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/blk-integrity.h>
14#include <linux/kmemleak.h>
15#include <linux/mm.h>
16#include <linux/init.h>
17#include <linux/slab.h>
18#include <linux/workqueue.h>
19#include <linux/smp.h>
20#include <linux/interrupt.h>
21#include <linux/llist.h>
22#include <linux/cpu.h>
23#include <linux/cache.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#include <linux/part_stat.h>
31
32#include <trace/events/block.h>
33
34#include <linux/t10-pi.h>
35#include "blk.h"
36#include "blk-mq.h"
37#include "blk-mq-debugfs.h"
38#include "blk-pm.h"
39#include "blk-stat.h"
40#include "blk-mq-sched.h"
41#include "blk-rq-qos.h"
42
43static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
44static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd);
45
46static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
47static void blk_mq_request_bypass_insert(struct request *rq,
48 blk_insert_t flags);
49static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
50 struct list_head *list);
51static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
52 struct io_comp_batch *iob, unsigned int flags);
53
54/*
55 * Check if any of the ctx, dispatch list or elevator
56 * have pending work in this hardware queue.
57 */
58static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
59{
60 return !list_empty_careful(&hctx->dispatch) ||
61 sbitmap_any_bit_set(&hctx->ctx_map) ||
62 blk_mq_sched_has_work(hctx);
63}
64
65/*
66 * Mark this ctx as having pending work in this hardware queue
67 */
68static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
69 struct blk_mq_ctx *ctx)
70{
71 const int bit = ctx->index_hw[hctx->type];
72
73 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
74 sbitmap_set_bit(&hctx->ctx_map, bit);
75}
76
77static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
78 struct blk_mq_ctx *ctx)
79{
80 const int bit = ctx->index_hw[hctx->type];
81
82 sbitmap_clear_bit(&hctx->ctx_map, bit);
83}
84
85struct mq_inflight {
86 struct block_device *part;
87 unsigned int inflight[2];
88};
89
90static bool blk_mq_check_inflight(struct request *rq, void *priv)
91{
92 struct mq_inflight *mi = priv;
93
94 if (rq->part && blk_do_io_stat(rq) &&
95 (!mi->part->bd_partno || rq->part == mi->part) &&
96 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
97 mi->inflight[rq_data_dir(rq)]++;
98
99 return true;
100}
101
102unsigned int blk_mq_in_flight(struct request_queue *q,
103 struct block_device *part)
104{
105 struct mq_inflight mi = { .part = part };
106
107 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
108
109 return mi.inflight[0] + mi.inflight[1];
110}
111
112void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
113 unsigned int inflight[2])
114{
115 struct mq_inflight mi = { .part = part };
116
117 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
118 inflight[0] = mi.inflight[0];
119 inflight[1] = mi.inflight[1];
120}
121
122void blk_freeze_queue_start(struct request_queue *q)
123{
124 mutex_lock(&q->mq_freeze_lock);
125 if (++q->mq_freeze_depth == 1) {
126 percpu_ref_kill(&q->q_usage_counter);
127 mutex_unlock(&q->mq_freeze_lock);
128 if (queue_is_mq(q))
129 blk_mq_run_hw_queues(q, false);
130 } else {
131 mutex_unlock(&q->mq_freeze_lock);
132 }
133}
134EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
135
136void blk_mq_freeze_queue_wait(struct request_queue *q)
137{
138 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
139}
140EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
141
142int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
143 unsigned long timeout)
144{
145 return wait_event_timeout(q->mq_freeze_wq,
146 percpu_ref_is_zero(&q->q_usage_counter),
147 timeout);
148}
149EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
150
151/*
152 * Guarantee no request is in use, so we can change any data structure of
153 * the queue afterward.
154 */
155void blk_freeze_queue(struct request_queue *q)
156{
157 /*
158 * In the !blk_mq case we are only calling this to kill the
159 * q_usage_counter, otherwise this increases the freeze depth
160 * and waits for it to return to zero. For this reason there is
161 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
162 * exported to drivers as the only user for unfreeze is blk_mq.
163 */
164 blk_freeze_queue_start(q);
165 blk_mq_freeze_queue_wait(q);
166}
167
168void blk_mq_freeze_queue(struct request_queue *q)
169{
170 /*
171 * ...just an alias to keep freeze and unfreeze actions balanced
172 * in the blk_mq_* namespace
173 */
174 blk_freeze_queue(q);
175}
176EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
177
178void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
179{
180 mutex_lock(&q->mq_freeze_lock);
181 if (force_atomic)
182 q->q_usage_counter.data->force_atomic = true;
183 q->mq_freeze_depth--;
184 WARN_ON_ONCE(q->mq_freeze_depth < 0);
185 if (!q->mq_freeze_depth) {
186 percpu_ref_resurrect(&q->q_usage_counter);
187 wake_up_all(&q->mq_freeze_wq);
188 }
189 mutex_unlock(&q->mq_freeze_lock);
190}
191
192void blk_mq_unfreeze_queue(struct request_queue *q)
193{
194 __blk_mq_unfreeze_queue(q, false);
195}
196EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
197
198/*
199 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
200 * mpt3sas driver such that this function can be removed.
201 */
202void blk_mq_quiesce_queue_nowait(struct request_queue *q)
203{
204 unsigned long flags;
205
206 spin_lock_irqsave(&q->queue_lock, flags);
207 if (!q->quiesce_depth++)
208 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
209 spin_unlock_irqrestore(&q->queue_lock, flags);
210}
211EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
212
213/**
214 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
215 * @set: tag_set to wait on
216 *
217 * Note: it is driver's responsibility for making sure that quiesce has
218 * been started on or more of the request_queues of the tag_set. This
219 * function only waits for the quiesce on those request_queues that had
220 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
221 */
222void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
223{
224 if (set->flags & BLK_MQ_F_BLOCKING)
225 synchronize_srcu(set->srcu);
226 else
227 synchronize_rcu();
228}
229EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
230
231/**
232 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
233 * @q: request queue.
234 *
235 * Note: this function does not prevent that the struct request end_io()
236 * callback function is invoked. Once this function is returned, we make
237 * sure no dispatch can happen until the queue is unquiesced via
238 * blk_mq_unquiesce_queue().
239 */
240void blk_mq_quiesce_queue(struct request_queue *q)
241{
242 blk_mq_quiesce_queue_nowait(q);
243 /* nothing to wait for non-mq queues */
244 if (queue_is_mq(q))
245 blk_mq_wait_quiesce_done(q->tag_set);
246}
247EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
248
249/*
250 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
251 * @q: request queue.
252 *
253 * This function recovers queue into the state before quiescing
254 * which is done by blk_mq_quiesce_queue.
255 */
256void blk_mq_unquiesce_queue(struct request_queue *q)
257{
258 unsigned long flags;
259 bool run_queue = false;
260
261 spin_lock_irqsave(&q->queue_lock, flags);
262 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
263 ;
264 } else if (!--q->quiesce_depth) {
265 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
266 run_queue = true;
267 }
268 spin_unlock_irqrestore(&q->queue_lock, flags);
269
270 /* dispatch requests which are inserted during quiescing */
271 if (run_queue)
272 blk_mq_run_hw_queues(q, true);
273}
274EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
275
276void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
277{
278 struct request_queue *q;
279
280 mutex_lock(&set->tag_list_lock);
281 list_for_each_entry(q, &set->tag_list, tag_set_list) {
282 if (!blk_queue_skip_tagset_quiesce(q))
283 blk_mq_quiesce_queue_nowait(q);
284 }
285 blk_mq_wait_quiesce_done(set);
286 mutex_unlock(&set->tag_list_lock);
287}
288EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
289
290void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
291{
292 struct request_queue *q;
293
294 mutex_lock(&set->tag_list_lock);
295 list_for_each_entry(q, &set->tag_list, tag_set_list) {
296 if (!blk_queue_skip_tagset_quiesce(q))
297 blk_mq_unquiesce_queue(q);
298 }
299 mutex_unlock(&set->tag_list_lock);
300}
301EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
302
303void blk_mq_wake_waiters(struct request_queue *q)
304{
305 struct blk_mq_hw_ctx *hctx;
306 unsigned long i;
307
308 queue_for_each_hw_ctx(q, hctx, i)
309 if (blk_mq_hw_queue_mapped(hctx))
310 blk_mq_tag_wakeup_all(hctx->tags, true);
311}
312
313void blk_rq_init(struct request_queue *q, struct request *rq)
314{
315 memset(rq, 0, sizeof(*rq));
316
317 INIT_LIST_HEAD(&rq->queuelist);
318 rq->q = q;
319 rq->__sector = (sector_t) -1;
320 INIT_HLIST_NODE(&rq->hash);
321 RB_CLEAR_NODE(&rq->rb_node);
322 rq->tag = BLK_MQ_NO_TAG;
323 rq->internal_tag = BLK_MQ_NO_TAG;
324 rq->start_time_ns = blk_time_get_ns();
325 rq->part = NULL;
326 blk_crypto_rq_set_defaults(rq);
327}
328EXPORT_SYMBOL(blk_rq_init);
329
330/* Set start and alloc time when the allocated request is actually used */
331static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
332{
333 if (blk_mq_need_time_stamp(rq))
334 rq->start_time_ns = blk_time_get_ns();
335 else
336 rq->start_time_ns = 0;
337
338#ifdef CONFIG_BLK_RQ_ALLOC_TIME
339 if (blk_queue_rq_alloc_time(rq->q))
340 rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;
341 else
342 rq->alloc_time_ns = 0;
343#endif
344}
345
346static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
347 struct blk_mq_tags *tags, unsigned int tag)
348{
349 struct blk_mq_ctx *ctx = data->ctx;
350 struct blk_mq_hw_ctx *hctx = data->hctx;
351 struct request_queue *q = data->q;
352 struct request *rq = tags->static_rqs[tag];
353
354 rq->q = q;
355 rq->mq_ctx = ctx;
356 rq->mq_hctx = hctx;
357 rq->cmd_flags = data->cmd_flags;
358
359 if (data->flags & BLK_MQ_REQ_PM)
360 data->rq_flags |= RQF_PM;
361 if (blk_queue_io_stat(q))
362 data->rq_flags |= RQF_IO_STAT;
363 rq->rq_flags = data->rq_flags;
364
365 if (data->rq_flags & RQF_SCHED_TAGS) {
366 rq->tag = BLK_MQ_NO_TAG;
367 rq->internal_tag = tag;
368 } else {
369 rq->tag = tag;
370 rq->internal_tag = BLK_MQ_NO_TAG;
371 }
372 rq->timeout = 0;
373
374 rq->part = NULL;
375 rq->io_start_time_ns = 0;
376 rq->stats_sectors = 0;
377 rq->nr_phys_segments = 0;
378#if defined(CONFIG_BLK_DEV_INTEGRITY)
379 rq->nr_integrity_segments = 0;
380#endif
381 rq->end_io = NULL;
382 rq->end_io_data = NULL;
383
384 blk_crypto_rq_set_defaults(rq);
385 INIT_LIST_HEAD(&rq->queuelist);
386 /* tag was already set */
387 WRITE_ONCE(rq->deadline, 0);
388 req_ref_set(rq, 1);
389
390 if (rq->rq_flags & RQF_USE_SCHED) {
391 struct elevator_queue *e = data->q->elevator;
392
393 INIT_HLIST_NODE(&rq->hash);
394 RB_CLEAR_NODE(&rq->rb_node);
395
396 if (e->type->ops.prepare_request)
397 e->type->ops.prepare_request(rq);
398 }
399
400 return rq;
401}
402
403static inline struct request *
404__blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
405{
406 unsigned int tag, tag_offset;
407 struct blk_mq_tags *tags;
408 struct request *rq;
409 unsigned long tag_mask;
410 int i, nr = 0;
411
412 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
413 if (unlikely(!tag_mask))
414 return NULL;
415
416 tags = blk_mq_tags_from_data(data);
417 for (i = 0; tag_mask; i++) {
418 if (!(tag_mask & (1UL << i)))
419 continue;
420 tag = tag_offset + i;
421 prefetch(tags->static_rqs[tag]);
422 tag_mask &= ~(1UL << i);
423 rq = blk_mq_rq_ctx_init(data, tags, tag);
424 rq_list_add(data->cached_rq, rq);
425 nr++;
426 }
427 if (!(data->rq_flags & RQF_SCHED_TAGS))
428 blk_mq_add_active_requests(data->hctx, nr);
429 /* caller already holds a reference, add for remainder */
430 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
431 data->nr_tags -= nr;
432
433 return rq_list_pop(data->cached_rq);
434}
435
436static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
437{
438 struct request_queue *q = data->q;
439 u64 alloc_time_ns = 0;
440 struct request *rq;
441 unsigned int tag;
442
443 /* alloc_time includes depth and tag waits */
444 if (blk_queue_rq_alloc_time(q))
445 alloc_time_ns = blk_time_get_ns();
446
447 if (data->cmd_flags & REQ_NOWAIT)
448 data->flags |= BLK_MQ_REQ_NOWAIT;
449
450 if (q->elevator) {
451 /*
452 * All requests use scheduler tags when an I/O scheduler is
453 * enabled for the queue.
454 */
455 data->rq_flags |= RQF_SCHED_TAGS;
456
457 /*
458 * Flush/passthrough requests are special and go directly to the
459 * dispatch list.
460 */
461 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
462 !blk_op_is_passthrough(data->cmd_flags)) {
463 struct elevator_mq_ops *ops = &q->elevator->type->ops;
464
465 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
466
467 data->rq_flags |= RQF_USE_SCHED;
468 if (ops->limit_depth)
469 ops->limit_depth(data->cmd_flags, data);
470 }
471 }
472
473retry:
474 data->ctx = blk_mq_get_ctx(q);
475 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
476 if (!(data->rq_flags & RQF_SCHED_TAGS))
477 blk_mq_tag_busy(data->hctx);
478
479 if (data->flags & BLK_MQ_REQ_RESERVED)
480 data->rq_flags |= RQF_RESV;
481
482 /*
483 * Try batched alloc if we want more than 1 tag.
484 */
485 if (data->nr_tags > 1) {
486 rq = __blk_mq_alloc_requests_batch(data);
487 if (rq) {
488 blk_mq_rq_time_init(rq, alloc_time_ns);
489 return rq;
490 }
491 data->nr_tags = 1;
492 }
493
494 /*
495 * Waiting allocations only fail because of an inactive hctx. In that
496 * case just retry the hctx assignment and tag allocation as CPU hotplug
497 * should have migrated us to an online CPU by now.
498 */
499 tag = blk_mq_get_tag(data);
500 if (tag == BLK_MQ_NO_TAG) {
501 if (data->flags & BLK_MQ_REQ_NOWAIT)
502 return NULL;
503 /*
504 * Give up the CPU and sleep for a random short time to
505 * ensure that thread using a realtime scheduling class
506 * are migrated off the CPU, and thus off the hctx that
507 * is going away.
508 */
509 msleep(3);
510 goto retry;
511 }
512
513 if (!(data->rq_flags & RQF_SCHED_TAGS))
514 blk_mq_inc_active_requests(data->hctx);
515 rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
516 blk_mq_rq_time_init(rq, alloc_time_ns);
517 return rq;
518}
519
520static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
521 struct blk_plug *plug,
522 blk_opf_t opf,
523 blk_mq_req_flags_t flags)
524{
525 struct blk_mq_alloc_data data = {
526 .q = q,
527 .flags = flags,
528 .cmd_flags = opf,
529 .nr_tags = plug->nr_ios,
530 .cached_rq = &plug->cached_rq,
531 };
532 struct request *rq;
533
534 if (blk_queue_enter(q, flags))
535 return NULL;
536
537 plug->nr_ios = 1;
538
539 rq = __blk_mq_alloc_requests(&data);
540 if (unlikely(!rq))
541 blk_queue_exit(q);
542 return rq;
543}
544
545static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
546 blk_opf_t opf,
547 blk_mq_req_flags_t flags)
548{
549 struct blk_plug *plug = current->plug;
550 struct request *rq;
551
552 if (!plug)
553 return NULL;
554
555 if (rq_list_empty(plug->cached_rq)) {
556 if (plug->nr_ios == 1)
557 return NULL;
558 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
559 if (!rq)
560 return NULL;
561 } else {
562 rq = rq_list_peek(&plug->cached_rq);
563 if (!rq || rq->q != q)
564 return NULL;
565
566 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
567 return NULL;
568 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
569 return NULL;
570
571 plug->cached_rq = rq_list_next(rq);
572 blk_mq_rq_time_init(rq, 0);
573 }
574
575 rq->cmd_flags = opf;
576 INIT_LIST_HEAD(&rq->queuelist);
577 return rq;
578}
579
580struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
581 blk_mq_req_flags_t flags)
582{
583 struct request *rq;
584
585 rq = blk_mq_alloc_cached_request(q, opf, flags);
586 if (!rq) {
587 struct blk_mq_alloc_data data = {
588 .q = q,
589 .flags = flags,
590 .cmd_flags = opf,
591 .nr_tags = 1,
592 };
593 int ret;
594
595 ret = blk_queue_enter(q, flags);
596 if (ret)
597 return ERR_PTR(ret);
598
599 rq = __blk_mq_alloc_requests(&data);
600 if (!rq)
601 goto out_queue_exit;
602 }
603 rq->__data_len = 0;
604 rq->__sector = (sector_t) -1;
605 rq->bio = rq->biotail = NULL;
606 return rq;
607out_queue_exit:
608 blk_queue_exit(q);
609 return ERR_PTR(-EWOULDBLOCK);
610}
611EXPORT_SYMBOL(blk_mq_alloc_request);
612
613struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
614 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
615{
616 struct blk_mq_alloc_data data = {
617 .q = q,
618 .flags = flags,
619 .cmd_flags = opf,
620 .nr_tags = 1,
621 };
622 u64 alloc_time_ns = 0;
623 struct request *rq;
624 unsigned int cpu;
625 unsigned int tag;
626 int ret;
627
628 /* alloc_time includes depth and tag waits */
629 if (blk_queue_rq_alloc_time(q))
630 alloc_time_ns = blk_time_get_ns();
631
632 /*
633 * If the tag allocator sleeps we could get an allocation for a
634 * different hardware context. No need to complicate the low level
635 * allocator for this for the rare use case of a command tied to
636 * a specific queue.
