blob: c661896e2465bd0f9248c85400be210b67efed76 [file] [log] [blame]
Jens Axboe320ae512013-10-24 09:20:05 +01001#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
30DEFINE_PER_CPU(struct llist_head, ipi_lists);
31
32static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
33 unsigned int cpu)
34{
35 return per_cpu_ptr(q->queue_ctx, cpu);
36}
37
38/*
39 * This assumes per-cpu software queueing queues. They could be per-node
40 * as well, for instance. For now this is hardcoded as-is. Note that we don't
41 * care about preemption, since we know the ctx's are persistent. This does
42 * mean that we can't rely on ctx always matching the currently running CPU.
43 */
44static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
45{
46 return __blk_mq_get_ctx(q, get_cpu());
47}
48
49static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
50{
51 put_cpu();
52}
53
54/*
55 * Check if any of the ctx's have pending work in this hardware queue
56 */
57static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
58{
59 unsigned int i;
60
61 for (i = 0; i < hctx->nr_ctx_map; i++)
62 if (hctx->ctx_map[i])
63 return true;
64
65 return false;
66}
67
68/*
69 * Mark this ctx as having pending work in this hardware queue
70 */
71static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
72 struct blk_mq_ctx *ctx)
73{
74 if (!test_bit(ctx->index_hw, hctx->ctx_map))
75 set_bit(ctx->index_hw, hctx->ctx_map);
76}
77
78static struct request *blk_mq_alloc_rq(struct blk_mq_hw_ctx *hctx, gfp_t gfp,
79 bool reserved)
80{
81 struct request *rq;
82 unsigned int tag;
83
84 tag = blk_mq_get_tag(hctx->tags, gfp, reserved);
85 if (tag != BLK_MQ_TAG_FAIL) {
86 rq = hctx->rqs[tag];
87 rq->tag = tag;
88
89 return rq;
90 }
91
92 return NULL;
93}
94
95static int blk_mq_queue_enter(struct request_queue *q)
96{
97 int ret;
98
99 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
100 smp_wmb();
101 /* we have problems to freeze the queue if it's initializing */
102 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
103 return 0;
104
105 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
106
107 spin_lock_irq(q->queue_lock);
108 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
109 !blk_queue_bypass(q), *q->queue_lock);
110 /* inc usage with lock hold to avoid freeze_queue runs here */
111 if (!ret)
112 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
113 spin_unlock_irq(q->queue_lock);
114
115 return ret;
116}
117
118static void blk_mq_queue_exit(struct request_queue *q)
119{
120 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
121}
122
123/*
124 * Guarantee no request is in use, so we can change any data structure of
125 * the queue afterward.
126 */
127static void blk_mq_freeze_queue(struct request_queue *q)
128{
129 bool drain;
130
131 spin_lock_irq(q->queue_lock);
132 drain = !q->bypass_depth++;
133 queue_flag_set(QUEUE_FLAG_BYPASS, q);
134 spin_unlock_irq(q->queue_lock);
135
136 if (!drain)
137 return;
138
139 while (true) {
140 s64 count;
141
142 spin_lock_irq(q->queue_lock);
143 count = percpu_counter_sum(&q->mq_usage_counter);
144 spin_unlock_irq(q->queue_lock);
145
146 if (count == 0)
147 break;
148 blk_mq_run_queues(q, false);
149 msleep(10);
150 }
151}
152
153static void blk_mq_unfreeze_queue(struct request_queue *q)
154{
155 bool wake = false;
156
157 spin_lock_irq(q->queue_lock);
158 if (!--q->bypass_depth) {
159 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
160 wake = true;
161 }
162 WARN_ON_ONCE(q->bypass_depth < 0);
163 spin_unlock_irq(q->queue_lock);
164 if (wake)
165 wake_up_all(&q->mq_freeze_wq);
166}
167
168bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
169{
170 return blk_mq_has_free_tags(hctx->tags);
171}
172EXPORT_SYMBOL(blk_mq_can_queue);
173
174static void blk_mq_rq_ctx_init(struct blk_mq_ctx *ctx, struct request *rq,
175 unsigned int rw_flags)
176{
177 rq->mq_ctx = ctx;
178 rq->cmd_flags = rw_flags;
179 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
180}
181
182static struct request *__blk_mq_alloc_request(struct blk_mq_hw_ctx *hctx,
183 gfp_t gfp, bool reserved)
184{
185 return blk_mq_alloc_rq(hctx, gfp, reserved);
186}
187
188static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
189 int rw, gfp_t gfp,
190 bool reserved)
191{
192 struct request *rq;
193
194 do {
195 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
196 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
197
198 rq = __blk_mq_alloc_request(hctx, gfp & ~__GFP_WAIT, reserved);
199 if (rq) {
200 blk_mq_rq_ctx_init(ctx, rq, rw);
201 break;
202 } else if (!(gfp & __GFP_WAIT))
203 break;
204
205 blk_mq_put_ctx(ctx);
206 __blk_mq_run_hw_queue(hctx);
207 blk_mq_wait_for_tags(hctx->tags);
208 } while (1);
209
210 return rq;
211}
212
Christoph Hellwig3228f482013-10-28 13:33:58 -0600213struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
214 gfp_t gfp, bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100215{
