blob: 1fa9dd153fde22a976483ef3b5da3ec39f0f458e [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
Jens Axboe320ae512013-10-24 09:20:05 +010030static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
31 unsigned int cpu)
32{
33 return per_cpu_ptr(q->queue_ctx, cpu);
34}
35
36/*
37 * This assumes per-cpu software queueing queues. They could be per-node
38 * as well, for instance. For now this is hardcoded as-is. Note that we don't
39 * care about preemption, since we know the ctx's are persistent. This does
40 * mean that we can't rely on ctx always matching the currently running CPU.
41 */
42static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
43{
44 return __blk_mq_get_ctx(q, get_cpu());
45}
46
47static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
48{
49 put_cpu();
50}
51
52/*
53 * Check if any of the ctx's have pending work in this hardware queue
54 */
55static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
56{
57 unsigned int i;
58
59 for (i = 0; i < hctx->nr_ctx_map; i++)
60 if (hctx->ctx_map[i])
61 return true;
62
63 return false;
64}
65
66/*
67 * Mark this ctx as having pending work in this hardware queue
68 */
69static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
70 struct blk_mq_ctx *ctx)
71{
72 if (!test_bit(ctx->index_hw, hctx->ctx_map))
73 set_bit(ctx->index_hw, hctx->ctx_map);
74}
75
76static struct request *blk_mq_alloc_rq(struct blk_mq_hw_ctx *hctx, gfp_t gfp,
77 bool reserved)
78{
79 struct request *rq;
80 unsigned int tag;
81
82 tag = blk_mq_get_tag(hctx->tags, gfp, reserved);
83 if (tag != BLK_MQ_TAG_FAIL) {
84 rq = hctx->rqs[tag];
85 rq->tag = tag;
86
87 return rq;
88 }
89
90 return NULL;
91}
92
93static int blk_mq_queue_enter(struct request_queue *q)
94{
95 int ret;
96
97 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
98 smp_wmb();
99 /* we have problems to freeze the queue if it's initializing */
100 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
101 return 0;
102
103 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
104
105 spin_lock_irq(q->queue_lock);
106 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
Ming Lei43a5e4e2013-12-26 21:31:35 +0800107 !blk_queue_bypass(q) || blk_queue_dying(q),
108 *q->queue_lock);
Jens Axboe320ae512013-10-24 09:20:05 +0100109 /* inc usage with lock hold to avoid freeze_queue runs here */
Ming Lei43a5e4e2013-12-26 21:31:35 +0800110 if (!ret && !blk_queue_dying(q))
Jens Axboe320ae512013-10-24 09:20:05 +0100111 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
Ming Lei43a5e4e2013-12-26 21:31:35 +0800112 else if (blk_queue_dying(q))
113 ret = -ENODEV;
Jens Axboe320ae512013-10-24 09:20:05 +0100114 spin_unlock_irq(q->queue_lock);
115
116 return ret;
117}
118
119static void blk_mq_queue_exit(struct request_queue *q)
120{
121 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
122}
123
Ming Lei43a5e4e2013-12-26 21:31:35 +0800124static void __blk_mq_drain_queue(struct request_queue *q)
125{
126 while (true) {
127 s64 count;
128
129 spin_lock_irq(q->queue_lock);
130 count = percpu_counter_sum(&q->mq_usage_counter);
131 spin_unlock_irq(q->queue_lock);
132
133 if (count == 0)
134 break;
135 blk_mq_run_queues(q, false);
136 msleep(10);
137 }
138}
139
Jens Axboe320ae512013-10-24 09:20:05 +0100140/*
141 * Guarantee no request is in use, so we can change any data structure of
142 * the queue afterward.