637 */
638 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
639 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
640 return ERR_PTR(-EINVAL);
641
642 if (hctx_idx >= q->nr_hw_queues)
643 return ERR_PTR(-EIO);
644
645 ret = blk_queue_enter(q, flags);
646 if (ret)
647 return ERR_PTR(ret);
648
649 /*
650 * Check if the hardware context is actually mapped to anything.
651 * If not tell the caller that it should skip this queue.
652 */
653 ret = -EXDEV;
654 data.hctx = xa_load(&q->hctx_table, hctx_idx);
655 if (!blk_mq_hw_queue_mapped(data.hctx))
656 goto out_queue_exit;
657 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
658 if (cpu >= nr_cpu_ids)
659 goto out_queue_exit;
660 data.ctx = __blk_mq_get_ctx(q, cpu);
661
662 if (q->elevator)
663 data.rq_flags |= RQF_SCHED_TAGS;
664 else
665 blk_mq_tag_busy(data.hctx);
666
667 if (flags & BLK_MQ_REQ_RESERVED)
668 data.rq_flags |= RQF_RESV;
669
670 ret = -EWOULDBLOCK;
671 tag = blk_mq_get_tag(&data);
672 if (tag == BLK_MQ_NO_TAG)
673 goto out_queue_exit;
674 if (!(data.rq_flags & RQF_SCHED_TAGS))
675 blk_mq_inc_active_requests(data.hctx);
676 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
677 blk_mq_rq_time_init(rq, alloc_time_ns);
678 rq->__data_len = 0;
679 rq->__sector = (sector_t) -1;
680 rq->bio = rq->biotail = NULL;
681 return rq;
682
683out_queue_exit:
684 blk_queue_exit(q);
685 return ERR_PTR(ret);
686}
687EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
688
689static void blk_mq_finish_request(struct request *rq)
690{
691 struct request_queue *q = rq->q;
692
693 if (rq->rq_flags & RQF_USE_SCHED) {
694 q->elevator->type->ops.finish_request(rq);
695 /*
696 * For postflush request that may need to be
697 * completed twice, we should clear this flag
698 * to avoid double finish_request() on the rq.
699 */
700 rq->rq_flags &= ~RQF_USE_SCHED;
701 }
702}
703
704static void __blk_mq_free_request(struct request *rq)
705{
706 struct request_queue *q = rq->q;
707 struct blk_mq_ctx *ctx = rq->mq_ctx;
708 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
709 const int sched_tag = rq->internal_tag;
710
711 blk_crypto_free_request(rq);
712 blk_pm_mark_last_busy(rq);
713 rq->mq_hctx = NULL;
714
715 if (rq->tag != BLK_MQ_NO_TAG) {
716 blk_mq_dec_active_requests(hctx);
717 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
718 }
719 if (sched_tag != BLK_MQ_NO_TAG)
720 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
721 blk_mq_sched_restart(hctx);
722 blk_queue_exit(q);
723}
724
725void blk_mq_free_request(struct request *rq)
726{
727 struct request_queue *q = rq->q;
728
729 blk_mq_finish_request(rq);
730
731 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
732 laptop_io_completion(q->disk->bdi);
733
734 rq_qos_done(q, rq);
735
736 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
737 if (req_ref_put_and_test(rq))
738 __blk_mq_free_request(rq);
739}
740EXPORT_SYMBOL_GPL(blk_mq_free_request);
741
742void blk_mq_free_plug_rqs(struct blk_plug *plug)
743{
744 struct request *rq;
745
746 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
747 blk_mq_free_request(rq);
748}
749
750void blk_dump_rq_flags(struct request *rq, char *msg)
751{
752 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
753 rq->q->disk ? rq->q->disk->disk_name : "?",
754 (__force unsigned long long) rq->cmd_flags);
755
756 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
757 (unsigned long long)blk_rq_pos(rq),
758 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
759 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
760 rq->bio, rq->biotail, blk_rq_bytes(rq));
761}
762EXPORT_SYMBOL(blk_dump_rq_flags);
763
764static void req_bio_endio(struct request *rq, struct bio *bio,
765 unsigned int nbytes, blk_status_t error)
766{
767 if (unlikely(error)) {
768 bio->bi_status = error;
769 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
770 /*
771 * Partial zone append completions cannot be supported as the
772 * BIO fragments may end up not being written sequentially.
773 */
774 if (bio->bi_iter.bi_size != nbytes)
775 bio->bi_status = BLK_STS_IOERR;
776 else
777 bio->bi_iter.bi_sector = rq->__sector;
778 }
779
780 bio_advance(bio, nbytes);
781
782 if (unlikely(rq->rq_flags & RQF_QUIET))
783 bio_set_flag(bio, BIO_QUIET);
784 /* don't actually finish bio if it's part of flush sequence */
785 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
786 bio_endio(bio);
787}
788
789static void blk_account_io_completion(struct request *req, unsigned int bytes)
790{
791 if (req->part && blk_do_io_stat(req)) {
792 const int sgrp = op_stat_group(req_op(req));
793
794 part_stat_lock();
795 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
796 part_stat_unlock();
797 }
798}
799
800static void blk_print_req_error(struct request *req, blk_status_t status)
801{
802 printk_ratelimited(KERN_ERR
803 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
804 "phys_seg %u prio class %u\n",
805 blk_status_to_str(status),
806 req->q->disk ? req->q->disk->disk_name : "?",
807 blk_rq_pos(req), (__force u32)req_op(req),
808 blk_op_str(req_op(req)),
809 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
810 req->nr_phys_segments,
811 IOPRIO_PRIO_CLASS(req->ioprio));
812}
813
814/*
815 * Fully end IO on a request. Does not support partial completions, or
816 * errors.
817 */
818static void blk_complete_request(struct request *req)
819{
820 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
821 int total_bytes = blk_rq_bytes(req);
822 struct bio *bio = req->bio;
823
824 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
825
826 if (!bio)
827 return;
828
829#ifdef CONFIG_BLK_DEV_INTEGRITY
830 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
831 req->q->integrity.profile->complete_fn(req, total_bytes);
832#endif
833
834 /*
835 * Upper layers may call blk_crypto_evict_key() anytime after the last
836 * bio_endio(). Therefore, the keyslot must be released before that.
837 */
838 blk_crypto_rq_put_keyslot(req);
839
840 blk_account_io_completion(req, total_bytes);
841
842 do {
843 struct bio *next = bio->bi_next;
844
845 /* Completion has already been traced */
846 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
847
848 if (req_op(req) == REQ_OP_ZONE_APPEND)
849 bio->bi_iter.bi_sector = req->__sector;
850
851 if (!is_flush)
852 bio_endio(bio);
853 bio = next;
854 } while (bio);
855
856 /*
857 * Reset counters so that the request stacking driver
858 * can find how many bytes remain in the request
859 * later.
860 */
861 if (!req->end_io) {
862 req->bio = NULL;
863 req->__data_len = 0;
864 }
865}
866
867/**
868 * blk_update_request - Complete multiple bytes without completing the request
869 * @req: the request being processed
870 * @error: block status code
871 * @nr_bytes: number of bytes to complete for @req
872 *
873 * Description:
874 * Ends I/O on a number of bytes attached to @req, but doesn't complete
875 * the request structure even if @req doesn't have leftover.
876 * If @req has leftover, sets it up for the next range of segments.
877 *
878 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
879 * %false return from this function.
880 *
881 * Note:
882 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
883 * except in the consistency check at the end of this function.
884 *
885 * Return:
886 * %false - this request doesn't have any more data
887 * %true - this request has more data
888 **/
889bool blk_update_request(struct request *req, blk_status_t error,
890 unsigned int nr_bytes)
891{
892 int total_bytes;
893
894 trace_block_rq_complete(req, error, nr_bytes);
895
896 if (!req->bio)
897 return false;
898
899#ifdef CONFIG_BLK_DEV_INTEGRITY
900 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
901 error == BLK_STS_OK)
902 req->q->integrity.profile->complete_fn(req, nr_bytes);
903#endif
904
905 /*
906 * Upper layers may call blk_crypto_evict_key() anytime after the last
907 * bio_endio(). Therefore, the keyslot must be released before that.
908 */
909 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
910 __blk_crypto_rq_put_keyslot(req);
911
912 if (unlikely(error && !blk_rq_is_passthrough(req) &&
913 !(req->rq_flags & RQF_QUIET)) &&
914 !test_bit(GD_DEAD, &req->q->disk->state)) {
915 blk_print_req_error(req, error);
916 trace_block_rq_error(req, error, nr_bytes);
917 }
918
919 blk_account_io_completion(req, nr_bytes);
920
921 total_bytes = 0;
922 while (req->bio) {
923 struct bio *bio = req->bio;
924 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
925
926 if (bio_bytes == bio->bi_iter.bi_size)
927 req->bio = bio->bi_next;
928
929 /* Completion has already been traced */
930 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
931 req_bio_endio(req, bio, bio_bytes, error);
932
933 total_bytes += bio_bytes;
934 nr_bytes -= bio_bytes;
935
936 if (!nr_bytes)
937 break;
938 }
939
940 /*
941 * completely done
942 */
943 if (!req->bio) {
944 /*
945 * Reset counters so that the request stacking driver
946 * can find how many bytes remain in the request
947 * later.
948 */
949 req->__data_len = 0;
950 return false;
951 }
952
953 req->__data_len -= total_bytes;
954
955 /* update sector only for requests with clear definition of sector */
956 if (!blk_rq_is_passthrough(req))
957 req->__sector += total_bytes >> 9;
958
959 /* mixed attributes always follow the first bio */
960 if (req->rq_flags & RQF_MIXED_MERGE) {
961 req->cmd_flags &= ~REQ_FAILFAST_MASK;
962 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
963 }
964
965 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
966 /*
967 * If total number of sectors is less than the first segment
968 * size, something has gone terribly wrong.
969 */
970 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
971 blk_dump_rq_flags(req, "request botched");
972 req->__data_len = blk_rq_cur_bytes(req);
973 }
974
975 /* recalculate the number of segments */
976 req->nr_phys_segments = blk_recalc_rq_segments(req);
977 }
978
979 return true;
980}
981EXPORT_SYMBOL_GPL(blk_update_request);
982
983static inline void blk_account_io_done(struct request *req, u64 now)
984{
985 trace_block_io_done(req);
986
987 /*
988 * Account IO completion. flush_rq isn't accounted as a
989 * normal IO on queueing nor completion. Accounting the
990 * containing request is enough.
991 */
992 if (blk_do_io_stat(req) && req->part &&
993 !(req->rq_flags & RQF_FLUSH_SEQ)) {
994 const int sgrp = op_stat_group(req_op(req));
995
996 part_stat_lock();
997 update_io_ticks(req->part, jiffies, true);
998 part_stat_inc(req->part, ios[sgrp]);
999 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
1000 part_stat_local_dec(req->part,
1001 in_flight[op_is_write(req_op(req))]);
1002 part_stat_unlock();
1003 }
1004}
1005
1006static inline void blk_account_io_start(struct request *req)
1007{
1008 trace_block_io_start(req);
1009
1010 if (blk_do_io_stat(req)) {
1011 /*
1012 * All non-passthrough requests are created from a bio with one
1013 * exception: when a flush command that is part of a flush sequence
1014 * generated by the state machine in blk-flush.c is cloned onto the
1015 * lower device by dm-multipath we can get here without a bio.
1016 */
1017 if (req->bio)
1018 req->part = req->bio->bi_bdev;
1019 else
1020 req->part = req->q->disk->part0;
1021
1022 part_stat_lock();
1023 update_io_ticks(req->part, jiffies, false);
1024 part_stat_local_inc(req->part,
1025 in_flight[op_is_write(req_op(req))]);
1026 part_stat_unlock();
1027 }
1028}
1029
1030static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1031{
1032 if (rq->rq_flags & RQF_STATS)
1033 blk_stat_add(rq, now);
1034
1035 blk_mq_sched_completed_request(rq, now);
1036 blk_account_io_done(rq, now);
1037}
1038
1039inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1040{
1041 if (blk_mq_need_time_stamp(rq))
1042 __blk_mq_end_request_acct(rq, blk_time_get_ns());
1043
1044 blk_mq_finish_request(rq);
1045
1046 if (rq->end_io) {
1047 rq_qos_done(rq->q, rq);
1048 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1049 blk_mq_free_request(rq);
1050 } else {
1051 blk_mq_free_request(rq);
1052 }
1053}
1054EXPORT_SYMBOL(__blk_mq_end_request);
1055
1056void blk_mq_end_request(struct request *rq, blk_status_t error)
1057{
1058 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1059 BUG();
1060 __blk_mq_end_request(rq, error);
1061}
1062EXPORT_SYMBOL(blk_mq_end_request);
1063
1064#define TAG_COMP_BATCH 32
1065
1066static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1067 int *tag_array, int nr_tags)
1068{
1069 struct request_queue *q = hctx->queue;
1070
1071 blk_mq_sub_active_requests(hctx, nr_tags);
1072
1073 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1074 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1075}
1076
1077void blk_mq_end_request_batch(struct io_comp_batch *iob)
1078{
1079 int tags[TAG_COMP_BATCH], nr_tags = 0;
1080 struct blk_mq_hw_ctx *cur_hctx = NULL;
1081 struct request *rq;
1082 u64 now = 0;
1083
1084 if (iob->need_ts)
1085 now = blk_time_get_ns();
1086
1087 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1088 prefetch(rq->bio);
1089 prefetch(rq->rq_next);
1090
1091 blk_complete_request(rq);
1092 if (iob->need_ts)
1093 __blk_mq_end_request_acct(rq, now);
1094
1095 blk_mq_finish_request(rq);
1096
1097 rq_qos_done(rq->q, rq);
1098
1099 /*
1100 * If end_io handler returns NONE, then it still has
1101 * ownership of the request.
1102 */
1103 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1104 continue;
1105
1106 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1107 if (!req_ref_put_and_test(rq))
1108 continue;
1109
1110 blk_crypto_free_request(rq);
1111 blk_pm_mark_last_busy(rq);
1112
1113 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1114 if (cur_hctx)
1115 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1116 nr_tags = 0;
1117 cur_hctx = rq->mq_hctx;
1118 }
1119 tags[nr_tags++] = rq->tag;
1120 }
1121
1122 if (nr_tags)
1123 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1124}
1125EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1126
1127static void blk_complete_reqs(struct llist_head *list)
1128{
1129 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1130 struct request *rq, *next;
1131
1132 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1133 rq->q->mq_ops->complete(rq);
1134}
1135
1136static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1137{
1138 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1139}
1140
1141static int blk_softirq_cpu_dead(unsigned int cpu)
1142{
1143 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1144 return 0;
1145}
1146
1147static void __blk_mq_complete_request_remote(void *data)
1148{
1149 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1150}
1151
1152static inline bool blk_mq_complete_need_ipi(struct request *rq)
1153{
1154 int cpu = raw_smp_processor_id();
1155
1156 if (!IS_ENABLED(CONFIG_SMP) ||
1157 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1158 return false;
1159 /*
1160 * With force threaded interrupts enabled, raising softirq from an SMP
1161 * function call will always result in waking the ksoftirqd thread.
1162 * This is probably worse than completing the request on a different
1163 * cache domain.
1164 */
1165 if (force_irqthreads())
1166 return false;
1167
1168 /* same CPU or cache domain and capacity? Complete locally */
1169 if (cpu == rq->mq_ctx->cpu ||
1170 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1171 cpus_share_cache(cpu, rq->mq_ctx->cpu) &&
1172 cpus_equal_capacity(cpu, rq->mq_ctx->cpu)))
1173 return false;
1174
1175 /* don't try to IPI to an offline CPU */
1176 return cpu_online(rq->mq_ctx->cpu);
1177}
1178
1179static void blk_mq_complete_send_ipi(struct request *rq)
1180{
1181 unsigned int cpu;
1182
1183 cpu = rq->mq_ctx->cpu;
1184 if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))
1185 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));
1186}
1187
1188static void blk_mq_raise_softirq(struct request *rq)
1189{
1190 struct llist_head *list;
1191
1192 preempt_disable();
1193 list = this_cpu_ptr(&blk_cpu_done);
1194 if (llist_add(&rq->ipi_list, list))
1195 raise_softirq(BLOCK_SOFTIRQ);
1196 preempt_enable();
1197}
1198
1199bool blk_mq_complete_request_remote(struct request *rq)
1200{
1201 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1202
1203 /*
1204 * For request which hctx has only one ctx mapping,
1205 * or a polled request, always complete locally,
1206 * it's pointless to redirect the completion.
1207 */
1208 if ((rq->mq_hctx->nr_ctx == 1 &&
1209 rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1210 rq->cmd_flags & REQ_POLLED)
1211 return false;
1212
1213 if (blk_mq_complete_need_ipi(rq)) {
1214 blk_mq_complete_send_ipi(rq);
1215 return true;
1216 }
1217
1218 if (rq->q->nr_hw_queues == 1) {
1219 blk_mq_raise_softirq(rq);
1220 return true;
1221 }
1222 return false;
1223}
1224EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1225
1226/**
1227 * blk_mq_complete_request - end I/O on a request
1228 * @rq: the request being processed
1229 *
1230 * Description:
1231 * Complete a request by scheduling the ->complete_rq operation.
1232 **/
1233void blk_mq_complete_request(struct request *rq)
1234{
1235 if (!blk_mq_complete_request_remote(rq))
1236 rq->q->mq_ops->complete(rq);
1237}
1238EXPORT_SYMBOL(blk_mq_complete_request);
1239
1240/**
1241 * blk_mq_start_request - Start processing a request
1242 * @rq: Pointer to request to be started
1243 *
1244 * Function used by device drivers to notify the block layer that a request
1245 * is going to be processed now, so blk layer can do proper initializations
1246 * such as starting the timeout timer.
