]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - block/blk-mq.c
Merge tag 'tty-4.4-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/tty
[mirror_ubuntu-artful-kernel.git] / block / blk-mq.c
1 /*
2 * Block multiqueue core code
3 *
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
6 */
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
13 #include <linux/mm.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/delay.h>
24 #include <linux/crash_dump.h>
25
26 #include <trace/events/block.h>
27
28 #include <linux/blk-mq.h>
29 #include "blk.h"
30 #include "blk-mq.h"
31 #include "blk-mq-tag.h"
32
33 static DEFINE_MUTEX(all_q_mutex);
34 static LIST_HEAD(all_q_list);
35
36 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
37
38 /*
39 * Check if any of the ctx's have pending work in this hardware queue
40 */
41 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
42 {
43 unsigned int i;
44
45 for (i = 0; i < hctx->ctx_map.size; i++)
46 if (hctx->ctx_map.map[i].word)
47 return true;
48
49 return false;
50 }
51
52 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
53 struct blk_mq_ctx *ctx)
54 {
55 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
56 }
57
58 #define CTX_TO_BIT(hctx, ctx) \
59 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
60
61 /*
62 * Mark this ctx as having pending work in this hardware queue
63 */
64 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
65 struct blk_mq_ctx *ctx)
66 {
67 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
68
69 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
70 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
71 }
72
73 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
74 struct blk_mq_ctx *ctx)
75 {
76 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
77
78 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
79 }
80
81 void blk_mq_freeze_queue_start(struct request_queue *q)
82 {
83 int freeze_depth;
84
85 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
86 if (freeze_depth == 1) {
87 percpu_ref_kill(&q->q_usage_counter);
88 blk_mq_run_hw_queues(q, false);
89 }
90 }
91 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
92
93 static void blk_mq_freeze_queue_wait(struct request_queue *q)
94 {
95 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
96 }
97
98 /*
99 * Guarantee no request is in use, so we can change any data structure of
100 * the queue afterward.
101 */
102 void blk_freeze_queue(struct request_queue *q)
103 {
104 /*
105 * In the !blk_mq case we are only calling this to kill the
106 * q_usage_counter, otherwise this increases the freeze depth
107 * and waits for it to return to zero. For this reason there is
108 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
109 * exported to drivers as the only user for unfreeze is blk_mq.
110 */
111 blk_mq_freeze_queue_start(q);
112 blk_mq_freeze_queue_wait(q);
113 }
114
115 void blk_mq_freeze_queue(struct request_queue *q)
116 {
117 /*
118 * ...just an alias to keep freeze and unfreeze actions balanced
119 * in the blk_mq_* namespace
120 */
121 blk_freeze_queue(q);
122 }
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
124
125 void blk_mq_unfreeze_queue(struct request_queue *q)
126 {
127 int freeze_depth;
128
129 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
130 WARN_ON_ONCE(freeze_depth < 0);
131 if (!freeze_depth) {
132 percpu_ref_reinit(&q->q_usage_counter);
133 wake_up_all(&q->mq_freeze_wq);
134 }
135 }
136 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
137
138 void blk_mq_wake_waiters(struct request_queue *q)
139 {
140 struct blk_mq_hw_ctx *hctx;
141 unsigned int i;
142
143 queue_for_each_hw_ctx(q, hctx, i)
144 if (blk_mq_hw_queue_mapped(hctx))
145 blk_mq_tag_wakeup_all(hctx->tags, true);
146
147 /*
148 * If we are called because the queue has now been marked as
149 * dying, we need to ensure that processes currently waiting on
150 * the queue are notified as well.
151 */
152 wake_up_all(&q->mq_freeze_wq);
153 }
154
155 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
156 {
157 return blk_mq_has_free_tags(hctx->tags);
158 }
159 EXPORT_SYMBOL(blk_mq_can_queue);
160
161 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
162 struct request *rq, unsigned int rw_flags)
163 {
164 if (blk_queue_io_stat(q))
165 rw_flags |= REQ_IO_STAT;
166
167 INIT_LIST_HEAD(&rq->queuelist);
168 /* csd/requeue_work/fifo_time is initialized before use */
169 rq->q = q;
170 rq->mq_ctx = ctx;
171 rq->cmd_flags |= rw_flags;
172 /* do not touch atomic flags, it needs atomic ops against the timer */
173 rq->cpu = -1;
174 INIT_HLIST_NODE(&rq->hash);
175 RB_CLEAR_NODE(&rq->rb_node);
176 rq->rq_disk = NULL;
177 rq->part = NULL;
178 rq->start_time = jiffies;
179 #ifdef CONFIG_BLK_CGROUP
180 rq->rl = NULL;
181 set_start_time_ns(rq);
182 rq->io_start_time_ns = 0;
183 #endif
184 rq->nr_phys_segments = 0;
185 #if defined(CONFIG_BLK_DEV_INTEGRITY)
186 rq->nr_integrity_segments = 0;
187 #endif
188 rq->special = NULL;
189 /* tag was already set */
190 rq->errors = 0;
191
192 rq->cmd = rq->__cmd;
193
194 rq->extra_len = 0;
195 rq->sense_len = 0;
196 rq->resid_len = 0;
197 rq->sense = NULL;
198
199 INIT_LIST_HEAD(&rq->timeout_list);
200 rq->timeout = 0;
201
202 rq->end_io = NULL;
203 rq->end_io_data = NULL;
204 rq->next_rq = NULL;
205
206 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
207 }
208
209 static struct request *
210 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
211 {
212 struct request *rq;
213 unsigned int tag;
214
215 tag = blk_mq_get_tag(data);
216 if (tag != BLK_MQ_TAG_FAIL) {
217 rq = data->hctx->tags->rqs[tag];
218
219 if (blk_mq_tag_busy(data->hctx)) {
220 rq->cmd_flags = REQ_MQ_INFLIGHT;
221 atomic_inc(&data->hctx->nr_active);
222 }
223
224 rq->tag = tag;
225 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
226 return rq;
227 }
228
229 return NULL;
230 }
231
232 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
233 bool reserved)
234 {
235 struct blk_mq_ctx *ctx;
236 struct blk_mq_hw_ctx *hctx;
237 struct request *rq;
238 struct blk_mq_alloc_data alloc_data;
239 int ret;
240
241 ret = blk_queue_enter(q, gfp);
242 if (ret)
243 return ERR_PTR(ret);
244
245 ctx = blk_mq_get_ctx(q);
246 hctx = q->mq_ops->map_queue(q, ctx->cpu);
247 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
248 reserved, ctx, hctx);
249
250 rq = __blk_mq_alloc_request(&alloc_data, rw);
251 if (!rq && (gfp & __GFP_WAIT)) {
252 __blk_mq_run_hw_queue(hctx);
253 blk_mq_put_ctx(ctx);
254
255 ctx = blk_mq_get_ctx(q);
256 hctx = q->mq_ops->map_queue(q, ctx->cpu);
257 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
258 hctx);
259 rq = __blk_mq_alloc_request(&alloc_data, rw);
260 ctx = alloc_data.ctx;
261 }
262 blk_mq_put_ctx(ctx);
263 if (!rq) {
264 blk_queue_exit(q);
265 return ERR_PTR(-EWOULDBLOCK);
266 }
267 return rq;
268 }
269 EXPORT_SYMBOL(blk_mq_alloc_request);
270
271 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
272 struct blk_mq_ctx *ctx, struct request *rq)
273 {
274 const int tag = rq->tag;
275 struct request_queue *q = rq->q;
276
277 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
278 atomic_dec(&hctx->nr_active);
279 rq->cmd_flags = 0;
280
281 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
282 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
283 blk_queue_exit(q);
284 }
285
286 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
287 {
288 struct blk_mq_ctx *ctx = rq->mq_ctx;
289
290 ctx->rq_completed[rq_is_sync(rq)]++;
291 __blk_mq_free_request(hctx, ctx, rq);
292
293 }
294 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
295
296 void blk_mq_free_request(struct request *rq)
297 {
298 struct blk_mq_hw_ctx *hctx;
299 struct request_queue *q = rq->q;
300
301 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
302 blk_mq_free_hctx_request(hctx, rq);
303 }
304 EXPORT_SYMBOL_GPL(blk_mq_free_request);
305
306 inline void __blk_mq_end_request(struct request *rq, int error)
307 {
308 blk_account_io_done(rq);
309
310 if (rq->end_io) {
311 rq->end_io(rq, error);
312 } else {
313 if (unlikely(blk_bidi_rq(rq)))
314 blk_mq_free_request(rq->next_rq);
315 blk_mq_free_request(rq);
316 }
317 }
318 EXPORT_SYMBOL(__blk_mq_end_request);
319
320 void blk_mq_end_request(struct request *rq, int error)
321 {
322 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
323 BUG();
324 __blk_mq_end_request(rq, error);
325 }
326 EXPORT_SYMBOL(blk_mq_end_request);
327
328 static void __blk_mq_complete_request_remote(void *data)
329 {
330 struct request *rq = data;
331
332 rq->q->softirq_done_fn(rq);
333 }
334
335 static void blk_mq_ipi_complete_request(struct request *rq)
336 {
337 struct blk_mq_ctx *ctx = rq->mq_ctx;
338 bool shared = false;
339 int cpu;
340
341 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
342 rq->q->softirq_done_fn(rq);
343 return;
344 }
345
346 cpu = get_cpu();
347 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
348 shared = cpus_share_cache(cpu, ctx->cpu);
349
350 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
351 rq->csd.func = __blk_mq_complete_request_remote;
352 rq->csd.info = rq;
353 rq->csd.flags = 0;
354 smp_call_function_single_async(ctx->cpu, &rq->csd);
355 } else {
356 rq->q->softirq_done_fn(rq);
357 }
358 put_cpu();
359 }
360
361 void __blk_mq_complete_request(struct request *rq)
362 {
363 struct request_queue *q = rq->q;
364
365 if (!q->softirq_done_fn)
366 blk_mq_end_request(rq, rq->errors);
367 else
368 blk_mq_ipi_complete_request(rq);
369 }
370
371 /**
372 * blk_mq_complete_request - end I/O on a request
373 * @rq: the request being processed
374 *
375 * Description:
376 * Ends all I/O on a request. It does not handle partial completions.
