h>
10#include <linux/backing-dev.h>
11#include <linux/bio.h>
12#include <linux/blkdev.h>
13#include <linux/kmemleak.h>
14#include <linux/mm.h>
15#include <linux/init.h>
16#include <linux/slab.h>
17#include <linux/workqueue.h>
18#include <linux/smp.h>
19#include <linux/llist.h>
20#include <linux/list_sort.h>
21#include <linux/cpu.h>
22#include <linux/cache.h>
23#include <linux/sched/sysctl.h>
24#include <linux/sched/topology.h>
25#include <linux/sched/signal.h>
26#include <linux/delay.h>
27#include <linux/crash_dump.h>
28#include <linux/prefetch.h>
29#include <linux/blk-crypto.h>
30
31#include <trace/events/block.h>
32
33#include <linux/blk-mq.h>
34#include <linux/t10-pi.h>
35#include "blk.h"
36#include "blk-mq.h"
37#include "blk-mq-debugfs.h"
38#include "blk-mq-tag.h"
39#include "blk-pm.h"
40#include "blk-stat.h"
41#include "blk-mq-sched.h"
42#include "blk-rq-qos.h"
43
44static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
45
46static void blk_mq_poll_stats_start(struct request_queue *q);
47static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
48
49static int blk_mq_poll_stats_bkt(const struct request *rq)
50{
51 int ddir, sectors, bucket;
52
53 ddir = rq_data_dir(rq);
54 sectors = blk_rq_stats_sectors(rq);
55
56 bucket = ddir + 2 * ilog2(sectors);
57
58 if (bucket < 0)
59 return -1;
60 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
61 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
62
63 return bucket;
64}
65
66/*
67 * Check if any of the ctx, dispatch list or elevator
68 * have pending work in this hardware queue.
69 */
70static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
71{
72 return !list_empty_careful(&hctx->dispatch) ||
73 sbitmap_any_bit_set(&hctx->ctx_map) ||
74 blk_mq_sched_has_work(hctx);
75}
76
77/*
78 * Mark this ctx as having pending work in this hardware queue
79 */
80static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
81 struct blk_mq_ctx *ctx)
82{
83 const int bit = ctx->index_hw[hctx->type];
84
85 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
86 sbitmap_set_bit(&hctx->ctx_map, bit);
87}
88
89static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
90 struct blk_mq_ctx *ctx)
91{
92 const int bit = ctx->index_hw[hctx->type];
93
94 sbitmap_clear_bit(&hctx->ctx_map, bit);
95}
96
97struct mq_inflight {
98 struct hd_struct *part;
99 unsigned int inflight[2];
100};
101
102static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
103 struct request *rq, void *priv,
104 bool reserved)
105{
106 struct mq_inflight *mi = priv;
107
108 if (rq->part == mi->part)
109 mi->inflight[rq_data_dir(rq)]++;
110
111 return true;
112}
113
114unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
115{
116 struct mq_inflight mi = { .part = part };
117
118 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
119
120 return mi.inflight[0] + mi.inflight[1];
121}
122
123void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
124 unsigned int inflight[2])
125{
126 struct mq_inflight mi = { .part = part };
127
128 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
129 inflight[0] = mi.inflight[0];
130 inflight[1] = mi.inflight[1];
131}
132
133void blk_freeze_queue_start(struct request_queue *q)
134{
135 mutex_lock(&q->mq_freeze_lock);
136 if (++q->mq_freeze_depth == 1) {
137 percpu_ref_kill(&q->q_usage_counter);
138 mutex_unlock(&q->mq_freeze_lock);
139 if (queue_is_mq(q))
140 blk_mq_run_hw_queues(q, false);
141 } else {
142 mutex_unlock(&q->mq_freeze_lock);
143 }
144}
145EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
146
147void blk_mq_freeze_queue_wait(struct request_queue *q)
148{
149 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
150}
151EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
152
153int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
154 unsigned long timeout)
155{
156 return wait_event_timeout(q->mq_freeze_wq,
157 percpu_ref_is_zero(&q->q_usage_counter),
158 timeout);
159}
160EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
161
162/*
163 * Guarantee no request is in use, so we can change any data structure of
164 * the queue afterward.
165 */
166void blk_freeze_queue(struct request_queue *q)
167{
168 /*
169 * In the !blk_mq case we are only calling this to kill the
170 * q_usage_counter, otherwise this increases the freeze depth
171 * and waits for it to return to zero. For this reason there is
172 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
173 * exported to drivers as the only user for unfreeze is blk_mq.
174 */
175 blk_freeze_queue_start(q);
176 blk_mq_freeze_queue_wait(q);
177}
178
179void blk_mq_freeze_queue(struct request_queue *q)
180{
181 /*
182 * ...just an alias to keep freeze and unfreeze actions balanced
183 * in the blk_mq_* namespace
184 */
185 blk_freeze_queue(q);
186}
187EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
188
189void blk_mq_unfreeze_queue(struct request_queue *q)
190{
191 mutex_lock(&q->mq_freeze_lock);
192 q->mq_freeze_depth--;
193 WARN_ON_ONCE(q->mq_freeze_depth < 0);
194 if (!q->mq_freeze_depth) {
195 percpu_ref_resurrect(&q->q_usage_counter);
196 wake_up_all(&q->mq_freeze_wq);
197 }
198 mutex_unlock(&q->mq_freeze_lock);
199}
200EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
201
202/*
203 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
204 * mpt3sas driver such that this function can be removed.
205 */
206void blk_mq_quiesce_queue_nowait(struct request_queue *q)
207{
208 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
209}
210EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
211
212/**
213 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
214 * @q: request queue.
215 *
216 * Note: this function does not prevent that the struct request end_io()
217 * callback function is invoked. Once this function is returned, we make
218 * sure no dispatch can happen until the queue is unquiesced via
219 * blk_mq_unquiesce_queue().
220 */
221void blk_mq_quiesce_queue(struct request_queue *q)
222{
223 struct blk_mq_hw_ctx *hctx;
224 unsigned int i;
225 bool rcu = false;
226
227 blk_mq_quiesce_queue_nowait(q);
228
229 queue_for_each_hw_ctx(q, hctx, i) {
230 if (hctx->flags & BLK_MQ_F_BLOCKING)
231 synchronize_srcu(hctx->srcu);
232 else
233 rcu = true;
234 }
235 if (rcu)
236 synchronize_rcu();
237}
238EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
239
240/*
241 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
242 * @q: request queue.
243 *
244 * This function recovers queue into the state before quiescing
245 * which is done by blk_mq_quiesce_queue.
246 */
247void blk_mq_unquiesce_queue(struct request_queue *q)
248{
249 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
250
251 /* dispatch requests which are inserted during quiescing */
252 blk_mq_run_hw_queues(q, true);
253}
254EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
255
256void blk_mq_wake_waiters(struct request_queue *q)
257{
258 struct blk_mq_hw_ctx *hctx;
259 unsigned int i;
260
261 queue_for_each_hw_ctx(q, hctx, i)
262 if (blk_mq_hw_queue_mapped(hctx))
263 blk_mq_tag_wakeup_all(hctx->tags, true);
264}
265
266/*
267 * Only need start/end time stamping if we have iostat or
268 * blk stats enabled, or using an IO scheduler.
269 */
270static inline bool blk_mq_need_time_stamp(struct request *rq)
271{
272 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
273}
274
275static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
276 unsigned int tag, u64 alloc_time_ns)
277{
278 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
279 struct request *rq = tags->static_rqs[tag];
280
281 if (data->q->elevator) {
282 rq->tag = BLK_MQ_NO_TAG;
283 rq->internal_tag = tag;
284 } else {
285 rq->tag = tag;
286 rq->internal_tag = BLK_MQ_NO_TAG;
287 }
288
289 /* csd/requeue_work/fifo_time is initialized before use */
290 rq->q = data->q;
291 rq->mq_ctx = data->ctx;
292 rq->mq_hctx = data->hctx;
293 rq->rq_flags = 0;
294 rq->cmd_flags = data->cmd_flags;
295 if (data->flags & BLK_MQ_REQ_PREEMPT)
296 rq->rq_flags |= RQF_PREEMPT;
297 if (blk_queue_io_stat(data->q))
298 rq->rq_flags |= RQF_IO_STAT;
299 INIT_LIST_HEAD(&rq->queuelist);
300 INIT_HLIST_NODE(&rq->hash);
301 RB_CLEAR_NODE(&rq->rb_node);
302 rq->rq_disk = NULL;
303 rq->part = NULL;
304#ifdef CONFIG_BLK_RQ_ALLOC_TIME
305 rq->alloc_time_ns = alloc_time_ns;
306#endif
307 if (blk_mq_need_time_stamp(rq))
308 rq->start_time_ns = ktime_get_ns();
309 else
310 rq->start_time_ns = 0;
311 rq->io_start_time_ns = 0;
312 rq->stats_sectors = 0;
313 rq->nr_phys_segments = 0;
314#if defined(CONFIG_BLK_DEV_INTEGRITY)
315 rq->nr_integrity_segments = 0;
316#endif
317 blk_crypto_rq_set_defaults(rq);
318 /* tag was already set */
319 WRITE_ONCE(rq->deadline, 0);
320
321 rq->timeout = 0;
322
323 rq->end_io = NULL;
324 rq->end_io_data = NULL;
325
326 data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
327 refcount_set(&rq->ref, 1);
328
329 if (!op_is_flush(data->cmd_flags)) {
330 struct elevator_queue *e = data->q->elevator;
331
332 rq->elv.icq = NULL;
333 if (e && e->type->ops.prepare_request) {
334 if (e->type->icq_cache)
335 blk_mq_sched_assign_ioc(rq);
336
337 e->type->ops.prepare_request(rq);
338 rq->rq_flags |= RQF_ELVPRIV;
339 }
340 }
341
342 data->hctx->queued++;
343 return rq;
344}
345
346static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
347{
348 struct request_queue *q = data->q;
349 struct elevator_queue *e = q->elevator;
350 u64 alloc_time_ns = 0;
351 unsigned int tag;
352
353 /* alloc_time includes depth and tag waits */
354 if (blk_queue_rq_alloc_time(q))
355 alloc_time_ns = ktime_get_ns();
356
357 if (data->cmd_flags & REQ_NOWAIT)
358 data->flags |= BLK_MQ_REQ_NOWAIT;
359
360 if (e) {
361 /*
362 * Flush requests are special and go directly to the
363 * dispatch list. Don't include reserved tags in the
364 * limiting, as it isn't useful.
365 */
366 if (!op_is_flush(data->cmd_flags) &&
367 e->type->ops.limit_depth &&
368 !(data->flags & BLK_MQ_REQ_RESERVED))
369 e->type->ops.limit_depth(data->cmd_flags, data);
370 }
371
372retry:
373 data->ctx = blk_mq_get_ctx(q);
374 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
375 if (!e)
376 blk_mq_tag_busy(data->hctx);
377
378 /*
379 * Waiting allocations only fail because of an inactive hctx. In that
380 * case just retry the hctx assignment and tag allocation as CPU hotplug
381 * should have migrated us to an online CPU by now.
382 */
383 tag = blk_mq_get_tag(data);
384 if (tag == BLK_MQ_NO_TAG) {
385 if (data->flags & BLK_MQ_REQ_NOWAIT)
386 return NULL;
387
388 /*
389 * Give up the CPU and sleep for a random short time to ensure
390 * that thread using a realtime scheduling class are migrated
391 * off the CPU, and thus off the hctx that is going away.
392 */
393 msleep(3);
394 goto retry;
395 }
396 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
397}
398
399struct 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 data = {
403 .q = q,
404 .flags = flags,
405 .cmd_flags = op,
406 };
407 struct request *rq;
408 int ret;
409
410 ret = blk_queue_enter(q, flags);
411 if (ret)
412 return ERR_PTR(ret);
413
414 rq = __blk_mq_alloc_request(&data);
415 if (!rq)
416 goto out_queue_exit;
417 rq->__data_len = 0;
418 rq->__sector = (sector_t) -1;
419 rq->bio = rq->biotail = NULL;
420 return rq;
421out_queue_exit:
422 blk_queue_exit(q);
423 return ERR_PTR(-EWOULDBLOCK);
424}
425EXPORT_SYMBOL(blk_mq_alloc_request);
426
427struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
428 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
429{
430 struct blk_mq_alloc_data data = {
431 .q = q,
432 .flags = flags,
433 .cmd_flags = op,
434 };
435 u64 alloc_time_ns = 0;
436 unsigned int cpu;
437 unsigned int tag;
438 int ret;
439
440 /* alloc_time includes depth and tag waits */
441 if (blk_queue_rq_alloc_time(q))
442 alloc_time_ns = ktime_get_ns();
443
444 /*
445 * If the tag allocator sleeps we could get an allocation for a
446 * different hardware context. No need to complicate the low level
447 * allocator for this for the rare use case of a command tied to
448 * a specific queue.
449 */
450 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
451 return ERR_PTR(-EINVAL);
452
453 if (hctx_idx >= q->nr_hw_queues)
454 return ERR_PTR(-EIO);
455
456 ret = blk_queue_enter(q, flags);
457 if (ret)
458 return ERR_PTR(ret);
459
460 /*
461 * Check if the hardware context is actually mapped to anything.
462 * If not tell the caller that it should skip this queue.