216 struct request *rq;
217
218 if (blk_mq_queue_enter(q))
219 return NULL;
220
Christoph Hellwig3228f482013-10-28 13:33:58 -0600221 rq = blk_mq_alloc_request_pinned(q, rw, gfp, reserved);
Jens Axboe320ae512013-10-24 09:20:05 +0100222 blk_mq_put_ctx(rq->mq_ctx);
223 return rq;
224}
225
226struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
227 gfp_t gfp)
228{
229 struct request *rq;
230
231 if (blk_mq_queue_enter(q))
232 return NULL;
233
234 rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
235 blk_mq_put_ctx(rq->mq_ctx);
236 return rq;
237}
238EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
239
240/*
241 * Re-init and set pdu, if we have it
242 */
243static void blk_mq_rq_init(struct blk_mq_hw_ctx *hctx, struct request *rq)
244{
245 blk_rq_init(hctx->queue, rq);
246
247 if (hctx->cmd_size)
248 rq->special = blk_mq_rq_to_pdu(rq);
249}
250
251static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
252 struct blk_mq_ctx *ctx, struct request *rq)
253{
254 const int tag = rq->tag;
255 struct request_queue *q = rq->q;
256
257 blk_mq_rq_init(hctx, rq);
258 blk_mq_put_tag(hctx->tags, tag);
259
260 blk_mq_queue_exit(q);
261}
262
263void blk_mq_free_request(struct request *rq)
264{
265 struct blk_mq_ctx *ctx = rq->mq_ctx;
266 struct blk_mq_hw_ctx *hctx;
267 struct request_queue *q = rq->q;
268
269 ctx->rq_completed[rq_is_sync(rq)]++;
270
271 hctx = q->mq_ops->map_queue(q, ctx->cpu);
272 __blk_mq_free_request(hctx, ctx, rq);
273}
274
275static void blk_mq_bio_endio(struct request *rq, struct bio *bio, int error)
276{
277 if (error)
278 clear_bit(BIO_UPTODATE, &bio->bi_flags);
279 else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
280 error = -EIO;
281
282 if (unlikely(rq->cmd_flags & REQ_QUIET))
283 set_bit(BIO_QUIET, &bio->bi_flags);
284
285 /* don't actually finish bio if it's part of flush sequence */
286 if (!(rq->cmd_flags & REQ_FLUSH_SEQ))
287 bio_endio(bio, error);
288}
289
290void blk_mq_complete_request(struct request *rq, int error)
291{
292 struct bio *bio = rq->bio;
293 unsigned int bytes = 0;
294
295 trace_block_rq_complete(rq->q, rq);
296
297 while (bio) {
298 struct bio *next = bio->bi_next;
299
300 bio->bi_next = NULL;
301 bytes += bio->bi_size;
302 blk_mq_bio_endio(rq, bio, error);
303 bio = next;
304 }
305
306 blk_account_io_completion(rq, bytes);
307
308 if (rq->end_io)
309 rq->end_io(rq, error);
310 else
311 blk_mq_free_request(rq);
312
313 blk_account_io_done(rq);
314}
315
316void __blk_mq_end_io(struct request *rq, int error)
317{
318 if (!blk_mark_rq_complete(rq))
319 blk_mq_complete_request(rq, error);
320}
321
Christoph Hellwig0a06ff02013-11-14 14:32:07 -0800322#if defined(CONFIG_SMP)
Jens Axboe320ae512013-10-24 09:20:05 +0100323
324/*
325 * Called with interrupts disabled.
326 */
327static void ipi_end_io(void *data)
328{
329 struct llist_head *list = &per_cpu(ipi_lists, smp_processor_id());
330 struct llist_node *entry, *next;
331 struct request *rq;
332
333 entry = llist_del_all(list);
334
335 while (entry) {
336 next = entry->next;
337 rq = llist_entry(entry, struct request, ll_list);
338 __blk_mq_end_io(rq, rq->errors);
339 entry = next;
340 }
341}
342
343static int ipi_remote_cpu(struct blk_mq_ctx *ctx, const int cpu,
344 struct request *rq, const int error)
345{
346 struct call_single_data *data = &rq->csd;
347
348 rq->errors = error;
349 rq->ll_list.next = NULL;
350
351 /*
352 * If the list is non-empty, an existing IPI must already
353 * be "in flight". If that is the case, we need not schedule
354 * a new one.
355 */
356 if (llist_add(&rq->ll_list, &per_cpu(ipi_lists, ctx->cpu))) {
357 data->func = ipi_end_io;
358 data->flags = 0;
359 __smp_call_function_single(ctx->cpu, data, 0);
360 }
361
362 return true;
363}
Christoph Hellwig0a06ff02013-11-14 14:32:07 -0800364#else /* CONFIG_SMP */
Jens Axboe320ae512013-10-24 09:20:05 +0100365static int ipi_remote_cpu(struct blk_mq_ctx *ctx, const int cpu,
366 struct request *rq, const int error)
367{
368 return false;
369}
370#endif
371
372/*
373 * End IO on this request on a multiqueue enabled driver. We'll either do
374 * it directly inline, or punt to a local IPI handler on the matching
375 * remote CPU.
376 */
377void blk_mq_end_io(struct request *rq, int error)
378{
379 struct blk_mq_ctx *ctx = rq->mq_ctx;
380 int cpu;
381
382 if (!ctx->ipi_redirect)
383 return __blk_mq_end_io(rq, error);
384
385 cpu = get_cpu();
386
387 if (cpu == ctx->cpu || !cpu_online(ctx->cpu) ||
388 !ipi_remote_cpu(ctx, cpu, rq, error))
389 __blk_mq_end_io(rq, error);
390
391 put_cpu();
392}
393EXPORT_SYMBOL(blk_mq_end_io);
394
395static void blk_mq_start_request(struct request *rq)
396{
397 struct request_queue *q = rq->q;
398
399 trace_block_rq_issue(q, rq);
400
401 /*
402 * Just mark start time and set the started bit. Due to memory
403 * ordering, we know we'll see the correct deadline as long as
404 * REQ_ATOMIC_STARTED is seen.