377 * The actual completion happens out-of-order, through a IPI handler.
378 **/
379 void blk_mq_complete_request(struct request *rq, int error)
380 {
381 struct request_queue *q = rq->q;
382
383 if (unlikely(blk_should_fake_timeout(q)))
384 return;
385 if (!blk_mark_rq_complete(rq)) {
386 rq->errors = error;
387 __blk_mq_complete_request(rq);
388 }
389 }
390 EXPORT_SYMBOL(blk_mq_complete_request);
391
392 int blk_mq_request_started(struct request *rq)
393 {
394 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
395 }
396 EXPORT_SYMBOL_GPL(blk_mq_request_started);
397
398 void blk_mq_start_request(struct request *rq)
399 {
400 struct request_queue *q = rq->q;
401
402 trace_block_rq_issue(q, rq);
403
404 rq->resid_len = blk_rq_bytes(rq);
405 if (unlikely(blk_bidi_rq(rq)))
406 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
407
408 blk_add_timer(rq);
409
410 /*
411 * Ensure that ->deadline is visible before set the started
412 * flag and clear the completed flag.
413 */
414 smp_mb__before_atomic();
415
416 /*
417 * Mark us as started and clear complete. Complete might have been
418 * set if requeue raced with timeout, which then marked it as
419 * complete. So be sure to clear complete again when we start
420 * the request, otherwise we'll ignore the completion event.
421 */
422 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
423 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
424 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
425 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
426
427 if (q->dma_drain_size && blk_rq_bytes(rq)) {
428 /*
429 * Make sure space for the drain appears. We know we can do
430 * this because max_hw_segments has been adjusted to be one
431 * fewer than the device can handle.
432 */
433 rq->nr_phys_segments++;
434 }
435 }
436 EXPORT_SYMBOL(blk_mq_start_request);
437
438 static void __blk_mq_requeue_request(struct request *rq)
439 {
440 struct request_queue *q = rq->q;
441
442 trace_block_rq_requeue(q, rq);
443
444 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
445 if (q->dma_drain_size && blk_rq_bytes(rq))
446 rq->nr_phys_segments--;
447 }
448 }
449
450 void blk_mq_requeue_request(struct request *rq)
451 {
452 __blk_mq_requeue_request(rq);
453
454 BUG_ON(blk_queued_rq(rq));
455 blk_mq_add_to_requeue_list(rq, true);
456 }
457 EXPORT_SYMBOL(blk_mq_requeue_request);
458
459 static void blk_mq_requeue_work(struct work_struct *work)
460 {
461 struct request_queue *q =
462 container_of(work, struct request_queue, requeue_work);
463 LIST_HEAD(rq_list);
464 struct request *rq, *next;
465 unsigned long flags;
466
467 spin_lock_irqsave(&q->requeue_lock, flags);
468 list_splice_init(&q->requeue_list, &rq_list);
469 spin_unlock_irqrestore(&q->requeue_lock, flags);
470
471 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
472 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
473 continue;
474
475 rq->cmd_flags &= ~REQ_SOFTBARRIER;
476 list_del_init(&rq->queuelist);
477 blk_mq_insert_request(rq, true, false, false);
478 }
479
480 while (!list_empty(&rq_list)) {
481 rq = list_entry(rq_list.next, struct request, queuelist);
482 list_del_init(&rq->queuelist);
483 blk_mq_insert_request(rq, false, false, false);
484 }
485
486 /*
487 * Use the start variant of queue running here, so that running
488 * the requeue work will kick stopped queues.
489 */
490 blk_mq_start_hw_queues(q);
491 }
492
493 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
494 {
495 struct request_queue *q = rq->q;
496 unsigned long flags;
497
498 /*
499 * We abuse this flag that is otherwise used by the I/O scheduler to
500 * request head insertation from the workqueue.
501 */
502 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
503
504 spin_lock_irqsave(&q->requeue_lock, flags);
505 if (at_head) {
506 rq->cmd_flags |= REQ_SOFTBARRIER;
507 list_add(&rq->queuelist, &q->requeue_list);
508 } else {
509 list_add_tail(&rq->queuelist, &q->requeue_list);
510 }
511 spin_unlock_irqrestore(&q->requeue_lock, flags);
512 }
513 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
514
515 void blk_mq_cancel_requeue_work(struct request_queue *q)
516 {
517 cancel_work_sync(&q->requeue_work);
518 }
519 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
520
521 void blk_mq_kick_requeue_list(struct request_queue *q)
522 {
523 kblockd_schedule_work(&q->requeue_work);
524 }
525 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
526
527 void blk_mq_abort_requeue_list(struct request_queue *q)
528 {
529 unsigned long flags;
530 LIST_HEAD(rq_list);
531
532 spin_lock_irqsave(&q->requeue_lock, flags);
533 list_splice_init(&q->requeue_list, &rq_list);
534 spin_unlock_irqrestore(&q->requeue_lock, flags);
535
536 while (!list_empty(&rq_list)) {
537 struct request *rq;
538
539 rq = list_first_entry(&rq_list, struct request, queuelist);
540 list_del_init(&rq->queuelist);
541 rq->errors = -EIO;
542 blk_mq_end_request(rq, rq->errors);
543 }
544 }
545 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
546
547 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
548 {
549 return tags->rqs[tag];
550 }
551 EXPORT_SYMBOL(blk_mq_tag_to_rq);
552
553 struct blk_mq_timeout_data {
554 unsigned long next;
555 unsigned int next_set;
556 };
557
558 void blk_mq_rq_timed_out(struct request *req, bool reserved)
559 {
560 struct blk_mq_ops *ops = req->q->mq_ops;
561 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
562
563 /*
564 * We know that complete is set at this point. If STARTED isn't set
565 * anymore, then the request isn't active and the "timeout" should
566 * just be ignored. This can happen due to the bitflag ordering.
567 * Timeout first checks if STARTED is set, and if it is, assumes
568 * the request is active. But if we race with completion, then
569 * we both flags will get cleared. So check here again, and ignore
570 * a timeout event with a request that isn't active.
571 */
572 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
573 return;
574
575 if (ops->timeout)
576 ret = ops->timeout(req, reserved);
577
578 switch (ret) {
579 case BLK_EH_HANDLED:
580 __blk_mq_complete_request(req);
581 break;
582 case BLK_EH_RESET_TIMER:
583 blk_add_timer(req);
584 blk_clear_rq_complete(req);
585 break;
586 case BLK_EH_NOT_HANDLED:
587 break;
588 default:
589 printk(KERN_ERR "block: bad eh return: %d\n", ret);
590 break;
591 }
592 }
593
594 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
595 struct request *rq, void *priv, bool reserved)
596 {
597 struct blk_mq_timeout_data *data = priv;
598
599 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
600 /*
601 * If a request wasn't started before the queue was
602 * marked dying, kill it here or it'll go unnoticed.