463 */
464 ret = -EXDEV;
465 data.hctx = q->queue_hw_ctx[hctx_idx];
466 if (!blk_mq_hw_queue_mapped(data.hctx))
467 goto out_queue_exit;
468 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
469 data.ctx = __blk_mq_get_ctx(q, cpu);
470
471 if (!q->elevator)
472 blk_mq_tag_busy(data.hctx);
473
474 ret = -EWOULDBLOCK;
475 tag = blk_mq_get_tag(&data);
476 if (tag == BLK_MQ_NO_TAG)
477 goto out_queue_exit;
478 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
479
480out_queue_exit:
481 blk_queue_exit(q);
482 return ERR_PTR(ret);
483}
484EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
485
486static void __blk_mq_free_request(struct request *rq)
487{
488 struct request_queue *q = rq->q;
489 struct blk_mq_ctx *ctx = rq->mq_ctx;
490 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
491 const int sched_tag = rq->internal_tag;
492
493 blk_crypto_free_request(rq);
494 blk_pm_mark_last_busy(rq);
495 rq->mq_hctx = NULL;
496 if (rq->tag != BLK_MQ_NO_TAG)
497 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
498 if (sched_tag != BLK_MQ_NO_TAG)
499 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
500 blk_mq_sched_restart(hctx);
501 blk_queue_exit(q);
502}
503
504void blk_mq_free_request(struct request *rq)
505{
506 struct request_queue *q = rq->q;
507 struct elevator_queue *e = q->elevator;
508 struct blk_mq_ctx *ctx = rq->mq_ctx;
509 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
510
511 if (rq->rq_flags & RQF_ELVPRIV) {
512 if (e && e->type->ops.finish_request)
513 e->type->ops.finish_request(rq);
514 if (rq->elv.icq) {
515 put_io_context(rq->elv.icq->ioc);
516 rq->elv.icq = NULL;
517 }
518 }
519
520 ctx->rq_completed[rq_is_sync(rq)]++;
521 if (rq->rq_flags & RQF_MQ_INFLIGHT)
522 atomic_dec(&hctx->nr_active);
523
524 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
525 laptop_io_completion(q->backing_dev_info);
526
527 rq_qos_done(q, rq);
528
529 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
530 if (refcount_dec_and_test(&rq->ref))
531 __blk_mq_free_request(rq);
532}
533EXPORT_SYMBOL_GPL(blk_mq_free_request);
534
535inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
536{
537 u64 now = 0;
538
539 if (blk_mq_need_time_stamp(rq))
540 now = ktime_get_ns();
541
542 if (rq->rq_flags & RQF_STATS) {
543 blk_mq_poll_stats_start(rq->q);
544 blk_stat_add(rq, now);
545 }
546
547 blk_mq_sched_completed_request(rq, now);
548
549 blk_account_io_done(rq, now);
550
551 if (rq->end_io) {
552 rq_qos_done(rq->q, rq);
553 rq->end_io(rq, error);
554 } else {
555 blk_mq_free_request(rq);
556 }
557}
558EXPORT_SYMBOL(__blk_mq_end_request);
559
560void blk_mq_end_request(struct request *rq, blk_status_t error)
561{
562 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
563 BUG();
564 __blk_mq_end_request(rq, error);
565}
566EXPORT_SYMBOL(blk_mq_end_request);
567
568/*
569 * Softirq action handler - move entries to local list and loop over them
570 * while passing them to the queue registered handler.
571 */
572static __latent_entropy void blk_done_softirq(struct softirq_action *h)
573{
574 struct list_head *cpu_list, local_list;
575
576 local_irq_disable();
577 cpu_list = this_cpu_ptr(&blk_cpu_done);
578 list_replace_init(cpu_list, &local_list);
579 local_irq_enable();
580
581 while (!list_empty(&local_list)) {
582 struct request *rq;
583
584 rq = list_entry(local_list.next, struct request, ipi_list);
585 list_del_init(&rq->ipi_list);
586 rq->q->mq_ops->complete(rq);
587 }
588}
589
590static void blk_mq_trigger_softirq(struct request *rq)
591{
592 struct list_head *list;
593 unsigned long flags;
594
595 local_irq_save(flags);
596 list = this_cpu_ptr(&blk_cpu_done);
597 list_add_tail(&rq->ipi_list, list);
598
599 /*
600 * If the list only contains our just added request, signal a raise of
601 * the softirq. If there are already entries there, someone already
602 * raised the irq but it hasn't run yet.
603 */
604 if (list->next == &rq->ipi_list)
605 raise_softirq_irqoff(BLOCK_SOFTIRQ);
606 local_irq_restore(flags);
607}
608
609static int blk_softirq_cpu_dead(unsigned int cpu)
610{
611 /*
612 * If a CPU goes away, splice its entries to the current CPU
613 * and trigger a run of the softirq
614 */
615 local_irq_disable();
616 list_splice_init(&per_cpu(blk_cpu_done, cpu),
617 this_cpu_ptr(&blk_cpu_done));
618 raise_softirq_irqoff(BLOCK_SOFTIRQ);
619 local_irq_enable();
620
621 return 0;
622}
623
624
625static void __blk_mq_complete_request_remote(void *data)
626{
627 struct request *rq = data;
628
629 /*
630 * For most of single queue controllers, there is only one irq vector
631 * for handling I/O completion, and the only irq's affinity is set
632 * to all possible CPUs. On most of ARCHs, this affinity means the irq
633 * is handled on one specific CPU.
634 *
635 * So complete I/O requests in softirq context in case of single queue
636 * devices to avoid degrading I/O performance due to irqsoff latency.
637 */
638 if (rq->q->nr_hw_queues == 1)
639 blk_mq_trigger_softirq(rq);
640 else
641 rq->q->mq_ops->complete(rq);
642}
643
644static inline bool blk_mq_complete_need_ipi(struct request *rq)
645{
646 int cpu = raw_smp_processor_id();
647
648 if (!IS_ENABLED(CONFIG_SMP) ||
649 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
650 return false;
651
652 /* same CPU or cache domain? Complete locally */
653 if (cpu == rq->mq_ctx->cpu ||
654 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
655 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
656 return false;
657
658 /* don't try to IPI to an offline CPU */
659 return cpu_online(rq->mq_ctx->cpu);
660}
661
662bool blk_mq_complete_request_remote(struct request *rq)
663{
664 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
665
666 /*
667 * For a polled request, always complete locallly, it's pointless
668 * to redirect the completion.
669 */
670 if (rq->cmd_flags & REQ_HIPRI)
671 return false;
672
673 if (blk_mq_complete_need_ipi(rq)) {
674 rq->csd.func = __blk_mq_complete_request_remote;
675 rq->csd.info = rq;
676 rq->csd.flags = 0;
677 smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
678 } else {
679 if (rq->q->nr_hw_queues > 1)
680 return false;
681 blk_mq_trigger_softirq(rq);
682 }
683
684 return true;
685}
686EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
687
688/**
689 * blk_mq_complete_request - end I/O on a request
690 * @rq: the request being processed
691 *
692 * Description:
693 * Complete a request by scheduling the ->complete_rq operation.
694 **/
695void blk_mq_complete_request(struct request *rq)
696{
697 if (!blk_mq_complete_request_remote(rq))
698 rq->q->mq_ops->complete(rq);
699}
700EXPORT_SYMBOL(blk_mq_complete_request);
701
702static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
703 __releases(hctx->srcu)
704{
705 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
706 rcu_read_unlock();
707 else
708 srcu_read_unlock(hctx->srcu, srcu_idx);
709}
710
711static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
712 __acquires(hctx->srcu)
713{
714 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
715 /* shut up gcc false positive */
716 *srcu_idx = 0;
717 rcu_read_lock();
718 } else
719 *srcu_idx = srcu_read_lock(hctx->srcu);
720}
721
722/**
723 * blk_mq_start_request - Start processing a request
724 * @rq: Pointer to request to be started
725 *
726 * Function used by device drivers to notify the block layer that a request
727 * is going to be processed now, so blk layer can do proper initializations
728 * such as starting the timeout timer.
729 */
730void blk_mq_start_request(struct request *rq)
731{
732 struct request_queue *q = rq->q;
733
734 trace_block_rq_issue(q, rq);
735
736 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
737 rq->io_start_time_ns = ktime_get_ns();
738 rq->stats_sectors = blk_rq_sectors(rq);
739 rq->rq_flags |= RQF_STATS;
740 rq_qos_issue(q, rq);
741 }
742
743 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
744
745 blk_add_timer(rq);
746 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
747
748#ifdef CONFIG_BLK_DEV_INTEGRITY
749 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
750 q->integrity.profile->prepare_fn(rq);
751#endif
752}
753EXPORT_SYMBOL(blk_mq_start_request);
754
755static void __blk_mq_requeue_request(struct request *rq)
756{
757 struct request_queue *q = rq->q;
758
759 blk_mq_put_driver_tag(rq);
760
761 trace_block_rq_requeue(q, rq);
762 rq_qos_requeue(q, rq);
763
764 if (blk_mq_request_started(rq)) {
765 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
766 rq->rq_flags &= ~RQF_TIMED_OUT;
767 }
768}
769
770void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
771{
772 __blk_mq_requeue_request(rq);
773
774 /* this request will be re-inserted to io scheduler queue */
775 blk_mq_sched_requeue_request(rq);
776
777 BUG_ON(!list_empty(&rq->queuelist));
778 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
779}
780EXPORT_SYMBOL(blk_mq_requeue_request);
781
782static void blk_mq_requeue_work(struct work_struct *work)
783{
784 struct request_queue *q =
785 container_of(work, struct request_queue, requeue_work.work);
786 LIST_HEAD(rq_list);
787 struct request *rq, *next;
788
789 spin_lock_irq(&q->requeue_lock);
790 list_splice_init(&q->requeue_list, &rq_list);
791 spin_unlock_irq(&q->requeue_lock);
792
793 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
794 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
795 continue;
796
797 rq->rq_flags &= ~RQF_SOFTBARRIER;
798 list_del_init(&rq->queuelist);
799 /*
800 * If RQF_DONTPREP, rq has contained some driver specific
801 * data, so insert it to hctx dispatch list to avoid any
802 * merge.
803 */
804 if (rq->rq_flags & RQF_DONTPREP)
805 blk_mq_request_bypass_insert(rq, false, false);
806 else
807 blk_mq_sched_insert_request(rq, true, false, false);
808 }
809
810 while (!list_empty(&rq_list)) {
811 rq = list_entry(rq_list.next, struct request, queuelist);
812 list_del_init(&rq->queuelist);
813 blk_mq_sched_insert_request(rq, false, false, false);
814 }
815
816 blk_mq_run_hw_queues(q, false);
817}
818
819void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
820 bool kick_requeue_list)
821{
822 struct request_queue *q = rq->q;
823 unsigned long flags;
824
825 /*
826 * We abuse this flag that is otherwise used by the I/O scheduler to
827 * request head insertion from the workqueue.
828 */
829 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
830
831 spin_lock_irqsave(&q->requeue_lock, flags);
832 if (at_head) {
833 rq->rq_flags |= RQF_SOFTBARRIER;
834 list_add(&rq->queuelist, &q->requeue_list);
835 } else {
836 list_add_tail(&rq->queuelist, &q->requeue_list);
837 }
838 spin_unlock_irqrestore(&q->requeue_lock, flags);
839
840 if (kick_requeue_list)
841 blk_mq_kick_requeue_list(q);
842}
843
844void blk_mq_kick_requeue_list(struct request_queue *q)
845{
846 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
847}
848EXPORT_SYMBOL(blk_mq_kick_requeue_list);
849
850void blk_mq_delay_kick_requeue_list(struct request_queue *q,
851 unsigned long msecs)
852{
853 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
854 msecs_to_jiffies(msecs));
855}
856EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
857
858struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
859{
860 if (tag < tags->nr_tags) {
861 prefetch(tags->rqs[tag]);
862 return tags->rqs[tag];
863 }
864
865 return NULL;
866}
867EXPORT_SYMBOL(blk_mq_tag_to_rq);
868
869static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
870 void *priv, bool reserved)
871{
872 /*
873 * If we find a request that isn't idle and the queue matches,
874 * we know the queue is busy. Return false to stop the iteration.
875 */
876 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
877 bool *busy = priv;
878
879 *busy = true;
880 return false;
881 }
882
883 return true;
884}
885
886bool blk_mq_queue_inflight(struct request_queue *q)
887{
888 bool busy = false;
889
890 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
891 return busy;
892}
893EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
894
895static void blk_mq_rq_timed_out(struct request *req, bool reserved)
896{
897 req->rq_flags |= RQF_TIMED_OUT;
898 if (req->q->mq_ops->timeout) {
899 enum blk_eh_timer_return ret;
900
901 ret = req->q->mq_ops->timeout(req, reserved);
902 if (ret == BLK_EH_DONE)
903 return;
904 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
905 }
906
907 blk_add_timer(req);
908}
909
910static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
911{
912 unsigned long deadline;
913
914 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
915 return false;
916 if (rq->rq_flags & RQF_TIMED_OUT)
917 return false;
918
919 deadline = READ_ONCE(rq->deadline);
920 if (time_after_eq(jiffies, deadline))
921 return true;
922
923 if (*next == 0)
924 *next = deadline;
925 else if (time_after(*next, deadline))
926 *next = deadline;
927 return false;
928}
929
930static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
931 struct request *rq, void *priv, bool reserved)
932{
933 unsigned long *next = priv;
934
935 /*
936 * Just do a quick check if it is expired before locking the request in
937 * so we're not unnecessarilly synchronizing across CPUs.
938 */
939 if (!blk_mq_req_expired(rq, next))
940 return true;
941
942 /*
943 * We have reason to believe the request may be expired. Take a
944 * reference on the request to lock this request lifetime into its
945 * currently allocated context to prevent it from being reallocated in
946 * the event the completion by-passes this timeout handler.
947 *
948 * If the reference was already released, then the driver beat the
949 * timeout handler to posting a natural completion.
950 */
951 if (!refcount_inc_not_zero(&rq->ref))
952 return true;
953
954 /*
955 * The request is now locked and cannot be reallocated underneath the
956 * timeout handler's processing. Re-verify this exact request is truly
957 * expired; if it is not expired, then the request was completed and
958 * reallocated as a new request.
959 */
960 if (blk_mq_req_expired(rq, next))
961 blk_mq_rq_timed_out(rq, reserved);
962
963 if (is_flush_rq(rq, hctx))
964 rq->end_io(rq, 0);
965 else if (refcount_dec_and_test(&rq->ref))
966 __blk_mq_free_request(rq);
967
968 return true;
969}
970
971static void blk_mq_timeout_work(struct work_struct *work)
972{
973 struct request_queue *q =
974 container_of(work, struct request_queue, timeout_work);
975 unsigned long next = 0;
976 struct blk_mq_hw_ctx *hctx;
977 int i;
978
979 /* A deadlock might occur if a request is stuck requiring a
980 * timeout at the same time a queue freeze is waiting
981 * completion, since the timeout code would not be able to
982 * acquire the queue reference here.