143 */
144static void blk_mq_freeze_queue(struct request_queue *q)
145{
146 bool drain;
147
148 spin_lock_irq(q->queue_lock);
149 drain = !q->bypass_depth++;
150 queue_flag_set(QUEUE_FLAG_BYPASS, q);
151 spin_unlock_irq(q->queue_lock);
152
Ming Lei43a5e4e2013-12-26 21:31:35 +0800153 if (drain)
154 __blk_mq_drain_queue(q);
155}
Jens Axboe320ae512013-10-24 09:20:05 +0100156
Ming Lei43a5e4e2013-12-26 21:31:35 +0800157void blk_mq_drain_queue(struct request_queue *q)
158{
159 __blk_mq_drain_queue(q);
Jens Axboe320ae512013-10-24 09:20:05 +0100160}
161
162static void blk_mq_unfreeze_queue(struct request_queue *q)
163{
164 bool wake = false;
165
166 spin_lock_irq(q->queue_lock);
167 if (!--q->bypass_depth) {
168 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
169 wake = true;
170 }
171 WARN_ON_ONCE(q->bypass_depth < 0);
172 spin_unlock_irq(q->queue_lock);
173 if (wake)
174 wake_up_all(&q->mq_freeze_wq);
175}
176
177bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
178{
179 return blk_mq_has_free_tags(hctx->tags);
180}
181EXPORT_SYMBOL(blk_mq_can_queue);
182
Jens Axboe94eddfb2013-11-19 09:25:07 -0700183static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
184 struct request *rq, unsigned int rw_flags)
Jens Axboe320ae512013-10-24 09:20:05 +0100185{
Jens Axboe94eddfb2013-11-19 09:25:07 -0700186 if (blk_queue_io_stat(q))
187 rw_flags |= REQ_IO_STAT;
188
Jens Axboe320ae512013-10-24 09:20:05 +0100189 rq->mq_ctx = ctx;
190 rq->cmd_flags = rw_flags;
Ming Lei0fec08b2014-01-03 10:00:08 -0700191 rq->start_time = jiffies;
192 set_start_time_ns(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100193 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
194}
195
196static struct request *__blk_mq_alloc_request(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig18741982014-02-10 09:29:00 -0700197 gfp_t gfp, bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +0100198{
Christoph Hellwig18741982014-02-10 09:29:00 -0700199 return blk_mq_alloc_rq(hctx, gfp, reserved);
Jens Axboe320ae512013-10-24 09:20:05 +0100200}
201
202static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
203 int rw, gfp_t gfp,
204 bool reserved)
205{
206 struct request *rq;
207
208 do {
209 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
210 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
211
Christoph Hellwig18741982014-02-10 09:29:00 -0700212 rq = __blk_mq_alloc_request(hctx, gfp & ~__GFP_WAIT, reserved);
Jens Axboe320ae512013-10-24 09:20:05 +0100213 if (rq) {
Jens Axboe94eddfb2013-11-19 09:25:07 -0700214 blk_mq_rq_ctx_init(q, ctx, rq, rw);
Jens Axboe320ae512013-10-24 09:20:05 +0100215 break;
Jeff Moyer959a35f2013-12-03 14:23:00 -0700216 }
Jens Axboe320ae512013-10-24 09:20:05 +0100217
218 blk_mq_put_ctx(ctx);
Jeff Moyer959a35f2013-12-03 14:23:00 -0700219 if (!(gfp & __GFP_WAIT))
220 break;
221
Jens Axboe320ae512013-10-24 09:20:05 +0100222 __blk_mq_run_hw_queue(hctx);
223 blk_mq_wait_for_tags(hctx->tags);
224 } while (1);
225
226 return rq;
227}
228
Christoph Hellwig18741982014-02-10 09:29:00 -0700229struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp)
Jens Axboe320ae512013-10-24 09:20:05 +0100230{
231 struct request *rq;
232
233 if (blk_mq_queue_enter(q))
234 return NULL;
235
Christoph Hellwig18741982014-02-10 09:29:00 -0700236 rq = blk_mq_alloc_request_pinned(q, rw, gfp, false);
Jeff Moyer959a35f2013-12-03 14:23:00 -0700237 if (rq)
238 blk_mq_put_ctx(rq->mq_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +0100239 return rq;
240}
241
242struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
243 gfp_t gfp)
244{
245 struct request *rq;
246
247 if (blk_mq_queue_enter(q))
248 return NULL;
249
250 rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
Jeff Moyer959a35f2013-12-03 14:23:00 -0700251 if (rq)
252 blk_mq_put_ctx(rq->mq_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +0100253 return rq;
254}
255EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
256
257/*
258 * Re-init and set pdu, if we have it
259 */