603 */
604 if (unlikely(blk_queue_dying(rq->q)))
605 blk_mq_complete_request(rq, -EIO);
606 return;
607 }
608 if (rq->cmd_flags & REQ_NO_TIMEOUT)
609 return;
610
611 if (time_after_eq(jiffies, rq->deadline)) {
612 if (!blk_mark_rq_complete(rq))
613 blk_mq_rq_timed_out(rq, reserved);
614 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
615 data->next = rq->deadline;
616 data->next_set = 1;
617 }
618 }
619
620 static void blk_mq_rq_timer(unsigned long priv)
621 {
622 struct request_queue *q = (struct request_queue *)priv;
623 struct blk_mq_timeout_data data = {
624 .next = 0,
625 .next_set = 0,
626 };
627 int i;
628
629 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
630
631 if (data.next_set) {
632 data.next = blk_rq_timeout(round_jiffies_up(data.next));
633 mod_timer(&q->timeout, data.next);
634 } else {
635 struct blk_mq_hw_ctx *hctx;
636
637 queue_for_each_hw_ctx(q, hctx, i) {
638 /* the hctx may be unmapped, so check it here */
639 if (blk_mq_hw_queue_mapped(hctx))
640 blk_mq_tag_idle(hctx);
641 }
642 }
643 }
644
645 /*
646 * Reverse check our software queue for entries that we could potentially
647 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
648 * too much time checking for merges.
649 */
650 static bool blk_mq_attempt_merge(struct request_queue *q,
651 struct blk_mq_ctx *ctx, struct bio *bio)
652 {
653 struct request *rq;
654 int checked = 8;
655
656 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
657 int el_ret;
658
659 if (!checked--)
660 break;
661
662 if (!blk_rq_merge_ok(rq, bio))
663 continue;
664
665 el_ret = blk_try_merge(rq, bio);
666 if (el_ret == ELEVATOR_BACK_MERGE) {
667 if (bio_attempt_back_merge(q, rq, bio)) {
668 ctx->rq_merged++;
669 return true;
670 }
671 break;
672 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
673 if (bio_attempt_front_merge(q, rq, bio)) {
674 ctx->rq_merged++;
675 return true;
676 }
677 break;
678 }
679 }
680
681 return false;
682 }
683
684 /*
685 * Process software queues that have been marked busy, splicing them
686 * to the for-dispatch
687 */
688 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
689 {
690 struct blk_mq_ctx *ctx;
691 int i;
692
693 for (i = 0; i < hctx->ctx_map.size; i++) {
694 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
695 unsigned int off, bit;
696
697 if (!bm->word)
698 continue;
699
700 bit = 0;
701 off = i * hctx->ctx_map.bits_per_word;
702 do {
703 bit = find_next_bit(&bm->word, bm->depth, bit);
704 if (bit >= bm->depth)
705 break;
706
707 ctx = hctx->ctxs[bit + off];
708 clear_bit(bit, &bm->word);
709 spin_lock(&ctx->lock);
710 list_splice_tail_init(&ctx->rq_list, list);
711 spin_unlock(&ctx->lock);
712
713 bit++;
714 } while (1);
715 }
716 }
717
718 /*
719 * Run this hardware queue, pulling any software queues mapped to it in.
720 * Note that this function currently has various problems around ordering
721 * of IO. In particular, we'd like FIFO behaviour on handling existing
722 * items on the hctx->dispatch list. Ignore that for now.
723 */
724 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
725 {
726 struct request_queue *q = hctx->queue;
727 struct request *rq;
728 LIST_HEAD(rq_list);
729 LIST_HEAD(driver_list);
730 struct list_head *dptr;
731 int queued;
732
733 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
734
735 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
736 return;
737
738 hctx->run++;
739
740 /*
741 * Touch any software queue that has pending entries.
742 */
743 flush_busy_ctxs(hctx, &rq_list);
744
745 /*
746 * If we have previous entries on our dispatch list, grab them
747 * and stuff them at the front for more fair dispatch.
748 */
749 if (!list_empty_careful(&hctx->dispatch)) {
750 spin_lock(&hctx->lock);
751 if (!list_empty(&hctx->dispatch))
752 list_splice_init(&hctx->dispatch, &rq_list);
753 spin_unlock(&hctx->lock);
754 }
755
756 /*
757 * Start off with dptr being NULL, so we start the first request
758 * immediately, even if we have more pending.
759 */
760 dptr = NULL;
761
762 /*
763 * Now process all the entries, sending them to the driver.
764 */
765 queued = 0;
766 while (!list_empty(&rq_list)) {
767 struct blk_mq_queue_data bd;
768 int ret;
769
770 rq = list_first_entry(&rq_list, struct request, queuelist);
771 list_del_init(&rq->queuelist);
772
773 bd.rq = rq;
774 bd.list = dptr;
775 bd.last = list_empty(&rq_list);
776
777 ret = q->mq_ops->queue_rq(hctx, &bd);
778 switch (ret) {
779 case BLK_MQ_RQ_QUEUE_OK:
780 queued++;
781 continue;
782 case BLK_MQ_RQ_QUEUE_BUSY:
783 list_add(&rq->queuelist, &rq_list);
784 __blk_mq_requeue_request(rq);
785 break;
786 default:
787 pr_err("blk-mq: bad return on queue: %d\n", ret);
788 case BLK_MQ_RQ_QUEUE_ERROR:
789 rq->errors = -EIO;
790 blk_mq_end_request(rq, rq->errors);
791 break;
792 }
793
794 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
795 break;
796
797 /*
798 * We've done the first request. If we have more than 1
799 * left in the list, set dptr to defer issue.
800 */
801 if (!dptr && rq_list.next != rq_list.prev)
802 dptr = &driver_list;
803 }
804
805 if (!queued)
806 hctx->dispatched[0]++;
807 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
808 hctx->dispatched[ilog2(queued) + 1]++;
809
810 /*
811 * Any items that need requeuing? Stuff them into hctx->dispatch,
812 * that is where we will continue on next queue run.
813 */
814 if (!list_empty(&rq_list)) {
815 spin_lock(&hctx->lock);
816 list_splice(&rq_list, &hctx->dispatch);
817 spin_unlock(&hctx->lock);
818 /*
819 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
820 * it's possible the queue is stopped and restarted again
821 * before this. Queue restart will dispatch requests. And since
822 * requests in rq_list aren't added into hctx->dispatch yet,
823 * the requests in rq_list might get lost.
824 *
825 * blk_mq_run_hw_queue() already checks the STOPPED bit
826 **/
827 blk_mq_run_hw_queue(hctx, true);
828 }
829 }
830
831 /*
832 * It'd be great if the workqueue API had a way to pass
833 * in a mask and had some smarts for more clever placement.
834 * For now we just round-robin here, switching for every
835 * BLK_MQ_CPU_WORK_BATCH queued items.