983 *
984 * That's why we don't use blk_queue_enter here; instead, we use
985 * percpu_ref_tryget directly, because we need to be able to
986 * obtain a reference even in the short window between the queue
987 * starting to freeze, by dropping the first reference in
988 * blk_freeze_queue_start, and the moment the last request is
989 * consumed, marked by the instant q_usage_counter reaches
990 * zero.
991 */
992 if (!percpu_ref_tryget(&q->q_usage_counter))
993 return;
994
995 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
996
997 if (next != 0) {
998 mod_timer(&q->timeout, next);
999 } else {
1000 /*
1001 * Request timeouts are handled as a forward rolling timer. If
1002 * we end up here it means that no requests are pending and
1003 * also that no request has been pending for a while. Mark
1004 * each hctx as idle.
1005 */
1006 queue_for_each_hw_ctx(q, hctx, i) {
1007 /* the hctx may be unmapped, so check it here */
1008 if (blk_mq_hw_queue_mapped(hctx))
1009 blk_mq_tag_idle(hctx);
1010 }
1011 }
1012 blk_queue_exit(q);
1013}
1014
1015struct flush_busy_ctx_data {
1016 struct blk_mq_hw_ctx *hctx;
1017 struct list_head *list;
1018};
1019
1020static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1021{
1022 struct flush_busy_ctx_data *flush_data = data;
1023 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1024 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1025 enum hctx_type type = hctx->type;
1026
1027 spin_lock(&ctx->lock);
1028 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1029 sbitmap_clear_bit(sb, bitnr);
1030 spin_unlock(&ctx->lock);
1031 return true;
1032}
1033
1034/*
1035 * Process software queues that have been marked busy, splicing them
1036 * to the for-dispatch
1037 */
1038void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1039{
1040 struct flush_busy_ctx_data data = {
1041 .hctx = hctx,
1042 .list = list,
1043 };
1044
1045 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1046}
1047EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1048
1049struct dispatch_rq_data {
1050 struct blk_mq_hw_ctx *hctx;
1051 struct request *rq;
1052};
1053
1054static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1055 void *data)
1056{
1057 struct dispatch_rq_data *dispatch_data = data;
1058 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1059 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1060 enum hctx_type type = hctx->type;
1061
1062 spin_lock(&ctx->lock);
1063 if (!list_empty(&ctx->rq_lists[type])) {
1064 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1065 list_del_init(&dispatch_data->rq->queuelist);
1066 if (list_empty(&ctx->rq_lists[type]))
1067 sbitmap_clear_bit(sb, bitnr);
1068 }
1069 spin_unlock(&ctx->lock);
1070
1071 return !dispatch_data->rq;
1072}
1073
1074struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1075 struct blk_mq_ctx *start)
1076{
1077 unsigned off = start ? start->index_hw[hctx->type] : 0;
1078 struct dispatch_rq_data data = {
1079 .hctx = hctx,
1080 .rq = NULL,
1081 };
1082
1083 __sbitmap_for_each_set(&hctx->ctx_map, off,
1084 dispatch_rq_from_ctx, &data);
1085
1086 return data.rq;
1087}
1088
1089static inline unsigned int queued_to_index(unsigned int queued)
1090{
1091 if (!queued)
1092 return 0;
1093
1094 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1095}
1096
1097static bool __blk_mq_get_driver_tag(struct request *rq)
1098{
1099 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1100 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1101 int tag;
1102
1103 blk_mq_tag_busy(rq->mq_hctx);
1104
1105 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1106 bt = &rq->mq_hctx->tags->breserved_tags;
1107 tag_offset = 0;
1108 }
1109
1110 if (!hctx_may_queue(rq->mq_hctx, bt))
1111 return false;
1112 tag = __sbitmap_queue_get(bt);
1113 if (tag == BLK_MQ_NO_TAG)
1114 return false;
1115
1116 rq->tag = tag + tag_offset;
1117 return true;
1118}
1119
1120static bool blk_mq_get_driver_tag(struct request *rq)
1121{
1122 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1123
1124 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1125 return false;
1126
1127 if ((hctx->flags & BLK_MQ_F_TAG_SHARED) &&
1128 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1129 rq->rq_flags |= RQF_MQ_INFLIGHT;
1130 atomic_inc(&hctx->nr_active);
1131 }
1132 hctx->tags->rqs[rq->tag] = rq;
1133 return true;
1134}
1135
1136static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1137 int flags, void *key)
1138{
1139 struct blk_mq_hw_ctx *hctx;
1140
1141 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1142
1143 spin_lock(&hctx->dispatch_wait_lock);
1144 if (!list_empty(&wait->entry)) {
1145 struct sbitmap_queue *sbq;
1146
1147 list_del_init(&wait->entry);
1148 sbq = &hctx->tags->bitmap_tags;
1149 atomic_dec(&sbq->ws_active);
1150 }
1151 spin_unlock(&hctx->dispatch_wait_lock);
1152
1153 blk_mq_run_hw_queue(hctx, true);
1154 return 1;
1155}
1156
1157/*
1158 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1159 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1160 * restart. For both cases, take care to check the condition again after
1161 * marking us as waiting.
1162 */
1163static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1164 struct request *rq)
1165{
1166 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1167 struct wait_queue_head *wq;
1168 wait_queue_entry_t *wait;
1169 bool ret;
1170
1171 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1172 blk_mq_sched_mark_restart_hctx(hctx);
1173
1174 /*
1175 * It's possible that a tag was freed in the window between the
1176 * allocation failure and adding the hardware queue to the wait
1177 * queue.
1178 *
1179 * Don't clear RESTART here, someone else could have set it.
1180 * At most this will cost an extra queue run.
1181 */
1182 return blk_mq_get_driver_tag(rq);
1183 }
1184
1185 wait = &hctx->dispatch_wait;
1186 if (!list_empty_careful(&wait->entry))
1187 return false;
1188
1189 wq = &bt_wait_ptr(sbq, hctx)->wait;
1190
1191 spin_lock_irq(&wq->lock);
1192 spin_lock(&hctx->dispatch_wait_lock);
1193 if (!list_empty(&wait->entry)) {
1194 spin_unlock(&hctx->dispatch_wait_lock);
1195 spin_unlock_irq(&wq->lock);
1196 return false;
1197 }
1198
1199 atomic_inc(&sbq->ws_active);
1200 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1201 __add_wait_queue(wq, wait);
1202
1203 /*
1204 * It's possible that a tag was freed in the window between the
1205 * allocation failure and adding the hardware queue to the wait
1206 * queue.
1207 */
1208 ret = blk_mq_get_driver_tag(rq);
1209 if (!ret) {
1210 spin_unlock(&hctx->dispatch_wait_lock);
1211 spin_unlock_irq(&wq->lock);
1212 return false;
1213 }
1214
1215 /*
1216 * We got a tag, remove ourselves from the wait queue to ensure
1217 * someone else gets the wakeup.
1218 */
1219 list_del_init(&wait->entry);
1220 atomic_dec(&sbq->ws_active);
1221 spin_unlock(&hctx->dispatch_wait_lock);
1222 spin_unlock_irq(&wq->lock);
1223
1224 return true;
1225}
1226
1227#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1228#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1229/*
1230 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1231 * - EWMA is one simple way to compute running average value
1232 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1233 * - take 4 as factor for avoiding to get too small(0) result, and this
1234 * factor doesn't matter because EWMA decreases exponentially
1235 */
1236static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1237{
1238 unsigned int ewma;
1239
1240 if (hctx->queue->elevator)
1241 return;
1242
1243 ewma = hctx->dispatch_busy;
1244
1245 if (!ewma && !busy)
1246 return;
1247
1248 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1249 if (busy)
1250 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1251 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1252
1253 hctx->dispatch_busy = ewma;
1254}
1255
1256#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1257
1258static void blk_mq_handle_dev_resource(struct request *rq,
1259 struct list_head *list)
1260{
1261 struct request *next =
1262 list_first_entry_or_null(list, struct request, queuelist);
1263
1264 /*
1265 * If an I/O scheduler has been configured and we got a driver tag for
1266 * the next request already, free it.
1267 */
1268 if (next)
1269 blk_mq_put_driver_tag(next);
1270
1271 list_add(&rq->queuelist, list);
1272 __blk_mq_requeue_request(rq);
1273}
1274
1275static void blk_mq_handle_zone_resource(struct request *rq,
1276 struct list_head *zone_list)
1277{
1278 /*
1279 * If we end up here it is because we cannot dispatch a request to a
1280 * specific zone due to LLD level zone-write locking or other zone
1281 * related resource not being available. In this case, set the request
1282 * aside in zone_list for retrying it later.
1283 */
1284 list_add(&rq->queuelist, zone_list);
1285 __blk_mq_requeue_request(rq);
1286}
1287
1288enum prep_dispatch {
1289 PREP_DISPATCH_OK,
1290 PREP_DISPATCH_NO_TAG,
1291 PREP_DISPATCH_NO_BUDGET,
1292};
1293
1294static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1295 bool need_budget)
1296{
1297 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1298
1299 if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) {
1300 blk_mq_put_driver_tag(rq);
1301 return PREP_DISPATCH_NO_BUDGET;
1302 }
1303
1304 if (!blk_mq_get_driver_tag(rq)) {
1305 /*
1306 * The initial allocation attempt failed, so we need to
1307 * rerun the hardware queue when a tag is freed. The
1308 * waitqueue takes care of that. If the queue is run
1309 * before we add this entry back on the dispatch list,
1310 * we'll re-run it below.
1311 */
1312 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1313 /*
1314 * All budgets not got from this function will be put
1315 * together during handling partial dispatch
1316 */
1317 if (need_budget)
1318 blk_mq_put_dispatch_budget(rq->q);
1319 return PREP_DISPATCH_NO_TAG;
1320 }
1321 }
1322
1323 return PREP_DISPATCH_OK;
1324}
1325
1326/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1327static void blk_mq_release_budgets(struct request_queue *q,
1328 unsigned int nr_budgets)
1329{
1330 int i;
1331
1332 for (i = 0; i < nr_budgets; i++)
1333 blk_mq_put_dispatch_budget(q);
1334}
1335
1336/*
1337 * Returns true if we did some work AND can potentially do more.
1338 */
1339bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1340 unsigned int nr_budgets)
1341{
1342 enum prep_dispatch prep;
1343 struct request_queue *q = hctx->queue;
1344 struct request *rq, *nxt;
1345 int errors, queued;
1346 blk_status_t ret = BLK_STS_OK;
1347 LIST_HEAD(zone_list);
1348
1349 if (list_empty(list))
1350 return false;
1351
1352 /*
1353 * Now process all the entries, sending them to the driver.
1354 */
1355 errors = queued = 0;
1356 do {
1357 struct blk_mq_queue_data bd;
1358
1359 rq = list_first_entry(list, struct request, queuelist);
1360
1361 WARN_ON_ONCE(hctx != rq->mq_hctx);
1362 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1363 if (prep != PREP_DISPATCH_OK)
1364 break;
1365
1366 list_del_init(&rq->queuelist);
1367
1368 bd.rq = rq;
1369
1370 /*
1371 * Flag last if we have no more requests, or if we have more
1372 * but can't assign a driver tag to it.
1373 */
1374 if (list_empty(list))
1375 bd.last = true;
1376 else {
1377 nxt = list_first_entry(list, struct request, queuelist);
1378 bd.last = !blk_mq_get_driver_tag(nxt);
1379 }
1380
1381 /*
1382 * once the request is queued to lld, no need to cover the
1383 * budget any more
1384 */
1385 if (nr_budgets)
1386 nr_budgets--;
1387 ret = q->mq_ops->queue_rq(hctx, &bd);
1388 switch (ret) {
1389 case BLK_STS_OK:
1390 queued++;
1391 break;
1392 case BLK_STS_RESOURCE:
1393 case BLK_STS_DEV_RESOURCE:
1394 blk_mq_handle_dev_resource(rq, list);
1395 goto out;
1396 case BLK_STS_ZONE_RESOURCE:
1397 /*
1398 * Move the request to zone_list and keep going through
1399 * the dispatch list to find more requests the drive can
1400 * accept.
1401 */
1402 blk_mq_handle_zone_resource(rq, &zone_list);
1403 break;
1404 default:
1405 errors++;
1406 blk_mq_end_request(rq, BLK_STS_IOERR);
1407 }
1408 } while (!list_empty(list));
1409out:
1410 if (!list_empty(&zone_list))
1411 list_splice_tail_init(&zone_list, list);
1412
1413 hctx->dispatched[queued_to_index(queued)]++;
1414
1415 /* If we didn't flush the entire list, we could have told the driver
1416 * there was more coming, but that turned out to be a lie.
1417 */
1418 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1419 q->mq_ops->commit_rqs(hctx);
1420 /*
1421 * Any items that need requeuing? Stuff them into hctx->dispatch,
1422 * that is where we will continue on next queue run.
1423 */
1424 if (!list_empty(list)) {
1425 bool needs_restart;
1426 /* For non-shared tags, the RESTART check will suffice */
1427 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1428 (hctx->flags & BLK_MQ_F_TAG_SHARED);
1429 bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
1430
1431 blk_mq_release_budgets(q, nr_budgets);
1432
1433 spin_lock(&hctx->lock);
1434 list_splice_tail_init(list, &hctx->dispatch);
1435 spin_unlock(&hctx->lock);
1436
1437 /*
1438 * Order adding requests to hctx->dispatch and checking
1439 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1440 * in blk_mq_sched_restart(). Avoid restart code path to
1441 * miss the new added requests to hctx->dispatch, meantime
1442 * SCHED_RESTART is observed here.
1443 */
1444 smp_mb();
1445
1446 /*
1447 * If SCHED_RESTART was set by the caller of this function and
1448 * it is no longer set that means that it was cleared by another
1449 * thread and hence that a queue rerun is needed.
1450 *
1451 * If 'no_tag' is set, that means that we failed getting
1452 * a driver tag with an I/O scheduler attached. If our dispatch
1453 * waitqueue is no longer active, ensure that we run the queue
1454 * AFTER adding our entries back to the list.