1247 */
1248void blk_mq_start_request(struct request *rq)
1249{
1250 struct request_queue *q = rq->q;
1251
1252 trace_block_rq_issue(rq);
1253
1254 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) &&
1255 !blk_rq_is_passthrough(rq)) {
1256 rq->io_start_time_ns = blk_time_get_ns();
1257 rq->stats_sectors = blk_rq_sectors(rq);
1258 rq->rq_flags |= RQF_STATS;
1259 rq_qos_issue(q, rq);
1260 }
1261
1262 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1263
1264 blk_add_timer(rq);
1265 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1266 rq->mq_hctx->tags->rqs[rq->tag] = rq;
1267
1268#ifdef CONFIG_BLK_DEV_INTEGRITY
1269 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1270 q->integrity.profile->prepare_fn(rq);
1271#endif
1272 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1273 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1274}
1275EXPORT_SYMBOL(blk_mq_start_request);
1276
1277/*
1278 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1279 * queues. This is important for md arrays to benefit from merging
1280 * requests.
1281 */
1282static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1283{
1284 if (plug->multiple_queues)
1285 return BLK_MAX_REQUEST_COUNT * 2;
1286 return BLK_MAX_REQUEST_COUNT;
1287}
1288
1289static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1290{
1291 struct request *last = rq_list_peek(&plug->mq_list);
1292
1293 if (!plug->rq_count) {
1294 trace_block_plug(rq->q);
1295 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1296 (!blk_queue_nomerges(rq->q) &&
1297 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1298 blk_mq_flush_plug_list(plug, false);
1299 last = NULL;
1300 trace_block_plug(rq->q);
1301 }
1302
1303 if (!plug->multiple_queues && last && last->q != rq->q)
1304 plug->multiple_queues = true;
1305 /*
1306 * Any request allocated from sched tags can't be issued to
1307 * ->queue_rqs() directly
1308 */
1309 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1310 plug->has_elevator = true;
1311 rq->rq_next = NULL;
1312 rq_list_add(&plug->mq_list, rq);
1313 plug->rq_count++;
1314}
1315
1316/**
1317 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1318 * @rq: request to insert
1319 * @at_head: insert request at head or tail of queue
1320 *
1321 * Description:
1322 * Insert a fully prepared request at the back of the I/O scheduler queue
1323 * for execution. Don't wait for completion.
1324 *
1325 * Note:
1326 * This function will invoke @done directly if the queue is dead.
1327 */
1328void blk_execute_rq_nowait(struct request *rq, bool at_head)
1329{
1330 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1331
1332 WARN_ON(irqs_disabled());
1333 WARN_ON(!blk_rq_is_passthrough(rq));
1334
1335 blk_account_io_start(rq);
1336
1337 /*
1338 * As plugging can be enabled for passthrough requests on a zoned
1339 * device, directly accessing the plug instead of using blk_mq_plug()
1340 * should not have any consequences.
1341 */
1342 if (current->plug && !at_head) {
1343 blk_add_rq_to_plug(current->plug, rq);
1344 return;
1345 }
1346
1347 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1348 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
1349}
1350EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1351
1352struct blk_rq_wait {
1353 struct completion done;
1354 blk_status_t ret;
1355};
1356
1357static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1358{
1359 struct blk_rq_wait *wait = rq->end_io_data;
1360
1361 wait->ret = ret;
1362 complete(&wait->done);
1363 return RQ_END_IO_NONE;
1364}
1365
1366bool blk_rq_is_poll(struct request *rq)
1367{
1368 if (!rq->mq_hctx)
1369 return false;
1370 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1371 return false;
1372 return true;
1373}
1374EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1375
1376static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1377{
1378 do {
1379 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
1380 cond_resched();
1381 } while (!completion_done(wait));
1382}
1383
1384/**
1385 * blk_execute_rq - insert a request into queue for execution
1386 * @rq: request to insert
1387 * @at_head: insert request at head or tail of queue
1388 *
1389 * Description:
1390 * Insert a fully prepared request at the back of the I/O scheduler queue
1391 * for execution and wait for completion.
1392 * Return: The blk_status_t result provided to blk_mq_end_request().
1393 */
1394blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1395{
1396 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1397 struct blk_rq_wait wait = {
1398 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1399 };
1400
1401 WARN_ON(irqs_disabled());
1402 WARN_ON(!blk_rq_is_passthrough(rq));
1403
1404 rq->end_io_data = &wait;
1405 rq->end_io = blk_end_sync_rq;
1406
1407 blk_account_io_start(rq);
1408 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1409 blk_mq_run_hw_queue(hctx, false);
1410
1411 if (blk_rq_is_poll(rq))
1412 blk_rq_poll_completion(rq, &wait.done);
1413 else
1414 blk_wait_io(&wait.done);
1415
1416 return wait.ret;
1417}
1418EXPORT_SYMBOL(blk_execute_rq);
1419
1420static void __blk_mq_requeue_request(struct request *rq)
1421{
1422 struct request_queue *q = rq->q;
1423
1424 blk_mq_put_driver_tag(rq);
1425
1426 trace_block_rq_requeue(rq);
1427 rq_qos_requeue(q, rq);
1428
1429 if (blk_mq_request_started(rq)) {
1430 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1431 rq->rq_flags &= ~RQF_TIMED_OUT;
1432 }
1433}
1434
1435void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1436{
1437 struct request_queue *q = rq->q;
1438 unsigned long flags;
1439
1440 __blk_mq_requeue_request(rq);
1441
1442 /* this request will be re-inserted to io scheduler queue */
1443 blk_mq_sched_requeue_request(rq);
1444
1445 spin_lock_irqsave(&q->requeue_lock, flags);
1446 list_add_tail(&rq->queuelist, &q->requeue_list);
1447 spin_unlock_irqrestore(&q->requeue_lock, flags);
1448
1449 if (kick_requeue_list)
1450 blk_mq_kick_requeue_list(q);
1451}
1452EXPORT_SYMBOL(blk_mq_requeue_request);
1453
1454static void blk_mq_requeue_work(struct work_struct *work)
1455{
1456 struct request_queue *q =
1457 container_of(work, struct request_queue, requeue_work.work);
1458 LIST_HEAD(rq_list);
1459 LIST_HEAD(flush_list);
1460 struct request *rq;
1461
1462 spin_lock_irq(&q->requeue_lock);
1463 list_splice_init(&q->requeue_list, &rq_list);
1464 list_splice_init(&q->flush_list, &flush_list);
1465 spin_unlock_irq(&q->requeue_lock);
1466
1467 while (!list_empty(&rq_list)) {
1468 rq = list_entry(rq_list.next, struct request, queuelist);
1469 /*
1470 * If RQF_DONTPREP ist set, the request has been started by the
1471 * driver already and might have driver-specific data allocated
1472 * already. Insert it into the hctx dispatch list to avoid
1473 * block layer merges for the request.
1474 */
1475 if (rq->rq_flags & RQF_DONTPREP) {
1476 list_del_init(&rq->queuelist);
1477 blk_mq_request_bypass_insert(rq, 0);
1478 } else {
1479 list_del_init(&rq->queuelist);
1480 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1481 }
1482 }
1483
1484 while (!list_empty(&flush_list)) {
1485 rq = list_entry(flush_list.next, struct request, queuelist);
1486 list_del_init(&rq->queuelist);
1487 blk_mq_insert_request(rq, 0);
1488 }
1489
1490 blk_mq_run_hw_queues(q, false);
1491}
1492
1493void blk_mq_kick_requeue_list(struct request_queue *q)
1494{
1495 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1496}
1497EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1498
1499void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1500 unsigned long msecs)
1501{
1502 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1503 msecs_to_jiffies(msecs));
1504}
1505EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1506
1507static bool blk_is_flush_data_rq(struct request *rq)
1508{
1509 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);
1510}
1511
1512static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1513{
1514 /*
1515 * If we find a request that isn't idle we know the queue is busy
1516 * as it's checked in the iter.
1517 * Return false to stop the iteration.
1518 *
1519 * In case of queue quiesce, if one flush data request is completed,
1520 * don't count it as inflight given the flush sequence is suspended,
1521 * and the original flush data request is invisible to driver, just
1522 * like other pending requests because of quiesce
1523 */
1524 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
1525 blk_is_flush_data_rq(rq) &&
1526 blk_mq_request_completed(rq))) {
1527 bool *busy = priv;
1528
1529 *busy = true;
1530 return false;
1531 }
1532
1533 return true;
1534}
1535
1536bool blk_mq_queue_inflight(struct request_queue *q)
1537{
1538 bool busy = false;
1539
1540 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1541 return busy;
1542}
1543EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1544
1545static void blk_mq_rq_timed_out(struct request *req)
1546{
1547 req->rq_flags |= RQF_TIMED_OUT;
1548 if (req->q->mq_ops->timeout) {
1549 enum blk_eh_timer_return ret;
1550
1551 ret = req->q->mq_ops->timeout(req);
1552 if (ret == BLK_EH_DONE)
1553 return;
1554 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1555 }
1556
1557 blk_add_timer(req);
1558}
1559
1560struct blk_expired_data {
1561 bool has_timedout_rq;
1562 unsigned long next;
1563 unsigned long timeout_start;
1564};
1565
1566static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1567{
1568 unsigned long deadline;
1569
1570 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1571 return false;
1572 if (rq->rq_flags & RQF_TIMED_OUT)
1573 return false;
1574
1575 deadline = READ_ONCE(rq->deadline);
1576 if (time_after_eq(expired->timeout_start, deadline))
1577 return true;
1578
1579 if (expired->next == 0)
1580 expired->next = deadline;
1581 else if (time_after(expired->next, deadline))
1582 expired->next = deadline;
1583 return false;
1584}
1585
1586void blk_mq_put_rq_ref(struct request *rq)
1587{
1588 if (is_flush_rq(rq)) {
1589 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1590 blk_mq_free_request(rq);
1591 } else if (req_ref_put_and_test(rq)) {
1592 __blk_mq_free_request(rq);
1593 }
1594}
1595
1596static bool blk_mq_check_expired(struct request *rq, void *priv)
1597{
1598 struct blk_expired_data *expired = priv;
1599
1600 /*
1601 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1602 * be reallocated underneath the timeout handler's processing, then
1603 * the expire check is reliable. If the request is not expired, then
1604 * it was completed and reallocated as a new request after returning
1605 * from blk_mq_check_expired().
1606 */
1607 if (blk_mq_req_expired(rq, expired)) {
1608 expired->has_timedout_rq = true;
1609 return false;
1610 }
1611 return true;
1612}
1613
1614static bool blk_mq_handle_expired(struct request *rq, void *priv)
1615{
1616 struct blk_expired_data *expired = priv;
1617
1618 if (blk_mq_req_expired(rq, expired))
1619 blk_mq_rq_timed_out(rq);
1620 return true;
1621}
1622
1623static void blk_mq_timeout_work(struct work_struct *work)
1624{
1625 struct request_queue *q =
1626 container_of(work, struct request_queue, timeout_work);
1627 struct blk_expired_data expired = {
1628 .timeout_start = jiffies,
1629 };
1630 struct blk_mq_hw_ctx *hctx;
1631 unsigned long i;
1632
1633 /* A deadlock might occur if a request is stuck requiring a
1634 * timeout at the same time a queue freeze is waiting
1635 * completion, since the timeout code would not be able to
1636 * acquire the queue reference here.
1637 *
1638 * That's why we don't use blk_queue_enter here; instead, we use
1639 * percpu_ref_tryget directly, because we need to be able to
1640 * obtain a reference even in the short window between the queue
1641 * starting to freeze, by dropping the first reference in
1642 * blk_freeze_queue_start, and the moment the last request is
1643 * consumed, marked by the instant q_usage_counter reaches
1644 * zero.
1645 */
1646 if (!percpu_ref_tryget(&q->q_usage_counter))
1647 return;
1648
1649 /* check if there is any timed-out request */
1650 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1651 if (expired.has_timedout_rq) {
1652 /*
1653 * Before walking tags, we must ensure any submit started
1654 * before the current time has finished. Since the submit
1655 * uses srcu or rcu, wait for a synchronization point to
1656 * ensure all running submits have finished
1657 */
1658 blk_mq_wait_quiesce_done(q->tag_set);
1659
1660 expired.next = 0;
1661 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1662 }
1663
1664 if (expired.next != 0) {
1665 mod_timer(&q->timeout, expired.next);
1666 } else {
1667 /*
1668 * Request timeouts are handled as a forward rolling timer. If
1669 * we end up here it means that no requests are pending and
1670 * also that no request has been pending for a while. Mark
1671 * each hctx as idle.
1672 */
1673 queue_for_each_hw_ctx(q, hctx, i) {
1674 /* the hctx may be unmapped, so check it here */
1675 if (blk_mq_hw_queue_mapped(hctx))
1676 blk_mq_tag_idle(hctx);
1677 }
1678 }
1679 blk_queue_exit(q);
1680}
1681
1682struct flush_busy_ctx_data {
1683 struct blk_mq_hw_ctx *hctx;
1684 struct list_head *list;
1685};
1686
1687static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1688{
1689 struct flush_busy_ctx_data *flush_data = data;
1690 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1691 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1692 enum hctx_type type = hctx->type;
1693
1694 spin_lock(&ctx->lock);
1695 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1696 sbitmap_clear_bit(sb, bitnr);
1697 spin_unlock(&ctx->lock);
1698 return true;
1699}
1700
1701/*
1702 * Process software queues that have been marked busy, splicing them
1703 * to the for-dispatch
1704 */
1705void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1706{
1707 struct flush_busy_ctx_data data = {
1708 .hctx = hctx,
1709 .list = list,
1710 };
1711
1712 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1713}
1714EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1715
1716struct dispatch_rq_data {
1717 struct blk_mq_hw_ctx *hctx;
1718 struct request *rq;
1719};
1720
1721static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1722 void *data)
1723{
1724 struct dispatch_rq_data *dispatch_data = data;
1725 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1726 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1727 enum hctx_type type = hctx->type;
1728
1729 spin_lock(&ctx->lock);
1730 if (!list_empty(&ctx->rq_lists[type])) {
1731 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1732 list_del_init(&dispatch_data->rq->queuelist);
1733 if (list_empty(&ctx->rq_lists[type]))
1734 sbitmap_clear_bit(sb, bitnr);
1735 }
1736 spin_unlock(&ctx->lock);
1737
1738 return !dispatch_data->rq;
1739}
1740
1741struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1742 struct blk_mq_ctx *start)
1743{
1744 unsigned off = start ? start->index_hw[hctx->type] : 0;
1745 struct dispatch_rq_data data = {
1746 .hctx = hctx,
1747 .rq = NULL,
1748 };
1749
1750 __sbitmap_for_each_set(&hctx->ctx_map, off,
1751 dispatch_rq_from_ctx, &data);
1752
1753 return data.rq;
1754}
1755
1756bool __blk_mq_alloc_driver_tag(struct request *rq)
1757{
1758 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1759 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1760 int tag;
1761
1762 blk_mq_tag_busy(rq->mq_hctx);
1763
1764 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1765 bt = &rq->mq_hctx->tags->breserved_tags;
1766 tag_offset = 0;
1767 } else {
1768 if (!hctx_may_queue(rq->mq_hctx, bt))
1769 return false;
1770 }
1771
1772 tag = __sbitmap_queue_get(bt);
1773 if (tag == BLK_MQ_NO_TAG)
1774 return false;
1775
1776 rq->tag = tag + tag_offset;
1777 blk_mq_inc_active_requests(rq->mq_hctx);
1778 return true;
1779}
1780
1781static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1782 int flags, void *key)
1783{
1784 struct blk_mq_hw_ctx *hctx;
1785
1786 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1787
1788 spin_lock(&hctx->dispatch_wait_lock);
1789 if (!list_empty(&wait->entry)) {
1790 struct sbitmap_queue *sbq;
1791
1792 list_del_init(&wait->entry);
1793 sbq = &hctx->tags->bitmap_tags;
1794 atomic_dec(&sbq->ws_active);
1795 }
1796 spin_unlock(&hctx->dispatch_wait_lock);
1797
1798 blk_mq_run_hw_queue(hctx, true);
1799 return 1;
1800}
1801
1802/*
1803 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1804 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1805 * restart. For both cases, take care to check the condition again after
1806 * marking us as waiting.
1807 */
1808static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1809 struct request *rq)
1810{
1811 struct sbitmap_queue *sbq;
1812 struct wait_queue_head *wq;
1813 wait_queue_entry_t *wait;
1814 bool ret;
1815
1816 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1817 !(blk_mq_is_shared_tags(hctx->flags))) {
1818 blk_mq_sched_mark_restart_hctx(hctx);
1819
1820 /*
1821 * It's possible that a tag was freed in the window between the
1822 * allocation failure and adding the hardware queue to the wait
1823 * queue.
1824 *
1825 * Don't clear RESTART here, someone else could have set it.
1826 * At most this will cost an extra queue run.
1827 */
1828 return blk_mq_get_driver_tag(rq);
1829 }
1830
1831 wait = &hctx->dispatch_wait;
1832 if (!list_empty_careful(&wait->entry))
1833 return false;
1834
1835 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1836 sbq = &hctx->tags->breserved_tags;
1837 else
1838 sbq = &hctx->tags->bitmap_tags;
1839 wq = &bt_wait_ptr(sbq, hctx)->wait;
1840
1841 spin_lock_irq(&wq->lock);
1842 spin_lock(&hctx->dispatch_wait_lock);
1843 if (!list_empty(&wait->entry)) {
1844 spin_unlock(&hctx->dispatch_wait_lock);
1845 spin_unlock_irq(&wq->lock);
1846 return false;
1847 }
1848
1849 atomic_inc(&sbq->ws_active);
1850 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1851 __add_wait_queue(wq, wait);
1852
1853 /*
1854 * Add one explicit barrier since blk_mq_get_driver_tag() may
1855 * not imply barrier in case of failure.
1856 *
1857 * Order adding us to wait queue and allocating driver tag.
1858 *
1859 * The pair is the one implied in sbitmap_queue_wake_up() which
1860 * orders clearing sbitmap tag bits and waitqueue_active() in
1861 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1862 *
1863 * Otherwise, re-order of adding wait queue and getting driver tag
1864 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1865 * the waitqueue_active() may not observe us in wait queue.