Christoph Hellwig18741982014-02-10 09:29:00 -0700260void blk_mq_rq_init(struct blk_mq_hw_ctx *hctx, struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100261{
262 blk_rq_init(hctx->queue, rq);
263
264 if (hctx->cmd_size)
265 rq->special = blk_mq_rq_to_pdu(rq);
266}
267
268static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
269 struct blk_mq_ctx *ctx, struct request *rq)
270{
271 const int tag = rq->tag;
272 struct request_queue *q = rq->q;
273
274 blk_mq_rq_init(hctx, rq);
275 blk_mq_put_tag(hctx->tags, tag);
276
277 blk_mq_queue_exit(q);
278}
279
280void blk_mq_free_request(struct request *rq)
281{
282 struct blk_mq_ctx *ctx = rq->mq_ctx;
283 struct blk_mq_hw_ctx *hctx;
284 struct request_queue *q = rq->q;
285
286 ctx->rq_completed[rq_is_sync(rq)]++;
287
288 hctx = q->mq_ops->map_queue(q, ctx->cpu);
289 __blk_mq_free_request(hctx, ctx, rq);
290}
291
292static void blk_mq_bio_endio(struct request *rq, struct bio *bio, int error)
293{
294 if (error)
295 clear_bit(BIO_UPTODATE, &bio->bi_flags);
296 else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
297 error = -EIO;
298
299 if (unlikely(rq->cmd_flags & REQ_QUIET))
300 set_bit(BIO_QUIET, &bio->bi_flags);
301
302 /* don't actually finish bio if it's part of flush sequence */
303 if (!(rq->cmd_flags & REQ_FLUSH_SEQ))
304 bio_endio(bio, error);
305}
306
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800307void blk_mq_end_io(struct request *rq, int error)
Jens Axboe320ae512013-10-24 09:20:05 +0100308{
309 struct bio *bio = rq->bio;
310 unsigned int bytes = 0;
311
312 trace_block_rq_complete(rq->q, rq);
313
314 while (bio) {
315 struct bio *next = bio->bi_next;
316
317 bio->bi_next = NULL;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700318 bytes += bio->bi_iter.bi_size;
Jens Axboe320ae512013-10-24 09:20:05 +0100319 blk_mq_bio_endio(rq, bio, error);
320 bio = next;
321 }
322
323 blk_account_io_completion(rq, bytes);
324
Ming Lei0d11e6a2013-12-05 10:50:39 -0700325 blk_account_io_done(rq);
326
Jens Axboe320ae512013-10-24 09:20:05 +0100327 if (rq->end_io)
328 rq->end_io(rq, error);
329 else
330 blk_mq_free_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100331}
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800332EXPORT_SYMBOL(blk_mq_end_io);
Jens Axboe320ae512013-10-24 09:20:05 +0100333
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800334static void __blk_mq_complete_request_remote(void *data)
Jens Axboe320ae512013-10-24 09:20:05 +0100335{
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800336 struct request *rq = data;
Jens Axboe320ae512013-10-24 09:20:05 +0100337
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800338 rq->q->softirq_done_fn(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100339}
340
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800341void __blk_mq_complete_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100342{
343 struct blk_mq_ctx *ctx = rq->mq_ctx;
344 int cpu;
345
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800346 if (!ctx->ipi_redirect) {
347 rq->q->softirq_done_fn(rq);
348 return;
349 }
Jens Axboe320ae512013-10-24 09:20:05 +0100350
351 cpu = get_cpu();
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800352 if (cpu != ctx->cpu && cpu_online(ctx->cpu)) {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800353 rq->csd.func = __blk_mq_complete_request_remote;
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800354 rq->csd.info = rq;
355 rq->csd.flags = 0;
356 __smp_call_function_single(ctx->cpu, &rq->csd, 0);
357 } else {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800358 rq->q->softirq_done_fn(rq);
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800359 }
Jens Axboe320ae512013-10-24 09:20:05 +0100360 put_cpu();
361}
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800362
363/**
364 * blk_mq_complete_request - end I/O on a request
365 * @rq: the request being processed
366 *
367 * Description:
368 * Ends all I/O on a request. It does not handle partial completions.
369 * The actual completion happens out-of-order, through a IPI handler.