1455 *
1456 * If no I/O scheduler has been configured it is possible that
1457 * the hardware queue got stopped and restarted before requests
1458 * were pushed back onto the dispatch list. Rerun the queue to
1459 * avoid starvation. Notes:
1460 * - blk_mq_run_hw_queue() checks whether or not a queue has
1461 * been stopped before rerunning a queue.
1462 * - Some but not all block drivers stop a queue before
1463 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1464 * and dm-rq.
1465 *
1466 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1467 * bit is set, run queue after a delay to avoid IO stalls
1468 * that could otherwise occur if the queue is idle. We'll do
1469 * similar if we couldn't get budget and SCHED_RESTART is set.
1470 */
1471 needs_restart = blk_mq_sched_needs_restart(hctx);
1472 if (!needs_restart ||
1473 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1474 blk_mq_run_hw_queue(hctx, true);
1475 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1476 no_budget_avail))
1477 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1478
1479 blk_mq_update_dispatch_busy(hctx, true);
1480 return false;
1481 } else
1482 blk_mq_update_dispatch_busy(hctx, false);
1483
1484 return (queued + errors) != 0;
1485}
1486
1487/**
1488 * __blk_mq_run_hw_queue - Run a hardware queue.
1489 * @hctx: Pointer to the hardware queue to run.
1490 *
1491 * Send pending requests to the hardware.
1492 */
1493static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1494{
1495 int srcu_idx;
1496
1497 /*
1498 * We should be running this queue from one of the CPUs that
1499 * are mapped to it.
1500 *
1501 * There are at least two related races now between setting
1502 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1503 * __blk_mq_run_hw_queue():
1504 *
1505 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1506 * but later it becomes online, then this warning is harmless
1507 * at all
1508 *
1509 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1510 * but later it becomes offline, then the warning can't be
1511 * triggered, and we depend on blk-mq timeout handler to
1512 * handle dispatched requests to this hctx
1513 */
1514 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1515 cpu_online(hctx->next_cpu)) {
1516 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1517 raw_smp_processor_id(),
1518 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1519 dump_stack();
1520 }
1521
1522 /*
1523 * We can't run the queue inline with ints disabled. Ensure that
1524 * we catch bad users of this early.
1525 */
1526 WARN_ON_ONCE(in_interrupt());
1527
1528 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1529
1530 hctx_lock(hctx, &srcu_idx);
1531 blk_mq_sched_dispatch_requests(hctx);
1532 hctx_unlock(hctx, srcu_idx);
1533}
1534
1535static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1536{
1537 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1538
1539 if (cpu >= nr_cpu_ids)
1540 cpu = cpumask_first(hctx->cpumask);
1541 return cpu;
1542}
1543
1544/*
1545 * It'd be great if the workqueue API had a way to pass
1546 * in a mask and had some smarts for more clever placement.
1547 * For now we just round-robin here, switching for every
1548 * BLK_MQ_CPU_WORK_BATCH queued items.
1549 */
1550static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1551{
1552 bool tried = false;
1553 int next_cpu = hctx->next_cpu;
1554
1555 if (hctx->queue->nr_hw_queues == 1)
1556 return WORK_CPU_UNBOUND;
1557
1558 if (--hctx->next_cpu_batch <= 0) {
1559select_cpu:
1560 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1561 cpu_online_mask);
1562 if (next_cpu >= nr_cpu_ids)
1563 next_cpu = blk_mq_first_mapped_cpu(hctx);
1564 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1565 }
1566
1567 /*
1568 * Do unbound schedule if we can't find a online CPU for this hctx,
1569 * and it should only happen in the path of handling CPU DEAD.
1570 */
1571 if (!cpu_online(next_cpu)) {
1572 if (!tried) {
1573 tried = true;
1574 goto select_cpu;
1575 }
1576
1577 /*
1578 * Make sure to re-select CPU next time once after CPUs
1579 * in hctx->cpumask become online again.
1580 */
1581 hctx->next_cpu = next_cpu;
1582 hctx->next_cpu_batch = 1;
1583 return WORK_CPU_UNBOUND;
1584 }
1585
1586 hctx->next_cpu = next_cpu;
1587 return next_cpu;
1588}
1589
1590/**
1591 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1592 * @hctx: Pointer to the hardware queue to run.
1593 * @async: If we want to run the queue asynchronously.
1594 * @msecs: Microseconds of delay to wait before running the queue.
1595 *
1596 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1597 * with a delay of @msecs.
1598 */
1599static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1600 unsigned long msecs)
1601{
1602 if (unlikely(blk_mq_hctx_stopped(hctx)))
1603 return;
1604
1605 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1606 int cpu = get_cpu();
1607 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1608 __blk_mq_run_hw_queue(hctx);
1609 put_cpu();
1610 return;
1611 }
1612
1613 put_cpu();
1614 }
1615
1616 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1617 msecs_to_jiffies(msecs));
1618}
1619
1620/**
1621 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1622 * @hctx: Pointer to the hardware queue to run.
1623 * @msecs: Microseconds of delay to wait before running the queue.
1624 *
1625 * Run a hardware queue asynchronously with a delay of @msecs.
1626 */
1627void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1628{
1629 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1630}
1631EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1632
1633/**
1634 * blk_mq_run_hw_queue - Start to run a hardware queue.
1635 * @hctx: Pointer to the hardware queue to run.
1636 * @async: If we want to run the queue asynchronously.
1637 *
1638 * Check if the request queue is not in a quiesced state and if there are
1639 * pending requests to be sent. If this is true, run the queue to send requests
1640 * to hardware.
1641 */
1642void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1643{
1644 int srcu_idx;
1645 bool need_run;
1646
1647 /*
1648 * When queue is quiesced, we may be switching io scheduler, or
1649 * updating nr_hw_queues, or other things, and we can't run queue
1650 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1651 *
1652 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1653 * quiesced.
1654 */
1655 hctx_lock(hctx, &srcu_idx);
1656 need_run = !blk_queue_quiesced(hctx->queue) &&
1657 blk_mq_hctx_has_pending(hctx);
1658 hctx_unlock(hctx, srcu_idx);
1659
1660 if (need_run)
1661 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1662}
1663EXPORT_SYMBOL(blk_mq_run_hw_queue);
1664
1665/**
1666 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1667 * @q: Pointer to the request queue to run.
1668 * @async: If we want to run the queue asynchronously.
1669 */
1670void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1671{
1672 struct blk_mq_hw_ctx *hctx;
1673 int i;
1674
1675 queue_for_each_hw_ctx(q, hctx, i) {
1676 if (blk_mq_hctx_stopped(hctx))
1677 continue;
1678
1679 blk_mq_run_hw_queue(hctx, async);
1680 }
1681}
1682EXPORT_SYMBOL(blk_mq_run_hw_queues);
1683
1684/**
1685 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1686 * @q: Pointer to the request queue to run.
1687 * @msecs: Microseconds of delay to wait before running the queues.
1688 */
1689void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1690{
1691 struct blk_mq_hw_ctx *hctx;
1692 int i;
1693
1694 queue_for_each_hw_ctx(q, hctx, i) {
1695 if (blk_mq_hctx_stopped(hctx))
1696 continue;
1697
1698 blk_mq_delay_run_hw_queue(hctx, msecs);
1699 }
1700}
1701EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1702
1703/**
1704 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1705 * @q: request queue.
1706 *
1707 * The caller is responsible for serializing this function against
1708 * blk_mq_{start,stop}_hw_queue().
1709 */
1710bool blk_mq_queue_stopped(struct request_queue *q)
1711{
1712 struct blk_mq_hw_ctx *hctx;
1713 int i;
1714
1715 queue_for_each_hw_ctx(q, hctx, i)
1716 if (blk_mq_hctx_stopped(hctx))
1717 return true;
1718
1719 return false;
1720}
1721EXPORT_SYMBOL(blk_mq_queue_stopped);
1722
1723/*
1724 * This function is often used for pausing .queue_rq() by driver when
1725 * there isn't enough resource or some conditions aren't satisfied, and
1726 * BLK_STS_RESOURCE is usually returned.
1727 *
1728 * We do not guarantee that dispatch can be drained or blocked
1729 * after blk_mq_stop_hw_queue() returns. Please use
1730 * blk_mq_quiesce_queue() for that requirement.
1731 */
1732void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1733{
1734 cancel_delayed_work(&hctx->run_work);
1735
1736 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1737}
1738EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1739
1740/*
1741 * This function is often used for pausing .queue_rq() by driver when
1742 * there isn't enough resource or some conditions aren't satisfied, and
1743 * BLK_STS_RESOURCE is usually returned.
1744 *
1745 * We do not guarantee that dispatch can be drained or blocked
1746 * after blk_mq_stop_hw_queues() returns. Please use
1747 * blk_mq_quiesce_queue() for that requirement.
1748 */
1749void blk_mq_stop_hw_queues(struct request_queue *q)
1750{
1751 struct blk_mq_hw_ctx *hctx;
1752 int i;
1753
1754 queue_for_each_hw_ctx(q, hctx, i)
1755 blk_mq_stop_hw_queue(hctx);
1756}
1757EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1758
1759void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1760{
1761 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1762
1763 blk_mq_run_hw_queue(hctx, false);
1764}
1765EXPORT_SYMBOL(blk_mq_start_hw_queue);
1766
1767void blk_mq_start_hw_queues(struct request_queue *q)
1768{
1769 struct blk_mq_hw_ctx *hctx;
1770 int i;
1771
1772 queue_for_each_hw_ctx(q, hctx, i)
1773 blk_mq_start_hw_queue(hctx);
1774}
1775EXPORT_SYMBOL(blk_mq_start_hw_queues);
1776
1777void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1778{
1779 if (!blk_mq_hctx_stopped(hctx))
1780 return;
1781
1782 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1783 blk_mq_run_hw_queue(hctx, async);
1784}
1785EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1786
1787void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1788{
1789 struct blk_mq_hw_ctx *hctx;
1790 int i;
1791
1792 queue_for_each_hw_ctx(q, hctx, i)
1793 blk_mq_start_stopped_hw_queue(hctx, async);
1794}
1795EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1796
1797static void blk_mq_run_work_fn(struct work_struct *work)
1798{
1799 struct blk_mq_hw_ctx *hctx;
1800
1801 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1802
1803 /*
1804 * If we are stopped, don't run the queue.
1805 */
1806 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1807 return;
1808
1809 __blk_mq_run_hw_queue(hctx);
1810}
1811
1812static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1813 struct request *rq,
1814 bool at_head)
1815{
1816 struct blk_mq_ctx *ctx = rq->mq_ctx;
1817 enum hctx_type type = hctx->type;
1818
1819 lockdep_assert_held(&ctx->lock);
1820
1821 trace_block_rq_insert(hctx->queue, rq);
1822
1823 if (at_head)
1824 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1825 else
1826 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1827}
1828
1829void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1830 bool at_head)
1831{
1832 struct blk_mq_ctx *ctx = rq->mq_ctx;
1833
1834 lockdep_assert_held(&ctx->lock);
1835
1836 __blk_mq_insert_req_list(hctx, rq, at_head);
1837 blk_mq_hctx_mark_pending(hctx, ctx);
1838}
1839
1840/**
1841 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1842 * @rq: Pointer to request to be inserted.
1843 * @at_head: true if the request should be inserted at the head of the list.
1844 * @run_queue: If we should run the hardware queue after inserting the request.
1845 *
1846 * Should only be used carefully, when the caller knows we want to
1847 * bypass a potential IO scheduler on the target device.
1848 */
1849void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1850 bool run_queue)
1851{
1852 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1853
1854 spin_lock(&hctx->lock);
1855 if (at_head)
1856 list_add(&rq->queuelist, &hctx->dispatch);
1857 else
1858 list_add_tail(&rq->queuelist, &hctx->dispatch);
1859 spin_unlock(&hctx->lock);
1860
1861 if (run_queue)
1862 blk_mq_run_hw_queue(hctx, false);
1863}
1864
1865void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1866 struct list_head *list)
1867
1868{
1869 struct request *rq;
1870 enum hctx_type type = hctx->type;
1871
1872 /*
1873 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1874 * offline now
1875 */
1876 list_for_each_entry(rq, list, queuelist) {
1877 BUG_ON(rq->mq_ctx != ctx);
1878 trace_block_rq_insert(hctx->queue, rq);
1879 }
1880
1881 spin_lock(&ctx->lock);
1882 list_splice_tail_init(list, &ctx->rq_lists[type]);
1883 blk_mq_hctx_mark_pending(hctx, ctx);
1884 spin_unlock(&ctx->lock);
1885}
1886
1887static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1888{
1889 struct request *rqa = container_of(a, struct request, queuelist);
1890 struct request *rqb = container_of(b, struct request, queuelist);
1891
1892 if (rqa->mq_ctx != rqb->mq_ctx)
1893 return rqa->mq_ctx > rqb->mq_ctx;
1894 if (rqa->mq_hctx != rqb->mq_hctx)
1895 return rqa->mq_hctx > rqb->mq_hctx;
1896
1897 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1898}
1899
1900void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1901{
1902 LIST_HEAD(list);
1903
1904 if (list_empty(&plug->mq_list))
1905 return;
1906 list_splice_init(&plug->mq_list, &list);
1907
1908 if (plug->rq_count > 2 && plug->multiple_queues)
1909 list_sort(NULL, &list, plug_rq_cmp);
1910
1911 plug->rq_count = 0;
1912
1913 do {
1914 struct list_head rq_list;
1915 struct request *rq, *head_rq = list_entry_rq(list.next);
1916 struct list_head *pos = &head_rq->queuelist; /* skip first */
1917 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1918 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1919 unsigned int depth = 1;
1920
1921 list_for_each_continue(pos, &list) {
1922 rq = list_entry_rq(pos);
1923 BUG_ON(!rq->q);
1924 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1925 break;
1926 depth++;
1927 }
1928
1929 list_cut_before(&rq_list, &list, pos);
1930 trace_block_unplug(head_rq->q, depth, !from_schedule);
1931 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1932 from_schedule);
1933 } while(!list_empty(&list));
1934}
1935
1936static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1937 unsigned int nr_segs)
1938{
1939 if (bio->bi_opf & REQ_RAHEAD)
1940 rq->cmd_flags |= REQ_FAILFAST_MASK;
1941
1942 rq->__sector = bio->bi_iter.bi_sector;
1943 rq->write_hint = bio->bi_write_hint;
1944 blk_rq_bio_prep(rq, bio, nr_segs);
1945 blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1946
1947 blk_account_io_start(rq);
1948}
1949
1950static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1951 struct request *rq,
1952 blk_qc_t *cookie, bool last)
1953{
1954 struct request_queue *q = rq->q;
1955 struct blk_mq_queue_data bd = {
1956 .rq = rq,
1957 .last = last,
1958 };
1959 blk_qc_t new_cookie;
1960 blk_status_t ret;
1961
1962 new_cookie = request_to_qc_t(hctx, rq);
1963
1964 /*
1965 * For OK queue, we are done. For error, caller may kill it.