1866 */
1867 smp_mb();
1868
1869 /*
1870 * It's possible that a tag was freed in the window between the
1871 * allocation failure and adding the hardware queue to the wait
1872 * queue.
1873 */
1874 ret = blk_mq_get_driver_tag(rq);
1875 if (!ret) {
1876 spin_unlock(&hctx->dispatch_wait_lock);
1877 spin_unlock_irq(&wq->lock);
1878 return false;
1879 }
1880
1881 /*
1882 * We got a tag, remove ourselves from the wait queue to ensure
1883 * someone else gets the wakeup.
1884 */
1885 list_del_init(&wait->entry);
1886 atomic_dec(&sbq->ws_active);
1887 spin_unlock(&hctx->dispatch_wait_lock);
1888 spin_unlock_irq(&wq->lock);
1889
1890 return true;
1891}
1892
1893#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1894#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1895/*
1896 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1897 * - EWMA is one simple way to compute running average value
1898 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1899 * - take 4 as factor for avoiding to get too small(0) result, and this
1900 * factor doesn't matter because EWMA decreases exponentially
1901 */
1902static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1903{
1904 unsigned int ewma;
1905
1906 ewma = hctx->dispatch_busy;
1907
1908 if (!ewma && !busy)
1909 return;
1910
1911 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1912 if (busy)
1913 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1914 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1915
1916 hctx->dispatch_busy = ewma;
1917}
1918
1919#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1920
1921static void blk_mq_handle_dev_resource(struct request *rq,
1922 struct list_head *list)
1923{
1924 list_add(&rq->queuelist, list);
1925 __blk_mq_requeue_request(rq);
1926}
1927
1928static void blk_mq_handle_zone_resource(struct request *rq,
1929 struct list_head *zone_list)
1930{
1931 /*
1932 * If we end up here it is because we cannot dispatch a request to a
1933 * specific zone due to LLD level zone-write locking or other zone
1934 * related resource not being available. In this case, set the request
1935 * aside in zone_list for retrying it later.
1936 */
1937 list_add(&rq->queuelist, zone_list);
1938 __blk_mq_requeue_request(rq);
1939}
1940
1941enum prep_dispatch {
1942 PREP_DISPATCH_OK,
1943 PREP_DISPATCH_NO_TAG,
1944 PREP_DISPATCH_NO_BUDGET,
1945};
1946
1947static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1948 bool need_budget)
1949{
1950 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1951 int budget_token = -1;
1952
1953 if (need_budget) {
1954 budget_token = blk_mq_get_dispatch_budget(rq->q);
1955 if (budget_token < 0) {
1956 blk_mq_put_driver_tag(rq);
1957 return PREP_DISPATCH_NO_BUDGET;
1958 }
1959 blk_mq_set_rq_budget_token(rq, budget_token);
1960 }
1961
1962 if (!blk_mq_get_driver_tag(rq)) {
1963 /*
1964 * The initial allocation attempt failed, so we need to
1965 * rerun the hardware queue when a tag is freed. The
1966 * waitqueue takes care of that. If the queue is run
1967 * before we add this entry back on the dispatch list,
1968 * we'll re-run it below.
1969 */
1970 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1971 /*
1972 * All budgets not got from this function will be put
1973 * together during handling partial dispatch
1974 */
1975 if (need_budget)
1976 blk_mq_put_dispatch_budget(rq->q, budget_token);
1977 return PREP_DISPATCH_NO_TAG;
1978 }
1979 }
1980
1981 return PREP_DISPATCH_OK;
1982}
1983
1984/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1985static void blk_mq_release_budgets(struct request_queue *q,
1986 struct list_head *list)
1987{
1988 struct request *rq;
1989
1990 list_for_each_entry(rq, list, queuelist) {
1991 int budget_token = blk_mq_get_rq_budget_token(rq);
1992
1993 if (budget_token >= 0)
1994 blk_mq_put_dispatch_budget(q, budget_token);
1995 }
1996}
1997
1998/*
1999 * blk_mq_commit_rqs will notify driver using bd->last that there is no
2000 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
2001 * details)
2002 * Attention, we should explicitly call this in unusual cases:
2003 * 1) did not queue everything initially scheduled to queue
2004 * 2) the last attempt to queue a request failed
2005 */
2006static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2007 bool from_schedule)
2008{
2009 if (hctx->queue->mq_ops->commit_rqs && queued) {
2010 trace_block_unplug(hctx->queue, queued, !from_schedule);
2011 hctx->queue->mq_ops->commit_rqs(hctx);
2012 }
2013}
2014
2015/*
2016 * Returns true if we did some work AND can potentially do more.
2017 */
2018bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2019 unsigned int nr_budgets)
2020{
2021 enum prep_dispatch prep;
2022 struct request_queue *q = hctx->queue;
2023 struct request *rq;
2024 int queued;
2025 blk_status_t ret = BLK_STS_OK;
2026 LIST_HEAD(zone_list);
2027 bool needs_resource = false;
2028
2029 if (list_empty(list))
2030 return false;
2031
2032 /*
2033 * Now process all the entries, sending them to the driver.
2034 */
2035 queued = 0;
2036 do {
2037 struct blk_mq_queue_data bd;
2038
2039 rq = list_first_entry(list, struct request, queuelist);
2040
2041 WARN_ON_ONCE(hctx != rq->mq_hctx);
2042 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2043 if (prep != PREP_DISPATCH_OK)
2044 break;
2045
2046 list_del_init(&rq->queuelist);
2047
2048 bd.rq = rq;
2049 bd.last = list_empty(list);
2050
2051 /*
2052 * once the request is queued to lld, no need to cover the
2053 * budget any more
2054 */
2055 if (nr_budgets)
2056 nr_budgets--;
2057 ret = q->mq_ops->queue_rq(hctx, &bd);
2058 switch (ret) {
2059 case BLK_STS_OK:
2060 queued++;
2061 break;
2062 case BLK_STS_RESOURCE:
2063 needs_resource = true;
2064 fallthrough;
2065 case BLK_STS_DEV_RESOURCE:
2066 blk_mq_handle_dev_resource(rq, list);
2067 goto out;
2068 case BLK_STS_ZONE_RESOURCE:
2069 /*
2070 * Move the request to zone_list and keep going through
2071 * the dispatch list to find more requests the drive can
2072 * accept.
2073 */
2074 blk_mq_handle_zone_resource(rq, &zone_list);
2075 needs_resource = true;
2076 break;
2077 default:
2078 blk_mq_end_request(rq, ret);
2079 }
2080 } while (!list_empty(list));
2081out:
2082 if (!list_empty(&zone_list))
2083 list_splice_tail_init(&zone_list, list);
2084
2085 /* If we didn't flush the entire list, we could have told the driver
2086 * there was more coming, but that turned out to be a lie.
2087 */
2088 if (!list_empty(list) || ret != BLK_STS_OK)
2089 blk_mq_commit_rqs(hctx, queued, false);
2090
2091 /*
2092 * Any items that need requeuing? Stuff them into hctx->dispatch,
2093 * that is where we will continue on next queue run.
2094 */
2095 if (!list_empty(list)) {
2096 bool needs_restart;
2097 /* For non-shared tags, the RESTART check will suffice */
2098 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2099 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2100 blk_mq_is_shared_tags(hctx->flags));
2101
2102 if (nr_budgets)
2103 blk_mq_release_budgets(q, list);
2104
2105 spin_lock(&hctx->lock);
2106 list_splice_tail_init(list, &hctx->dispatch);
2107 spin_unlock(&hctx->lock);
2108
2109 /*
2110 * Order adding requests to hctx->dispatch and checking
2111 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2112 * in blk_mq_sched_restart(). Avoid restart code path to
2113 * miss the new added requests to hctx->dispatch, meantime
2114 * SCHED_RESTART is observed here.
2115 */
2116 smp_mb();
2117
2118 /*
2119 * If SCHED_RESTART was set by the caller of this function and
2120 * it is no longer set that means that it was cleared by another
2121 * thread and hence that a queue rerun is needed.
2122 *
2123 * If 'no_tag' is set, that means that we failed getting
2124 * a driver tag with an I/O scheduler attached. If our dispatch
2125 * waitqueue is no longer active, ensure that we run the queue
2126 * AFTER adding our entries back to the list.
2127 *
2128 * If no I/O scheduler has been configured it is possible that
2129 * the hardware queue got stopped and restarted before requests
2130 * were pushed back onto the dispatch list. Rerun the queue to
2131 * avoid starvation. Notes:
2132 * - blk_mq_run_hw_queue() checks whether or not a queue has
2133 * been stopped before rerunning a queue.
2134 * - Some but not all block drivers stop a queue before
2135 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2136 * and dm-rq.
2137 *
2138 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2139 * bit is set, run queue after a delay to avoid IO stalls
2140 * that could otherwise occur if the queue is idle. We'll do
2141 * similar if we couldn't get budget or couldn't lock a zone
2142 * and SCHED_RESTART is set.
2143 */
2144 needs_restart = blk_mq_sched_needs_restart(hctx);
2145 if (prep == PREP_DISPATCH_NO_BUDGET)
2146 needs_resource = true;
2147 if (!needs_restart ||
2148 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2149 blk_mq_run_hw_queue(hctx, true);
2150 else if (needs_resource)
2151 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2152
2153 blk_mq_update_dispatch_busy(hctx, true);
2154 return false;
2155 }
2156
2157 blk_mq_update_dispatch_busy(hctx, false);
2158 return true;
2159}
2160
2161static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2162{
2163 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2164
2165 if (cpu >= nr_cpu_ids)
2166 cpu = cpumask_first(hctx->cpumask);
2167 return cpu;
2168}
2169
2170/*
2171 * It'd be great if the workqueue API had a way to pass
2172 * in a mask and had some smarts for more clever placement.
2173 * For now we just round-robin here, switching for every
2174 * BLK_MQ_CPU_WORK_BATCH queued items.
2175 */
2176static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2177{
2178 bool tried = false;
2179 int next_cpu = hctx->next_cpu;
2180
2181 if (hctx->queue->nr_hw_queues == 1)
2182 return WORK_CPU_UNBOUND;
2183
2184 if (--hctx->next_cpu_batch <= 0) {
2185select_cpu:
2186 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2187 cpu_online_mask);
2188 if (next_cpu >= nr_cpu_ids)
2189 next_cpu = blk_mq_first_mapped_cpu(hctx);
2190 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2191 }
2192
2193 /*
2194 * Do unbound schedule if we can't find a online CPU for this hctx,
2195 * and it should only happen in the path of handling CPU DEAD.
2196 */
2197 if (!cpu_online(next_cpu)) {
2198 if (!tried) {
2199 tried = true;
2200 goto select_cpu;
2201 }
2202
2203 /*
2204 * Make sure to re-select CPU next time once after CPUs
2205 * in hctx->cpumask become online again.
2206 */
2207 hctx->next_cpu = next_cpu;
2208 hctx->next_cpu_batch = 1;
2209 return WORK_CPU_UNBOUND;
2210 }
2211
2212 hctx->next_cpu = next_cpu;
2213 return next_cpu;
2214}
2215
2216/**
2217 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2218 * @hctx: Pointer to the hardware queue to run.
2219 * @msecs: Milliseconds of delay to wait before running the queue.
2220 *
2221 * Run a hardware queue asynchronously with a delay of @msecs.
2222 */
2223void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2224{
2225 if (unlikely(blk_mq_hctx_stopped(hctx)))
2226 return;
2227 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2228 msecs_to_jiffies(msecs));
2229}
2230EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2231
2232/**
2233 * blk_mq_run_hw_queue - Start to run a hardware queue.
2234 * @hctx: Pointer to the hardware queue to run.
2235 * @async: If we want to run the queue asynchronously.
2236 *
2237 * Check if the request queue is not in a quiesced state and if there are
2238 * pending requests to be sent. If this is true, run the queue to send requests
2239 * to hardware.
2240 */
2241void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2242{
2243 bool need_run;
2244
2245 /*
2246 * We can't run the queue inline with interrupts disabled.
2247 */
2248 WARN_ON_ONCE(!async && in_interrupt());
2249
2250 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2251
2252 /*
2253 * When queue is quiesced, we may be switching io scheduler, or
2254 * updating nr_hw_queues, or other things, and we can't run queue
2255 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2256 *
2257 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2258 * quiesced.
2259 */
2260 __blk_mq_run_dispatch_ops(hctx->queue, false,
2261 need_run = !blk_queue_quiesced(hctx->queue) &&
2262 blk_mq_hctx_has_pending(hctx));
2263
2264 if (!need_run)
2265 return;
2266
2267 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2268 blk_mq_delay_run_hw_queue(hctx, 0);
2269 return;
2270 }
2271
2272 blk_mq_run_dispatch_ops(hctx->queue,
2273 blk_mq_sched_dispatch_requests(hctx));
2274}
2275EXPORT_SYMBOL(blk_mq_run_hw_queue);
2276
2277/*
2278 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2279 * scheduler.
2280 */
2281static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2282{
2283 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2284 /*
2285 * If the IO scheduler does not respect hardware queues when
2286 * dispatching, we just don't bother with multiple HW queues and
2287 * dispatch from hctx for the current CPU since running multiple queues
2288 * just causes lock contention inside the scheduler and pointless cache
2289 * bouncing.
2290 */
2291 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2292
2293 if (!blk_mq_hctx_stopped(hctx))
2294 return hctx;
2295 return NULL;
2296}
2297
2298/**
2299 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2300 * @q: Pointer to the request queue to run.
2301 * @async: If we want to run the queue asynchronously.
2302 */
2303void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2304{
2305 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2306 unsigned long i;
2307
2308 sq_hctx = NULL;
2309 if (blk_queue_sq_sched(q))
2310 sq_hctx = blk_mq_get_sq_hctx(q);
2311 queue_for_each_hw_ctx(q, hctx, i) {
2312 if (blk_mq_hctx_stopped(hctx))
2313 continue;
2314 /*
2315 * Dispatch from this hctx either if there's no hctx preferred
2316 * by IO scheduler or if it has requests that bypass the
2317 * scheduler.
2318 */
2319 if (!sq_hctx || sq_hctx == hctx ||
2320 !list_empty_careful(&hctx->dispatch))
2321 blk_mq_run_hw_queue(hctx, async);
2322 }
2323}
2324EXPORT_SYMBOL(blk_mq_run_hw_queues);
2325
2326/**
2327 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2328 * @q: Pointer to the request queue to run.
2329 * @msecs: Milliseconds of delay to wait before running the queues.
2330 */
2331void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2332{
2333 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2334 unsigned long i;
2335
2336 sq_hctx = NULL;
2337 if (blk_queue_sq_sched(q))
2338 sq_hctx = blk_mq_get_sq_hctx(q);
2339 queue_for_each_hw_ctx(q, hctx, i) {
2340 if (blk_mq_hctx_stopped(hctx))
2341 continue;
2342 /*
2343 * If there is already a run_work pending, leave the
2344 * pending delay untouched. Otherwise, a hctx can stall
2345 * if another hctx is re-delaying the other's work
2346 * before the work executes.
2347 */
2348 if (delayed_work_pending(&hctx->run_work))
2349 continue;
2350 /*
2351 * Dispatch from this hctx either if there's no hctx preferred
2352 * by IO scheduler or if it has requests that bypass the
2353 * scheduler.
2354 */
2355 if (!sq_hctx || sq_hctx == hctx ||
2356 !list_empty_careful(&hctx->dispatch))
2357 blk_mq_delay_run_hw_queue(hctx, msecs);
2358 }
2359}
2360EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2361
2362/*
2363 * This function is often used for pausing .queue_rq() by driver when
2364 * there isn't enough resource or some conditions aren't satisfied, and
2365 * BLK_STS_RESOURCE is usually returned.
2366 *
2367 * We do not guarantee that dispatch can be drained or blocked
2368 * after blk_mq_stop_hw_queue() returns. Please use
2369 * blk_mq_quiesce_queue() for that requirement.
2370 */
2371void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2372{
2373 cancel_delayed_work(&hctx->run_work);
2374
2375 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2376}
2377EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2378
2379/*
2380 * This function is often used for pausing .queue_rq() by driver when
2381 * there isn't enough resource or some conditions aren't satisfied, and
2382 * BLK_STS_RESOURCE is usually returned.
2383 *
2384 * We do not guarantee that dispatch can be drained or blocked
2385 * after blk_mq_stop_hw_queues() returns. Please use
2386 * blk_mq_quiesce_queue() for that requirement.
2387 */
2388void blk_mq_stop_hw_queues(struct request_queue *q)
2389{
2390 struct blk_mq_hw_ctx *hctx;
2391 unsigned long i;
2392
2393 queue_for_each_hw_ctx(q, hctx, i)
2394 blk_mq_stop_hw_queue(hctx);
2395}
2396EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2397
2398void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2399{
2400 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2401
2402 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
2403}
2404EXPORT_SYMBOL(blk_mq_start_hw_queue);
2405
2406void blk_mq_start_hw_queues(struct request_queue *q)
2407{
2408 struct blk_mq_hw_ctx *hctx;
2409 unsigned long i;
2410
2411 queue_for_each_hw_ctx(q, hctx, i)
2412 blk_mq_start_hw_queue(hctx);
2413}
2414EXPORT_SYMBOL(blk_mq_start_hw_queues);
2415
2416void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2417{
2418 if (!blk_mq_hctx_stopped(hctx))
2419 return;
2420
2421 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2422 blk_mq_run_hw_queue(hctx, async);
2423}
2424EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2425
2426void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2427{
2428 struct blk_mq_hw_ctx *hctx;
2429 unsigned long i;
2430
2431 queue_for_each_hw_ctx(q, hctx, i)
2432 blk_mq_start_stopped_hw_queue(hctx, async ||
2433 (hctx->flags & BLK_MQ_F_BLOCKING));
2434}
2435EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2436
2437static void blk_mq_run_work_fn(struct work_struct *work)
2438{
2439 struct blk_mq_hw_ctx *hctx =
2440 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2441
2442 blk_mq_run_dispatch_ops(hctx->queue,
2443 blk_mq_sched_dispatch_requests(hctx));
2444}
2445
2446/**
2447 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2448 * @rq: Pointer to request to be inserted.