405 */
406 rq->deadline = jiffies + q->rq_timeout;
407 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
408}
409
410static void blk_mq_requeue_request(struct request *rq)
411{
412 struct request_queue *q = rq->q;
413
414 trace_block_rq_requeue(q, rq);
415 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
416}
417
418struct blk_mq_timeout_data {
419 struct blk_mq_hw_ctx *hctx;
420 unsigned long *next;
421 unsigned int *next_set;
422};
423
424static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
425{
426 struct blk_mq_timeout_data *data = __data;
427 struct blk_mq_hw_ctx *hctx = data->hctx;
428 unsigned int tag;
429
430 /* It may not be in flight yet (this is where
431 * the REQ_ATOMIC_STARTED flag comes in). The requests are
432 * statically allocated, so we know it's always safe to access the
433 * memory associated with a bit offset into ->rqs[].
434 */
435 tag = 0;
436 do {
437 struct request *rq;
438
439 tag = find_next_zero_bit(free_tags, hctx->queue_depth, tag);
440 if (tag >= hctx->queue_depth)
441 break;
442
443 rq = hctx->rqs[tag++];
444
445 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
446 continue;
447
448 blk_rq_check_expired(rq, data->next, data->next_set);
449 } while (1);
450}
451
452static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
453 unsigned long *next,
454 unsigned int *next_set)
455{
456 struct blk_mq_timeout_data data = {
457 .hctx = hctx,
458 .next = next,
459 .next_set = next_set,
460 };
461
462 /*
463 * Ask the tagging code to iterate busy requests, so we can
464 * check them for timeout.
465 */
466 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
467}
468
469static void blk_mq_rq_timer(unsigned long data)
470{
471 struct request_queue *q = (struct request_queue *) data;
472 struct blk_mq_hw_ctx *hctx;
473 unsigned long next = 0;
474 int i, next_set = 0;
475
476 queue_for_each_hw_ctx(q, hctx, i)
477 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
478
479 if (next_set)
480 mod_timer(&q->timeout, round_jiffies_up(next));
481}
482
483/*
484 * Reverse check our software queue for entries that we could potentially
485 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
486 * too much time checking for merges.
487 */
488static bool blk_mq_attempt_merge(struct request_queue *q,
489 struct blk_mq_ctx *ctx, struct bio *bio)
490{
491 struct request *rq;
492 int checked = 8;
493
494 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
495 int el_ret;
496
497 if (!checked--)
498 break;
499
500 if (!blk_rq_merge_ok(rq, bio))
501 continue;
502
503 el_ret = blk_try_merge(rq, bio);
504 if (el_ret == ELEVATOR_BACK_MERGE) {
505 if (bio_attempt_back_merge(q, rq, bio)) {
506 ctx->rq_merged++;
507 return true;
508 }
509 break;
510 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
511 if (bio_attempt_front_merge(q, rq, bio)) {
512 ctx->rq_merged++;
513 return true;
514 }
515 break;
516 }
517 }
518
519 return false;
520}
521
522void blk_mq_add_timer(struct request *rq)
523{
524 __blk_add_timer(rq, NULL);
525}
526
527/*
528 * Run this hardware queue, pulling any software queues mapped to it in.
529 * Note that this function currently has various problems around ordering
530 * of IO. In particular, we'd like FIFO behaviour on handling existing
531 * items on the hctx->dispatch list. Ignore that for now.
532 */
533static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
534{
535 struct request_queue *q = hctx->queue;
536 struct blk_mq_ctx *ctx;
537 struct request *rq;
538 LIST_HEAD(rq_list);
539 int bit, queued;
540
541 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
542 return;
543
544 hctx->run++;
545
546 /*
547 * Touch any software queue that has pending entries.
548 */
549 for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
550 clear_bit(bit, hctx->ctx_map);
551 ctx = hctx->ctxs[bit];
552 BUG_ON(bit != ctx->index_hw);
553
554 spin_lock(&ctx->lock);
555 list_splice_tail_init(&ctx->rq_list, &rq_list);
556 spin_unlock(&ctx->lock);
557 }
558
559 /*
560 * If we have previous entries on our dispatch list, grab them
561 * and stuff them at the front for more fair dispatch.
562 */
563 if (!list_empty_careful(&hctx->dispatch)) {
564 spin_lock(&hctx->lock);
565 if (!list_empty(&hctx->dispatch))
566 list_splice_init(&hctx->dispatch, &rq_list);
567 spin_unlock(&hctx->lock);
568 }
569
570 /*
571 * Delete and return all entries from our dispatch list
572 */
573 queued = 0;
574
575 /*
576 * Now process all the entries, sending them to the driver.
577 */
578 while (!list_empty(&rq_list)) {
579 int ret;
580
581 rq = list_first_entry(&rq_list, struct request, queuelist);
582 list_del_init(&rq->queuelist);
583 blk_mq_start_request(rq);
584
585 /*
586 * Last request in the series. Flag it as such, this
587 * enables drivers to know when IO should be kicked off,
588 * if they don't do it on a per-request basis.
589 *
590 * Note: the flag isn't the only condition drivers
591 * should do kick off. If drive is busy, the last
592 * request might not have the bit set.
593 */
594 if (list_empty(&rq_list))
595 rq->cmd_flags |= REQ_END;
596
597 ret = q->mq_ops->queue_rq(hctx, rq);
598 switch (ret) {
599 case BLK_MQ_RQ_QUEUE_OK:
600 queued++;
601 continue;
602 case BLK_MQ_RQ_QUEUE_BUSY:
603 /*
604 * FIXME: we should have a mechanism to stop the queue
605 * like blk_stop_queue, otherwise we will waste cpu
606 * time
607 */
608 list_add(&rq->queuelist, &rq_list);
609 blk_mq_requeue_request(rq);
610 break;
611 default:
612 pr_err("blk-mq: bad return on queue: %d\n", ret);
613 rq->errors = -EIO;
614 case BLK_MQ_RQ_QUEUE_ERROR:
615 blk_mq_end_io(rq, rq->errors);
616 break;
617 }
618
619 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
620 break;
621 }
622
623 if (!queued)
624 hctx->dispatched[0]++;
625 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
626 hctx->dispatched[ilog2(queued) + 1]++;
627
628 /*
629 * Any items that need requeuing? Stuff them into hctx->dispatch,
630 * that is where we will continue on next queue run.