836 */
837 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
838 {
839 if (hctx->queue->nr_hw_queues == 1)
840 return WORK_CPU_UNBOUND;
841
842 if (--hctx->next_cpu_batch <= 0) {
843 int cpu = hctx->next_cpu, next_cpu;
844
845 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
846 if (next_cpu >= nr_cpu_ids)
847 next_cpu = cpumask_first(hctx->cpumask);
848
849 hctx->next_cpu = next_cpu;
850 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
851
852 return cpu;
853 }
854
855 return hctx->next_cpu;
856 }
857
858 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
859 {
860 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
861 !blk_mq_hw_queue_mapped(hctx)))
862 return;
863
864 if (!async) {
865 int cpu = get_cpu();
866 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
867 __blk_mq_run_hw_queue(hctx);
868 put_cpu();
869 return;
870 }
871
872 put_cpu();
873 }
874
875 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
876 &hctx->run_work, 0);
877 }
878
879 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
880 {
881 struct blk_mq_hw_ctx *hctx;
882 int i;
883
884 queue_for_each_hw_ctx(q, hctx, i) {
885 if ((!blk_mq_hctx_has_pending(hctx) &&
886 list_empty_careful(&hctx->dispatch)) ||
887 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
888 continue;
889
890 blk_mq_run_hw_queue(hctx, async);
891 }
892 }
893 EXPORT_SYMBOL(blk_mq_run_hw_queues);
894
895 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
896 {
897 cancel_delayed_work(&hctx->run_work);
898 cancel_delayed_work(&hctx->delay_work);
899 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
900 }
901 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
902
903 void blk_mq_stop_hw_queues(struct request_queue *q)
904 {
905 struct blk_mq_hw_ctx *hctx;
906 int i;
907
908 queue_for_each_hw_ctx(q, hctx, i)
909 blk_mq_stop_hw_queue(hctx);
910 }
911 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
912
913 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
914 {
915 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
916
917 blk_mq_run_hw_queue(hctx, false);
918 }
919 EXPORT_SYMBOL(blk_mq_start_hw_queue);
920
921 void blk_mq_start_hw_queues(struct request_queue *q)
922 {
923 struct blk_mq_hw_ctx *hctx;
924 int i;
925
926 queue_for_each_hw_ctx(q, hctx, i)
927 blk_mq_start_hw_queue(hctx);
928 }
929 EXPORT_SYMBOL(blk_mq_start_hw_queues);
930
931 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
932 {
933 struct blk_mq_hw_ctx *hctx;
934 int i;
935
936 queue_for_each_hw_ctx(q, hctx, i) {
937 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
938 continue;
939
940 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
941 blk_mq_run_hw_queue(hctx, async);
942 }
943 }
944 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
945
946 static void blk_mq_run_work_fn(struct work_struct *work)
947 {
948 struct blk_mq_hw_ctx *hctx;
949
950 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
951
952 __blk_mq_run_hw_queue(hctx);
953 }
954
955 static void blk_mq_delay_work_fn(struct work_struct *work)
956 {
957 struct blk_mq_hw_ctx *hctx;
958
959 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
960
961 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
962 __blk_mq_run_hw_queue(hctx);
963 }
964
965 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
966 {
967 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
968 return;
969
970 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
971 &hctx->delay_work, msecs_to_jiffies(msecs));
972 }
973 EXPORT_SYMBOL(blk_mq_delay_queue);
974
975 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
976 struct blk_mq_ctx *ctx,
977 struct request *rq,
978 bool at_head)
979 {
980 trace_block_rq_insert(hctx->queue, rq);
981
982 if (at_head)
983 list_add(&rq->queuelist, &ctx->rq_list);
984 else
985 list_add_tail(&rq->queuelist, &ctx->rq_list);
986 }
987
988 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
989 struct request *rq, bool at_head)
990 {
991 struct blk_mq_ctx *ctx = rq->mq_ctx;
992
993 __blk_mq_insert_req_list(hctx, ctx, rq, at_head);
994 blk_mq_hctx_mark_pending(hctx, ctx);
995 }
996
997 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
998 bool async)
999 {
1000 struct request_queue *q = rq->q;
1001 struct blk_mq_hw_ctx *hctx;
1002 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1003
1004 current_ctx = blk_mq_get_ctx(q);
1005 if (!cpu_online(ctx->cpu))
1006 rq->mq_ctx = ctx = current_ctx;
1007
1008 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1009
1010 spin_lock(&ctx->lock);
1011 __blk_mq_insert_request(hctx, rq, at_head);
1012 spin_unlock(&ctx->lock);
1013
1014 if (run_queue)
1015 blk_mq_run_hw_queue(hctx, async);
1016
1017 blk_mq_put_ctx(current_ctx);
1018 }
1019
1020 static void blk_mq_insert_requests(struct request_queue *q,
1021 struct blk_mq_ctx *ctx,
1022 struct list_head *list,
1023 int depth,
1024 bool from_schedule)
1025
1026 {
1027 struct blk_mq_hw_ctx *hctx;
1028 struct blk_mq_ctx *current_ctx;
1029
1030 trace_block_unplug(q, depth, !from_schedule);
1031
1032 current_ctx = blk_mq_get_ctx(q);
1033
1034 if (!cpu_online(ctx->cpu))
1035 ctx = current_ctx;
1036 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1037
1038 /*
1039 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1040 * offline now
1041 */
1042 spin_lock(&ctx->lock);
1043 while (!list_empty(list)) {
1044 struct request *rq;
1045
1046 rq = list_first_entry(list, struct request, queuelist);
1047 list_del_init(&rq->queuelist);
1048 rq->mq_ctx = ctx;
1049 __blk_mq_insert_req_list(hctx, ctx, rq, false);
1050 }
1051 blk_mq_hctx_mark_pending(hctx, ctx);
1052 spin_unlock(&ctx->lock);
1053
1054 blk_mq_run_hw_queue(hctx, from_schedule);
1055 blk_mq_put_ctx(current_ctx);
1056 }
1057
1058 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1059 {
1060 struct request *rqa = container_of(a, struct request, queuelist);
1061 struct request *rqb = container_of(b, struct request, queuelist);
1062
1063 return !(rqa->mq_ctx < rqb->mq_ctx ||
1064 (rqa->mq_ctx == rqb->mq_ctx &&
1065 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1066 }
1067
1068 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1069 {
1070 struct blk_mq_ctx *this_ctx;
1071 struct request_queue *this_q;
1072 struct request *rq;
1073 LIST_HEAD(list);
1074 LIST_HEAD(ctx_list);
1075 unsigned int depth;
1076
1077 list_splice_init(&plug->mq_list, &list);
1078
1079 list_sort(NULL, &list, plug_ctx_cmp);
1080
1081 this_q = NULL;
1082 this_ctx = NULL;
1083 depth = 0;
1084
1085 while (!list_empty(&list)) {
1086 rq = list_entry_rq(list.next);
1087 list_del_init(&rq->queuelist);
1088 BUG_ON(!rq->q);
1089 if (rq->mq_ctx != this_ctx) {
1090 if (this_ctx) {
1091 blk_mq_insert_requests(this_q, this_ctx,
1092 &ctx_list, depth,
1093 from_schedule);
1094 }
1095
1096 this_ctx = rq->mq_ctx;
1097 this_q = rq->q;
1098 depth = 0;
1099 }
1100
1101 depth++;
1102 list_add_tail(&rq->queuelist, &ctx_list);
1103 }
1104
1105 /*
1106 * If 'this_ctx' is set, we know we have entries to complete
1107 * on 'ctx_list'. Do those.
1108 */
1109 if (this_ctx) {
1110 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1111 from_schedule);
1112 }
1113 }
1114
1115 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1116 {
1117 init_request_from_bio(rq, bio);
1118
1119 if (blk_do_io_stat(rq))
1120 blk_account_io_start(rq, 1);
1121 }
1122
1123 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1124 {
1125 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1126 !blk_queue_nomerges(hctx->queue);
1127 }
1128
1129 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1130 struct blk_mq_ctx *ctx,
1131 struct request *rq, struct bio *bio)
1132 {
1133 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1134 blk_mq_bio_to_request(rq, bio);
1135 spin_lock(&ctx->lock);
1136 insert_rq:
1137 __blk_mq_insert_request(hctx, rq, false);
1138 spin_unlock(&ctx->lock);
1139 return false;
1140 } else {
1141 struct request_queue *q = hctx->queue;
1142
1143 spin_lock(&ctx->lock);
1144 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1145 blk_mq_bio_to_request(rq, bio);
1146 goto insert_rq;
1147 }
1148
1149 spin_unlock(&ctx->lock);
1150 __blk_mq_free_request(hctx, ctx, rq);
1151 return true;
1152 }
1153 }
1154
1155 struct blk_map_ctx {
1156 struct blk_mq_hw_ctx *hctx;
1157 struct blk_mq_ctx *ctx;
1158 };
1159
1160 static struct request *blk_mq_map_request(struct request_queue *q,
1161 struct bio *bio,
1162 struct blk_map_ctx *data)
1163 {
1164 struct blk_mq_hw_ctx *hctx;
1165 struct blk_mq_ctx *ctx;
1166 struct request *rq;
1167 int rw = bio_data_dir(bio);
1168 struct blk_mq_alloc_data alloc_data;
1169
1170 blk_queue_enter_live(q);
1171 ctx = blk_mq_get_ctx(q);
1172 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1173
1174 if (rw_is_sync(bio->bi_rw))
1175 rw |= REQ_SYNC;
1176
1177 trace_block_getrq(q, bio, rw);
1178 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1179 hctx);
1180 rq = __blk_mq_alloc_request(&alloc_data, rw);
1181 if (unlikely(!rq)) {
1182 __blk_mq_run_hw_queue(hctx);
1183 blk_mq_put_ctx(ctx);
1184 trace_block_sleeprq(q, bio, rw);
1185
1186 ctx = blk_mq_get_ctx(q);
1187 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1188 blk_mq_set_alloc_data(&alloc_data, q,
1189 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1190 rq = __blk_mq_alloc_request(&alloc_data, rw);
1191 ctx = alloc_data.ctx;
1192 hctx = alloc_data.hctx;
1193 }
1194
1195 hctx->queued++;
1196 data->hctx = hctx;
1197 data->ctx = ctx;
1198 return rq;
1199 }
1200
1201 static int blk_mq_direct_issue_request(struct request *rq)
1202 {
1203 int ret;
1204 struct request_queue *q = rq->q;
1205 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1206 rq->mq_ctx->cpu);
1207 struct blk_mq_queue_data bd = {
1208 .rq = rq,
1209 .list = NULL,
1210 .last = 1
1211 };
1212
1213 /*
1214 * For OK queue, we are done. For error, kill it. Any other
1215 * error (busy), just add it to our list as we previously
1216 * would have done
1217 */
1218 ret = q->mq_ops->queue_rq(hctx, &bd);
1219 if (ret == BLK_MQ_RQ_QUEUE_OK)
1220 return 0;
1221 else {
1222 __blk_mq_requeue_request(rq);
1223
1224 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1225 rq->errors = -EIO;
1226 blk_mq_end_request(rq, rq->errors);
1227 return 0;
1228 }
1229 return -1;
1230 }
1231 }
1232
1233 /*
1234 * Multiple hardware queue variant. This will not use per-process plugs,
1235 * but will attempt to bypass the hctx queueing if we can go straight to
1236 * hardware for SYNC IO.