2449 * @flags: BLK_MQ_INSERT_*
2450 *
2451 * Should only be used carefully, when the caller knows we want to
2452 * bypass a potential IO scheduler on the target device.
2453 */
2454static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2455{
2456 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2457
2458 spin_lock(&hctx->lock);
2459 if (flags & BLK_MQ_INSERT_AT_HEAD)
2460 list_add(&rq->queuelist, &hctx->dispatch);
2461 else
2462 list_add_tail(&rq->queuelist, &hctx->dispatch);
2463 spin_unlock(&hctx->lock);
2464}
2465
2466static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2467 struct blk_mq_ctx *ctx, struct list_head *list,
2468 bool run_queue_async)
2469{
2470 struct request *rq;
2471 enum hctx_type type = hctx->type;
2472
2473 /*
2474 * Try to issue requests directly if the hw queue isn't busy to save an
2475 * extra enqueue & dequeue to the sw queue.
2476 */
2477 if (!hctx->dispatch_busy && !run_queue_async) {
2478 blk_mq_run_dispatch_ops(hctx->queue,
2479 blk_mq_try_issue_list_directly(hctx, list));
2480 if (list_empty(list))
2481 goto out;
2482 }
2483
2484 /*
2485 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2486 * offline now
2487 */
2488 list_for_each_entry(rq, list, queuelist) {
2489 BUG_ON(rq->mq_ctx != ctx);
2490 trace_block_rq_insert(rq);
2491 if (rq->cmd_flags & REQ_NOWAIT)
2492 run_queue_async = true;
2493 }
2494
2495 spin_lock(&ctx->lock);
2496 list_splice_tail_init(list, &ctx->rq_lists[type]);
2497 blk_mq_hctx_mark_pending(hctx, ctx);
2498 spin_unlock(&ctx->lock);
2499out:
2500 blk_mq_run_hw_queue(hctx, run_queue_async);
2501}
2502
2503static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2504{
2505 struct request_queue *q = rq->q;
2506 struct blk_mq_ctx *ctx = rq->mq_ctx;
2507 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2508
2509 if (blk_rq_is_passthrough(rq)) {
2510 /*
2511 * Passthrough request have to be added to hctx->dispatch
2512 * directly. The device may be in a situation where it can't
2513 * handle FS request, and always returns BLK_STS_RESOURCE for
2514 * them, which gets them added to hctx->dispatch.
2515 *
2516 * If a passthrough request is required to unblock the queues,
2517 * and it is added to the scheduler queue, there is no chance to
2518 * dispatch it given we prioritize requests in hctx->dispatch.
2519 */
2520 blk_mq_request_bypass_insert(rq, flags);
2521 } else if (req_op(rq) == REQ_OP_FLUSH) {
2522 /*
2523 * Firstly normal IO request is inserted to scheduler queue or
2524 * sw queue, meantime we add flush request to dispatch queue(
2525 * hctx->dispatch) directly and there is at most one in-flight
2526 * flush request for each hw queue, so it doesn't matter to add
2527 * flush request to tail or front of the dispatch queue.
2528 *
2529 * Secondly in case of NCQ, flush request belongs to non-NCQ
2530 * command, and queueing it will fail when there is any
2531 * in-flight normal IO request(NCQ command). When adding flush
2532 * rq to the front of hctx->dispatch, it is easier to introduce
2533 * extra time to flush rq's latency because of S_SCHED_RESTART
2534 * compared with adding to the tail of dispatch queue, then
2535 * chance of flush merge is increased, and less flush requests
2536 * will be issued to controller. It is observed that ~10% time
2537 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2538 * drive when adding flush rq to the front of hctx->dispatch.
2539 *
2540 * Simply queue flush rq to the front of hctx->dispatch so that
2541 * intensive flush workloads can benefit in case of NCQ HW.
2542 */
2543 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2544 } else if (q->elevator) {
2545 LIST_HEAD(list);
2546
2547 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2548
2549 list_add(&rq->queuelist, &list);
2550 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2551 } else {
2552 trace_block_rq_insert(rq);
2553
2554 spin_lock(&ctx->lock);
2555 if (flags & BLK_MQ_INSERT_AT_HEAD)
2556 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2557 else
2558 list_add_tail(&rq->queuelist,
2559 &ctx->rq_lists[hctx->type]);
2560 blk_mq_hctx_mark_pending(hctx, ctx);
2561 spin_unlock(&ctx->lock);
2562 }
2563}
2564
2565static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2566 unsigned int nr_segs)
2567{
2568 int err;
2569
2570 if (bio->bi_opf & REQ_RAHEAD)
2571 rq->cmd_flags |= REQ_FAILFAST_MASK;
2572
2573 rq->__sector = bio->bi_iter.bi_sector;
2574 rq->write_hint = bio->bi_write_hint;
2575 blk_rq_bio_prep(rq, bio, nr_segs);
2576
2577 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2578 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2579 WARN_ON_ONCE(err);
2580
2581 blk_account_io_start(rq);
2582}
2583
2584static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2585 struct request *rq, bool last)
2586{
2587 struct request_queue *q = rq->q;
2588 struct blk_mq_queue_data bd = {
2589 .rq = rq,
2590 .last = last,
2591 };
2592 blk_status_t ret;
2593
2594 /*
2595 * For OK queue, we are done. For error, caller may kill it.
2596 * Any other error (busy), just add it to our list as we
2597 * previously would have done.
2598 */
2599 ret = q->mq_ops->queue_rq(hctx, &bd);
2600 switch (ret) {
2601 case BLK_STS_OK:
2602 blk_mq_update_dispatch_busy(hctx, false);
2603 break;
2604 case BLK_STS_RESOURCE:
2605 case BLK_STS_DEV_RESOURCE:
2606 blk_mq_update_dispatch_busy(hctx, true);
2607 __blk_mq_requeue_request(rq);
2608 break;
2609 default:
2610 blk_mq_update_dispatch_busy(hctx, false);
2611 break;
2612 }
2613
2614 return ret;
2615}
2616
2617static bool blk_mq_get_budget_and_tag(struct request *rq)
2618{
2619 int budget_token;
2620
2621 budget_token = blk_mq_get_dispatch_budget(rq->q);
2622 if (budget_token < 0)
2623 return false;
2624 blk_mq_set_rq_budget_token(rq, budget_token);
2625 if (!blk_mq_get_driver_tag(rq)) {
2626 blk_mq_put_dispatch_budget(rq->q, budget_token);
2627 return false;
2628 }
2629 return true;
2630}
2631
2632/**
2633 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2634 * @hctx: Pointer of the associated hardware queue.
2635 * @rq: Pointer to request to be sent.
2636 *
2637 * If the device has enough resources to accept a new request now, send the
2638 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2639 * we can try send it another time in the future. Requests inserted at this
2640 * queue have higher priority.
2641 */
2642static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2643 struct request *rq)
2644{
2645 blk_status_t ret;
2646
2647 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2648 blk_mq_insert_request(rq, 0);
2649 return;
2650 }
2651
2652 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2653 blk_mq_insert_request(rq, 0);
2654 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
2655 return;
2656 }
2657
2658 ret = __blk_mq_issue_directly(hctx, rq, true);
2659 switch (ret) {
2660 case BLK_STS_OK:
2661 break;
2662 case BLK_STS_RESOURCE:
2663 case BLK_STS_DEV_RESOURCE:
2664 blk_mq_request_bypass_insert(rq, 0);
2665 blk_mq_run_hw_queue(hctx, false);
2666 break;
2667 default:
2668 blk_mq_end_request(rq, ret);
2669 break;
2670 }
2671}
2672
2673static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2674{
2675 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2676
2677 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2678 blk_mq_insert_request(rq, 0);
2679 return BLK_STS_OK;
2680 }
2681
2682 if (!blk_mq_get_budget_and_tag(rq))
2683 return BLK_STS_RESOURCE;
2684 return __blk_mq_issue_directly(hctx, rq, last);
2685}
2686
2687static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2688{
2689 struct blk_mq_hw_ctx *hctx = NULL;
2690 struct request *rq;
2691 int queued = 0;
2692 blk_status_t ret = BLK_STS_OK;
2693
2694 while ((rq = rq_list_pop(&plug->mq_list))) {
2695 bool last = rq_list_empty(plug->mq_list);
2696
2697 if (hctx != rq->mq_hctx) {
2698 if (hctx) {
2699 blk_mq_commit_rqs(hctx, queued, false);
2700 queued = 0;
2701 }
2702 hctx = rq->mq_hctx;
2703 }
2704
2705 ret = blk_mq_request_issue_directly(rq, last);
2706 switch (ret) {
2707 case BLK_STS_OK:
2708 queued++;
2709 break;
2710 case BLK_STS_RESOURCE:
2711 case BLK_STS_DEV_RESOURCE:
2712 blk_mq_request_bypass_insert(rq, 0);
2713 blk_mq_run_hw_queue(hctx, false);
2714 goto out;
2715 default:
2716 blk_mq_end_request(rq, ret);
2717 break;
2718 }
2719 }
2720
2721out:
2722 if (ret != BLK_STS_OK)
2723 blk_mq_commit_rqs(hctx, queued, false);
2724}
2725
2726static void __blk_mq_flush_plug_list(struct request_queue *q,
2727 struct blk_plug *plug)
2728{
2729 if (blk_queue_quiesced(q))
2730 return;
2731 q->mq_ops->queue_rqs(&plug->mq_list);
2732}
2733
2734static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2735{
2736 struct blk_mq_hw_ctx *this_hctx = NULL;
2737 struct blk_mq_ctx *this_ctx = NULL;
2738 struct request *requeue_list = NULL;
2739 struct request **requeue_lastp = &requeue_list;
2740 unsigned int depth = 0;
2741 bool is_passthrough = false;
2742 LIST_HEAD(list);
2743
2744 do {
2745 struct request *rq = rq_list_pop(&plug->mq_list);
2746
2747 if (!this_hctx) {
2748 this_hctx = rq->mq_hctx;
2749 this_ctx = rq->mq_ctx;
2750 is_passthrough = blk_rq_is_passthrough(rq);
2751 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2752 is_passthrough != blk_rq_is_passthrough(rq)) {
2753 rq_list_add_tail(&requeue_lastp, rq);
2754 continue;
2755 }
2756 list_add(&rq->queuelist, &list);
2757 depth++;
2758 } while (!rq_list_empty(plug->mq_list));
2759
2760 plug->mq_list = requeue_list;
2761 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2762
2763 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2764 /* passthrough requests should never be issued to the I/O scheduler */
2765 if (is_passthrough) {
2766 spin_lock(&this_hctx->lock);
2767 list_splice_tail_init(&list, &this_hctx->dispatch);
2768 spin_unlock(&this_hctx->lock);
2769 blk_mq_run_hw_queue(this_hctx, from_sched);
2770 } else if (this_hctx->queue->elevator) {
2771 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2772 &list, 0);
2773 blk_mq_run_hw_queue(this_hctx, from_sched);
2774 } else {
2775 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2776 }
2777 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2778}
2779
2780void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2781{
2782 struct request *rq;
2783
2784 /*
2785 * We may have been called recursively midway through handling
2786 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2787 * To avoid mq_list changing under our feet, clear rq_count early and
2788 * bail out specifically if rq_count is 0 rather than checking
2789 * whether the mq_list is empty.
2790 */
2791 if (plug->rq_count == 0)
2792 return;
2793 plug->rq_count = 0;
2794
2795 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2796 struct request_queue *q;
2797
2798 rq = rq_list_peek(&plug->mq_list);
2799 q = rq->q;
2800
2801 /*
2802 * Peek first request and see if we have a ->queue_rqs() hook.
2803 * If we do, we can dispatch the whole plug list in one go. We
2804 * already know at this point that all requests belong to the
2805 * same queue, caller must ensure that's the case.
2806 */
2807 if (q->mq_ops->queue_rqs) {
2808 blk_mq_run_dispatch_ops(q,
2809 __blk_mq_flush_plug_list(q, plug));
2810 if (rq_list_empty(plug->mq_list))
2811 return;
2812 }
2813
2814 blk_mq_run_dispatch_ops(q,
2815 blk_mq_plug_issue_direct(plug));
2816 if (rq_list_empty(plug->mq_list))
2817 return;
2818 }
2819
2820 do {
2821 blk_mq_dispatch_plug_list(plug, from_schedule);
2822 } while (!rq_list_empty(plug->mq_list));
2823}
2824
2825static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2826 struct list_head *list)
2827{
2828 int queued = 0;
2829 blk_status_t ret = BLK_STS_OK;
2830
2831 while (!list_empty(list)) {
2832 struct request *rq = list_first_entry(list, struct request,
2833 queuelist);
2834
2835 list_del_init(&rq->queuelist);
2836 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2837 switch (ret) {
2838 case BLK_STS_OK:
2839 queued++;
2840 break;
2841 case BLK_STS_RESOURCE:
2842 case BLK_STS_DEV_RESOURCE:
2843 blk_mq_request_bypass_insert(rq, 0);
2844 if (list_empty(list))
2845 blk_mq_run_hw_queue(hctx, false);
2846 goto out;
2847 default:
2848 blk_mq_end_request(rq, ret);
2849 break;
2850 }
2851 }
2852
2853out:
2854 if (ret != BLK_STS_OK)
2855 blk_mq_commit_rqs(hctx, queued, false);
2856}
2857
2858static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2859 struct bio *bio, unsigned int nr_segs)
2860{
2861 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2862 if (blk_attempt_plug_merge(q, bio, nr_segs))
2863 return true;
2864 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2865 return true;
2866 }
2867 return false;
2868}
2869
2870static struct request *blk_mq_get_new_requests(struct request_queue *q,
2871 struct blk_plug *plug,
2872 struct bio *bio,
2873 unsigned int nsegs)
2874{
2875 struct blk_mq_alloc_data data = {
2876 .q = q,
2877 .nr_tags = 1,
2878 .cmd_flags = bio->bi_opf,
2879 };
2880 struct request *rq;
2881
2882 rq_qos_throttle(q, bio);
2883
2884 if (plug) {
2885 data.nr_tags = plug->nr_ios;
2886 plug->nr_ios = 1;
2887 data.cached_rq = &plug->cached_rq;
2888 }
2889
2890 rq = __blk_mq_alloc_requests(&data);
2891 if (rq)
2892 return rq;
2893 rq_qos_cleanup(q, bio);
2894 if (bio->bi_opf & REQ_NOWAIT)
2895 bio_wouldblock_error(bio);
2896 return NULL;
2897}
2898
2899/*
2900 * Check if there is a suitable cached request and return it.
2901 */
2902static struct request *blk_mq_peek_cached_request(struct blk_plug *plug,
2903 struct request_queue *q, blk_opf_t opf)
2904{
2905 enum hctx_type type = blk_mq_get_hctx_type(opf);
2906 struct request *rq;
2907
2908 if (!plug)
2909 return NULL;
2910 rq = rq_list_peek(&plug->cached_rq);
2911 if (!rq || rq->q != q)
2912 return NULL;
2913 if (type != rq->mq_hctx->type &&
2914 (type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT))
2915 return NULL;
2916 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
2917 return NULL;
2918 return rq;
2919}
2920
2921static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug,
2922 struct bio *bio)
2923{
2924 WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq);
2925
2926 /*
2927 * If any qos ->throttle() end up blocking, we will have flushed the
2928 * plug and hence killed the cached_rq list as well. Pop this entry
2929 * before we throttle.
2930 */
2931 plug->cached_rq = rq_list_next(rq);
2932 rq_qos_throttle(rq->q, bio);
2933
2934 blk_mq_rq_time_init(rq, 0);
2935 rq->cmd_flags = bio->bi_opf;
2936 INIT_LIST_HEAD(&rq->queuelist);
2937}
2938
2939/**
2940 * blk_mq_submit_bio - Create and send a request to block device.
2941 * @bio: Bio pointer.
2942 *
2943 * Builds up a request structure from @q and @bio and send to the device. The
2944 * request may not be queued directly to hardware if:
2945 * * This request can be merged with another one
2946 * * We want to place request at plug queue for possible future merging
2947 * * There is an IO scheduler active at this queue
2948 *
2949 * It will not queue the request if there is an error with the bio, or at the
2950 * request creation.
2951 */
2952void blk_mq_submit_bio(struct bio *bio)
2953{
2954 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2955 struct blk_plug *plug = blk_mq_plug(bio);
2956 const int is_sync = op_is_sync(bio->bi_opf);
2957 struct blk_mq_hw_ctx *hctx;
2958 unsigned int nr_segs = 1;
2959 struct request *rq;
2960 blk_status_t ret;
2961
2962 bio = blk_queue_bounce(bio, q);
2963
2964 /*
2965 * If the plug has a cached request for this queue, try use it.
2966 *
2967 * The cached request already holds a q_usage_counter reference and we
2968 * don't have to acquire a new one if we use it.
2969 */
2970 rq = blk_mq_peek_cached_request(plug, q, bio->bi_opf);
2971 if (!rq) {
2972 if (unlikely(bio_queue_enter(bio)))
2973 return;
2974 }
2975
2976 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2977 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2978 if (!bio)
2979 goto queue_exit;
2980 }
2981 if (!bio_integrity_prep(bio))
2982 goto queue_exit;
2983
2984 if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
2985 goto queue_exit;
2986
2987 if (!rq) {
2988 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2989 if (unlikely(!rq))
2990 goto queue_exit;
2991 } else {
2992 blk_mq_use_cached_rq(rq, plug, bio);
2993 }
2994
2995 trace_block_getrq(bio);
2996
2997 rq_qos_track(q, rq, bio);
2998
2999 blk_mq_bio_to_request(rq, bio, nr_segs);
3000
3001 ret = blk_crypto_rq_get_keyslot(rq);
3002 if (ret != BLK_STS_OK) {
3003 bio->bi_status = ret;
3004 bio_endio(bio);
3005 blk_mq_free_request(rq);
3006 return;
3007 }
3008
3009 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
3010 return;
3011
3012 if (plug) {
3013 blk_add_rq_to_plug(plug, rq);
3014 return;
3015 }
3016
3017 hctx = rq->mq_hctx;
3018 if ((rq->rq_flags & RQF_USE_SCHED) ||
3019 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3020 blk_mq_insert_request(rq, 0);
3021 blk_mq_run_hw_queue(hctx, true);
3022 } else {
3023 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3024 }
3025 return;
3026
3027queue_exit:
3028 /*
3029 * Don't drop the queue reference if we were trying to use a cached
3030 * request and thus didn't acquire one.