631 */
632 if (!list_empty(&rq_list)) {
633 spin_lock(&hctx->lock);
634 list_splice(&rq_list, &hctx->dispatch);
635 spin_unlock(&hctx->lock);
636 }
637}
638
639void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
640{
641 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
642 return;
643
644 if (!async)
645 __blk_mq_run_hw_queue(hctx);
646 else {
647 struct request_queue *q = hctx->queue;
648
649 kblockd_schedule_delayed_work(q, &hctx->delayed_work, 0);
650 }
651}
652
653void blk_mq_run_queues(struct request_queue *q, bool async)
654{
655 struct blk_mq_hw_ctx *hctx;
656 int i;
657
658 queue_for_each_hw_ctx(q, hctx, i) {
659 if ((!blk_mq_hctx_has_pending(hctx) &&
660 list_empty_careful(&hctx->dispatch)) ||
661 test_bit(BLK_MQ_S_STOPPED, &hctx->flags))
662 continue;
663
664 blk_mq_run_hw_queue(hctx, async);
665 }
666}
667EXPORT_SYMBOL(blk_mq_run_queues);
668
669void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
670{
671 cancel_delayed_work(&hctx->delayed_work);
672 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
673}
674EXPORT_SYMBOL(blk_mq_stop_hw_queue);
675
Christoph Hellwig280d45f2013-10-25 14:45:58 +0100676void blk_mq_stop_hw_queues(struct request_queue *q)
677{
678 struct blk_mq_hw_ctx *hctx;
679 int i;
680
681 queue_for_each_hw_ctx(q, hctx, i)
682 blk_mq_stop_hw_queue(hctx);
683}
684EXPORT_SYMBOL(blk_mq_stop_hw_queues);
685
Jens Axboe320ae512013-10-24 09:20:05 +0100686void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
687{
688 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
689 __blk_mq_run_hw_queue(hctx);
690}
691EXPORT_SYMBOL(blk_mq_start_hw_queue);
692
693void blk_mq_start_stopped_hw_queues(struct request_queue *q)
694{
695 struct blk_mq_hw_ctx *hctx;
696 int i;
697
698 queue_for_each_hw_ctx(q, hctx, i) {
699 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
700 continue;
701
702 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
703 blk_mq_run_hw_queue(hctx, true);
704 }
705}
706EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
707
708static void blk_mq_work_fn(struct work_struct *work)
709{
710 struct blk_mq_hw_ctx *hctx;
711
712 hctx = container_of(work, struct blk_mq_hw_ctx, delayed_work.work);
713 __blk_mq_run_hw_queue(hctx);
714}
715
716static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
717 struct request *rq)
718{
719 struct blk_mq_ctx *ctx = rq->mq_ctx;
720
721 list_add_tail(&rq->queuelist, &ctx->rq_list);
722 blk_mq_hctx_mark_pending(hctx, ctx);
723
724 /*
725 * We do this early, to ensure we are on the right CPU.
726 */
727 blk_mq_add_timer(rq);
728}
729
730void blk_mq_insert_request(struct request_queue *q, struct request *rq,
731 bool run_queue)
732{
733 struct blk_mq_hw_ctx *hctx;
734 struct blk_mq_ctx *ctx, *current_ctx;
735
736 ctx = rq->mq_ctx;
737 hctx = q->mq_ops->map_queue(q, ctx->cpu);
738
739 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA)) {
740 blk_insert_flush(rq);
741 } else {
742 current_ctx = blk_mq_get_ctx(q);
743
744 if (!cpu_online(ctx->cpu)) {
745 ctx = current_ctx;
746 hctx = q->mq_ops->map_queue(q, ctx->cpu);
747 rq->mq_ctx = ctx;
748 }
749 spin_lock(&ctx->lock);
750 __blk_mq_insert_request(hctx, rq);
751 spin_unlock(&ctx->lock);
752
753 blk_mq_put_ctx(current_ctx);
754 }
755
756 if (run_queue)
757 __blk_mq_run_hw_queue(hctx);
758}
759EXPORT_SYMBOL(blk_mq_insert_request);
760
761/*
762 * This is a special version of blk_mq_insert_request to bypass FLUSH request
763 * check. Should only be used internally.