1237 */
1238 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1239 {
1240 const int is_sync = rw_is_sync(bio->bi_rw);
1241 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1242 struct blk_map_ctx data;
1243 struct request *rq;
1244 unsigned int request_count = 0;
1245 struct blk_plug *plug;
1246 struct request *same_queue_rq = NULL;
1247
1248 blk_queue_bounce(q, &bio);
1249
1250 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1251 bio_io_error(bio);
1252 return;
1253 }
1254
1255 blk_queue_split(q, &bio, q->bio_split);
1256
1257 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1258 if (blk_attempt_plug_merge(q, bio, &request_count,
1259 &same_queue_rq))
1260 return;
1261 } else
1262 request_count = blk_plug_queued_count(q);
1263
1264 rq = blk_mq_map_request(q, bio, &data);
1265 if (unlikely(!rq))
1266 return;
1267
1268 if (unlikely(is_flush_fua)) {
1269 blk_mq_bio_to_request(rq, bio);
1270 blk_insert_flush(rq);
1271 goto run_queue;
1272 }
1273
1274 plug = current->plug;
1275 /*
1276 * If the driver supports defer issued based on 'last', then
1277 * queue it up like normal since we can potentially save some
1278 * CPU this way.
1279 */
1280 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1281 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1282 struct request *old_rq = NULL;
1283
1284 blk_mq_bio_to_request(rq, bio);
1285
1286 /*
1287 * we do limited pluging. If bio can be merged, do merge.
1288 * Otherwise the existing request in the plug list will be
1289 * issued. So the plug list will have one request at most
1290 */
1291 if (plug) {
1292 /*
1293 * The plug list might get flushed before this. If that
1294 * happens, same_queue_rq is invalid and plug list is empty
1295 **/
1296 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1297 old_rq = same_queue_rq;
1298 list_del_init(&old_rq->queuelist);
1299 }
1300 list_add_tail(&rq->queuelist, &plug->mq_list);
1301 } else /* is_sync */
1302 old_rq = rq;
1303 blk_mq_put_ctx(data.ctx);
1304 if (!old_rq)
1305 return;
1306 if (!blk_mq_direct_issue_request(old_rq))
1307 return;
1308 blk_mq_insert_request(old_rq, false, true, true);
1309 return;
1310 }
1311
1312 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1313 /*
1314 * For a SYNC request, send it to the hardware immediately. For
1315 * an ASYNC request, just ensure that we run it later on. The
1316 * latter allows for merging opportunities and more efficient
1317 * dispatching.
1318 */
1319 run_queue:
1320 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1321 }
1322 blk_mq_put_ctx(data.ctx);
1323 }
1324
1325 /*
1326 * Single hardware queue variant. This will attempt to use any per-process
1327 * plug for merging and IO deferral.
1328 */
1329 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1330 {
1331 const int is_sync = rw_is_sync(bio->bi_rw);
1332 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1333 struct blk_plug *plug;
1334 unsigned int request_count = 0;
1335 struct blk_map_ctx data;
1336 struct request *rq;
1337
1338 blk_queue_bounce(q, &bio);
1339
1340 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1341 bio_io_error(bio);
1342 return;
1343 }
1344
1345 blk_queue_split(q, &bio, q->bio_split);
1346
1347 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1348 blk_attempt_plug_merge(q, bio, &request_count, NULL))
1349 return;
1350
1351 rq = blk_mq_map_request(q, bio, &data);
1352 if (unlikely(!rq))
1353 return;
1354
1355 if (unlikely(is_flush_fua)) {
1356 blk_mq_bio_to_request(rq, bio);
1357 blk_insert_flush(rq);
1358 goto run_queue;
1359 }
1360
1361 /*
1362 * A task plug currently exists. Since this is completely lockless,
1363 * utilize that to temporarily store requests until the task is
1364 * either done or scheduled away.
1365 */
1366 plug = current->plug;
1367 if (plug) {
1368 blk_mq_bio_to_request(rq, bio);
1369 if (!request_count)
1370 trace_block_plug(q);
1371 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1372 blk_flush_plug_list(plug, false);
1373 trace_block_plug(q);
1374 }
1375 list_add_tail(&rq->queuelist, &plug->mq_list);
1376 blk_mq_put_ctx(data.ctx);
1377 return;
1378 }
1379
1380 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1381 /*
1382 * For a SYNC request, send it to the hardware immediately. For
1383 * an ASYNC request, just ensure that we run it later on. The
1384 * latter allows for merging opportunities and more efficient
1385 * dispatching.
1386 */
1387 run_queue:
1388 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1389 }
1390
1391 blk_mq_put_ctx(data.ctx);
1392 }
1393
1394 /*
1395 * Default mapping to a software queue, since we use one per CPU.
1396 */
1397 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1398 {
1399 return q->queue_hw_ctx[q->mq_map[cpu]];
1400 }
1401 EXPORT_SYMBOL(blk_mq_map_queue);
1402
1403 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1404 struct blk_mq_tags *tags, unsigned int hctx_idx)
1405 {
1406 struct page *page;
1407
1408 if (tags->rqs && set->ops->exit_request) {
1409 int i;
1410
1411 for (i = 0; i < tags->nr_tags; i++) {
1412 if (!tags->rqs[i])
1413 continue;
1414 set->ops->exit_request(set->driver_data, tags->rqs[i],
1415 hctx_idx, i);
1416 tags->rqs[i] = NULL;
1417 }
1418 }
1419
1420 while (!list_empty(&tags->page_list)) {
1421 page = list_first_entry(&tags->page_list, struct page, lru);
1422 list_del_init(&page->lru);
1423 /*
1424 * Remove kmemleak object previously allocated in
1425 * blk_mq_init_rq_map().
1426 */
1427 kmemleak_free(page_address(page));
1428 __free_pages(page, page->private);
1429 }
1430
1431 kfree(tags->rqs);
1432
1433 blk_mq_free_tags(tags);
1434 }
1435
1436 static size_t order_to_size(unsigned int order)
1437 {
1438 return (size_t)PAGE_SIZE << order;
1439 }
1440
1441 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1442 unsigned int hctx_idx)
1443 {
1444 struct blk_mq_tags *tags;
1445 unsigned int i, j, entries_per_page, max_order = 4;
1446 size_t rq_size, left;
1447
1448 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1449 set->numa_node,
1450 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1451 if (!tags)
1452 return NULL;
1453
1454 INIT_LIST_HEAD(&tags->page_list);
1455
1456 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1457 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1458 set->numa_node);
1459 if (!tags->rqs) {
1460 blk_mq_free_tags(tags);
1461 return NULL;
1462 }
1463
1464 /*
1465 * rq_size is the size of the request plus driver payload, rounded
1466 * to the cacheline size
1467 */
1468 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1469 cache_line_size());
1470 left = rq_size * set->queue_depth;
1471
1472 for (i = 0; i < set->queue_depth; ) {
1473 int this_order = max_order;
1474 struct page *page;
1475 int to_do;
1476 void *p;
1477
1478 while (left < order_to_size(this_order - 1) && this_order)
1479 this_order--;
1480
1481 do {
1482 page = alloc_pages_node(set->numa_node,
1483 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1484 this_order);
1485 if (page)
1486 break;
1487 if (!this_order--)
1488 break;
1489 if (order_to_size(this_order) < rq_size)
1490 break;
1491 } while (1);
1492
1493 if (!page)
1494 goto fail;
1495
1496 page->private = this_order;
1497 list_add_tail(&page->lru, &tags->page_list);
1498
1499 p = page_address(page);
1500 /*
1501 * Allow kmemleak to scan these pages as they contain pointers
1502 * to additional allocations like via ops->init_request().