1966 * Any other error (busy), just add it to our list as we
1967 * previously would have done.
1968 */
1969 ret = q->mq_ops->queue_rq(hctx, &bd);
1970 switch (ret) {
1971 case BLK_STS_OK:
1972 blk_mq_update_dispatch_busy(hctx, false);
1973 *cookie = new_cookie;
1974 break;
1975 case BLK_STS_RESOURCE:
1976 case BLK_STS_DEV_RESOURCE:
1977 blk_mq_update_dispatch_busy(hctx, true);
1978 __blk_mq_requeue_request(rq);
1979 break;
1980 default:
1981 blk_mq_update_dispatch_busy(hctx, false);
1982 *cookie = BLK_QC_T_NONE;
1983 break;
1984 }
1985
1986 return ret;
1987}
1988
1989static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1990 struct request *rq,
1991 blk_qc_t *cookie,
1992 bool bypass_insert, bool last)
1993{
1994 struct request_queue *q = rq->q;
1995 bool run_queue = true;
1996
1997 /*
1998 * RCU or SRCU read lock is needed before checking quiesced flag.
1999 *
2000 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2001 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2002 * and avoid driver to try to dispatch again.
2003 */
2004 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2005 run_queue = false;
2006 bypass_insert = false;
2007 goto insert;
2008 }
2009
2010 if (q->elevator && !bypass_insert)
2011 goto insert;
2012
2013 if (!blk_mq_get_dispatch_budget(q))
2014 goto insert;
2015
2016 if (!blk_mq_get_driver_tag(rq)) {
2017 blk_mq_put_dispatch_budget(q);
2018 goto insert;
2019 }
2020
2021 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2022insert:
2023 if (bypass_insert)
2024 return BLK_STS_RESOURCE;
2025
2026 blk_mq_sched_insert_request(rq, false, run_queue, false);
2027
2028 return BLK_STS_OK;
2029}
2030
2031/**
2032 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2033 * @hctx: Pointer of the associated hardware queue.
2034 * @rq: Pointer to request to be sent.
2035 * @cookie: Request queue cookie.
2036 *
2037 * If the device has enough resources to accept a new request now, send the
2038 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2039 * we can try send it another time in the future. Requests inserted at this
2040 * queue have higher priority.
2041 */
2042static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2043 struct request *rq, blk_qc_t *cookie)
2044{
2045 blk_status_t ret;
2046 int srcu_idx;
2047
2048 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2049
2050 hctx_lock(hctx, &srcu_idx);
2051
2052 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2053 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2054 blk_mq_request_bypass_insert(rq, false, true);
2055 else if (ret != BLK_STS_OK)
2056 blk_mq_end_request(rq, ret);
2057
2058 hctx_unlock(hctx, srcu_idx);
2059}
2060
2061blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2062{
2063 blk_status_t ret;
2064 int srcu_idx;
2065 blk_qc_t unused_cookie;
2066 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2067
2068 hctx_lock(hctx, &srcu_idx);
2069 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2070 hctx_unlock(hctx, srcu_idx);
2071
2072 return ret;
2073}
2074
2075void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2076 struct list_head *list)
2077{
2078 int queued = 0;
2079 int errors = 0;
2080
2081 while (!list_empty(list)) {
2082 blk_status_t ret;
2083 struct request *rq = list_first_entry(list, struct request,
2084 queuelist);
2085
2086 list_del_init(&rq->queuelist);
2087 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2088 if (ret != BLK_STS_OK) {
2089 if (ret == BLK_STS_RESOURCE ||
2090 ret == BLK_STS_DEV_RESOURCE) {
2091 blk_mq_request_bypass_insert(rq, false,
2092 list_empty(list));
2093 break;
2094 }
2095 blk_mq_end_request(rq, ret);
2096 errors++;
2097 } else
2098 queued++;
2099 }
2100
2101 /*
2102 * If we didn't flush the entire list, we could have told
2103 * the driver there was more coming, but that turned out to
2104 * be a lie.
2105 */
2106 if ((!list_empty(list) || errors) &&
2107 hctx->queue->mq_ops->commit_rqs && queued)
2108 hctx->queue->mq_ops->commit_rqs(hctx);
2109}
2110
2111static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2112{
2113 list_add_tail(&rq->queuelist, &plug->mq_list);
2114 plug->rq_count++;
2115 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2116 struct request *tmp;
2117
2118 tmp = list_first_entry(&plug->mq_list, struct request,
2119 queuelist);
2120 if (tmp->q != rq->q)
2121 plug->multiple_queues = true;
2122 }
2123}
2124
2125/**
2126 * blk_mq_submit_bio - Create and send a request to block device.
2127 * @bio: Bio pointer.
2128 *
2129 * Builds up a request structure from @q and @bio and send to the device. The
2130 * request may not be queued directly to hardware if:
2131 * * This request can be merged with another one
2132 * * We want to place request at plug queue for possible future merging
2133 * * There is an IO scheduler active at this queue
2134 *
2135 * It will not queue the request if there is an error with the bio, or at the
2136 * request creation.
2137 *
2138 * Returns: Request queue cookie.
2139 */
2140blk_qc_t blk_mq_submit_bio(struct bio *bio)
2141{
2142 struct request_queue *q = bio->bi_disk->queue;
2143 const int is_sync = op_is_sync(bio->bi_opf);
2144 const int is_flush_fua = op_is_flush(bio->bi_opf);
2145 struct blk_mq_alloc_data data = {
2146 .q = q,
2147 };
2148 struct request *rq;
2149 struct blk_plug *plug;
2150 struct request *same_queue_rq = NULL;
2151 unsigned int nr_segs;
2152 blk_qc_t cookie;
2153 blk_status_t ret;
2154
2155 blk_queue_bounce(q, &bio);
2156 __blk_queue_split(&bio, &nr_segs);
2157
2158 if (!bio_integrity_prep(bio))
2159 goto queue_exit;
2160
2161 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2162 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2163 goto queue_exit;
2164
2165 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2166 goto queue_exit;
2167
2168 rq_qos_throttle(q, bio);
2169
2170 data.cmd_flags = bio->bi_opf;
2171 rq = __blk_mq_alloc_request(&data);
2172 if (unlikely(!rq)) {
2173 rq_qos_cleanup(q, bio);
2174 if (bio->bi_opf & REQ_NOWAIT)
2175 bio_wouldblock_error(bio);
2176 goto queue_exit;
2177 }
2178
2179 trace_block_getrq(q, bio, bio->bi_opf);
2180
2181 rq_qos_track(q, rq, bio);
2182
2183 cookie = request_to_qc_t(data.hctx, rq);
2184
2185 blk_mq_bio_to_request(rq, bio, nr_segs);
2186
2187 ret = blk_crypto_init_request(rq);
2188 if (ret != BLK_STS_OK) {
2189 bio->bi_status = ret;
2190 bio_endio(bio);
2191 blk_mq_free_request(rq);
2192 return BLK_QC_T_NONE;
2193 }
2194
2195 plug = blk_mq_plug(q, bio);
2196 if (unlikely(is_flush_fua)) {
2197 /* Bypass scheduler for flush requests */
2198 blk_insert_flush(rq);
2199 blk_mq_run_hw_queue(data.hctx, true);
2200 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2201 !blk_queue_nonrot(q))) {
2202 /*
2203 * Use plugging if we have a ->commit_rqs() hook as well, as
2204 * we know the driver uses bd->last in a smart fashion.
2205 *
2206 * Use normal plugging if this disk is slow HDD, as sequential
2207 * IO may benefit a lot from plug merging.
2208 */
2209 unsigned int request_count = plug->rq_count;
2210 struct request *last = NULL;
2211
2212 if (!request_count)
2213 trace_block_plug(q);
2214 else
2215 last = list_entry_rq(plug->mq_list.prev);
2216
2217 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2218 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2219 blk_flush_plug_list(plug, false);
2220 trace_block_plug(q);
2221 }
2222
2223 blk_add_rq_to_plug(plug, rq);
2224 } else if (q->elevator) {
2225 /* Insert the request at the IO scheduler queue */
2226 blk_mq_sched_insert_request(rq, false, true, true);
2227 } else if (plug && !blk_queue_nomerges(q)) {
2228 /*
2229 * We do limited plugging. If the bio can be merged, do that.
2230 * Otherwise the existing request in the plug list will be
2231 * issued. So the plug list will have one request at most
2232 * The plug list might get flushed before this. If that happens,
2233 * the plug list is empty, and same_queue_rq is invalid.
2234 */
2235 if (list_empty(&plug->mq_list))
2236 same_queue_rq = NULL;
2237 if (same_queue_rq) {
2238 list_del_init(&same_queue_rq->queuelist);
2239 plug->rq_count--;
2240 }
2241 blk_add_rq_to_plug(plug, rq);
2242 trace_block_plug(q);
2243
2244 if (same_queue_rq) {
2245 data.hctx = same_queue_rq->mq_hctx;
2246 trace_block_unplug(q, 1, true);
2247 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2248 &cookie);
2249 }
2250 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2251 !data.hctx->dispatch_busy) {
2252 /*
2253 * There is no scheduler and we can try to send directly
2254 * to the hardware.
2255 */
2256 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2257 } else {
2258 /* Default case. */
2259 blk_mq_sched_insert_request(rq, false, true, true);
2260 }
2261
2262 return cookie;
2263queue_exit:
2264 blk_queue_exit(q);
2265 return BLK_QC_T_NONE;
2266}
2267EXPORT_SYMBOL_GPL(blk_mq_submit_bio); /* only for request based dm */
2268
2269void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2270 unsigned int hctx_idx)
2271{
2272 struct page *page;
2273
2274 if (tags->rqs && set->ops->exit_request) {
2275 int i;
2276
2277 for (i = 0; i < tags->nr_tags; i++) {
2278 struct request *rq = tags->static_rqs[i];
2279
2280 if (!rq)
2281 continue;
2282 set->ops->exit_request(set, rq, hctx_idx);
2283 tags->static_rqs[i] = NULL;
2284 }
2285 }
2286
2287 while (!list_empty(&tags->page_list)) {
2288 page = list_first_entry(&tags->page_list, struct page, lru);
2289 list_del_init(&page->lru);
2290 /*
2291 * Remove kmemleak object previously allocated in
2292 * blk_mq_alloc_rqs().
2293 */
2294 kmemleak_free(page_address(page));
2295 __free_pages(page, page->private);
2296 }
2297}
2298
2299void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2300{
2301 kfree(tags->rqs);
2302 tags->rqs = NULL;
2303 kfree(tags->static_rqs);
2304 tags->static_rqs = NULL;
2305
2306 blk_mq_free_tags(tags);
2307}
2308
2309struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2310 unsigned int hctx_idx,
2311 unsigned int nr_tags,
2312 unsigned int reserved_tags)
2313{
2314 struct blk_mq_tags *tags;
2315 int node;
2316
2317 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2318 if (node == NUMA_NO_NODE)
2319 node = set->numa_node;
2320
2321 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2322 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2323 if (!tags)
2324 return NULL;
2325
2326 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2327 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2328 node);
2329 if (!tags->rqs) {
2330 blk_mq_free_tags(tags);
2331 return NULL;
2332 }
2333
2334 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2335 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2336 node);
2337 if (!tags->static_rqs) {
2338 kfree(tags->rqs);
2339 blk_mq_free_tags(tags);
2340 return NULL;
2341 }
2342
2343 return tags;
2344}
2345
2346static size_t order_to_size(unsigned int order)
2347{
2348 return (size_t)PAGE_SIZE << order;
2349}
2350
2351static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2352 unsigned int hctx_idx, int node)
2353{
2354 int ret;
2355
2356 if (set->ops->init_request) {
2357 ret = set->ops->init_request(set, rq, hctx_idx, node);
2358 if (ret)
2359 return ret;
2360 }
2361
2362 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2363 return 0;
2364}
2365
2366int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2367 unsigned int hctx_idx, unsigned int depth)
2368{
2369 unsigned int i, j, entries_per_page, max_order = 4;
2370 size_t rq_size, left;
2371 int node;
2372
2373 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2374 if (node == NUMA_NO_NODE)
2375 node = set->numa_node;
2376
2377 INIT_LIST_HEAD(&tags->page_list);
2378
2379 /*
2380 * rq_size is the size of the request plus driver payload, rounded
2381 * to the cacheline size
2382 */
2383 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2384 cache_line_size());
2385 left = rq_size * depth;
2386
2387 for (i = 0; i < depth; ) {
2388 int this_order = max_order;
2389 struct page *page;
2390 int to_do;
2391 void *p;
2392
2393 while (this_order && left < order_to_size(this_order - 1))
2394 this_order--;
2395
2396 do {
2397 page = alloc_pages_node(node,
2398 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2399 this_order);
2400 if (page)
2401 break;
2402 if (!this_order--)
2403 break;
2404 if (order_to_size(this_order) < rq_size)
2405 break;
2406 } while (1);
2407
2408 if (!page)
2409 goto fail;
2410
2411 page->private = this_order;
2412 list_add_tail(&page->lru, &tags->page_list);
2413
2414 p = page_address(page);
2415 /*
2416 * Allow kmemleak to scan these pages as they contain pointers
2417 * to additional allocations like via ops->init_request().