3031 */
3032 if (!rq)
3033 blk_queue_exit(q);
3034}
3035
3036#ifdef CONFIG_BLK_MQ_STACKING
3037/**
3038 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3039 * @rq: the request being queued
3040 */
3041blk_status_t blk_insert_cloned_request(struct request *rq)
3042{
3043 struct request_queue *q = rq->q;
3044 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3045 unsigned int max_segments = blk_rq_get_max_segments(rq);
3046 blk_status_t ret;
3047
3048 if (blk_rq_sectors(rq) > max_sectors) {
3049 /*
3050 * SCSI device does not have a good way to return if
3051 * Write Same/Zero is actually supported. If a device rejects
3052 * a non-read/write command (discard, write same,etc.) the
3053 * low-level device driver will set the relevant queue limit to
3054 * 0 to prevent blk-lib from issuing more of the offending
3055 * operations. Commands queued prior to the queue limit being
3056 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3057 * errors being propagated to upper layers.
3058 */
3059 if (max_sectors == 0)
3060 return BLK_STS_NOTSUPP;
3061
3062 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3063 __func__, blk_rq_sectors(rq), max_sectors);
3064 return BLK_STS_IOERR;
3065 }
3066
3067 /*
3068 * The queue settings related to segment counting may differ from the
3069 * original queue.
3070 */
3071 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3072 if (rq->nr_phys_segments > max_segments) {
3073 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3074 __func__, rq->nr_phys_segments, max_segments);
3075 return BLK_STS_IOERR;
3076 }
3077
3078 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3079 return BLK_STS_IOERR;
3080
3081 ret = blk_crypto_rq_get_keyslot(rq);
3082 if (ret != BLK_STS_OK)
3083 return ret;
3084
3085 blk_account_io_start(rq);
3086
3087 /*
3088 * Since we have a scheduler attached on the top device,
3089 * bypass a potential scheduler on the bottom device for
3090 * insert.
3091 */
3092 blk_mq_run_dispatch_ops(q,
3093 ret = blk_mq_request_issue_directly(rq, true));
3094 if (ret)
3095 blk_account_io_done(rq, blk_time_get_ns());
3096 return ret;
3097}
3098EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3099
3100/**
3101 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3102 * @rq: the clone request to be cleaned up
3103 *
3104 * Description:
3105 * Free all bios in @rq for a cloned request.
3106 */
3107void blk_rq_unprep_clone(struct request *rq)
3108{
3109 struct bio *bio;
3110
3111 while ((bio = rq->bio) != NULL) {
3112 rq->bio = bio->bi_next;
3113
3114 bio_put(bio);
3115 }
3116}
3117EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3118
3119/**
3120 * blk_rq_prep_clone - Helper function to setup clone request
3121 * @rq: the request to be setup
3122 * @rq_src: original request to be cloned
3123 * @bs: bio_set that bios for clone are allocated from
3124 * @gfp_mask: memory allocation mask for bio
3125 * @bio_ctr: setup function to be called for each clone bio.
3126 * Returns %0 for success, non %0 for failure.
3127 * @data: private data to be passed to @bio_ctr
3128 *
3129 * Description:
3130 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3131 * Also, pages which the original bios are pointing to are not copied
3132 * and the cloned bios just point same pages.
3133 * So cloned bios must be completed before original bios, which means
3134 * the caller must complete @rq before @rq_src.
3135 */
3136int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3137 struct bio_set *bs, gfp_t gfp_mask,
3138 int (*bio_ctr)(struct bio *, struct bio *, void *),
3139 void *data)
3140{
3141 struct bio *bio, *bio_src;
3142
3143 if (!bs)
3144 bs = &fs_bio_set;
3145
3146 __rq_for_each_bio(bio_src, rq_src) {
3147 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3148 bs);
3149 if (!bio)
3150 goto free_and_out;
3151
3152 if (bio_ctr && bio_ctr(bio, bio_src, data))
3153 goto free_and_out;
3154
3155 if (rq->bio) {
3156 rq->biotail->bi_next = bio;
3157 rq->biotail = bio;
3158 } else {
3159 rq->bio = rq->biotail = bio;
3160 }
3161 bio = NULL;
3162 }
3163
3164 /* Copy attributes of the original request to the clone request. */
3165 rq->__sector = blk_rq_pos(rq_src);
3166 rq->__data_len = blk_rq_bytes(rq_src);
3167 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3168 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3169 rq->special_vec = rq_src->special_vec;
3170 }
3171 rq->nr_phys_segments = rq_src->nr_phys_segments;
3172 rq->ioprio = rq_src->ioprio;
3173 rq->write_hint = rq_src->write_hint;
3174
3175 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3176 goto free_and_out;
3177
3178 return 0;
3179
3180free_and_out:
3181 if (bio)
3182 bio_put(bio);
3183 blk_rq_unprep_clone(rq);
3184
3185 return -ENOMEM;
3186}
3187EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3188#endif /* CONFIG_BLK_MQ_STACKING */
3189
3190/*
3191 * Steal bios from a request and add them to a bio list.
3192 * The request must not have been partially completed before.
3193 */
3194void blk_steal_bios(struct bio_list *list, struct request *rq)
3195{
3196 if (rq->bio) {
3197 if (list->tail)
3198 list->tail->bi_next = rq->bio;
3199 else
3200 list->head = rq->bio;
3201 list->tail = rq->biotail;
3202
3203 rq->bio = NULL;
3204 rq->biotail = NULL;
3205 }
3206
3207 rq->__data_len = 0;
3208}
3209EXPORT_SYMBOL_GPL(blk_steal_bios);
3210
3211static size_t order_to_size(unsigned int order)
3212{
3213 return (size_t)PAGE_SIZE << order;
3214}
3215
3216/* called before freeing request pool in @tags */
3217static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3218 struct blk_mq_tags *tags)
3219{
3220 struct page *page;
3221 unsigned long flags;
3222
3223 /*
3224 * There is no need to clear mapping if driver tags is not initialized
3225 * or the mapping belongs to the driver tags.
3226 */
3227 if (!drv_tags || drv_tags == tags)
3228 return;
3229
3230 list_for_each_entry(page, &tags->page_list, lru) {
3231 unsigned long start = (unsigned long)page_address(page);
3232 unsigned long end = start + order_to_size(page->private);
3233 int i;
3234
3235 for (i = 0; i < drv_tags->nr_tags; i++) {
3236 struct request *rq = drv_tags->rqs[i];
3237 unsigned long rq_addr = (unsigned long)rq;
3238
3239 if (rq_addr >= start && rq_addr < end) {
3240 WARN_ON_ONCE(req_ref_read(rq) != 0);
3241 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3242 }
3243 }
3244 }
3245
3246 /*
3247 * Wait until all pending iteration is done.
3248 *
3249 * Request reference is cleared and it is guaranteed to be observed
3250 * after the ->lock is released.
3251 */
3252 spin_lock_irqsave(&drv_tags->lock, flags);
3253 spin_unlock_irqrestore(&drv_tags->lock, flags);
3254}
3255
3256void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3257 unsigned int hctx_idx)
3258{
3259 struct blk_mq_tags *drv_tags;
3260 struct page *page;
3261
3262 if (list_empty(&tags->page_list))
3263 return;
3264
3265 if (blk_mq_is_shared_tags(set->flags))
3266 drv_tags = set->shared_tags;
3267 else
3268 drv_tags = set->tags[hctx_idx];
3269
3270 if (tags->static_rqs && set->ops->exit_request) {
3271 int i;
3272
3273 for (i = 0; i < tags->nr_tags; i++) {
3274 struct request *rq = tags->static_rqs[i];
3275
3276 if (!rq)
3277 continue;
3278 set->ops->exit_request(set, rq, hctx_idx);
3279 tags->static_rqs[i] = NULL;
3280 }
3281 }
3282
3283 blk_mq_clear_rq_mapping(drv_tags, tags);
3284
3285 while (!list_empty(&tags->page_list)) {
3286 page = list_first_entry(&tags->page_list, struct page, lru);
3287 list_del_init(&page->lru);
3288 /*
3289 * Remove kmemleak object previously allocated in
3290 * blk_mq_alloc_rqs().
3291 */
3292 kmemleak_free(page_address(page));
3293 __free_pages(page, page->private);
3294 }
3295}
3296
3297void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3298{
3299 kfree(tags->rqs);
3300 tags->rqs = NULL;
3301 kfree(tags->static_rqs);
3302 tags->static_rqs = NULL;
3303
3304 blk_mq_free_tags(tags);
3305}
3306
3307static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3308 unsigned int hctx_idx)
3309{
3310 int i;
3311
3312 for (i = 0; i < set->nr_maps; i++) {
3313 unsigned int start = set->map[i].queue_offset;
3314 unsigned int end = start + set->map[i].nr_queues;
3315
3316 if (hctx_idx >= start && hctx_idx < end)
3317 break;
3318 }
3319
3320 if (i >= set->nr_maps)
3321 i = HCTX_TYPE_DEFAULT;
3322
3323 return i;
3324}
3325
3326static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3327 unsigned int hctx_idx)
3328{
3329 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3330
3331 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3332}
3333
3334static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3335 unsigned int hctx_idx,
3336 unsigned int nr_tags,
3337 unsigned int reserved_tags)
3338{
3339 int node = blk_mq_get_hctx_node(set, hctx_idx);
3340 struct blk_mq_tags *tags;
3341
3342 if (node == NUMA_NO_NODE)
3343 node = set->numa_node;
3344
3345 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3346 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3347 if (!tags)
3348 return NULL;
3349
3350 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3351 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3352 node);
3353 if (!tags->rqs)
3354 goto err_free_tags;
3355
3356 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3357 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3358 node);
3359 if (!tags->static_rqs)
3360 goto err_free_rqs;
3361
3362 return tags;
3363
3364err_free_rqs:
3365 kfree(tags->rqs);
3366err_free_tags:
3367 blk_mq_free_tags(tags);
3368 return NULL;
3369}
3370
3371static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3372 unsigned int hctx_idx, int node)
3373{
3374 int ret;
3375
3376 if (set->ops->init_request) {
3377 ret = set->ops->init_request(set, rq, hctx_idx, node);
3378 if (ret)
3379 return ret;
3380 }
3381
3382 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3383 return 0;
3384}
3385
3386static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3387 struct blk_mq_tags *tags,
3388 unsigned int hctx_idx, unsigned int depth)
3389{
3390 unsigned int i, j, entries_per_page, max_order = 4;
3391 int node = blk_mq_get_hctx_node(set, hctx_idx);
3392 size_t rq_size, left;
3393
3394 if (node == NUMA_NO_NODE)
3395 node = set->numa_node;
3396
3397 INIT_LIST_HEAD(&tags->page_list);
3398
3399 /*
3400 * rq_size is the size of the request plus driver payload, rounded
3401 * to the cacheline size
3402 */
3403 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3404 cache_line_size());
3405 left = rq_size * depth;
3406
3407 for (i = 0; i < depth; ) {
3408 int this_order = max_order;
3409 struct page *page;
3410 int to_do;
3411 void *p;
3412
3413 while (this_order && left < order_to_size(this_order - 1))
3414 this_order--;
3415
3416 do {
3417 page = alloc_pages_node(node,
3418 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3419 this_order);
3420 if (page)
3421 break;
3422 if (!this_order--)
3423 break;
3424 if (order_to_size(this_order) < rq_size)
3425 break;
3426 } while (1);
3427
3428 if (!page)
3429 goto fail;
3430
3431 page->private = this_order;
3432 list_add_tail(&page->lru, &tags->page_list);
3433
3434 p = page_address(page);
3435 /*
3436 * Allow kmemleak to scan these pages as they contain pointers
3437 * to additional allocations like via ops->init_request().
3438 */
3439 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3440 entries_per_page = order_to_size(this_order) / rq_size;
3441 to_do = min(entries_per_page, depth - i);
3442 left -= to_do * rq_size;
3443 for (j = 0; j < to_do; j++) {
3444 struct request *rq = p;
3445
3446 tags->static_rqs[i] = rq;
3447 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3448 tags->static_rqs[i] = NULL;
3449 goto fail;
3450 }
3451
3452 p += rq_size;
3453 i++;
3454 }
3455 }
3456 return 0;
3457
3458fail:
3459 blk_mq_free_rqs(set, tags, hctx_idx);
3460 return -ENOMEM;
3461}
3462
3463struct rq_iter_data {
3464 struct blk_mq_hw_ctx *hctx;
3465 bool has_rq;
3466};
3467
3468static bool blk_mq_has_request(struct request *rq, void *data)
3469{
3470 struct rq_iter_data *iter_data = data;
3471
3472 if (rq->mq_hctx != iter_data->hctx)
3473 return true;
3474 iter_data->has_rq = true;
3475 return false;
3476}
3477
3478static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3479{
3480 struct blk_mq_tags *tags = hctx->sched_tags ?
3481 hctx->sched_tags : hctx->tags;
3482 struct rq_iter_data data = {
3483 .hctx = hctx,
3484 };
3485
3486 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3487 return data.has_rq;
3488}
3489
3490static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3491 struct blk_mq_hw_ctx *hctx)
3492{
3493 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3494 return false;
3495 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3496 return false;
3497 return true;
3498}
3499
3500static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3501{
3502 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3503 struct blk_mq_hw_ctx, cpuhp_online);
3504
3505 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3506 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3507 return 0;
3508
3509 /*
3510 * Prevent new request from being allocated on the current hctx.
3511 *
3512 * The smp_mb__after_atomic() Pairs with the implied barrier in
3513 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3514 * seen once we return from the tag allocator.
3515 */
3516 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3517 smp_mb__after_atomic();
3518
3519 /*
3520 * Try to grab a reference to the queue and wait for any outstanding
3521 * requests. If we could not grab a reference the queue has been
3522 * frozen and there are no requests.
3523 */
3524 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3525 while (blk_mq_hctx_has_requests(hctx))
3526 msleep(5);
3527 percpu_ref_put(&hctx->queue->q_usage_counter);
3528 }
3529
3530 return 0;
3531}
3532
3533static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3534{
3535 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3536 struct blk_mq_hw_ctx, cpuhp_online);
3537
3538 if (cpumask_test_cpu(cpu, hctx->cpumask))
3539 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3540 return 0;
3541}
3542
3543/*
3544 * 'cpu' is going away. splice any existing rq_list entries from this
3545 * software queue to the hw queue dispatch list, and ensure that it
3546 * gets run.
3547 */
3548static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3549{
3550 struct blk_mq_hw_ctx *hctx;
3551 struct blk_mq_ctx *ctx;
3552 LIST_HEAD(tmp);
3553 enum hctx_type type;
3554
3555 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3556 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3557 return 0;
3558
3559 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3560 type = hctx->type;
3561
3562 spin_lock(&ctx->lock);
3563 if (!list_empty(&ctx->rq_lists[type])) {
3564 list_splice_init(&ctx->rq_lists[type], &tmp);
3565 blk_mq_hctx_clear_pending(hctx, ctx);
3566 }
3567 spin_unlock(&ctx->lock);
3568
3569 if (list_empty(&tmp))
3570 return 0;
3571
3572 spin_lock(&hctx->lock);
3573 list_splice_tail_init(&tmp, &hctx->dispatch);
3574 spin_unlock(&hctx->lock);
3575
3576 blk_mq_run_hw_queue(hctx, true);
3577 return 0;
3578}
3579
3580static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3581{
3582 if (!(hctx->flags & BLK_MQ_F_STACKING))
3583 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3584 &hctx->cpuhp_online);
3585 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3586 &hctx->cpuhp_dead);
3587}
3588
3589/*
3590 * Before freeing hw queue, clearing the flush request reference in
3591 * tags->rqs[] for avoiding potential UAF.
3592 */
3593static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3594 unsigned int queue_depth, struct request *flush_rq)
3595{
3596 int i;
3597 unsigned long flags;
3598
3599 /* The hw queue may not be mapped yet */
3600 if (!tags)
3601 return;
3602
3603 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3604
3605 for (i = 0; i < queue_depth; i++)
3606 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3607
3608 /*
3609 * Wait until all pending iteration is done.
3610 *
3611 * Request reference is cleared and it is guaranteed to be observed
3612 * after the ->lock is released.