1503 */
1504 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1505 entries_per_page = order_to_size(this_order) / rq_size;
1506 to_do = min(entries_per_page, set->queue_depth - i);
1507 left -= to_do * rq_size;
1508 for (j = 0; j < to_do; j++) {
1509 tags->rqs[i] = p;
1510 if (set->ops->init_request) {
1511 if (set->ops->init_request(set->driver_data,
1512 tags->rqs[i], hctx_idx, i,
1513 set->numa_node)) {
1514 tags->rqs[i] = NULL;
1515 goto fail;
1516 }
1517 }
1518
1519 p += rq_size;
1520 i++;
1521 }
1522 }
1523 return tags;
1524
1525 fail:
1526 blk_mq_free_rq_map(set, tags, hctx_idx);
1527 return NULL;
1528 }
1529
1530 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1531 {
1532 kfree(bitmap->map);
1533 }
1534
1535 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1536 {
1537 unsigned int bpw = 8, total, num_maps, i;
1538
1539 bitmap->bits_per_word = bpw;
1540
1541 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1542 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1543 GFP_KERNEL, node);
1544 if (!bitmap->map)
1545 return -ENOMEM;
1546
1547 total = nr_cpu_ids;
1548 for (i = 0; i < num_maps; i++) {
1549 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1550 total -= bitmap->map[i].depth;
1551 }
1552
1553 return 0;
1554 }
1555
1556 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1557 {
1558 struct request_queue *q = hctx->queue;
1559 struct blk_mq_ctx *ctx;
1560 LIST_HEAD(tmp);
1561
1562 /*
1563 * Move ctx entries to new CPU, if this one is going away.
1564 */
1565 ctx = __blk_mq_get_ctx(q, cpu);
1566
1567 spin_lock(&ctx->lock);
1568 if (!list_empty(&ctx->rq_list)) {
1569 list_splice_init(&ctx->rq_list, &tmp);
1570 blk_mq_hctx_clear_pending(hctx, ctx);
1571 }
1572 spin_unlock(&ctx->lock);
1573
1574 if (list_empty(&tmp))
1575 return NOTIFY_OK;
1576
1577 ctx = blk_mq_get_ctx(q);
1578 spin_lock(&ctx->lock);
1579
1580 while (!list_empty(&tmp)) {
1581 struct request *rq;
1582
1583 rq = list_first_entry(&tmp, struct request, queuelist);
1584 rq->mq_ctx = ctx;
1585 list_move_tail(&rq->queuelist, &ctx->rq_list);
1586 }
1587
1588 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1589 blk_mq_hctx_mark_pending(hctx, ctx);
1590
1591 spin_unlock(&ctx->lock);
1592
1593 blk_mq_run_hw_queue(hctx, true);
1594 blk_mq_put_ctx(ctx);
1595 return NOTIFY_OK;
1596 }
1597
1598 static int blk_mq_hctx_notify(void *data, unsigned long action,
1599 unsigned int cpu)
1600 {
1601 struct blk_mq_hw_ctx *hctx = data;
1602
1603 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1604 return blk_mq_hctx_cpu_offline(hctx, cpu);
1605
1606 /*
1607 * In case of CPU online, tags may be reallocated
1608 * in blk_mq_map_swqueue() after mapping is updated.
1609 */
1610
1611 return NOTIFY_OK;
1612 }
1613
1614 /* hctx->ctxs will be freed in queue's release handler */
1615 static void blk_mq_exit_hctx(struct request_queue *q,
1616 struct blk_mq_tag_set *set,
1617 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1618 {
1619 unsigned flush_start_tag = set->queue_depth;
1620
1621 blk_mq_tag_idle(hctx);
1622
1623 if (set->ops->exit_request)
1624 set->ops->exit_request(set->driver_data,
1625 hctx->fq->flush_rq, hctx_idx,
1626 flush_start_tag + hctx_idx);
1627
1628 if (set->ops->exit_hctx)
1629 set->ops->exit_hctx(hctx, hctx_idx);
1630
1631 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1632 blk_free_flush_queue(hctx->fq);
1633 blk_mq_free_bitmap(&hctx->ctx_map);
1634 }
1635
1636 static void blk_mq_exit_hw_queues(struct request_queue *q,
1637 struct blk_mq_tag_set *set, int nr_queue)
1638 {
1639 struct blk_mq_hw_ctx *hctx;
1640 unsigned int i;
1641
1642 queue_for_each_hw_ctx(q, hctx, i) {
1643 if (i == nr_queue)
1644 break;
1645 blk_mq_exit_hctx(q, set, hctx, i);
1646 }
1647 }
1648
1649 static void blk_mq_free_hw_queues(struct request_queue *q,
1650 struct blk_mq_tag_set *set)
1651 {
1652 struct blk_mq_hw_ctx *hctx;
1653 unsigned int i;
1654
1655 queue_for_each_hw_ctx(q, hctx, i)
1656 free_cpumask_var(hctx->cpumask);
1657 }
1658
1659 static int blk_mq_init_hctx(struct request_queue *q,
1660 struct blk_mq_tag_set *set,
1661 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1662 {
1663 int node;
1664 unsigned flush_start_tag = set->queue_depth;
1665
1666 node = hctx->numa_node;
1667 if (node == NUMA_NO_NODE)
1668 node = hctx->numa_node = set->numa_node;
1669
1670 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1671 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1672 spin_lock_init(&hctx->lock);
1673 INIT_LIST_HEAD(&hctx->dispatch);
1674 hctx->queue = q;
1675 hctx->queue_num = hctx_idx;
1676 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1677
1678 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1679 blk_mq_hctx_notify, hctx);
1680 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1681
1682 hctx->tags = set->tags[hctx_idx];
1683
1684 /*
1685 * Allocate space for all possible cpus to avoid allocation at
1686 * runtime
1687 */
1688 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1689 GFP_KERNEL, node);
1690 if (!hctx->ctxs)
1691 goto unregister_cpu_notifier;
1692
1693 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1694 goto free_ctxs;
1695
1696 hctx->nr_ctx = 0;
1697
1698 if (set->ops->init_hctx &&
1699 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1700 goto free_bitmap;
1701
1702 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1703 if (!hctx->fq)
1704 goto exit_hctx;
1705
1706 if (set->ops->init_request &&
1707 set->ops->init_request(set->driver_data,
1708 hctx->fq->flush_rq, hctx_idx,
1709 flush_start_tag + hctx_idx, node))
1710 goto free_fq;
1711
1712 return 0;
1713
1714 free_fq:
1715 kfree(hctx->fq);
1716 exit_hctx:
1717 if (set->ops->exit_hctx)
1718 set->ops->exit_hctx(hctx, hctx_idx);
1719 free_bitmap:
1720 blk_mq_free_bitmap(&hctx->ctx_map);
1721 free_ctxs:
1722 kfree(hctx->ctxs);
1723 unregister_cpu_notifier:
1724 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1725
1726 return -1;
1727 }
1728
1729 static int blk_mq_init_hw_queues(struct request_queue *q,
1730 struct blk_mq_tag_set *set)
1731 {
1732 struct blk_mq_hw_ctx *hctx;
1733 unsigned int i;
1734
1735 /*
1736 * Initialize hardware queues
1737 */
1738 queue_for_each_hw_ctx(q, hctx, i) {
1739 if (blk_mq_init_hctx(q, set, hctx, i))
1740 break;
1741 }
1742
1743 if (i == q->nr_hw_queues)
1744 return 0;
1745
1746 /*
1747 * Init failed
1748 */
1749 blk_mq_exit_hw_queues(q, set, i);
1750
1751 return 1;
1752 }
1753
1754 static void blk_mq_init_cpu_queues(struct request_queue *q,
1755 unsigned int nr_hw_queues)
1756 {
1757 unsigned int i;
1758
1759 for_each_possible_cpu(i) {
1760 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1761 struct blk_mq_hw_ctx *hctx;
1762
1763 memset(__ctx, 0, sizeof(*__ctx));
1764 __ctx->cpu = i;
1765 spin_lock_init(&__ctx->lock);
1766 INIT_LIST_HEAD(&__ctx->rq_list);
1767 __ctx->queue = q;
1768
1769 /* If the cpu isn't online, the cpu is mapped to first hctx */
1770 if (!cpu_online(i))
1771 continue;
1772
1773 hctx = q->mq_ops->map_queue(q, i);
1774
1775 /*
1776 * Set local node, IFF we have more than one hw queue. If
1777 * not, we remain on the home node of the device
1778 */
1779 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1780 hctx->numa_node = cpu_to_node(i);
1781 }
1782 }
1783
1784 static void blk_mq_map_swqueue(struct request_queue *q,
1785 const struct cpumask *online_mask)
1786 {
1787 unsigned int i;
1788 struct blk_mq_hw_ctx *hctx;
1789 struct blk_mq_ctx *ctx;
1790 struct blk_mq_tag_set *set = q->tag_set;
1791
1792 /*
1793 * Avoid others reading imcomplete hctx->cpumask through sysfs
1794 */
1795 mutex_lock(&q->sysfs_lock);
1796
1797 queue_for_each_hw_ctx(q, hctx, i) {
1798 cpumask_clear(hctx->cpumask);
1799 hctx->nr_ctx = 0;
1800 }
1801
1802 /*
1803 * Map software to hardware queues
1804 */
1805 queue_for_each_ctx(q, ctx, i) {
1806 /* If the cpu isn't online, the cpu is mapped to first hctx */
1807 if (!cpumask_test_cpu(i, online_mask))
1808 continue;
1809
1810 hctx = q->mq_ops->map_queue(q, i);
1811 cpumask_set_cpu(i, hctx->cpumask);
1812 ctx->index_hw = hctx->nr_ctx;
1813 hctx->ctxs[hctx->nr_ctx++] = ctx;
1814 }
1815
1816 mutex_unlock(&q->sysfs_lock);
1817
1818 queue_for_each_hw_ctx(q, hctx, i) {
1819 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1820
1821 /*
1822 * If no software queues are mapped to this hardware queue,
1823 * disable it and free the request entries.