764 */
765void blk_mq_run_request(struct request *rq, bool run_queue, bool async)
766{
767 struct request_queue *q = rq->q;
768 struct blk_mq_hw_ctx *hctx;
769 struct blk_mq_ctx *ctx, *current_ctx;
770
771 current_ctx = blk_mq_get_ctx(q);
772
773 ctx = rq->mq_ctx;
774 if (!cpu_online(ctx->cpu)) {
775 ctx = current_ctx;
776 rq->mq_ctx = ctx;
777 }
778 hctx = q->mq_ops->map_queue(q, ctx->cpu);
779
780 /* ctx->cpu might be offline */
781 spin_lock(&ctx->lock);
782 __blk_mq_insert_request(hctx, rq);
783 spin_unlock(&ctx->lock);
784
785 blk_mq_put_ctx(current_ctx);
786
787 if (run_queue)
788 blk_mq_run_hw_queue(hctx, async);
789}
790
791static void blk_mq_insert_requests(struct request_queue *q,
792 struct blk_mq_ctx *ctx,
793 struct list_head *list,
794 int depth,
795 bool from_schedule)
796
797{
798 struct blk_mq_hw_ctx *hctx;
799 struct blk_mq_ctx *current_ctx;
800
801 trace_block_unplug(q, depth, !from_schedule);
802
803 current_ctx = blk_mq_get_ctx(q);
804
805 if (!cpu_online(ctx->cpu))
806 ctx = current_ctx;
807 hctx = q->mq_ops->map_queue(q, ctx->cpu);
808
809 /*
810 * preemption doesn't flush plug list, so it's possible ctx->cpu is
811 * offline now
812 */
813 spin_lock(&ctx->lock);
814 while (!list_empty(list)) {
815 struct request *rq;
816
817 rq = list_first_entry(list, struct request, queuelist);
818 list_del_init(&rq->queuelist);
819 rq->mq_ctx = ctx;
820 __blk_mq_insert_request(hctx, rq);
821 }
822 spin_unlock(&ctx->lock);
823
824 blk_mq_put_ctx(current_ctx);
825
826 blk_mq_run_hw_queue(hctx, from_schedule);
827}
828
829static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
830{
831 struct request *rqa = container_of(a, struct request, queuelist);
832 struct request *rqb = container_of(b, struct request, queuelist);
833
834 return !(rqa->mq_ctx < rqb->mq_ctx ||
835 (rqa->mq_ctx == rqb->mq_ctx &&
836 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
837}
838
839void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
840{
841 struct blk_mq_ctx *this_ctx;
842 struct request_queue *this_q;
843 struct request *rq;
844 LIST_HEAD(list);
845 LIST_HEAD(ctx_list);
846 unsigned int depth;
847
848 list_splice_init(&plug->mq_list, &list);
849
850 list_sort(NULL, &list, plug_ctx_cmp);
851
852 this_q = NULL;
853 this_ctx = NULL;
854 depth = 0;
855
856 while (!list_empty(&list)) {
857 rq = list_entry_rq(list.next);
858 list_del_init(&rq->queuelist);
859 BUG_ON(!rq->q);
860 if (rq->mq_ctx != this_ctx) {
861 if (this_ctx) {
862 blk_mq_insert_requests(this_q, this_ctx,
863 &ctx_list, depth,
864 from_schedule);
865 }
866
867 this_ctx = rq->mq_ctx;
868 this_q = rq->q;
869 depth = 0;
870 }
871
872 depth++;
873 list_add_tail(&rq->queuelist, &ctx_list);
874 }
875
876 /*
877 * If 'this_ctx' is set, we know we have entries to complete
878 * on 'ctx_list'. Do those.
879 */
880 if (this_ctx) {
881 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
882 from_schedule);
883 }
884}
885
886static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
887{
888 init_request_from_bio(rq, bio);
889 blk_account_io_start(rq, 1);
890}
891
892static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
893{
894 struct blk_mq_hw_ctx *hctx;
895 struct blk_mq_ctx *ctx;
896 const int is_sync = rw_is_sync(bio->bi_rw);
897 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
898 int rw = bio_data_dir(bio);
899 struct request *rq;
900 unsigned int use_plug, request_count = 0;
901
902 /*
903 * If we have multiple hardware queues, just go directly to
904 * one of those for sync IO.
905 */
906 use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
907
908 blk_queue_bounce(q, &bio);
909
910 if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
911 return;
912
913 if (blk_mq_queue_enter(q)) {
914 bio_endio(bio, -EIO);
915 return;
916 }
917
918 ctx = blk_mq_get_ctx(q);
919 hctx = q->mq_ops->map_queue(q, ctx->cpu);
920
921 trace_block_getrq(q, bio, rw);
922 rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
923 if (likely(rq))
924 blk_mq_rq_ctx_init(ctx, rq, rw);
925 else {
926 blk_mq_put_ctx(ctx);
927 trace_block_sleeprq(q, bio, rw);
928 rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
929 false);
930 ctx = rq->mq_ctx;
931 hctx = q->mq_ops->map_queue(q, ctx->cpu);
932 }
933
934 hctx->queued++;
935
936 if (unlikely(is_flush_fua)) {
937 blk_mq_bio_to_request(rq, bio);
938 blk_mq_put_ctx(ctx);
939 blk_insert_flush(rq);
940 goto run_queue;
941 }
942
943 /*
944 * A task plug currently exists. Since this is completely lockless,
945 * utilize that to temporarily store requests until the task is
946 * either done or scheduled away.
947 */
948 if (use_plug) {
949 struct blk_plug *plug = current->plug;
950
951 if (plug) {
952 blk_mq_bio_to_request(rq, bio);
Shaohua Li92f399c2013-10-29 12:01:03 -0600953 if (list_empty(&plug->mq_list))
Jens Axboe320ae512013-10-24 09:20:05 +0100954 trace_block_plug(q);
955 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
956 blk_flush_plug_list(plug, false);
957 trace_block_plug(q);
958 }
959 list_add_tail(&rq->queuelist, &plug->mq_list);
960 blk_mq_put_ctx(ctx);
961 return;
962 }
963 }
964
965 spin_lock(&ctx->lock);
966
967 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
968 blk_mq_attempt_merge(q, ctx, bio))
969 __blk_mq_free_request(hctx, ctx, rq);
970 else {
971 blk_mq_bio_to_request(rq, bio);
972 __blk_mq_insert_request(hctx, rq);
973 }
974
975 spin_unlock(&ctx->lock);
976 blk_mq_put_ctx(ctx);
977
978 /*
979 * For a SYNC request, send it to the hardware immediately. For an
980 * ASYNC request, just ensure that we run it later on. The latter
981 * allows for merging opportunities and more efficient dispatching.