3613 */
3614 spin_lock_irqsave(&tags->lock, flags);
3615 spin_unlock_irqrestore(&tags->lock, flags);
3616}
3617
3618/* hctx->ctxs will be freed in queue's release handler */
3619static void blk_mq_exit_hctx(struct request_queue *q,
3620 struct blk_mq_tag_set *set,
3621 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3622{
3623 struct request *flush_rq = hctx->fq->flush_rq;
3624
3625 if (blk_mq_hw_queue_mapped(hctx))
3626 blk_mq_tag_idle(hctx);
3627
3628 if (blk_queue_init_done(q))
3629 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3630 set->queue_depth, flush_rq);
3631 if (set->ops->exit_request)
3632 set->ops->exit_request(set, flush_rq, hctx_idx);
3633
3634 if (set->ops->exit_hctx)
3635 set->ops->exit_hctx(hctx, hctx_idx);
3636
3637 blk_mq_remove_cpuhp(hctx);
3638
3639 xa_erase(&q->hctx_table, hctx_idx);
3640
3641 spin_lock(&q->unused_hctx_lock);
3642 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3643 spin_unlock(&q->unused_hctx_lock);
3644}
3645
3646static void blk_mq_exit_hw_queues(struct request_queue *q,
3647 struct blk_mq_tag_set *set, int nr_queue)
3648{
3649 struct blk_mq_hw_ctx *hctx;
3650 unsigned long i;
3651
3652 queue_for_each_hw_ctx(q, hctx, i) {
3653 if (i == nr_queue)
3654 break;
3655 blk_mq_exit_hctx(q, set, hctx, i);
3656 }
3657}
3658
3659static int blk_mq_init_hctx(struct request_queue *q,
3660 struct blk_mq_tag_set *set,
3661 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3662{
3663 hctx->queue_num = hctx_idx;
3664
3665 if (!(hctx->flags & BLK_MQ_F_STACKING))
3666 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3667 &hctx->cpuhp_online);
3668 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3669
3670 hctx->tags = set->tags[hctx_idx];
3671
3672 if (set->ops->init_hctx &&
3673 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3674 goto unregister_cpu_notifier;
3675
3676 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3677 hctx->numa_node))
3678 goto exit_hctx;
3679
3680 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3681 goto exit_flush_rq;
3682
3683 return 0;
3684
3685 exit_flush_rq:
3686 if (set->ops->exit_request)
3687 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3688 exit_hctx:
3689 if (set->ops->exit_hctx)
3690 set->ops->exit_hctx(hctx, hctx_idx);
3691 unregister_cpu_notifier:
3692 blk_mq_remove_cpuhp(hctx);
3693 return -1;
3694}
3695
3696static struct blk_mq_hw_ctx *
3697blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3698 int node)
3699{
3700 struct blk_mq_hw_ctx *hctx;
3701 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3702
3703 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3704 if (!hctx)
3705 goto fail_alloc_hctx;
3706
3707 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3708 goto free_hctx;
3709
3710 atomic_set(&hctx->nr_active, 0);
3711 if (node == NUMA_NO_NODE)
3712 node = set->numa_node;
3713 hctx->numa_node = node;
3714
3715 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3716 spin_lock_init(&hctx->lock);
3717 INIT_LIST_HEAD(&hctx->dispatch);
3718 hctx->queue = q;
3719 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3720
3721 INIT_LIST_HEAD(&hctx->hctx_list);
3722
3723 /*
3724 * Allocate space for all possible cpus to avoid allocation at
3725 * runtime
3726 */
3727 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3728 gfp, node);
3729 if (!hctx->ctxs)
3730 goto free_cpumask;
3731
3732 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3733 gfp, node, false, false))
3734 goto free_ctxs;
3735 hctx->nr_ctx = 0;
3736
3737 spin_lock_init(&hctx->dispatch_wait_lock);
3738 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3739 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3740
3741 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3742 if (!hctx->fq)
3743 goto free_bitmap;
3744
3745 blk_mq_hctx_kobj_init(hctx);
3746
3747 return hctx;
3748
3749 free_bitmap:
3750 sbitmap_free(&hctx->ctx_map);
3751 free_ctxs:
3752 kfree(hctx->ctxs);
3753 free_cpumask:
3754 free_cpumask_var(hctx->cpumask);
3755 free_hctx:
3756 kfree(hctx);
3757 fail_alloc_hctx:
3758 return NULL;
3759}
3760
3761static void blk_mq_init_cpu_queues(struct request_queue *q,
3762 unsigned int nr_hw_queues)
3763{
3764 struct blk_mq_tag_set *set = q->tag_set;
3765 unsigned int i, j;
3766
3767 for_each_possible_cpu(i) {
3768 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3769 struct blk_mq_hw_ctx *hctx;
3770 int k;
3771
3772 __ctx->cpu = i;
3773 spin_lock_init(&__ctx->lock);
3774 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3775 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3776
3777 __ctx->queue = q;
3778
3779 /*
3780 * Set local node, IFF we have more than one hw queue. If
3781 * not, we remain on the home node of the device
3782 */
3783 for (j = 0; j < set->nr_maps; j++) {
3784 hctx = blk_mq_map_queue_type(q, j, i);
3785 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3786 hctx->numa_node = cpu_to_node(i);
3787 }
3788 }
3789}
3790
3791struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3792 unsigned int hctx_idx,
3793 unsigned int depth)
3794{
3795 struct blk_mq_tags *tags;
3796 int ret;
3797
3798 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3799 if (!tags)
3800 return NULL;
3801
3802 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3803 if (ret) {
3804 blk_mq_free_rq_map(tags);
3805 return NULL;
3806 }
3807
3808 return tags;
3809}
3810
3811static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3812 int hctx_idx)
3813{
3814 if (blk_mq_is_shared_tags(set->flags)) {
3815 set->tags[hctx_idx] = set->shared_tags;
3816
3817 return true;
3818 }
3819
3820 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3821 set->queue_depth);
3822
3823 return set->tags[hctx_idx];
3824}
3825
3826void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3827 struct blk_mq_tags *tags,
3828 unsigned int hctx_idx)
3829{
3830 if (tags) {
3831 blk_mq_free_rqs(set, tags, hctx_idx);
3832 blk_mq_free_rq_map(tags);
3833 }
3834}
3835
3836static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3837 unsigned int hctx_idx)
3838{
3839 if (!blk_mq_is_shared_tags(set->flags))
3840 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3841
3842 set->tags[hctx_idx] = NULL;
3843}
3844
3845static void blk_mq_map_swqueue(struct request_queue *q)
3846{
3847 unsigned int j, hctx_idx;
3848 unsigned long i;
3849 struct blk_mq_hw_ctx *hctx;
3850 struct blk_mq_ctx *ctx;
3851 struct blk_mq_tag_set *set = q->tag_set;
3852
3853 queue_for_each_hw_ctx(q, hctx, i) {
3854 cpumask_clear(hctx->cpumask);
3855 hctx->nr_ctx = 0;
3856 hctx->dispatch_from = NULL;
3857 }
3858
3859 /*
3860 * Map software to hardware queues.
3861 *
3862 * If the cpu isn't present, the cpu is mapped to first hctx.
3863 */
3864 for_each_possible_cpu(i) {
3865
3866 ctx = per_cpu_ptr(q->queue_ctx, i);
3867 for (j = 0; j < set->nr_maps; j++) {
3868 if (!set->map[j].nr_queues) {
3869 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3870 HCTX_TYPE_DEFAULT, i);
3871 continue;
3872 }
3873 hctx_idx = set->map[j].mq_map[i];
3874 /* unmapped hw queue can be remapped after CPU topo changed */
3875 if (!set->tags[hctx_idx] &&
3876 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3877 /*
3878 * If tags initialization fail for some hctx,
3879 * that hctx won't be brought online. In this
3880 * case, remap the current ctx to hctx[0] which
3881 * is guaranteed to always have tags allocated
3882 */
3883 set->map[j].mq_map[i] = 0;
3884 }
3885
3886 hctx = blk_mq_map_queue_type(q, j, i);
3887 ctx->hctxs[j] = hctx;
3888 /*
3889 * If the CPU is already set in the mask, then we've
3890 * mapped this one already. This can happen if
3891 * devices share queues across queue maps.
3892 */
3893 if (cpumask_test_cpu(i, hctx->cpumask))
3894 continue;
3895
3896 cpumask_set_cpu(i, hctx->cpumask);
3897 hctx->type = j;
3898 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3899 hctx->ctxs[hctx->nr_ctx++] = ctx;
3900
3901 /*
3902 * If the nr_ctx type overflows, we have exceeded the
3903 * amount of sw queues we can support.
3904 */
3905 BUG_ON(!hctx->nr_ctx);
3906 }
3907
3908 for (; j < HCTX_MAX_TYPES; j++)
3909 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3910 HCTX_TYPE_DEFAULT, i);
3911 }
3912
3913 queue_for_each_hw_ctx(q, hctx, i) {
3914 /*
3915 * If no software queues are mapped to this hardware queue,
3916 * disable it and free the request entries.
3917 */
3918 if (!hctx->nr_ctx) {
3919 /* Never unmap queue 0. We need it as a
3920 * fallback in case of a new remap fails
3921 * allocation
3922 */
3923 if (i)
3924 __blk_mq_free_map_and_rqs(set, i);
3925
3926 hctx->tags = NULL;
3927 continue;
3928 }
3929
3930 hctx->tags = set->tags[i];
3931 WARN_ON(!hctx->tags);
3932
3933 /*
3934 * Set the map size to the number of mapped software queues.
3935 * This is more accurate and more efficient than looping
3936 * over all possibly mapped software queues.
3937 */
3938 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3939
3940 /*
3941 * Initialize batch roundrobin counts
3942 */
3943 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3944 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3945 }
3946}
3947
3948/*
3949 * Caller needs to ensure that we're either frozen/quiesced, or that
3950 * the queue isn't live yet.
3951 */
3952static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3953{
3954 struct blk_mq_hw_ctx *hctx;
3955 unsigned long i;
3956
3957 queue_for_each_hw_ctx(q, hctx, i) {
3958 if (shared) {
3959 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3960 } else {
3961 blk_mq_tag_idle(hctx);
3962 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3963 }
3964 }
3965}
3966
3967static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3968 bool shared)
3969{
3970 struct request_queue *q;
3971
3972 lockdep_assert_held(&set->tag_list_lock);
3973
3974 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3975 blk_mq_freeze_queue(q);
3976 queue_set_hctx_shared(q, shared);
3977 blk_mq_unfreeze_queue(q);
3978 }
3979}
3980
3981static void blk_mq_del_queue_tag_set(struct request_queue *q)
3982{
3983 struct blk_mq_tag_set *set = q->tag_set;
3984
3985 mutex_lock(&set->tag_list_lock);
3986 list_del(&q->tag_set_list);
3987 if (list_is_singular(&set->tag_list)) {
3988 /* just transitioned to unshared */
3989 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3990 /* update existing queue */
3991 blk_mq_update_tag_set_shared(set, false);
3992 }
3993 mutex_unlock(&set->tag_list_lock);
3994 INIT_LIST_HEAD(&q->tag_set_list);
3995}
3996
3997static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3998 struct request_queue *q)
3999{
4000 mutex_lock(&set->tag_list_lock);
4001
4002 /*
4003 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4004 */
4005 if (!list_empty(&set->tag_list) &&
4006 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
4007 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
4008 /* update existing queue */
4009 blk_mq_update_tag_set_shared(set, true);
4010 }
4011 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
4012 queue_set_hctx_shared(q, true);
4013 list_add_tail(&q->tag_set_list, &set->tag_list);
4014
4015 mutex_unlock(&set->tag_list_lock);
4016}
4017
4018/* All allocations will be freed in release handler of q->mq_kobj */
4019static int blk_mq_alloc_ctxs(struct request_queue *q)
4020{
4021 struct blk_mq_ctxs *ctxs;
4022 int cpu;
4023
4024 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4025 if (!ctxs)
4026 return -ENOMEM;
4027
4028 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4029 if (!ctxs->queue_ctx)
4030 goto fail;
4031
4032 for_each_possible_cpu(cpu) {
4033 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4034 ctx->ctxs = ctxs;
4035 }
4036
4037 q->mq_kobj = &ctxs->kobj;
4038 q->queue_ctx = ctxs->queue_ctx;
4039
4040 return 0;
4041 fail:
4042 kfree(ctxs);
4043 return -ENOMEM;
4044}
4045
4046/*
4047 * It is the actual release handler for mq, but we do it from
4048 * request queue's release handler for avoiding use-after-free
4049 * and headache because q->mq_kobj shouldn't have been introduced,
4050 * but we can't group ctx/kctx kobj without it.
4051 */
4052void blk_mq_release(struct request_queue *q)
4053{
4054 struct blk_mq_hw_ctx *hctx, *next;
4055 unsigned long i;
4056
4057 queue_for_each_hw_ctx(q, hctx, i)
4058 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4059
4060 /* all hctx are in .unused_hctx_list now */
4061 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4062 list_del_init(&hctx->hctx_list);
4063 kobject_put(&hctx->kobj);
4064 }
4065
4066 xa_destroy(&q->hctx_table);
4067
4068 /*
4069 * release .mq_kobj and sw queue's kobject now because
4070 * both share lifetime with request queue.
4071 */
4072 blk_mq_sysfs_deinit(q);
4073}
4074
4075struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set,
4076 struct queue_limits *lim, void *queuedata)
4077{
4078 struct queue_limits default_lim = { };
4079 struct request_queue *q;
4080 int ret;
4081
4082 q = blk_alloc_queue(lim ? lim : &default_lim, set->numa_node);
4083 if (IS_ERR(q))
4084 return q;
4085 q->queuedata = queuedata;
4086 ret = blk_mq_init_allocated_queue(set, q);
4087 if (ret) {
4088 blk_put_queue(q);
4089 return ERR_PTR(ret);
4090 }
4091 return q;
4092}
4093EXPORT_SYMBOL(blk_mq_alloc_queue);
4094
4095/**
4096 * blk_mq_destroy_queue - shutdown a request queue
4097 * @q: request queue to shutdown
4098 *
4099 * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future
4100 * requests will be failed with -ENODEV. The caller is responsible for dropping
4101 * the reference from blk_mq_alloc_queue() by calling blk_put_queue().
4102 *
4103 * Context: can sleep
4104 */
4105void blk_mq_destroy_queue(struct request_queue *q)
4106{
4107 WARN_ON_ONCE(!queue_is_mq(q));
4108 WARN_ON_ONCE(blk_queue_registered(q));
4109
4110 might_sleep();
4111
4112 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4113 blk_queue_start_drain(q);
4114 blk_mq_freeze_queue_wait(q);
4115
4116 blk_sync_queue(q);
4117 blk_mq_cancel_work_sync(q);
4118 blk_mq_exit_queue(q);
4119}
4120EXPORT_SYMBOL(blk_mq_destroy_queue);
4121
4122struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set,
4123 struct queue_limits *lim, void *queuedata,
4124 struct lock_class_key *lkclass)
4125{
4126 struct request_queue *q;
4127 struct gendisk *disk;
4128
4129 q = blk_mq_alloc_queue(set, lim, queuedata);
4130 if (IS_ERR(q))
4131 return ERR_CAST(q);
4132
4133 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4134 if (!disk) {
4135 blk_mq_destroy_queue(q);
4136 blk_put_queue(q);
4137 return ERR_PTR(-ENOMEM);
4138 }
4139 set_bit(GD_OWNS_QUEUE, &disk->state);
4140 return disk;
4141}
4142EXPORT_SYMBOL(__blk_mq_alloc_disk);
4143
4144struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4145 struct lock_class_key *lkclass)
4146{
4147 struct gendisk *disk;
4148
4149 if (!blk_get_queue(q))
4150 return NULL;
4151 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4152 if (!disk)
4153 blk_put_queue(q);
4154 return disk;
4155}
4156EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4157
4158static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4159 struct blk_mq_tag_set *set, struct request_queue *q,
4160 int hctx_idx, int node)
4161{
4162 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4163
4164 /* reuse dead hctx first */
4165 spin_lock(&q->unused_hctx_lock);
4166 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4167 if (tmp->numa_node == node) {
4168 hctx = tmp;
4169 break;
4170 }
4171 }
4172 if (hctx)
4173 list_del_init(&hctx->hctx_list);
4174 spin_unlock(&q->unused_hctx_lock);
4175
4176 if (!hctx)
4177 hctx = blk_mq_alloc_hctx(q, set, node);
4178 if (!hctx)
4179 goto fail;
4180
4181 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4182 goto free_hctx;
4183
4184 return hctx;
4185
4186 free_hctx:
4187 kobject_put(&hctx->kobj);
4188 fail:
4189 return NULL;
4190}
4191
4192static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4193 struct request_queue *q)
4194{
4195 struct blk_mq_hw_ctx *hctx;
4196 unsigned long i, j;
4197
4198 /* protect against switching io scheduler */
4199 mutex_lock(&q->sysfs_lock);
4200 for (i = 0; i < set->nr_hw_queues; i++) {
4201 int old_node;
4202 int node = blk_mq_get_hctx_node(set, i);
4203 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4204
4205 if (old_hctx) {
4206 old_node = old_hctx->numa_node;
4207 blk_mq_exit_hctx(q, set, old_hctx, i);
4208 }
4209
4210 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4211 if (!old_hctx)
4212 break;
4213 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4214 node, old_node);
4215 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4216 WARN_ON_ONCE(!hctx);
4217 }
4218 }
4219 /*
4220 * Increasing nr_hw_queues fails. Free the newly allocated
4221 * hctxs and keep the previous q->nr_hw_queues.
4222 */
4223 if (i != set->nr_hw_queues) {
4224 j = q->nr_hw_queues;
4225 } else {
4226 j = i;
4227 q->nr_hw_queues = set->nr_hw_queues;
4228 }
4229
4230 xa_for_each_start(&q->hctx_table, j, hctx, j)
4231 blk_mq_exit_hctx(q, set, hctx, j);
4232 mutex_unlock(&q->sysfs_lock);
4233}
4234
4235static void blk_mq_update_poll_flag(struct request_queue *q)
4236{
4237 struct blk_mq_tag_set *set = q->tag_set;
4238
4239 if (set->nr_maps > HCTX_TYPE_POLL &&
4240 set->map[HCTX_TYPE_POLL].nr_queues)
4241 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4242 else
4243 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4244}
4245
4246int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4247 struct request_queue *q)
4248{
4249 /* mark the queue as mq asap */
4250 q->mq_ops = set->ops;
4251
4252 if (blk_mq_alloc_ctxs(q))
4253 goto err_exit;
4254
4255 /* init q->mq_kobj and sw queues' kobjects */
4256 blk_mq_sysfs_init(q);
4257
4258 INIT_LIST_HEAD(&q->unused_hctx_list);
4259 spin_lock_init(&q->unused_hctx_lock);
4260
4261 xa_init(&q->hctx_table);
4262
4263 blk_mq_realloc_hw_ctxs(set, q);
4264 if (!q->nr_hw_queues)
4265 goto err_hctxs;
4266
4267 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4268 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4269
4270 q->tag_set = set;
4271
4272 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4273 blk_mq_update_poll_flag(q);
4274
4275 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4276 INIT_LIST_HEAD(&q->flush_list);
4277 INIT_LIST_HEAD(&q->requeue_list);
4278 spin_lock_init(&q->requeue_lock);
4279
4280 q->nr_requests = set->queue_depth;
4281
4282 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4283 blk_mq_add_queue_tag_set(set, q);
4284 blk_mq_map_swqueue(q);
4285 return 0;
4286
4287err_hctxs:
4288 blk_mq_release(q);
4289err_exit:
4290 q->mq_ops = NULL;
4291 return -ENOMEM;
4292}
4293EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4294
4295/* tags can _not_ be used after returning from blk_mq_exit_queue */
4296void blk_mq_exit_queue(struct request_queue *q)
4297{
4298 struct blk_mq_tag_set *set = q->tag_set;
4299
4300 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4301 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4302 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4303 blk_mq_del_queue_tag_set(q);
4304}
4305
4306static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4307{
4308 int i;
4309
4310 if (blk_mq_is_shared_tags(set->flags)) {
4311 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4312 BLK_MQ_NO_HCTX_IDX,
4313 set->queue_depth);
4314 if (!set->shared_tags)
4315 return -ENOMEM;
4316 }
4317
4318 for (i = 0; i < set->nr_hw_queues; i++) {
4319 if (!__blk_mq_alloc_map_and_rqs(set, i))
4320 goto out_unwind;
4321 cond_resched();
4322 }
4323
4324 return 0;
4325
4326out_unwind:
4327 while (--i >= 0)
4328 __blk_mq_free_map_and_rqs(set, i);
4329
4330 if (blk_mq_is_shared_tags(set->flags)) {
4331 blk_mq_free_map_and_rqs(set, set->shared_tags,
4332 BLK_MQ_NO_HCTX_IDX);
4333 }
4334
4335 return -ENOMEM;
4336}
4337
4338/*
4339 * Allocate the request maps associated with this tag_set. Note that this
4340 * may reduce the depth asked for, if memory is tight. set->queue_depth
4341 * will be updated to reflect the allocated depth.