370 **/
371void blk_mq_complete_request(struct request *rq)
372{
373 if (unlikely(blk_should_fake_timeout(rq->q)))
374 return;
375 if (!blk_mark_rq_complete(rq))
376 __blk_mq_complete_request(rq);
377}
378EXPORT_SYMBOL(blk_mq_complete_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100379
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800380static void blk_mq_start_request(struct request *rq, bool last)
Jens Axboe320ae512013-10-24 09:20:05 +0100381{
382 struct request_queue *q = rq->q;
383
384 trace_block_rq_issue(q, rq);
385
386 /*
387 * Just mark start time and set the started bit. Due to memory
388 * ordering, we know we'll see the correct deadline as long as
389 * REQ_ATOMIC_STARTED is seen.
390 */
391 rq->deadline = jiffies + q->rq_timeout;
392 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800393
394 if (q->dma_drain_size && blk_rq_bytes(rq)) {
395 /*
396 * Make sure space for the drain appears. We know we can do
397 * this because max_hw_segments has been adjusted to be one
398 * fewer than the device can handle.
399 */
400 rq->nr_phys_segments++;
401 }
402
403 /*
404 * Flag the last request in the series so that drivers know when IO
405 * should be kicked off, if they don't do it on a per-request basis.
406 *
407 * Note: the flag isn't the only condition drivers should do kick off.
408 * If drive is busy, the last request might not have the bit set.
409 */
410 if (last)
411 rq->cmd_flags |= REQ_END;
Jens Axboe320ae512013-10-24 09:20:05 +0100412}
413
414static void blk_mq_requeue_request(struct request *rq)
415{
416 struct request_queue *q = rq->q;
417
418 trace_block_rq_requeue(q, rq);
419 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800420
421 rq->cmd_flags &= ~REQ_END;
422
423 if (q->dma_drain_size && blk_rq_bytes(rq))
424 rq->nr_phys_segments--;
Jens Axboe320ae512013-10-24 09:20:05 +0100425}
426
427struct blk_mq_timeout_data {
428 struct blk_mq_hw_ctx *hctx;
429 unsigned long *next;
430 unsigned int *next_set;
431};
432
433static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
434{
435 struct blk_mq_timeout_data *data = __data;
436 struct blk_mq_hw_ctx *hctx = data->hctx;
437 unsigned int tag;
438
439 /* It may not be in flight yet (this is where
440 * the REQ_ATOMIC_STARTED flag comes in). The requests are
441 * statically allocated, so we know it's always safe to access the
442 * memory associated with a bit offset into ->rqs[].
443 */
444 tag = 0;
445 do {
446 struct request *rq;
447
448 tag = find_next_zero_bit(free_tags, hctx->queue_depth, tag);
449 if (tag >= hctx->queue_depth)
450 break;
451
452 rq = hctx->rqs[tag++];
453
454 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
455 continue;
456
457 blk_rq_check_expired(rq, data->next, data->next_set);
458 } while (1);
459}
460
461static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
462 unsigned long *next,
463 unsigned int *next_set)
464{
465 struct blk_mq_timeout_data data = {
466 .hctx = hctx,
467 .next = next,
468 .next_set = next_set,
469 };
470
471 /*
472 * Ask the tagging code to iterate busy requests, so we can
473 * check them for timeout.
474 */
475 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
476}
477
478static void blk_mq_rq_timer(unsigned long data)
479{
480 struct request_queue *q = (struct request_queue *) data;
481 struct blk_mq_hw_ctx *hctx;
482 unsigned long next = 0;
483 int i, next_set = 0;
484
485 queue_for_each_hw_ctx(q, hctx, i)
486 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
487
488 if (next_set)
489 mod_timer(&q->timeout, round_jiffies_up(next));
490}
491
492/*
493 * Reverse check our software queue for entries that we could potentially
494 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
495 * too much time checking for merges.