2418 */
2419 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2420 entries_per_page = order_to_size(this_order) / rq_size;
2421 to_do = min(entries_per_page, depth - i);
2422 left -= to_do * rq_size;
2423 for (j = 0; j < to_do; j++) {
2424 struct request *rq = p;
2425
2426 tags->static_rqs[i] = rq;
2427 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2428 tags->static_rqs[i] = NULL;
2429 goto fail;
2430 }
2431
2432 p += rq_size;
2433 i++;
2434 }
2435 }
2436 return 0;
2437
2438fail:
2439 blk_mq_free_rqs(set, tags, hctx_idx);
2440 return -ENOMEM;
2441}
2442
2443struct rq_iter_data {
2444 struct blk_mq_hw_ctx *hctx;
2445 bool has_rq;
2446};
2447
2448static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2449{
2450 struct rq_iter_data *iter_data = data;
2451
2452 if (rq->mq_hctx != iter_data->hctx)
2453 return true;
2454 iter_data->has_rq = true;
2455 return false;
2456}
2457
2458static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2459{
2460 struct blk_mq_tags *tags = hctx->sched_tags ?
2461 hctx->sched_tags : hctx->tags;
2462 struct rq_iter_data data = {
2463 .hctx = hctx,
2464 };
2465
2466 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2467 return data.has_rq;
2468}
2469
2470static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2471 struct blk_mq_hw_ctx *hctx)
2472{
2473 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2474 return false;
2475 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2476 return false;
2477 return true;
2478}
2479
2480static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2481{
2482 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2483 struct blk_mq_hw_ctx, cpuhp_online);
2484
2485 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2486 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2487 return 0;
2488
2489 /*
2490 * Prevent new request from being allocated on the current hctx.
2491 *
2492 * The smp_mb__after_atomic() Pairs with the implied barrier in
2493 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2494 * seen once we return from the tag allocator.
2495 */
2496 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2497 smp_mb__after_atomic();
2498
2499 /*
2500 * Try to grab a reference to the queue and wait for any outstanding
2501 * requests. If we could not grab a reference the queue has been
2502 * frozen and there are no requests.
2503 */
2504 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2505 while (blk_mq_hctx_has_requests(hctx))
2506 msleep(5);
2507 percpu_ref_put(&hctx->queue->q_usage_counter);
2508 }
2509
2510 return 0;
2511}
2512
2513static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2514{
2515 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2516 struct blk_mq_hw_ctx, cpuhp_online);
2517
2518 if (cpumask_test_cpu(cpu, hctx->cpumask))
2519 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2520 return 0;
2521}
2522
2523/*
2524 * 'cpu' is going away. splice any existing rq_list entries from this
2525 * software queue to the hw queue dispatch list, and ensure that it
2526 * gets run.
2527 */
2528static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2529{
2530 struct blk_mq_hw_ctx *hctx;
2531 struct blk_mq_ctx *ctx;
2532 LIST_HEAD(tmp);
2533 enum hctx_type type;
2534
2535 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2536 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2537 return 0;
2538
2539 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2540 type = hctx->type;
2541
2542 spin_lock(&ctx->lock);
2543 if (!list_empty(&ctx->rq_lists[type])) {
2544 list_splice_init(&ctx->rq_lists[type], &tmp);
2545 blk_mq_hctx_clear_pending(hctx, ctx);
2546 }
2547 spin_unlock(&ctx->lock);
2548
2549 if (list_empty(&tmp))
2550 return 0;
2551
2552 spin_lock(&hctx->lock);
2553 list_splice_tail_init(&tmp, &hctx->dispatch);
2554 spin_unlock(&hctx->lock);
2555
2556 blk_mq_run_hw_queue(hctx, true);
2557 return 0;
2558}
2559
2560static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2561{
2562 if (!(hctx->flags & BLK_MQ_F_STACKING))
2563 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2564 &hctx->cpuhp_online);
2565 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2566 &hctx->cpuhp_dead);
2567}
2568
2569/* hctx->ctxs will be freed in queue's release handler */
2570static void blk_mq_exit_hctx(struct request_queue *q,
2571 struct blk_mq_tag_set *set,
2572 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2573{
2574 if (blk_mq_hw_queue_mapped(hctx))
2575 blk_mq_tag_idle(hctx);
2576
2577 if (set->ops->exit_request)
2578 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2579
2580 if (set->ops->exit_hctx)
2581 set->ops->exit_hctx(hctx, hctx_idx);
2582
2583 blk_mq_remove_cpuhp(hctx);
2584
2585 spin_lock(&q->unused_hctx_lock);
2586 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2587 spin_unlock(&q->unused_hctx_lock);
2588}
2589
2590static void blk_mq_exit_hw_queues(struct request_queue *q,
2591 struct blk_mq_tag_set *set, int nr_queue)
2592{
2593 struct blk_mq_hw_ctx *hctx;
2594 unsigned int i;
2595
2596 queue_for_each_hw_ctx(q, hctx, i) {
2597 if (i == nr_queue)
2598 break;
2599 blk_mq_debugfs_unregister_hctx(hctx);
2600 blk_mq_exit_hctx(q, set, hctx, i);
2601 }
2602}
2603
2604static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2605{
2606 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2607
2608 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2609 __alignof__(struct blk_mq_hw_ctx)) !=
2610 sizeof(struct blk_mq_hw_ctx));
2611
2612 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2613 hw_ctx_size += sizeof(struct srcu_struct);
2614
2615 return hw_ctx_size;
2616}
2617
2618static int blk_mq_init_hctx(struct request_queue *q,
2619 struct blk_mq_tag_set *set,
2620 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2621{
2622 hctx->queue_num = hctx_idx;
2623
2624 if (!(hctx->flags & BLK_MQ_F_STACKING))
2625 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2626 &hctx->cpuhp_online);
2627 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2628
2629 hctx->tags = set->tags[hctx_idx];
2630
2631 if (set->ops->init_hctx &&
2632 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2633 goto unregister_cpu_notifier;
2634
2635 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2636 hctx->numa_node))
2637 goto exit_hctx;
2638 return 0;
2639
2640 exit_hctx:
2641 if (set->ops->exit_hctx)
2642 set->ops->exit_hctx(hctx, hctx_idx);
2643 unregister_cpu_notifier:
2644 blk_mq_remove_cpuhp(hctx);
2645 return -1;
2646}
2647
2648static struct blk_mq_hw_ctx *
2649blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2650 int node)
2651{
2652 struct blk_mq_hw_ctx *hctx;
2653 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2654
2655 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2656 if (!hctx)
2657 goto fail_alloc_hctx;
2658
2659 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2660 goto free_hctx;
2661
2662 atomic_set(&hctx->nr_active, 0);
2663 if (node == NUMA_NO_NODE)
2664 node = set->numa_node;
2665 hctx->numa_node = node;
2666
2667 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2668 spin_lock_init(&hctx->lock);
2669 INIT_LIST_HEAD(&hctx->dispatch);
2670 hctx->queue = q;
2671 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2672
2673 INIT_LIST_HEAD(&hctx->hctx_list);
2674
2675 /*
2676 * Allocate space for all possible cpus to avoid allocation at
2677 * runtime
2678 */
2679 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2680 gfp, node);
2681 if (!hctx->ctxs)
2682 goto free_cpumask;
2683
2684 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2685 gfp, node))
2686 goto free_ctxs;
2687 hctx->nr_ctx = 0;
2688
2689 spin_lock_init(&hctx->dispatch_wait_lock);
2690 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2691 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2692
2693 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2694 if (!hctx->fq)
2695 goto free_bitmap;
2696
2697 if (hctx->flags & BLK_MQ_F_BLOCKING)
2698 init_srcu_struct(hctx->srcu);
2699 blk_mq_hctx_kobj_init(hctx);
2700
2701 return hctx;
2702
2703 free_bitmap:
2704 sbitmap_free(&hctx->ctx_map);
2705 free_ctxs:
2706 kfree(hctx->ctxs);
2707 free_cpumask:
2708 free_cpumask_var(hctx->cpumask);
2709 free_hctx:
2710 kfree(hctx);
2711 fail_alloc_hctx:
2712 return NULL;
2713}
2714
2715static void blk_mq_init_cpu_queues(struct request_queue *q,
2716 unsigned int nr_hw_queues)
2717{
2718 struct blk_mq_tag_set *set = q->tag_set;
2719 unsigned int i, j;
2720
2721 for_each_possible_cpu(i) {
2722 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2723 struct blk_mq_hw_ctx *hctx;
2724 int k;
2725
2726 __ctx->cpu = i;
2727 spin_lock_init(&__ctx->lock);
2728 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2729 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2730
2731 __ctx->queue = q;
2732
2733 /*
2734 * Set local node, IFF we have more than one hw queue. If
2735 * not, we remain on the home node of the device
2736 */
2737 for (j = 0; j < set->nr_maps; j++) {
2738 hctx = blk_mq_map_queue_type(q, j, i);
2739 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2740 hctx->numa_node = local_memory_node(cpu_to_node(i));
2741 }
2742 }
2743}
2744
2745static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2746 int hctx_idx)
2747{
2748 int ret = 0;
2749
2750 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2751 set->queue_depth, set->reserved_tags);
2752 if (!set->tags[hctx_idx])
2753 return false;
2754
2755 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2756 set->queue_depth);
2757 if (!ret)
2758 return true;
2759
2760 blk_mq_free_rq_map(set->tags[hctx_idx]);
2761 set->tags[hctx_idx] = NULL;
2762 return false;
2763}
2764
2765static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2766 unsigned int hctx_idx)
2767{
2768 if (set->tags && set->tags[hctx_idx]) {
2769 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2770 blk_mq_free_rq_map(set->tags[hctx_idx]);
2771 set->tags[hctx_idx] = NULL;
2772 }
2773}
2774
2775static void blk_mq_map_swqueue(struct request_queue *q)
2776{
2777 unsigned int i, j, hctx_idx;
2778 struct blk_mq_hw_ctx *hctx;
2779 struct blk_mq_ctx *ctx;
2780 struct blk_mq_tag_set *set = q->tag_set;
2781
2782 queue_for_each_hw_ctx(q, hctx, i) {
2783 cpumask_clear(hctx->cpumask);
2784 hctx->nr_ctx = 0;
2785 hctx->dispatch_from = NULL;
2786 }
2787
2788 /*
2789 * Map software to hardware queues.
2790 *
2791 * If the cpu isn't present, the cpu is mapped to first hctx.
2792 */
2793 for_each_possible_cpu(i) {
2794
2795 ctx = per_cpu_ptr(q->queue_ctx, i);
2796 for (j = 0; j < set->nr_maps; j++) {
2797 if (!set->map[j].nr_queues) {
2798 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2799 HCTX_TYPE_DEFAULT, i);
2800 continue;
2801 }
2802 hctx_idx = set->map[j].mq_map[i];
2803 /* unmapped hw queue can be remapped after CPU topo changed */
2804 if (!set->tags[hctx_idx] &&
2805 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2806 /*
2807 * If tags initialization fail for some hctx,
2808 * that hctx won't be brought online. In this
2809 * case, remap the current ctx to hctx[0] which
2810 * is guaranteed to always have tags allocated
2811 */
2812 set->map[j].mq_map[i] = 0;
2813 }
2814
2815 hctx = blk_mq_map_queue_type(q, j, i);
2816 ctx->hctxs[j] = hctx;
2817 /*
2818 * If the CPU is already set in the mask, then we've
2819 * mapped this one already. This can happen if
2820 * devices share queues across queue maps.
2821 */
2822 if (cpumask_test_cpu(i, hctx->cpumask))
2823 continue;
2824
2825 cpumask_set_cpu(i, hctx->cpumask);
2826 hctx->type = j;
2827 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2828 hctx->ctxs[hctx->nr_ctx++] = ctx;
2829
2830 /*
2831 * If the nr_ctx type overflows, we have exceeded the
2832 * amount of sw queues we can support.
2833 */
2834 BUG_ON(!hctx->nr_ctx);
2835 }
2836
2837 for (; j < HCTX_MAX_TYPES; j++)
2838 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2839 HCTX_TYPE_DEFAULT, i);
2840 }
2841
2842 queue_for_each_hw_ctx(q, hctx, i) {
2843 /*
2844 * If no software queues are mapped to this hardware queue,
2845 * disable it and free the request entries.
2846 */
2847 if (!hctx->nr_ctx) {
2848 /* Never unmap queue 0. We need it as a
2849 * fallback in case of a new remap fails
2850 * allocation
2851 */
2852 if (i && set->tags[i])
2853 blk_mq_free_map_and_requests(set, i);
2854
2855 hctx->tags = NULL;
2856 continue;
2857 }
2858
2859 hctx->tags = set->tags[i];
2860 WARN_ON(!hctx->tags);
2861
2862 /*
2863 * Set the map size to the number of mapped software queues.
2864 * This is more accurate and more efficient than looping
2865 * over all possibly mapped software queues.
2866 */
2867 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2868
2869 /*
2870 * Initialize batch roundrobin counts
2871 */
2872 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2873 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2874 }
2875}
2876
2877/*
2878 * Caller needs to ensure that we're either frozen/quiesced, or that
2879 * the queue isn't live yet.