982 */
983run_queue:
984 blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
985}
986
987/*
988 * Default mapping to a software queue, since we use one per CPU.
989 */
990struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
991{
992 return q->queue_hw_ctx[q->mq_map[cpu]];
993}
994EXPORT_SYMBOL(blk_mq_map_queue);
995
996struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_reg *reg,
997 unsigned int hctx_index)
998{
999 return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
1000 GFP_KERNEL | __GFP_ZERO, reg->numa_node);
1001}
1002EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
1003
1004void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
1005 unsigned int hctx_index)
1006{
1007 kfree(hctx);
1008}
1009EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
1010
1011static void blk_mq_hctx_notify(void *data, unsigned long action,
1012 unsigned int cpu)
1013{
1014 struct blk_mq_hw_ctx *hctx = data;
1015 struct blk_mq_ctx *ctx;
1016 LIST_HEAD(tmp);
1017
1018 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1019 return;
1020
1021 /*
1022 * Move ctx entries to new CPU, if this one is going away.
1023 */
1024 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1025
1026 spin_lock(&ctx->lock);
1027 if (!list_empty(&ctx->rq_list)) {
1028 list_splice_init(&ctx->rq_list, &tmp);
1029 clear_bit(ctx->index_hw, hctx->ctx_map);
1030 }
1031 spin_unlock(&ctx->lock);
1032
1033 if (list_empty(&tmp))
1034 return;
1035
1036 ctx = blk_mq_get_ctx(hctx->queue);
1037 spin_lock(&ctx->lock);
1038
1039 while (!list_empty(&tmp)) {
1040 struct request *rq;
1041
1042 rq = list_first_entry(&tmp, struct request, queuelist);
1043 rq->mq_ctx = ctx;
1044 list_move_tail(&rq->queuelist, &ctx->rq_list);
1045 }
1046
1047 blk_mq_hctx_mark_pending(hctx, ctx);
1048
1049 spin_unlock(&ctx->lock);
1050 blk_mq_put_ctx(ctx);
1051}
1052
1053static void blk_mq_init_hw_commands(struct blk_mq_hw_ctx *hctx,
1054 void (*init)(void *, struct blk_mq_hw_ctx *,
1055 struct request *, unsigned int),
1056 void *data)
1057{
1058 unsigned int i;
1059
1060 for (i = 0; i < hctx->queue_depth; i++) {
1061 struct request *rq = hctx->rqs[i];
1062
1063 init(data, hctx, rq, i);
1064 }
1065}
1066
1067void blk_mq_init_commands(struct request_queue *q,
1068 void (*init)(void *, struct blk_mq_hw_ctx *,
1069 struct request *, unsigned int),
1070 void *data)
1071{
1072 struct blk_mq_hw_ctx *hctx;
1073 unsigned int i;
1074
1075 queue_for_each_hw_ctx(q, hctx, i)
1076 blk_mq_init_hw_commands(hctx, init, data);
1077}
1078EXPORT_SYMBOL(blk_mq_init_commands);
1079
1080static void blk_mq_free_rq_map(struct blk_mq_hw_ctx *hctx)
1081{
1082 struct page *page;
1083
1084 while (!list_empty(&hctx->page_list)) {
1085 page = list_first_entry(&hctx->page_list, struct page, list);
1086 list_del_init(&page->list);
1087 __free_pages(page, page->private);
1088 }
1089
1090 kfree(hctx->rqs);
1091
1092 if (hctx->tags)
1093 blk_mq_free_tags(hctx->tags);
1094}
1095
1096static size_t order_to_size(unsigned int order)
1097{
1098 size_t ret = PAGE_SIZE;
1099
1100 while (order--)
1101 ret *= 2;
1102
1103 return ret;
1104}
1105
1106static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
1107 unsigned int reserved_tags, int node)
1108{
1109 unsigned int i, j, entries_per_page, max_order = 4;
1110 size_t rq_size, left;
1111
1112 INIT_LIST_HEAD(&hctx->page_list);
1113
1114 hctx->rqs = kmalloc_node(hctx->queue_depth * sizeof(struct request *),
1115 GFP_KERNEL, node);
1116 if (!hctx->rqs)
1117 return -ENOMEM;
1118
1119 /*
1120 * rq_size is the size of the request plus driver payload, rounded
1121 * to the cacheline size
1122 */
1123 rq_size = round_up(sizeof(struct request) + hctx->cmd_size,
1124 cache_line_size());
1125 left = rq_size * hctx->queue_depth;
1126
1127 for (i = 0; i < hctx->queue_depth;) {
1128 int this_order = max_order;
1129 struct page *page;
1130 int to_do;
1131 void *p;
1132
1133 while (left < order_to_size(this_order - 1) && this_order)
1134 this_order--;
1135
1136 do {
1137 page = alloc_pages_node(node, GFP_KERNEL, this_order);
1138 if (page)
1139 break;
1140 if (!this_order--)
1141 break;
1142 if (order_to_size(this_order) < rq_size)
1143 break;
1144 } while (1);
1145
1146 if (!page)
1147 break;
1148
1149 page->private = this_order;
1150 list_add_tail(&page->list, &hctx->page_list);
1151
1152 p = page_address(page);
1153 entries_per_page = order_to_size(this_order) / rq_size;
1154 to_do = min(entries_per_page, hctx->queue_depth - i);
1155 left -= to_do * rq_size;
1156 for (j = 0; j < to_do; j++) {
1157 hctx->rqs[i] = p;
1158 blk_mq_rq_init(hctx, hctx->rqs[i]);