1824 */
1825 if (!hctx->nr_ctx) {
1826 if (set->tags[i]) {
1827 blk_mq_free_rq_map(set, set->tags[i], i);
1828 set->tags[i] = NULL;
1829 }
1830 hctx->tags = NULL;
1831 continue;
1832 }
1833
1834 /* unmapped hw queue can be remapped after CPU topo changed */
1835 if (!set->tags[i])
1836 set->tags[i] = blk_mq_init_rq_map(set, i);
1837 hctx->tags = set->tags[i];
1838 WARN_ON(!hctx->tags);
1839
1840 /*
1841 * Set the map size to the number of mapped software queues.
1842 * This is more accurate and more efficient than looping
1843 * over all possibly mapped software queues.
1844 */
1845 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1846
1847 /*
1848 * Initialize batch roundrobin counts
1849 */
1850 hctx->next_cpu = cpumask_first(hctx->cpumask);
1851 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1852 }
1853
1854 queue_for_each_ctx(q, ctx, i) {
1855 if (!cpumask_test_cpu(i, online_mask))
1856 continue;
1857
1858 hctx = q->mq_ops->map_queue(q, i);
1859 cpumask_set_cpu(i, hctx->tags->cpumask);
1860 }
1861 }
1862
1863 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1864 {
1865 struct blk_mq_hw_ctx *hctx;
1866 int i;
1867
1868 queue_for_each_hw_ctx(q, hctx, i) {
1869 if (shared)
1870 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1871 else
1872 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1873 }
1874 }
1875
1876 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1877 {
1878 struct request_queue *q;
1879
1880 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1881 blk_mq_freeze_queue(q);
1882 queue_set_hctx_shared(q, shared);
1883 blk_mq_unfreeze_queue(q);
1884 }
1885 }
1886
1887 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1888 {
1889 struct blk_mq_tag_set *set = q->tag_set;
1890
1891 mutex_lock(&set->tag_list_lock);
1892 list_del_init(&q->tag_set_list);
1893 if (list_is_singular(&set->tag_list)) {
1894 /* just transitioned to unshared */
1895 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1896 /* update existing queue */
1897 blk_mq_update_tag_set_depth(set, false);
1898 }
1899 mutex_unlock(&set->tag_list_lock);
1900 }
1901
1902 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1903 struct request_queue *q)
1904 {
1905 q->tag_set = set;
1906
1907 mutex_lock(&set->tag_list_lock);
1908
1909 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1910 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1911 set->flags |= BLK_MQ_F_TAG_SHARED;
1912 /* update existing queue */
1913 blk_mq_update_tag_set_depth(set, true);
1914 }
1915 if (set->flags & BLK_MQ_F_TAG_SHARED)
1916 queue_set_hctx_shared(q, true);
1917 list_add_tail(&q->tag_set_list, &set->tag_list);
1918
1919 mutex_unlock(&set->tag_list_lock);
1920 }
1921
1922 /*
1923 * It is the actual release handler for mq, but we do it from
1924 * request queue's release handler for avoiding use-after-free
1925 * and headache because q->mq_kobj shouldn't have been introduced,
1926 * but we can't group ctx/kctx kobj without it.
1927 */
1928 void blk_mq_release(struct request_queue *q)
1929 {
1930 struct blk_mq_hw_ctx *hctx;
1931 unsigned int i;
1932
1933 /* hctx kobj stays in hctx */
1934 queue_for_each_hw_ctx(q, hctx, i) {
1935 if (!hctx)
1936 continue;
1937 kfree(hctx->ctxs);
1938 kfree(hctx);
1939 }
1940
1941 kfree(q->mq_map);
1942 q->mq_map = NULL;
1943
1944 kfree(q->queue_hw_ctx);
1945
1946 /* ctx kobj stays in queue_ctx */
1947 free_percpu(q->queue_ctx);
1948 }
1949
1950 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1951 {
1952 struct request_queue *uninit_q, *q;
1953
1954 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1955 if (!uninit_q)
1956 return ERR_PTR(-ENOMEM);
1957
1958 q = blk_mq_init_allocated_queue(set, uninit_q);
1959 if (IS_ERR(q))
1960 blk_cleanup_queue(uninit_q);
1961
1962 return q;
1963 }
1964 EXPORT_SYMBOL(blk_mq_init_queue);
1965
1966 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1967 struct request_queue *q)
1968 {
1969 struct blk_mq_hw_ctx **hctxs;
1970 struct blk_mq_ctx __percpu *ctx;
1971 unsigned int *map;
1972 int i;
1973
1974 ctx = alloc_percpu(struct blk_mq_ctx);
1975 if (!ctx)
1976 return ERR_PTR(-ENOMEM);
1977
1978 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1979 set->numa_node);
1980
1981 if (!hctxs)
1982 goto err_percpu;
1983
1984 map = blk_mq_make_queue_map(set);
1985 if (!map)
1986 goto err_map;
1987
1988 for (i = 0; i < set->nr_hw_queues; i++) {
1989 int node = blk_mq_hw_queue_to_node(map, i);
1990
1991 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1992 GFP_KERNEL, node);
1993 if (!hctxs[i])
1994 goto err_hctxs;
1995
1996 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1997 node))
1998 goto err_hctxs;
1999
2000 atomic_set(&hctxs[i]->nr_active, 0);
2001 hctxs[i]->numa_node = node;
2002 hctxs[i]->queue_num = i;
2003 }
2004
2005 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
2006 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2007
2008 q->nr_queues = nr_cpu_ids;
2009 q->nr_hw_queues = set->nr_hw_queues;
2010 q->mq_map = map;
2011
2012 q->queue_ctx = ctx;
2013 q->queue_hw_ctx = hctxs;
2014
2015 q->mq_ops = set->ops;
2016 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2017
2018 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2019 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2020
2021 q->sg_reserved_size = INT_MAX;
2022
2023 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2024 INIT_LIST_HEAD(&q->requeue_list);
2025 spin_lock_init(&q->requeue_lock);
2026
2027 if (q->nr_hw_queues > 1)
2028 blk_queue_make_request(q, blk_mq_make_request);
2029 else
2030 blk_queue_make_request(q, blk_sq_make_request);
2031
2032 /*
2033 * Do this after blk_queue_make_request() overrides it...