2880 */
2881static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2882{
2883 struct blk_mq_hw_ctx *hctx;
2884 int i;
2885
2886 queue_for_each_hw_ctx(q, hctx, i) {
2887 if (shared)
2888 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2889 else
2890 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2891 }
2892}
2893
2894static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2895 bool shared)
2896{
2897 struct request_queue *q;
2898
2899 lockdep_assert_held(&set->tag_list_lock);
2900
2901 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2902 blk_mq_freeze_queue(q);
2903 queue_set_hctx_shared(q, shared);
2904 blk_mq_unfreeze_queue(q);
2905 }
2906}
2907
2908static void blk_mq_del_queue_tag_set(struct request_queue *q)
2909{
2910 struct blk_mq_tag_set *set = q->tag_set;
2911
2912 mutex_lock(&set->tag_list_lock);
2913 list_del(&q->tag_set_list);
2914 if (list_is_singular(&set->tag_list)) {
2915 /* just transitioned to unshared */
2916 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2917 /* update existing queue */
2918 blk_mq_update_tag_set_depth(set, false);
2919 }
2920 mutex_unlock(&set->tag_list_lock);
2921 INIT_LIST_HEAD(&q->tag_set_list);
2922}
2923
2924static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2925 struct request_queue *q)
2926{
2927 mutex_lock(&set->tag_list_lock);
2928
2929 /*
2930 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2931 */
2932 if (!list_empty(&set->tag_list) &&
2933 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2934 set->flags |= BLK_MQ_F_TAG_SHARED;
2935 /* update existing queue */
2936 blk_mq_update_tag_set_depth(set, true);
2937 }
2938 if (set->flags & BLK_MQ_F_TAG_SHARED)
2939 queue_set_hctx_shared(q, true);
2940 list_add_tail(&q->tag_set_list, &set->tag_list);
2941
2942 mutex_unlock(&set->tag_list_lock);
2943}
2944
2945/* All allocations will be freed in release handler of q->mq_kobj */
2946static int blk_mq_alloc_ctxs(struct request_queue *q)
2947{
2948 struct blk_mq_ctxs *ctxs;
2949 int cpu;
2950
2951 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2952 if (!ctxs)
2953 return -ENOMEM;
2954
2955 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2956 if (!ctxs->queue_ctx)
2957 goto fail;
2958
2959 for_each_possible_cpu(cpu) {
2960 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2961 ctx->ctxs = ctxs;
2962 }
2963
2964 q->mq_kobj = &ctxs->kobj;
2965 q->queue_ctx = ctxs->queue_ctx;
2966
2967 return 0;
2968 fail:
2969 kfree(ctxs);
2970 return -ENOMEM;
2971}
2972
2973/*
2974 * It is the actual release handler for mq, but we do it from
2975 * request queue's release handler for avoiding use-after-free
2976 * and headache because q->mq_kobj shouldn't have been introduced,
2977 * but we can't group ctx/kctx kobj without it.
2978 */
2979void blk_mq_release(struct request_queue *q)
2980{
2981 struct blk_mq_hw_ctx *hctx, *next;
2982 int i;
2983
2984 queue_for_each_hw_ctx(q, hctx, i)
2985 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2986
2987 /* all hctx are in .unused_hctx_list now */
2988 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2989 list_del_init(&hctx->hctx_list);
2990 kobject_put(&hctx->kobj);
2991 }
2992
2993 kfree(q->queue_hw_ctx);
2994
2995 /*
2996 * release .mq_kobj and sw queue's kobject now because
2997 * both share lifetime with request queue.
2998 */
2999 blk_mq_sysfs_deinit(q);
3000}
3001
3002struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3003 void *queuedata)
3004{
3005 struct request_queue *uninit_q, *q;
3006
3007 uninit_q = blk_alloc_queue(set->numa_node);
3008 if (!uninit_q)
3009 return ERR_PTR(-ENOMEM);
3010 uninit_q->queuedata = queuedata;
3011
3012 /*
3013 * Initialize the queue without an elevator. device_add_disk() will do
3014 * the initialization.
3015 */
3016 q = blk_mq_init_allocated_queue(set, uninit_q, false);
3017 if (IS_ERR(q))
3018 blk_cleanup_queue(uninit_q);
3019
3020 return q;
3021}
3022EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
3023
3024struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3025{
3026 return blk_mq_init_queue_data(set, NULL);
3027}
3028EXPORT_SYMBOL(blk_mq_init_queue);
3029
3030/*
3031 * Helper for setting up a queue with mq ops, given queue depth, and
3032 * the passed in mq ops flags.
3033 */
3034struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
3035 const struct blk_mq_ops *ops,
3036 unsigned int queue_depth,
3037 unsigned int set_flags)
3038{
3039 struct request_queue *q;
3040 int ret;
3041
3042 memset(set, 0, sizeof(*set));
3043 set->ops = ops;
3044 set->nr_hw_queues = 1;
3045 set->nr_maps = 1;
3046 set->queue_depth = queue_depth;
3047 set->numa_node = NUMA_NO_NODE;
3048 set->flags = set_flags;
3049
3050 ret = blk_mq_alloc_tag_set(set);
3051 if (ret)
3052 return ERR_PTR(ret);
3053
3054 q = blk_mq_init_queue(set);
3055 if (IS_ERR(q)) {
3056 blk_mq_free_tag_set(set);
3057 return q;
3058 }
3059
3060 return q;
3061}
3062EXPORT_SYMBOL(blk_mq_init_sq_queue);
3063
3064static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3065 struct blk_mq_tag_set *set, struct request_queue *q,
3066 int hctx_idx, int node)
3067{
3068 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3069
3070 /* reuse dead hctx first */
3071 spin_lock(&q->unused_hctx_lock);
3072 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3073 if (tmp->numa_node == node) {
3074 hctx = tmp;
3075 break;
3076 }
3077 }
3078 if (hctx)
3079 list_del_init(&hctx->hctx_list);
3080 spin_unlock(&q->unused_hctx_lock);
3081
3082 if (!hctx)
3083 hctx = blk_mq_alloc_hctx(q, set, node);
3084 if (!hctx)
3085 goto fail;
3086
3087 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3088 goto free_hctx;
3089
3090 return hctx;
3091
3092 free_hctx:
3093 kobject_put(&hctx->kobj);
3094 fail:
3095 return NULL;
3096}
3097
3098static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3099 struct request_queue *q)
3100{
3101 int i, j, end;
3102 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3103
3104 if (q->nr_hw_queues < set->nr_hw_queues) {
3105 struct blk_mq_hw_ctx **new_hctxs;
3106
3107 new_hctxs = kcalloc_node(set->nr_hw_queues,
3108 sizeof(*new_hctxs), GFP_KERNEL,
3109 set->numa_node);
3110 if (!new_hctxs)
3111 return;
3112 if (hctxs)
3113 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3114 sizeof(*hctxs));
3115 q->queue_hw_ctx = new_hctxs;
3116 kfree(hctxs);
3117 hctxs = new_hctxs;
3118 }
3119
3120 /* protect against switching io scheduler */
3121 mutex_lock(&q->sysfs_lock);
3122 for (i = 0; i < set->nr_hw_queues; i++) {
3123 int node;
3124 struct blk_mq_hw_ctx *hctx;
3125
3126 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3127 /*
3128 * If the hw queue has been mapped to another numa node,
3129 * we need to realloc the hctx. If allocation fails, fallback
3130 * to use the previous one.
3131 */
3132 if (hctxs[i] && (hctxs[i]->numa_node == node))
3133 continue;
3134
3135 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3136 if (hctx) {
3137 if (hctxs[i])
3138 blk_mq_exit_hctx(q, set, hctxs[i], i);
3139 hctxs[i] = hctx;
3140 } else {
3141 if (hctxs[i])
3142 pr_warn("Allocate new hctx on node %d fails,\
3143 fallback to previous one on node %d\n",
3144 node, hctxs[i]->numa_node);
3145 else
3146 break;
3147 }
3148 }
3149 /*
3150 * Increasing nr_hw_queues fails. Free the newly allocated
3151 * hctxs and keep the previous q->nr_hw_queues.
3152 */
3153 if (i != set->nr_hw_queues) {
3154 j = q->nr_hw_queues;
3155 end = i;
3156 } else {
3157 j = i;
3158 end = q->nr_hw_queues;
3159 q->nr_hw_queues = set->nr_hw_queues;
3160 }
3161
3162 for (; j < end; j++) {
3163 struct blk_mq_hw_ctx *hctx = hctxs[j];
3164
3165 if (hctx) {
3166 if (hctx->tags)
3167 blk_mq_free_map_and_requests(set, j);
3168 blk_mq_exit_hctx(q, set, hctx, j);
3169 hctxs[j] = NULL;
3170 }
3171 }
3172 mutex_unlock(&q->sysfs_lock);
3173}
3174
3175struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3176 struct request_queue *q,
3177 bool elevator_init)
3178{
3179 /* mark the queue as mq asap */
3180 q->mq_ops = set->ops;
3181
3182 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3183 blk_mq_poll_stats_bkt,
3184 BLK_MQ_POLL_STATS_BKTS, q);
3185 if (!q->poll_cb)
3186 goto err_exit;
3187
3188 if (blk_mq_alloc_ctxs(q))
3189 goto err_poll;
3190
3191 /* init q->mq_kobj and sw queues' kobjects */
3192 blk_mq_sysfs_init(q);
3193
3194 INIT_LIST_HEAD(&q->unused_hctx_list);
3195 spin_lock_init(&q->unused_hctx_lock);
3196
3197 blk_mq_realloc_hw_ctxs(set, q);
3198 if (!q->nr_hw_queues)
3199 goto err_hctxs;
3200
3201 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3202 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3203
3204 q->tag_set = set;
3205
3206 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3207 if (set->nr_maps > HCTX_TYPE_POLL &&
3208 set->map[HCTX_TYPE_POLL].nr_queues)
3209 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3210
3211 q->sg_reserved_size = INT_MAX;
3212
3213 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3214 INIT_LIST_HEAD(&q->requeue_list);
3215 spin_lock_init(&q->requeue_lock);
3216
3217 q->nr_requests = set->queue_depth;
3218
3219 /*
3220 * Default to classic polling
3221 */
3222 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3223
3224 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3225 blk_mq_add_queue_tag_set(set, q);
3226 blk_mq_map_swqueue(q);
3227
3228 if (elevator_init)
3229 elevator_init_mq(q);
3230
3231 return q;
3232
3233err_hctxs:
3234 kfree(q->queue_hw_ctx);
3235 q->nr_hw_queues = 0;
3236 blk_mq_sysfs_deinit(q);
3237err_poll:
3238 blk_stat_free_callback(q->poll_cb);
3239 q->poll_cb = NULL;
3240err_exit:
3241 q->mq_ops = NULL;
3242 return ERR_PTR(-ENOMEM);
3243}
3244EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3245
3246/* tags can _not_ be used after returning from blk_mq_exit_queue */
3247void blk_mq_exit_queue(struct request_queue *q)
3248{
3249 struct blk_mq_tag_set *set = q->tag_set;
3250
3251 blk_mq_del_queue_tag_set(q);
3252 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3253}
3254
3255static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3256{
3257 int i;
3258
3259 for (i = 0; i < set->nr_hw_queues; i++)
3260 if (!__blk_mq_alloc_map_and_request(set, i))
3261 goto out_unwind;
3262
3263 return 0;
3264
3265out_unwind:
3266 while (--i >= 0)
3267 blk_mq_free_map_and_requests(set, i);
3268
3269 return -ENOMEM;
3270}
3271
3272/*
3273 * Allocate the request maps associated with this tag_set. Note that this
3274 * may reduce the depth asked for, if memory is tight. set->queue_depth
3275 * will be updated to reflect the allocated depth.
3276 */
3277static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3278{
3279 unsigned int depth;
3280 int err;
3281
3282 depth = set->queue_depth;
3283 do {
3284 err = __blk_mq_alloc_rq_maps(set);
3285 if (!err)
3286 break;
3287
3288 set->queue_depth >>= 1;
3289 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3290 err = -ENOMEM;
3291 break;
3292 }
3293 } while (set->queue_depth);
3294
3295 if (!set->queue_depth || err) {
3296 pr_err("blk-mq: failed to allocate request map\n");
3297 return -ENOMEM;
3298 }
3299
3300 if (depth != set->queue_depth)
3301 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3302 depth, set->queue_depth);
3303
3304 return 0;
3305}
3306
3307static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3308{
3309 /*
3310 * blk_mq_map_queues() and multiple .map_queues() implementations
3311 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3312 * number of hardware queues.
3313 */
3314 if (set->nr_maps == 1)
3315 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3316
3317 if (set->ops->map_queues && !is_kdump_kernel()) {
3318 int i;
3319
3320 /*
3321 * transport .map_queues is usually done in the following
3322 * way:
3323 *
3324 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3325 * mask = get_cpu_mask(queue)
3326 * for_each_cpu(cpu, mask)
3327 * set->map[x].mq_map[cpu] = queue;
3328 * }
3329 *
3330 * When we need to remap, the table has to be cleared for
3331 * killing stale mapping since one CPU may not be mapped
3332 * to any hw queue.
3333 */
3334 for (i = 0; i < set->nr_maps; i++)
3335 blk_mq_clear_mq_map(&set->map[i]);
3336
3337 return set->ops->map_queues(set);
3338 } else {
3339 BUG_ON(set->nr_maps > 1);
3340 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3341 }
3342}
3343
3344static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3345 int cur_nr_hw_queues, int new_nr_hw_queues)
3346{
3347 struct blk_mq_tags **new_tags;
3348
3349 if (cur_nr_hw_queues >= new_nr_hw_queues)
3350 return 0;
3351
3352 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3353 GFP_KERNEL, set->numa_node);
3354 if (!new_tags)
3355 return -ENOMEM;
3356
3357 if (set->tags)
3358 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3359 sizeof(*set->tags));
3360 kfree(set->tags);
3361 set->tags = new_tags;
3362 set->nr_hw_queues = new_nr_hw_queues;
3363
3364 return 0;
3365}
3366
3367/*
3368 * Alloc a tag set to be associated with one or more request queues.
3369 * May fail with EINVAL for various error conditions. May adjust the
3370 * requested depth down, if it's too large. In that case, the set
3371 * value will be stored in set->queue_depth.
3372 */
3373int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3374{
3375 int i, ret;
3376
3377 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3378
3379 if (!set->nr_hw_queues)
3380 return -EINVAL;
3381 if (!set->queue_depth)
3382 return -EINVAL;
3383 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3384 return -EINVAL;
3385
3386 if (!set->ops->queue_rq)
3387 return -EINVAL;
3388
3389 if (!set->ops->get_budget ^ !set->ops->put_budget)
3390 return -EINVAL;
3391
3392 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3393 pr_info("blk-mq: reduced tag depth to %u\n",
3394 BLK_MQ_MAX_DEPTH);
3395 set->queue_depth = BLK_MQ_MAX_DEPTH;
3396 }
3397
3398 if (!set->nr_maps)
3399 set->nr_maps = 1;
3400 else if (set->nr_maps > HCTX_MAX_TYPES)
3401 return -EINVAL;
3402
3403 /*
3404 * If a crashdump is active, then we are potentially in a very
3405 * memory constrained environment. Limit us to 1 queue and
3406 * 64 tags to prevent using too much memory.