1159 p += rq_size;
1160 i++;
1161 }
1162 }
1163
1164 if (i < (reserved_tags + BLK_MQ_TAG_MIN))
1165 goto err_rq_map;
1166 else if (i != hctx->queue_depth) {
1167 hctx->queue_depth = i;
1168 pr_warn("%s: queue depth set to %u because of low memory\n",
1169 __func__, i);
1170 }
1171
1172 hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node);
1173 if (!hctx->tags) {
1174err_rq_map:
1175 blk_mq_free_rq_map(hctx);
1176 return -ENOMEM;
1177 }
1178
1179 return 0;
1180}
1181
1182static int blk_mq_init_hw_queues(struct request_queue *q,
1183 struct blk_mq_reg *reg, void *driver_data)
1184{
1185 struct blk_mq_hw_ctx *hctx;
1186 unsigned int i, j;
1187
1188 /*
1189 * Initialize hardware queues
1190 */
1191 queue_for_each_hw_ctx(q, hctx, i) {
1192 unsigned int num_maps;
1193 int node;
1194
1195 node = hctx->numa_node;
1196 if (node == NUMA_NO_NODE)
1197 node = hctx->numa_node = reg->numa_node;
1198
1199 INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn);
1200 spin_lock_init(&hctx->lock);
1201 INIT_LIST_HEAD(&hctx->dispatch);
1202 hctx->queue = q;
1203 hctx->queue_num = i;
1204 hctx->flags = reg->flags;
1205 hctx->queue_depth = reg->queue_depth;
1206 hctx->cmd_size = reg->cmd_size;
1207
1208 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1209 blk_mq_hctx_notify, hctx);
1210 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1211
1212 if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node))
1213 break;
1214
1215 /*
1216 * Allocate space for all possible cpus to avoid allocation in
1217 * runtime
1218 */
1219 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1220 GFP_KERNEL, node);
1221 if (!hctx->ctxs)
1222 break;
1223
1224 num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1225 hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1226 GFP_KERNEL, node);
1227 if (!hctx->ctx_map)
1228 break;
1229
1230 hctx->nr_ctx_map = num_maps;
1231 hctx->nr_ctx = 0;
1232
1233 if (reg->ops->init_hctx &&
1234 reg->ops->init_hctx(hctx, driver_data, i))
1235 break;
1236 }
1237
1238 if (i == q->nr_hw_queues)
1239 return 0;
1240
1241 /*
1242 * Init failed
1243 */
1244 queue_for_each_hw_ctx(q, hctx, j) {
1245 if (i == j)
1246 break;
1247
1248 if (reg->ops->exit_hctx)
1249 reg->ops->exit_hctx(hctx, j);
1250
1251 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1252 blk_mq_free_rq_map(hctx);
1253 kfree(hctx->ctxs);
1254 }
1255
1256 return 1;
1257}
1258
1259static void blk_mq_init_cpu_queues(struct request_queue *q,
1260 unsigned int nr_hw_queues)
1261{
1262 unsigned int i;
1263
1264 for_each_possible_cpu(i) {
1265 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1266 struct blk_mq_hw_ctx *hctx;
1267
1268 memset(__ctx, 0, sizeof(*__ctx));
1269 __ctx->cpu = i;
1270 spin_lock_init(&__ctx->lock);
1271 INIT_LIST_HEAD(&__ctx->rq_list);
1272 __ctx->queue = q;
1273
1274 /* If the cpu isn't online, the cpu is mapped to first hctx */
1275 hctx = q->mq_ops->map_queue(q, i);
1276 hctx->nr_ctx++;
1277
1278 if (!cpu_online(i))
1279 continue;
1280
1281 /*
1282 * Set local node, IFF we have more than one hw queue. If
1283 * not, we remain on the home node of the device
1284 */
1285 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1286 hctx->numa_node = cpu_to_node(i);
1287 }
1288}
1289
1290static void blk_mq_map_swqueue(struct request_queue *q)
1291{
1292 unsigned int i;
1293 struct blk_mq_hw_ctx *hctx;
1294 struct blk_mq_ctx *ctx;
1295
1296 queue_for_each_hw_ctx(q, hctx, i) {
1297 hctx->nr_ctx = 0;
1298 }
1299
1300 /*
1301 * Map software to hardware queues
1302 */
1303 queue_for_each_ctx(q, ctx, i) {
1304 /* If the cpu isn't online, the cpu is mapped to first hctx */
1305 hctx = q->mq_ops->map_queue(q, i);
1306 ctx->index_hw = hctx->nr_ctx;
1307 hctx->ctxs[hctx->nr_ctx++] = ctx;
1308 }
1309}
1310
1311struct request_queue *blk_mq_init_queue(struct blk_mq_reg *reg,
1312 void *driver_data)
1313{
1314 struct blk_mq_hw_ctx **hctxs;
1315 struct blk_mq_ctx *ctx;
1316 struct request_queue *q;
1317 int i;
1318
1319 if (!reg->nr_hw_queues ||
1320 !reg->ops->queue_rq || !reg->ops->map_queue ||
1321 !reg->ops->alloc_hctx || !reg->ops->free_hctx)
1322 return ERR_PTR(-EINVAL);
1323
1324 if (!reg->queue_depth)
1325 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1326 else if (reg->queue_depth > BLK_MQ_MAX_DEPTH) {
1327 pr_err("blk-mq: queuedepth too large (%u)\n", reg->queue_depth);
1328 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1329 }
1330
Christoph Hellwig3228f482013-10-28 13:33:58 -06001331 /*
1332 * Set aside a tag for flush requests. It will only be used while
1333 * another flush request is in progress but outside the driver.