4342 */
4343static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4344{
4345 unsigned int depth;
4346 int err;
4347
4348 depth = set->queue_depth;
4349 do {
4350 err = __blk_mq_alloc_rq_maps(set);
4351 if (!err)
4352 break;
4353
4354 set->queue_depth >>= 1;
4355 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4356 err = -ENOMEM;
4357 break;
4358 }
4359 } while (set->queue_depth);
4360
4361 if (!set->queue_depth || err) {
4362 pr_err("blk-mq: failed to allocate request map\n");
4363 return -ENOMEM;
4364 }
4365
4366 if (depth != set->queue_depth)
4367 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4368 depth, set->queue_depth);
4369
4370 return 0;
4371}
4372
4373static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4374{
4375 /*
4376 * blk_mq_map_queues() and multiple .map_queues() implementations
4377 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4378 * number of hardware queues.
4379 */
4380 if (set->nr_maps == 1)
4381 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4382
4383 if (set->ops->map_queues) {
4384 int i;
4385
4386 /*
4387 * transport .map_queues is usually done in the following
4388 * way:
4389 *
4390 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4391 * mask = get_cpu_mask(queue)
4392 * for_each_cpu(cpu, mask)
4393 * set->map[x].mq_map[cpu] = queue;
4394 * }
4395 *
4396 * When we need to remap, the table has to be cleared for
4397 * killing stale mapping since one CPU may not be mapped
4398 * to any hw queue.
4399 */
4400 for (i = 0; i < set->nr_maps; i++)
4401 blk_mq_clear_mq_map(&set->map[i]);
4402
4403 set->ops->map_queues(set);
4404 } else {
4405 BUG_ON(set->nr_maps > 1);
4406 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4407 }
4408}
4409
4410static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4411 int new_nr_hw_queues)
4412{
4413 struct blk_mq_tags **new_tags;
4414 int i;
4415
4416 if (set->nr_hw_queues >= new_nr_hw_queues)
4417 goto done;
4418
4419 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4420 GFP_KERNEL, set->numa_node);
4421 if (!new_tags)
4422 return -ENOMEM;
4423
4424 if (set->tags)
4425 memcpy(new_tags, set->tags, set->nr_hw_queues *
4426 sizeof(*set->tags));
4427 kfree(set->tags);
4428 set->tags = new_tags;
4429
4430 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4431 if (!__blk_mq_alloc_map_and_rqs(set, i)) {
4432 while (--i >= set->nr_hw_queues)
4433 __blk_mq_free_map_and_rqs(set, i);
4434 return -ENOMEM;
4435 }
4436 cond_resched();
4437 }
4438
4439done:
4440 set->nr_hw_queues = new_nr_hw_queues;
4441 return 0;
4442}
4443
4444/*
4445 * Alloc a tag set to be associated with one or more request queues.
4446 * May fail with EINVAL for various error conditions. May adjust the
4447 * requested depth down, if it's too large. In that case, the set
4448 * value will be stored in set->queue_depth.
4449 */
4450int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4451{
4452 int i, ret;
4453
4454 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4455
4456 if (!set->nr_hw_queues)
4457 return -EINVAL;
4458 if (!set->queue_depth)
4459 return -EINVAL;
4460 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4461 return -EINVAL;
4462
4463 if (!set->ops->queue_rq)
4464 return -EINVAL;
4465
4466 if (!set->ops->get_budget ^ !set->ops->put_budget)
4467 return -EINVAL;
4468
4469 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4470 pr_info("blk-mq: reduced tag depth to %u\n",
4471 BLK_MQ_MAX_DEPTH);
4472 set->queue_depth = BLK_MQ_MAX_DEPTH;
4473 }
4474
4475 if (!set->nr_maps)
4476 set->nr_maps = 1;
4477 else if (set->nr_maps > HCTX_MAX_TYPES)
4478 return -EINVAL;
4479
4480 /*
4481 * If a crashdump is active, then we are potentially in a very
4482 * memory constrained environment. Limit us to 64 tags to prevent
4483 * using too much memory.
4484 */
4485 if (is_kdump_kernel())
4486 set->queue_depth = min(64U, set->queue_depth);
4487
4488 /*
4489 * There is no use for more h/w queues than cpus if we just have
4490 * a single map
4491 */
4492 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4493 set->nr_hw_queues = nr_cpu_ids;
4494
4495 if (set->flags & BLK_MQ_F_BLOCKING) {
4496 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4497 if (!set->srcu)
4498 return -ENOMEM;
4499 ret = init_srcu_struct(set->srcu);
4500 if (ret)
4501 goto out_free_srcu;
4502 }
4503
4504 ret = -ENOMEM;
4505 set->tags = kcalloc_node(set->nr_hw_queues,
4506 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4507 set->numa_node);
4508 if (!set->tags)
4509 goto out_cleanup_srcu;
4510
4511 for (i = 0; i < set->nr_maps; i++) {
4512 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4513 sizeof(set->map[i].mq_map[0]),
4514 GFP_KERNEL, set->numa_node);
4515 if (!set->map[i].mq_map)
4516 goto out_free_mq_map;
4517 set->map[i].nr_queues = set->nr_hw_queues;
4518 }
4519
4520 blk_mq_update_queue_map(set);
4521
4522 ret = blk_mq_alloc_set_map_and_rqs(set);
4523 if (ret)
4524 goto out_free_mq_map;
4525
4526 mutex_init(&set->tag_list_lock);
4527 INIT_LIST_HEAD(&set->tag_list);
4528
4529 return 0;
4530
4531out_free_mq_map:
4532 for (i = 0; i < set->nr_maps; i++) {
4533 kfree(set->map[i].mq_map);
4534 set->map[i].mq_map = NULL;
4535 }
4536 kfree(set->tags);
4537 set->tags = NULL;
4538out_cleanup_srcu:
4539 if (set->flags & BLK_MQ_F_BLOCKING)
4540 cleanup_srcu_struct(set->srcu);
4541out_free_srcu:
4542 if (set->flags & BLK_MQ_F_BLOCKING)
4543 kfree(set->srcu);
4544 return ret;
4545}
4546EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4547
4548/* allocate and initialize a tagset for a simple single-queue device */
4549int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4550 const struct blk_mq_ops *ops, unsigned int queue_depth,
4551 unsigned int set_flags)
4552{
4553 memset(set, 0, sizeof(*set));
4554 set->ops = ops;
4555 set->nr_hw_queues = 1;
4556 set->nr_maps = 1;
4557 set->queue_depth = queue_depth;
4558 set->numa_node = NUMA_NO_NODE;
4559 set->flags = set_flags;
4560 return blk_mq_alloc_tag_set(set);
4561}
4562EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4563
4564void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4565{
4566 int i, j;
4567
4568 for (i = 0; i < set->nr_hw_queues; i++)
4569 __blk_mq_free_map_and_rqs(set, i);
4570
4571 if (blk_mq_is_shared_tags(set->flags)) {
4572 blk_mq_free_map_and_rqs(set, set->shared_tags,
4573 BLK_MQ_NO_HCTX_IDX);
4574 }
4575
4576 for (j = 0; j < set->nr_maps; j++) {
4577 kfree(set->map[j].mq_map);
4578 set->map[j].mq_map = NULL;
4579 }
4580
4581 kfree(set->tags);
4582 set->tags = NULL;
4583 if (set->flags & BLK_MQ_F_BLOCKING) {
4584 cleanup_srcu_struct(set->srcu);
4585 kfree(set->srcu);
4586 }
4587}
4588EXPORT_SYMBOL(blk_mq_free_tag_set);
4589
4590int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4591{
4592 struct blk_mq_tag_set *set = q->tag_set;
4593 struct blk_mq_hw_ctx *hctx;
4594 int ret;
4595 unsigned long i;
4596
4597 if (!set)
4598 return -EINVAL;
4599
4600 if (q->nr_requests == nr)
4601 return 0;
4602
4603 blk_mq_freeze_queue(q);
4604 blk_mq_quiesce_queue(q);
4605
4606 ret = 0;
4607 queue_for_each_hw_ctx(q, hctx, i) {
4608 if (!hctx->tags)
4609 continue;
4610 /*
4611 * If we're using an MQ scheduler, just update the scheduler
4612 * queue depth. This is similar to what the old code would do.
4613 */
4614 if (hctx->sched_tags) {
4615 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4616 nr, true);
4617 } else {
4618 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4619 false);
4620 }
4621 if (ret)
4622 break;
4623 if (q->elevator && q->elevator->type->ops.depth_updated)
4624 q->elevator->type->ops.depth_updated(hctx);
4625 }
4626 if (!ret) {
4627 q->nr_requests = nr;
4628 if (blk_mq_is_shared_tags(set->flags)) {
4629 if (q->elevator)
4630 blk_mq_tag_update_sched_shared_tags(q);
4631 else
4632 blk_mq_tag_resize_shared_tags(set, nr);
4633 }
4634 }
4635
4636 blk_mq_unquiesce_queue(q);
4637 blk_mq_unfreeze_queue(q);
4638
4639 return ret;
4640}
4641
4642/*
4643 * request_queue and elevator_type pair.
4644 * It is just used by __blk_mq_update_nr_hw_queues to cache
4645 * the elevator_type associated with a request_queue.
4646 */
4647struct blk_mq_qe_pair {
4648 struct list_head node;
4649 struct request_queue *q;
4650 struct elevator_type *type;
4651};
4652
4653/*
4654 * Cache the elevator_type in qe pair list and switch the
4655 * io scheduler to 'none'
4656 */
4657static bool blk_mq_elv_switch_none(struct list_head *head,
4658 struct request_queue *q)
4659{
4660 struct blk_mq_qe_pair *qe;
4661
4662 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4663 if (!qe)
4664 return false;
4665
4666 /* q->elevator needs protection from ->sysfs_lock */
4667 mutex_lock(&q->sysfs_lock);
4668
4669 /* the check has to be done with holding sysfs_lock */
4670 if (!q->elevator) {
4671 kfree(qe);
4672 goto unlock;
4673 }
4674
4675 INIT_LIST_HEAD(&qe->node);
4676 qe->q = q;
4677 qe->type = q->elevator->type;
4678 /* keep a reference to the elevator module as we'll switch back */
4679 __elevator_get(qe->type);
4680 list_add(&qe->node, head);
4681 elevator_disable(q);
4682unlock:
4683 mutex_unlock(&q->sysfs_lock);
4684
4685 return true;
4686}
4687
4688static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4689 struct request_queue *q)
4690{
4691 struct blk_mq_qe_pair *qe;
4692
4693 list_for_each_entry(qe, head, node)
4694 if (qe->q == q)
4695 return qe;
4696
4697 return NULL;
4698}
4699
4700static void blk_mq_elv_switch_back(struct list_head *head,
4701 struct request_queue *q)
4702{
4703 struct blk_mq_qe_pair *qe;
4704 struct elevator_type *t;
4705
4706 qe = blk_lookup_qe_pair(head, q);
4707 if (!qe)
4708 return;
4709 t = qe->type;
4710 list_del(&qe->node);
4711 kfree(qe);
4712
4713 mutex_lock(&q->sysfs_lock);
4714 elevator_switch(q, t);
4715 /* drop the reference acquired in blk_mq_elv_switch_none */
4716 elevator_put(t);
4717 mutex_unlock(&q->sysfs_lock);
4718}
4719
4720static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4721 int nr_hw_queues)
4722{
4723 struct request_queue *q;
4724 LIST_HEAD(head);
4725 int prev_nr_hw_queues = set->nr_hw_queues;
4726 int i;
4727
4728 lockdep_assert_held(&set->tag_list_lock);
4729
4730 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4731 nr_hw_queues = nr_cpu_ids;
4732 if (nr_hw_queues < 1)
4733 return;
4734 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4735 return;
4736
4737 list_for_each_entry(q, &set->tag_list, tag_set_list)
4738 blk_mq_freeze_queue(q);
4739 /*
4740 * Switch IO scheduler to 'none', cleaning up the data associated
4741 * with the previous scheduler. We will switch back once we are done
4742 * updating the new sw to hw queue mappings.
4743 */
4744 list_for_each_entry(q, &set->tag_list, tag_set_list)
4745 if (!blk_mq_elv_switch_none(&head, q))
4746 goto switch_back;
4747
4748 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4749 blk_mq_debugfs_unregister_hctxs(q);
4750 blk_mq_sysfs_unregister_hctxs(q);
4751 }
4752
4753 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4754 goto reregister;
4755
4756fallback:
4757 blk_mq_update_queue_map(set);
4758 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4759 blk_mq_realloc_hw_ctxs(set, q);
4760 blk_mq_update_poll_flag(q);
4761 if (q->nr_hw_queues != set->nr_hw_queues) {
4762 int i = prev_nr_hw_queues;
4763
4764 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4765 nr_hw_queues, prev_nr_hw_queues);
4766 for (; i < set->nr_hw_queues; i++)
4767 __blk_mq_free_map_and_rqs(set, i);
4768
4769 set->nr_hw_queues = prev_nr_hw_queues;
4770 goto fallback;
4771 }
4772 blk_mq_map_swqueue(q);
4773 }
4774
4775reregister:
4776 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4777 blk_mq_sysfs_register_hctxs(q);
4778 blk_mq_debugfs_register_hctxs(q);
4779 }
4780
4781switch_back:
4782 list_for_each_entry(q, &set->tag_list, tag_set_list)
4783 blk_mq_elv_switch_back(&head, q);
4784
4785 list_for_each_entry(q, &set->tag_list, tag_set_list)
4786 blk_mq_unfreeze_queue(q);
4787
4788 /* Free the excess tags when nr_hw_queues shrink. */
4789 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
4790 __blk_mq_free_map_and_rqs(set, i);
4791}
4792
4793void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4794{
4795 mutex_lock(&set->tag_list_lock);
4796 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4797 mutex_unlock(&set->tag_list_lock);
4798}
4799EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4800
4801static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4802 struct io_comp_batch *iob, unsigned int flags)
4803{
4804 long state = get_current_state();
4805 int ret;
4806
4807 do {
4808 ret = q->mq_ops->poll(hctx, iob);
4809 if (ret > 0) {
4810 __set_current_state(TASK_RUNNING);
4811 return ret;
4812 }
4813
4814 if (signal_pending_state(state, current))
4815 __set_current_state(TASK_RUNNING);
4816 if (task_is_running(current))
4817 return 1;
4818
4819 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4820 break;
4821 cpu_relax();
4822 } while (!need_resched());
4823
4824 __set_current_state(TASK_RUNNING);
4825 return 0;
4826}
4827
4828int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4829 struct io_comp_batch *iob, unsigned int flags)
4830{
4831 struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4832
4833 return blk_hctx_poll(q, hctx, iob, flags);
4834}
4835
4836int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4837 unsigned int poll_flags)
4838{
4839 struct request_queue *q = rq->q;
4840 int ret;
4841
4842 if (!blk_rq_is_poll(rq))
4843 return 0;
4844 if (!percpu_ref_tryget(&q->q_usage_counter))
4845 return 0;
4846
4847 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4848 blk_queue_exit(q);
4849
4850 return ret;
4851}
4852EXPORT_SYMBOL_GPL(blk_rq_poll);
4853
4854unsigned int blk_mq_rq_cpu(struct request *rq)
4855{
4856 return rq->mq_ctx->cpu;
4857}
4858EXPORT_SYMBOL(blk_mq_rq_cpu);
4859
4860void blk_mq_cancel_work_sync(struct request_queue *q)
4861{
4862 struct blk_mq_hw_ctx *hctx;
4863 unsigned long i;
4864
4865 cancel_delayed_work_sync(&q->requeue_work);
4866
4867 queue_for_each_hw_ctx(q, hctx, i)
4868 cancel_delayed_work_sync(&hctx->run_work);
4869}
4870
4871static int __init blk_mq_init(void)
4872{
4873 int i;
4874
4875 for_each_possible_cpu(i)
4876 init_llist_head(&per_cpu(blk_cpu_done, i));
4877 for_each_possible_cpu(i)
4878 INIT_CSD(&per_cpu(blk_cpu_csd, i),
4879 __blk_mq_complete_request_remote, NULL);
4880 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4881
4882 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4883 "block/softirq:dead", NULL,
4884 blk_softirq_cpu_dead);
4885 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4886 blk_mq_hctx_notify_dead);
4887 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4888 blk_mq_hctx_notify_online,
4889 blk_mq_hctx_notify_offline);
4890 return 0;
4891}
4892subsys_initcall(blk_mq_init);