496 */
497static bool blk_mq_attempt_merge(struct request_queue *q,
498 struct blk_mq_ctx *ctx, struct bio *bio)
499{
500 struct request *rq;
501 int checked = 8;
502
503 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
504 int el_ret;
505
506 if (!checked--)
507 break;
508
509 if (!blk_rq_merge_ok(rq, bio))
510 continue;
511
512 el_ret = blk_try_merge(rq, bio);
513 if (el_ret == ELEVATOR_BACK_MERGE) {
514 if (bio_attempt_back_merge(q, rq, bio)) {
515 ctx->rq_merged++;
516 return true;
517 }
518 break;
519 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
520 if (bio_attempt_front_merge(q, rq, bio)) {
521 ctx->rq_merged++;
522 return true;
523 }
524 break;
525 }
526 }
527
528 return false;
529}
530
531void blk_mq_add_timer(struct request *rq)
532{
533 __blk_add_timer(rq, NULL);
534}
535
536/*
537 * Run this hardware queue, pulling any software queues mapped to it in.
538 * Note that this function currently has various problems around ordering
539 * of IO. In particular, we'd like FIFO behaviour on handling existing
540 * items on the hctx->dispatch list. Ignore that for now.
541 */
542static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
543{
544 struct request_queue *q = hctx->queue;
545 struct blk_mq_ctx *ctx;
546 struct request *rq;
547 LIST_HEAD(rq_list);
548 int bit, queued;
549
550 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
551 return;
552
553 hctx->run++;
554
555 /*
556 * Touch any software queue that has pending entries.
557 */
558 for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
559 clear_bit(bit, hctx->ctx_map);
560 ctx = hctx->ctxs[bit];
561 BUG_ON(bit != ctx->index_hw);
562
563 spin_lock(&ctx->lock);
564 list_splice_tail_init(&ctx->rq_list, &rq_list);
565 spin_unlock(&ctx->lock);
566 }
567
568 /*
569 * If we have previous entries on our dispatch list, grab them
570 * and stuff them at the front for more fair dispatch.
571 */
572 if (!list_empty_careful(&hctx->dispatch)) {
573 spin_lock(&hctx->lock);
574 if (!list_empty(&hctx->dispatch))
575 list_splice_init(&hctx->dispatch, &rq_list);
576 spin_unlock(&hctx->lock);
577 }
578
579 /*
580 * Delete and return all entries from our dispatch list
581 */
582 queued = 0;
583
584 /*
585 * Now process all the entries, sending them to the driver.
586 */
587 while (!list_empty(&rq_list)) {
588 int ret;
589
590 rq = list_first_entry(&rq_list, struct request, queuelist);
591 list_del_init(&rq->queuelist);
Jens Axboe320ae512013-10-24 09:20:05 +0100592
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800593 blk_mq_start_request(rq, list_empty(&rq_list));
Jens Axboe320ae512013-10-24 09:20:05 +0100594
595 ret = q->mq_ops->queue_rq(hctx, rq);
596 switch (ret) {
597 case BLK_MQ_RQ_QUEUE_OK:
598 queued++;
599 continue;
600 case BLK_MQ_RQ_QUEUE_BUSY:
601 /*
602 * FIXME: we should have a mechanism to stop the queue
603 * like blk_stop_queue, otherwise we will waste cpu
604 * time
605 */
606 list_add(&rq->queuelist, &rq_list);
607 blk_mq_requeue_request(rq);
608 break;
609 default:
610 pr_err("blk-mq: bad return on queue: %d\n", ret);
Jens Axboe320ae512013-10-24 09:20:05 +0100611 case BLK_MQ_RQ_QUEUE_ERROR:
Christoph Hellwig1e93b8c2014-02-11 08:27:13 -0800612 rq->errors = -EIO;
Jens Axboe320ae512013-10-24 09:20:05 +0100613 blk_mq_end_io(rq, rq->errors);
614 break;
615 }
616
617 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
618 break;
619 }
620
621 if (!queued)
622 hctx->dispatched[0]++;
623 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
624 hctx->dispatched[ilog2(queued) + 1]++;
625
626 /*
627 * Any items that need requeuing? Stuff them into hctx->dispatch,
628 * that is where we will continue on next queue run.