2034 */
2035 q->nr_requests = set->queue_depth;
2036
2037 if (set->ops->complete)
2038 blk_queue_softirq_done(q, set->ops->complete);
2039
2040 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2041
2042 if (blk_mq_init_hw_queues(q, set))
2043 goto err_hctxs;
2044
2045 get_online_cpus();
2046 mutex_lock(&all_q_mutex);
2047
2048 list_add_tail(&q->all_q_node, &all_q_list);
2049 blk_mq_add_queue_tag_set(set, q);
2050 blk_mq_map_swqueue(q, cpu_online_mask);
2051
2052 mutex_unlock(&all_q_mutex);
2053 put_online_cpus();
2054
2055 return q;
2056
2057 err_hctxs:
2058 kfree(map);
2059 for (i = 0; i < set->nr_hw_queues; i++) {
2060 if (!hctxs[i])
2061 break;
2062 free_cpumask_var(hctxs[i]->cpumask);
2063 kfree(hctxs[i]);
2064 }
2065 err_map:
2066 kfree(hctxs);
2067 err_percpu:
2068 free_percpu(ctx);
2069 return ERR_PTR(-ENOMEM);
2070 }
2071 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2072
2073 void blk_mq_free_queue(struct request_queue *q)
2074 {
2075 struct blk_mq_tag_set *set = q->tag_set;
2076
2077 mutex_lock(&all_q_mutex);
2078 list_del_init(&q->all_q_node);
2079 mutex_unlock(&all_q_mutex);
2080
2081 blk_mq_del_queue_tag_set(q);
2082
2083 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2084 blk_mq_free_hw_queues(q, set);
2085 }
2086
2087 /* Basically redo blk_mq_init_queue with queue frozen */
2088 static void blk_mq_queue_reinit(struct request_queue *q,
2089 const struct cpumask *online_mask)
2090 {
2091 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2092
2093 blk_mq_sysfs_unregister(q);
2094
2095 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2096
2097 /*
2098 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2099 * we should change hctx numa_node according to new topology (this
2100 * involves free and re-allocate memory, worthy doing?)
2101 */
2102
2103 blk_mq_map_swqueue(q, online_mask);
2104
2105 blk_mq_sysfs_register(q);
2106 }
2107
2108 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2109 unsigned long action, void *hcpu)
2110 {
2111 struct request_queue *q;
2112 int cpu = (unsigned long)hcpu;
2113 /*
2114 * New online cpumask which is going to be set in this hotplug event.
2115 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2116 * one-by-one and dynamically allocating this could result in a failure.
2117 */
2118 static struct cpumask online_new;
2119
2120 /*
2121 * Before hotadded cpu starts handling requests, new mappings must
2122 * be established. Otherwise, these requests in hw queue might
2123 * never be dispatched.
2124 *
2125 * For example, there is a single hw queue (hctx) and two CPU queues
2126 * (ctx0 for CPU0, and ctx1 for CPU1).
2127 *
2128 * Now CPU1 is just onlined and a request is inserted into
2129 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2130 * still zero.
2131 *
2132 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2133 * set in pending bitmap and tries to retrieve requests in
2134 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2135 * so the request in ctx1->rq_list is ignored.
2136 */
2137 switch (action & ~CPU_TASKS_FROZEN) {
2138 case CPU_DEAD:
2139 case CPU_UP_CANCELED:
2140 cpumask_copy(&online_new, cpu_online_mask);
2141 break;
2142 case CPU_UP_PREPARE:
2143 cpumask_copy(&online_new, cpu_online_mask);
2144 cpumask_set_cpu(cpu, &online_new);
2145 break;
2146 default:
2147 return NOTIFY_OK;
2148 }
2149
2150 mutex_lock(&all_q_mutex);
2151
2152 /*
2153 * We need to freeze and reinit all existing queues. Freezing
2154 * involves synchronous wait for an RCU grace period and doing it
2155 * one by one may take a long time. Start freezing all queues in
2156 * one swoop and then wait for the completions so that freezing can
2157 * take place in parallel.
2158 */
2159 list_for_each_entry(q, &all_q_list, all_q_node)
2160 blk_mq_freeze_queue_start(q);
2161 list_for_each_entry(q, &all_q_list, all_q_node) {
2162 blk_mq_freeze_queue_wait(q);
2163
2164 /*
2165 * timeout handler can't touch hw queue during the
2166 * reinitialization
2167 */
2168 del_timer_sync(&q->timeout);
2169 }
2170
2171 list_for_each_entry(q, &all_q_list, all_q_node)
2172 blk_mq_queue_reinit(q, &online_new);
2173
2174 list_for_each_entry(q, &all_q_list, all_q_node)
2175 blk_mq_unfreeze_queue(q);
2176
2177 mutex_unlock(&all_q_mutex);
2178 return NOTIFY_OK;
2179 }
2180
2181 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2182 {
2183 int i;
2184
2185 for (i = 0; i < set->nr_hw_queues; i++) {
2186 set->tags[i] = blk_mq_init_rq_map(set, i);
2187 if (!set->tags[i])
2188 goto out_unwind;
2189 }
2190
2191 return 0;
2192
2193 out_unwind:
2194 while (--i >= 0)
2195 blk_mq_free_rq_map(set, set->tags[i], i);
2196
2197 return -ENOMEM;
2198 }
2199
2200 /*
2201 * Allocate the request maps associated with this tag_set. Note that this
2202 * may reduce the depth asked for, if memory is tight. set->queue_depth
2203 * will be updated to reflect the allocated depth.
2204 */
2205 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2206 {
2207 unsigned int depth;
2208 int err;
2209
2210 depth = set->queue_depth;
2211 do {
2212 err = __blk_mq_alloc_rq_maps(set);
2213 if (!err)
2214 break;
2215
2216 set->queue_depth >>= 1;
2217 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2218 err = -ENOMEM;
2219 break;
2220 }
2221 } while (set->queue_depth);
2222
2223 if (!set->queue_depth || err) {
2224 pr_err("blk-mq: failed to allocate request map\n");
2225 return -ENOMEM;
2226 }
2227
2228 if (depth != set->queue_depth)
2229 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2230 depth, set->queue_depth);
2231
2232 return 0;
2233 }
2234
2235 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2236 {
2237 return tags->cpumask;
2238 }
2239 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2240
2241 /*
2242 * Alloc a tag set to be associated with one or more request queues.
2243 * May fail with EINVAL for various error conditions. May adjust the
2244 * requested depth down, if if it too large. In that case, the set
2245 * value will be stored in set->queue_depth.
2246 */
2247 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2248 {
2249 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2250
2251 if (!set->nr_hw_queues)
2252 return -EINVAL;
2253 if (!set->queue_depth)
2254 return -EINVAL;
2255 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2256 return -EINVAL;
2257
2258 if (!set->ops->queue_rq || !set->ops->map_queue)
2259 return -EINVAL;
2260
2261 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2262 pr_info("blk-mq: reduced tag depth to %u\n",
2263 BLK_MQ_MAX_DEPTH);
2264 set->queue_depth = BLK_MQ_MAX_DEPTH;
2265 }
2266
2267 /*
2268 * If a crashdump is active, then we are potentially in a very
2269 * memory constrained environment. Limit us to 1 queue and
2270 * 64 tags to prevent using too much memory.
2271 */
2272 if (is_kdump_kernel()) {
2273 set->nr_hw_queues = 1;
2274 set->queue_depth = min(64U, set->queue_depth);
2275 }
2276
2277 set->tags = kmalloc_node(set->nr_hw_queues *
2278 sizeof(struct blk_mq_tags *),
2279 GFP_KERNEL, set->numa_node);
2280 if (!set->tags)
2281 return -ENOMEM;
2282
2283 if (blk_mq_alloc_rq_maps(set))
2284 goto enomem;
2285
2286 mutex_init(&set->tag_list_lock);
2287 INIT_LIST_HEAD(&set->tag_list);
2288
2289 return 0;
2290 enomem:
2291 kfree(set->tags);
2292 set->tags = NULL;
2293 return -ENOMEM;
2294 }
2295 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2296
2297 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2298 {
2299 int i;
2300
2301 for (i = 0; i < set->nr_hw_queues; i++) {
2302 if (set->tags[i])
2303 blk_mq_free_rq_map(set, set->tags[i], i);
2304 }
2305
2306 kfree(set->tags);
2307 set->tags = NULL;
2308 }
2309 EXPORT_SYMBOL(blk_mq_free_tag_set);
2310
2311 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2312 {
2313 struct blk_mq_tag_set *set = q->tag_set;
2314 struct blk_mq_hw_ctx *hctx;
2315 int i, ret;
2316
2317 if (!set || nr > set->queue_depth)
2318 return -EINVAL;
2319
2320 ret = 0;
2321 queue_for_each_hw_ctx(q, hctx, i) {
2322 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2323 if (ret)
2324 break;
2325 }
2326
2327 if (!ret)
2328 q->nr_requests = nr;
2329
2330 return ret;
2331 }
2332
2333 void blk_mq_disable_hotplug(void)
2334 {
2335 mutex_lock(&all_q_mutex);
2336 }
2337
2338 void blk_mq_enable_hotplug(void)
2339 {
2340 mutex_unlock(&all_q_mutex);
2341 }
2342
2343 static int __init blk_mq_init(void)
2344 {
2345 blk_mq_cpu_init();
2346
2347 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2348
2349 return 0;
2350 }
2351 subsys_initcall(blk_mq_init);