1334 *
1335 * TODO: only allocate if flushes are supported
1336 */
1337 reg->queue_depth++;
1338 reg->reserved_tags++;
1339
Jens Axboe320ae512013-10-24 09:20:05 +01001340 if (reg->queue_depth < (reg->reserved_tags + BLK_MQ_TAG_MIN))
1341 return ERR_PTR(-EINVAL);
1342
1343 ctx = alloc_percpu(struct blk_mq_ctx);
1344 if (!ctx)
1345 return ERR_PTR(-ENOMEM);
1346
1347 hctxs = kmalloc_node(reg->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1348 reg->numa_node);
1349
1350 if (!hctxs)
1351 goto err_percpu;
1352
1353 for (i = 0; i < reg->nr_hw_queues; i++) {
1354 hctxs[i] = reg->ops->alloc_hctx(reg, i);
1355 if (!hctxs[i])
1356 goto err_hctxs;
1357
1358 hctxs[i]->numa_node = NUMA_NO_NODE;
1359 hctxs[i]->queue_num = i;
1360 }
1361
1362 q = blk_alloc_queue_node(GFP_KERNEL, reg->numa_node);
1363 if (!q)
1364 goto err_hctxs;
1365
1366 q->mq_map = blk_mq_make_queue_map(reg);
1367 if (!q->mq_map)
1368 goto err_map;
1369
1370 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1371 blk_queue_rq_timeout(q, 30000);
1372
1373 q->nr_queues = nr_cpu_ids;
1374 q->nr_hw_queues = reg->nr_hw_queues;
1375
1376 q->queue_ctx = ctx;
1377 q->queue_hw_ctx = hctxs;
1378
1379 q->mq_ops = reg->ops;
1380
1381 blk_queue_make_request(q, blk_mq_make_request);
1382 blk_queue_rq_timed_out(q, reg->ops->timeout);
1383 if (reg->timeout)
1384 blk_queue_rq_timeout(q, reg->timeout);
1385
1386 blk_mq_init_flush(q);
1387 blk_mq_init_cpu_queues(q, reg->nr_hw_queues);
1388
1389 if (blk_mq_init_hw_queues(q, reg, driver_data))
1390 goto err_hw;
1391
1392 blk_mq_map_swqueue(q);
1393
1394 mutex_lock(&all_q_mutex);
1395 list_add_tail(&q->all_q_node, &all_q_list);
1396 mutex_unlock(&all_q_mutex);
1397
1398 return q;
1399err_hw:
1400 kfree(q->mq_map);
1401err_map:
1402 blk_cleanup_queue(q);
1403err_hctxs:
1404 for (i = 0; i < reg->nr_hw_queues; i++) {
1405 if (!hctxs[i])
1406 break;
1407 reg->ops->free_hctx(hctxs[i], i);
1408 }
1409 kfree(hctxs);
1410err_percpu:
1411 free_percpu(ctx);
1412 return ERR_PTR(-ENOMEM);
1413}
1414EXPORT_SYMBOL(blk_mq_init_queue);
1415
1416void blk_mq_free_queue(struct request_queue *q)
1417{
1418 struct blk_mq_hw_ctx *hctx;
1419 int i;
1420
1421 queue_for_each_hw_ctx(q, hctx, i) {
1422 cancel_delayed_work_sync(&hctx->delayed_work);
1423 kfree(hctx->ctx_map);
1424 kfree(hctx->ctxs);
1425 blk_mq_free_rq_map(hctx);
1426 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1427 if (q->mq_ops->exit_hctx)
1428 q->mq_ops->exit_hctx(hctx, i);
1429 q->mq_ops->free_hctx(hctx, i);
1430 }
1431
1432 free_percpu(q->queue_ctx);
1433 kfree(q->queue_hw_ctx);
1434 kfree(q->mq_map);
1435
1436 q->queue_ctx = NULL;
1437 q->queue_hw_ctx = NULL;
1438 q->mq_map = NULL;
1439
1440 mutex_lock(&all_q_mutex);
1441 list_del_init(&q->all_q_node);
1442 mutex_unlock(&all_q_mutex);
1443}
1444EXPORT_SYMBOL(blk_mq_free_queue);
1445
1446/* Basically redo blk_mq_init_queue with queue frozen */
1447static void __cpuinit blk_mq_queue_reinit(struct request_queue *q)
1448{
1449 blk_mq_freeze_queue(q);
1450
1451 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1452
1453 /*
1454 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1455 * we should change hctx numa_node according to new topology (this
1456 * involves free and re-allocate memory, worthy doing?)
1457 */
1458
1459 blk_mq_map_swqueue(q);
1460
1461 blk_mq_unfreeze_queue(q);
1462}
1463
1464static int __cpuinit blk_mq_queue_reinit_notify(struct notifier_block *nb,
1465 unsigned long action, void *hcpu)
1466{
1467 struct request_queue *q;
1468
1469 /*
1470 * Before new mapping is established, hotadded cpu might already start
1471 * handling requests. This doesn't break anything as we map offline
1472 * CPUs to first hardware queue. We will re-init queue below to get
1473 * optimal settings.
1474 */
1475 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1476 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1477 return NOTIFY_OK;
1478
1479 mutex_lock(&all_q_mutex);
1480 list_for_each_entry(q, &all_q_list, all_q_node)
1481 blk_mq_queue_reinit(q);
1482 mutex_unlock(&all_q_mutex);
1483 return NOTIFY_OK;
1484}
1485
1486static int __init blk_mq_init(void)
1487{
1488 unsigned int i;
1489
1490 for_each_possible_cpu(i)
1491 init_llist_head(&per_cpu(ipi_lists, i));
1492
1493 blk_mq_cpu_init();
1494
1495 /* Must be called after percpu_counter_hotcpu_callback() */
1496 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1497
1498 return 0;
1499}
1500subsys_initcall(blk_mq_init);