629 */
630 if (!list_empty(&rq_list)) {
631 spin_lock(&hctx->lock);
632 list_splice(&rq_list, &hctx->dispatch);
633 spin_unlock(&hctx->lock);
634 }
635}
636
637void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
638{
639 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
640 return;
641
642 if (!async)
643 __blk_mq_run_hw_queue(hctx);
644 else {
645 struct request_queue *q = hctx->queue;
646
647 kblockd_schedule_delayed_work(q, &hctx->delayed_work, 0);
648 }
649}
650
651void blk_mq_run_queues(struct request_queue *q, bool async)
652{
653 struct blk_mq_hw_ctx *hctx;
654 int i;
655
656 queue_for_each_hw_ctx(q, hctx, i) {
657 if ((!blk_mq_hctx_has_pending(hctx) &&
658 list_empty_careful(&hctx->dispatch)) ||
659 test_bit(BLK_MQ_S_STOPPED, &hctx->flags))
660 continue;
661
662 blk_mq_run_hw_queue(hctx, async);
663 }
664}
665EXPORT_SYMBOL(blk_mq_run_queues);
666
667void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
668{
669 cancel_delayed_work(&hctx->delayed_work);
670 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
671}
672EXPORT_SYMBOL(blk_mq_stop_hw_queue);
673
Christoph Hellwig280d45f2013-10-25 14:45:58 +0100674void blk_mq_stop_hw_queues(struct request_queue *q)
675{
676 struct blk_mq_hw_ctx *hctx;
677 int i;
678
679 queue_for_each_hw_ctx(q, hctx, i)
680 blk_mq_stop_hw_queue(hctx);
681}
682EXPORT_SYMBOL(blk_mq_stop_hw_queues);
683
Jens Axboe320ae512013-10-24 09:20:05 +0100684void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
685{
686 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
687 __blk_mq_run_hw_queue(hctx);
688}
689EXPORT_SYMBOL(blk_mq_start_hw_queue);
690
691void blk_mq_start_stopped_hw_queues(struct request_queue *q)
692{
693 struct blk_mq_hw_ctx *hctx;
694 int i;
695
696 queue_for_each_hw_ctx(q, hctx, i) {
697 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
698 continue;
699
700 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
701 blk_mq_run_hw_queue(hctx, true);
702 }
703}
704EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
705
706static void blk_mq_work_fn(struct work_struct *work)
707{
708 struct blk_mq_hw_ctx *hctx;
709
710 hctx = container_of(work, struct blk_mq_hw_ctx, delayed_work.work);
711 __blk_mq_run_hw_queue(hctx);
712}
713
714static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig72a0a362014-02-07 10:22:36 -0800715 struct request *rq, bool at_head)
Jens Axboe320ae512013-10-24 09:20:05 +0100716{
717 struct blk_mq_ctx *ctx = rq->mq_ctx;
718
Jens Axboe01b983c2013-11-19 18:59:10 -0700719 trace_block_rq_insert(hctx->queue, rq);
720
Christoph Hellwig72a0a362014-02-07 10:22:36 -0800721 if (at_head)
722 list_add(&rq->queuelist, &ctx->rq_list);
723 else
724 list_add_tail(&rq->queuelist, &ctx->rq_list);
Jens Axboe320ae512013-10-24 09:20:05 +0100725 blk_mq_hctx_mark_pending(hctx, ctx);
726
727 /*
728 * We do this early, to ensure we are on the right CPU.
729 */
730 blk_mq_add_timer(rq);
731}
732
733void blk_mq_insert_request(struct request_queue *q, struct request *rq,
Christoph Hellwig72a0a362014-02-07 10:22:36 -0800734 bool at_head, bool run_queue)
Jens Axboe320ae512013-10-24 09:20:05 +0100735{
736 struct blk_mq_hw_ctx *hctx;
737 struct blk_mq_ctx *ctx, *current_ctx;
738
739 ctx = rq->mq_ctx;
740 hctx = q->mq_ops->map_queue(q, ctx->cpu);
741
742 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA)) {
743 blk_insert_flush(rq);
744 } else {
745 current_ctx = blk_mq_get_ctx(q);
746
747 if (!cpu_online(ctx->cpu)) {
748 ctx = current_ctx;
749 hctx = q->mq_ops->map_queue(q, ctx->cpu);
750 rq->mq_ctx = ctx;
751 }
752 spin_lock(&ctx->lock);
Christoph Hellwig72a0a362014-02-07 10:22:36 -0800753 __blk_mq_insert_request(hctx, rq, at_head);
Jens Axboe320ae512013-10-24 09:20:05 +0100754 spin_unlock(&ctx->lock);
755
756 blk_mq_put_ctx(current_ctx);
757 }
758
759 if (run_queue)
760 __blk_mq_run_hw_queue(hctx);
761}
762EXPORT_SYMBOL(blk_mq_insert_request);
763
764/*
765 * This is a special version of blk_mq_insert_request to bypass FLUSH request
766 * check. Should only be used internally.