3407 */
3408 if (is_kdump_kernel()) {
3409 set->nr_hw_queues = 1;
3410 set->nr_maps = 1;
3411 set->queue_depth = min(64U, set->queue_depth);
3412 }
3413 /*
3414 * There is no use for more h/w queues than cpus if we just have
3415 * a single map
3416 */
3417 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3418 set->nr_hw_queues = nr_cpu_ids;
3419
3420 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3421 return -ENOMEM;
3422
3423 ret = -ENOMEM;
3424 for (i = 0; i < set->nr_maps; i++) {
3425 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3426 sizeof(set->map[i].mq_map[0]),
3427 GFP_KERNEL, set->numa_node);
3428 if (!set->map[i].mq_map)
3429 goto out_free_mq_map;
3430 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3431 }
3432
3433 ret = blk_mq_update_queue_map(set);
3434 if (ret)
3435 goto out_free_mq_map;
3436
3437 ret = blk_mq_alloc_map_and_requests(set);
3438 if (ret)
3439 goto out_free_mq_map;
3440
3441 mutex_init(&set->tag_list_lock);
3442 INIT_LIST_HEAD(&set->tag_list);
3443
3444 return 0;
3445
3446out_free_mq_map:
3447 for (i = 0; i < set->nr_maps; i++) {
3448 kfree(set->map[i].mq_map);
3449 set->map[i].mq_map = NULL;
3450 }
3451 kfree(set->tags);
3452 set->tags = NULL;
3453 return ret;
3454}
3455EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3456
3457void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3458{
3459 int i, j;
3460
3461 for (i = 0; i < set->nr_hw_queues; i++)
3462 blk_mq_free_map_and_requests(set, i);
3463
3464 for (j = 0; j < set->nr_maps; j++) {
3465 kfree(set->map[j].mq_map);
3466 set->map[j].mq_map = NULL;
3467 }
3468
3469 kfree(set->tags);
3470 set->tags = NULL;
3471}
3472EXPORT_SYMBOL(blk_mq_free_tag_set);
3473
3474int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3475{
3476 struct blk_mq_tag_set *set = q->tag_set;
3477 struct blk_mq_hw_ctx *hctx;
3478 int i, ret;
3479
3480 if (!set)
3481 return -EINVAL;
3482
3483 if (q->nr_requests == nr)
3484 return 0;
3485
3486 blk_mq_freeze_queue(q);
3487 blk_mq_quiesce_queue(q);
3488
3489 ret = 0;
3490 queue_for_each_hw_ctx(q, hctx, i) {
3491 if (!hctx->tags)
3492 continue;
3493 /*
3494 * If we're using an MQ scheduler, just update the scheduler
3495 * queue depth. This is similar to what the old code would do.
3496 */
3497 if (!hctx->sched_tags) {
3498 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3499 false);
3500 } else {
3501 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3502 nr, true);
3503 }
3504 if (ret)
3505 break;
3506 if (q->elevator && q->elevator->type->ops.depth_updated)
3507 q->elevator->type->ops.depth_updated(hctx);
3508 }
3509
3510 if (!ret)
3511 q->nr_requests = nr;
3512
3513 blk_mq_unquiesce_queue(q);
3514 blk_mq_unfreeze_queue(q);
3515
3516 return ret;
3517}
3518
3519/*
3520 * request_queue and elevator_type pair.
3521 * It is just used by __blk_mq_update_nr_hw_queues to cache
3522 * the elevator_type associated with a request_queue.
3523 */
3524struct blk_mq_qe_pair {
3525 struct list_head node;
3526 struct request_queue *q;
3527 struct elevator_type *type;
3528};
3529
3530/*
3531 * Cache the elevator_type in qe pair list and switch the
3532 * io scheduler to 'none'
3533 */
3534static bool blk_mq_elv_switch_none(struct list_head *head,
3535 struct request_queue *q)
3536{
3537 struct blk_mq_qe_pair *qe;
3538
3539 if (!q->elevator)
3540 return true;
3541
3542 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3543 if (!qe)
3544 return false;
3545
3546 INIT_LIST_HEAD(&qe->node);
3547 qe->q = q;
3548 qe->type = q->elevator->type;
3549 list_add(&qe->node, head);
3550
3551 mutex_lock(&q->sysfs_lock);
3552 /*
3553 * After elevator_switch_mq, the previous elevator_queue will be
3554 * released by elevator_release. The reference of the io scheduler
3555 * module get by elevator_get will also be put. So we need to get
3556 * a reference of the io scheduler module here to prevent it to be
3557 * removed.
3558 */
3559 __module_get(qe->type->elevator_owner);
3560 elevator_switch_mq(q, NULL);
3561 mutex_unlock(&q->sysfs_lock);
3562
3563 return true;
3564}
3565
3566static void blk_mq_elv_switch_back(struct list_head *head,
3567 struct request_queue *q)
3568{
3569 struct blk_mq_qe_pair *qe;
3570 struct elevator_type *t = NULL;
3571
3572 list_for_each_entry(qe, head, node)
3573 if (qe->q == q) {
3574 t = qe->type;
3575 break;
3576 }
3577
3578 if (!t)
3579 return;
3580
3581 list_del(&qe->node);
3582 kfree(qe);
3583
3584 mutex_lock(&q->sysfs_lock);
3585 elevator_switch_mq(q, t);
3586 mutex_unlock(&q->sysfs_lock);
3587}
3588
3589static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3590 int nr_hw_queues)
3591{
3592 struct request_queue *q;
3593 LIST_HEAD(head);
3594 int prev_nr_hw_queues;
3595
3596 lockdep_assert_held(&set->tag_list_lock);
3597
3598 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3599 nr_hw_queues = nr_cpu_ids;
3600 if (nr_hw_queues < 1)
3601 return;
3602 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3603 return;
3604
3605 list_for_each_entry(q, &set->tag_list, tag_set_list)
3606 blk_mq_freeze_queue(q);
3607 /*
3608 * Switch IO scheduler to 'none', cleaning up the data associated
3609 * with the previous scheduler. We will switch back once we are done
3610 * updating the new sw to hw queue mappings.
3611 */
3612 list_for_each_entry(q, &set->tag_list, tag_set_list)
3613 if (!blk_mq_elv_switch_none(&head, q))
3614 goto switch_back;
3615
3616 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3617 blk_mq_debugfs_unregister_hctxs(q);
3618 blk_mq_sysfs_unregister(q);
3619 }
3620
3621 prev_nr_hw_queues = set->nr_hw_queues;
3622 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3623 0)
3624 goto reregister;
3625
3626 set->nr_hw_queues = nr_hw_queues;
3627fallback:
3628 blk_mq_update_queue_map(set);
3629 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3630 blk_mq_realloc_hw_ctxs(set, q);
3631 if (q->nr_hw_queues != set->nr_hw_queues) {
3632 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3633 nr_hw_queues, prev_nr_hw_queues);
3634 set->nr_hw_queues = prev_nr_hw_queues;
3635 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3636 goto fallback;
3637 }
3638 blk_mq_map_swqueue(q);
3639 }
3640
3641reregister:
3642 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3643 blk_mq_sysfs_register(q);
3644 blk_mq_debugfs_register_hctxs(q);
3645 }
3646
3647switch_back:
3648 list_for_each_entry(q, &set->tag_list, tag_set_list)
3649 blk_mq_elv_switch_back(&head, q);
3650
3651 list_for_each_entry(q, &set->tag_list, tag_set_list)
3652 blk_mq_unfreeze_queue(q);
3653}
3654
3655void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3656{
3657 mutex_lock(&set->tag_list_lock);
3658 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3659 mutex_unlock(&set->tag_list_lock);
3660}
3661EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3662
3663/* Enable polling stats and return whether they were already enabled. */
3664static bool blk_poll_stats_enable(struct request_queue *q)
3665{
3666 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3667 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3668 return true;
3669 blk_stat_add_callback(q, q->poll_cb);
3670 return false;
3671}
3672
3673static void blk_mq_poll_stats_start(struct request_queue *q)
3674{
3675 /*
3676 * We don't arm the callback if polling stats are not enabled or the
3677 * callback is already active.
3678 */
3679 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3680 blk_stat_is_active(q->poll_cb))
3681 return;
3682
3683 blk_stat_activate_msecs(q->poll_cb, 100);
3684}
3685
3686static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3687{
3688 struct request_queue *q = cb->data;
3689 int bucket;
3690
3691 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3692 if (cb->stat[bucket].nr_samples)
3693 q->poll_stat[bucket] = cb->stat[bucket];
3694 }
3695}
3696
3697static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3698 struct request *rq)
3699{
3700 unsigned long ret = 0;
3701 int bucket;
3702
3703 /*
3704 * If stats collection isn't on, don't sleep but turn it on for
3705 * future users
3706 */
3707 if (!blk_poll_stats_enable(q))
3708 return 0;
3709
3710 /*
3711 * As an optimistic guess, use half of the mean service time
3712 * for this type of request. We can (and should) make this smarter.
3713 * For instance, if the completion latencies are tight, we can
3714 * get closer than just half the mean. This is especially
3715 * important on devices where the completion latencies are longer
3716 * than ~10 usec. We do use the stats for the relevant IO size
3717 * if available which does lead to better estimates.
3718 */
3719 bucket = blk_mq_poll_stats_bkt(rq);
3720 if (bucket < 0)
3721 return ret;
3722
3723 if (q->poll_stat[bucket].nr_samples)
3724 ret = (q->poll_stat[bucket].mean + 1) / 2;
3725
3726 return ret;
3727}
3728
3729static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3730 struct request *rq)
3731{
3732 struct hrtimer_sleeper hs;
3733 enum hrtimer_mode mode;
3734 unsigned int nsecs;
3735 ktime_t kt;
3736
3737 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3738 return false;
3739
3740 /*
3741 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3742 *
3743 * 0: use half of prev avg
3744 * >0: use this specific value
3745 */
3746 if (q->poll_nsec > 0)
3747 nsecs = q->poll_nsec;
3748 else
3749 nsecs = blk_mq_poll_nsecs(q, rq);
3750
3751 if (!nsecs)
3752 return false;
3753
3754 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3755
3756 /*
3757 * This will be replaced with the stats tracking code, using
3758 * 'avg_completion_time / 2' as the pre-sleep target.
3759 */
3760 kt = nsecs;
3761
3762 mode = HRTIMER_MODE_REL;
3763 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3764 hrtimer_set_expires(&hs.timer, kt);
3765
3766 do {
3767 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3768 break;
3769 set_current_state(TASK_UNINTERRUPTIBLE);
3770 hrtimer_sleeper_start_expires(&hs, mode);
3771 if (hs.task)
3772 io_schedule();
3773 hrtimer_cancel(&hs.timer);
3774 mode = HRTIMER_MODE_ABS;
3775 } while (hs.task && !signal_pending(current));
3776
3777 __set_current_state(TASK_RUNNING);
3778 destroy_hrtimer_on_stack(&hs.timer);
3779 return true;
3780}
3781
3782static bool blk_mq_poll_hybrid(struct request_queue *q,
3783 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3784{
3785 struct request *rq;
3786
3787 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3788 return false;
3789
3790 if (!blk_qc_t_is_internal(cookie))
3791 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3792 else {
3793 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3794 /*
3795 * With scheduling, if the request has completed, we'll
3796 * get a NULL return here, as we clear the sched tag when
3797 * that happens. The request still remains valid, like always,
3798 * so we should be safe with just the NULL check.
3799 */
3800 if (!rq)
3801 return false;
3802 }
3803
3804 return blk_mq_poll_hybrid_sleep(q, rq);
3805}
3806
3807/**
3808 * blk_poll - poll for IO completions
3809 * @q: the queue
3810 * @cookie: cookie passed back at IO submission time
3811 * @spin: whether to spin for completions
3812 *
3813 * Description:
3814 * Poll for completions on the passed in queue. Returns number of
3815 * completed entries found. If @spin is true, then blk_poll will continue
3816 * looping until at least one completion is found, unless the task is
3817 * otherwise marked running (or we need to reschedule).
3818 */
3819int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3820{
3821 struct blk_mq_hw_ctx *hctx;
3822 long state;
3823
3824 if (!blk_qc_t_valid(cookie) ||
3825 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3826 return 0;
3827
3828 if (current->plug)
3829 blk_flush_plug_list(current->plug, false);
3830
3831 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3832
3833 /*
3834 * If we sleep, have the caller restart the poll loop to reset
3835 * the state. Like for the other success return cases, the
3836 * caller is responsible for checking if the IO completed. If
3837 * the IO isn't complete, we'll get called again and will go
3838 * straight to the busy poll loop.
3839 */
3840 if (blk_mq_poll_hybrid(q, hctx, cookie))
3841 return 1;
3842
3843 hctx->poll_considered++;
3844
3845 state = current->state;
3846 do {
3847 int ret;
3848
3849 hctx->poll_invoked++;
3850
3851 ret = q->mq_ops->poll(hctx);
3852 if (ret > 0) {
3853 hctx->poll_success++;
3854 __set_current_state(TASK_RUNNING);
3855 return ret;
3856 }
3857
3858 if (signal_pending_state(state, current))
3859 __set_current_state(TASK_RUNNING);
3860
3861 if (current->state == TASK_RUNNING)
3862 return 1;
3863 if (ret < 0 || !spin)
3864 break;
3865 cpu_relax();
3866 } while (!need_resched());
3867
3868 __set_current_state(TASK_RUNNING);
3869 return 0;
3870}
3871EXPORT_SYMBOL_GPL(blk_poll);
3872
3873unsigned int blk_mq_rq_cpu(struct request *rq)
3874{
3875 return rq->mq_ctx->cpu;
3876}
3877EXPORT_SYMBOL(blk_mq_rq_cpu);
3878
3879static int __init blk_mq_init(void)
3880{
3881 int i;
3882
3883 for_each_possible_cpu(i)
3884 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3885 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
3886
3887 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
3888 "block/softirq:dead", NULL,
3889 blk_softirq_cpu_dead);
3890 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3891 blk_mq_hctx_notify_dead);
3892 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
3893 blk_mq_hctx_notify_online,
3894 blk_mq_hctx_notify_offline);
3895 return 0;
3896}
3897subsys_initcall(blk_mq_init);