767 */
768void blk_mq_run_request(struct request *rq, bool run_queue, bool async)
769{
770 struct request_queue *q = rq->q;
771 struct blk_mq_hw_ctx *hctx;
772 struct blk_mq_ctx *ctx, *current_ctx;
773
774 current_ctx = blk_mq_get_ctx(q);
775
776 ctx = rq->mq_ctx;
777 if (!cpu_online(ctx->cpu)) {
778 ctx = current_ctx;
779 rq->mq_ctx = ctx;
780 }
781 hctx = q->mq_ops->map_queue(q, ctx->cpu);
782
783 /* ctx->cpu might be offline */
784 spin_lock(&ctx->lock);
Christoph Hellwig72a0a362014-02-07 10:22:36 -0800785 __blk_mq_insert_request(hctx, rq, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100786 spin_unlock(&ctx->lock);
787
788 blk_mq_put_ctx(current_ctx);
789
790 if (run_queue)
791 blk_mq_run_hw_queue(hctx, async);
792}
793
794static void blk_mq_insert_requests(struct request_queue *q,
795 struct blk_mq_ctx *ctx,
796 struct list_head *list,
797 int depth,
798 bool from_schedule)
799
800{
801 struct blk_mq_hw_ctx *hctx;
802 struct blk_mq_ctx *current_ctx;
803
804 trace_block_unplug(q, depth, !from_schedule);
805
806 current_ctx = blk_mq_get_ctx(q);
807
808 if (!cpu_online(ctx->cpu))
809 ctx = current_ctx;
810 hctx = q->mq_ops->map_queue(q, ctx->cpu);
811
812 /*
813 * preemption doesn't flush plug list, so it's possible ctx->cpu is
814 * offline now
815 */
816 spin_lock(&ctx->lock);
817 while (!list_empty(list)) {
818 struct request *rq;
819
820 rq = list_first_entry(list, struct request, queuelist);
821 list_del_init(&rq->queuelist);
822 rq->mq_ctx = ctx;
Christoph Hellwig72a0a362014-02-07 10:22:36 -0800823 __blk_mq_insert_request(hctx, rq, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100824 }
825 spin_unlock(&ctx->lock);
826
827 blk_mq_put_ctx(current_ctx);
828
829 blk_mq_run_hw_queue(hctx, from_schedule);
830}
831
832static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
833{
834 struct request *rqa = container_of(a, struct request, queuelist);
835 struct request *rqb = container_of(b, struct request, queuelist);
836
837 return !(rqa->mq_ctx < rqb->mq_ctx ||
838 (rqa->mq_ctx == rqb->mq_ctx &&
839 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
840}
841
842void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
843{
844 struct blk_mq_ctx *this_ctx;
845 struct request_queue *this_q;
846 struct request *rq;
847 LIST_HEAD(list);
848 LIST_HEAD(ctx_list);
849 unsigned int depth;
850
851 list_splice_init(&plug->mq_list, &list);
852
853 list_sort(NULL, &list, plug_ctx_cmp);
854
855 this_q = NULL;
856 this_ctx = NULL;
857 depth = 0;
858
859 while (!list_empty(&list)) {
860 rq = list_entry_rq(list.next);
861 list_del_init(&rq->queuelist);
862 BUG_ON(!rq->q);
863 if (rq->mq_ctx != this_ctx) {
864 if (this_ctx) {
865 blk_mq_insert_requests(this_q, this_ctx,
866 &ctx_list, depth,
867 from_schedule);
868 }
869
870 this_ctx = rq->mq_ctx;
871 this_q = rq->q;
872 depth = 0;
873 }
874
875 depth++;
876 list_add_tail(&rq->queuelist, &ctx_list);
877 }
878
879 /*
880 * If 'this_ctx' is set, we know we have entries to complete
881 * on 'ctx_list'. Do those.
882 */
883 if (this_ctx) {
884 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
885 from_schedule);
886 }
887}
888
889static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
890{
891 init_request_from_bio(rq, bio);
892 blk_account_io_start(rq, 1);
893}
894
895static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
896{
897 struct blk_mq_hw_ctx *hctx;
898 struct blk_mq_ctx *ctx;
899 const int is_sync = rw_is_sync(bio->bi_rw);
900 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
901 int rw = bio_data_dir(bio);
902 struct request *rq;
903 unsigned int use_plug, request_count = 0;
904
905 /*
906 * If we have multiple hardware queues, just go directly to
907 * one of those for sync IO.
908 */
909 use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
910
911 blk_queue_bounce(q, &bio);
912
Nicholas Bellinger14ec77f2014-02-07 13:45:39 -0700913 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
914 bio_endio(bio, -EIO);
915 return;
916 }
917
Jens Axboe320ae512013-10-24 09:20:05 +0100918 if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
919 return;
920
921 if (blk_mq_queue_enter(q)) {
922 bio_endio(bio, -EIO);
923 return;
924 }
925
926 ctx = blk_mq_get_ctx(q);
927 hctx = q->mq_ops->map_queue(q, ctx->cpu);
928
929 trace_block_getrq(q, bio, rw);
Christoph Hellwig18741982014-02-10 09:29:00 -0700930 rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100931 if (likely(rq))
Christoph Hellwig18741982014-02-10 09:29:00 -0700932 blk_mq_rq_ctx_init(q, ctx, rq, rw);
Jens Axboe320ae512013-10-24 09:20:05 +0100933 else {
934 blk_mq_put_ctx(ctx);
935 trace_block_sleeprq(q, bio, rw);
Christoph Hellwig18741982014-02-10 09:29:00 -0700936 rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
937 false);
Jens Axboe320ae512013-10-24 09:20:05 +0100938 ctx = rq->mq_ctx;
939 hctx = q->mq_ops->map_queue(q, ctx->cpu);
940 }
941
942 hctx->queued++;
943
944 if (unlikely(is_flush_fua)) {
945 blk_mq_bio_to_request(rq, bio);
946 blk_mq_put_ctx(ctx);
947 blk_insert_flush(rq);
948 goto run_queue;
949 }
950
951 /*
952 * A task plug currently exists. Since this is completely lockless,
953 * utilize that to temporarily store requests until the task is
954 * either done or scheduled away.
955 */
956 if (use_plug) {
957 struct blk_plug *plug = current->plug;
958
959 if (plug) {
960 blk_mq_bio_to_request(rq, bio);
Shaohua Li92f399c2013-10-29 12:01:03 -0600961 if (list_empty(&plug->mq_list))
Jens Axboe320ae512013-10-24 09:20:05 +0100962 trace_block_plug(q);
963 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
964 blk_flush_plug_list(plug, false);
965 trace_block_plug(q);
966 }
967 list_add_tail(&rq->queuelist, &plug->mq_list);
968 blk_mq_put_ctx(ctx);
969 return;
970 }
971 }
972
973 spin_lock(&ctx->lock);
974
975 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
976 blk_mq_attempt_merge(q, ctx, bio))
977 __blk_mq_free_request(hctx, ctx, rq);
978 else {
979 blk_mq_bio_to_request(rq, bio);
Christoph Hellwig72a0a362014-02-07 10:22:36 -0800980 __blk_mq_insert_request(hctx, rq, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100981 }
982
983 spin_unlock(&ctx->lock);
984 blk_mq_put_ctx(ctx);
985
986 /*
987 * For a SYNC request, send it to the hardware immediately. For an
988 * ASYNC request, just ensure that we run it later on. The latter
989 * allows for merging opportunities and more efficient dispatching.
990 */
991run_queue:
992 blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
993}
994
995/*
996 * Default mapping to a software queue, since we use one per CPU.
997 */
998struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
999{
1000 return q->queue_hw_ctx[q->mq_map[cpu]];
1001}
1002EXPORT_SYMBOL(blk_mq_map_queue);
1003
1004struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_reg *reg,
1005 unsigned int hctx_index)
1006{
1007 return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
1008 GFP_KERNEL | __GFP_ZERO, reg->numa_node);
1009}
1010EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
1011
1012void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
1013 unsigned int hctx_index)
1014{
1015 kfree(hctx);
1016}
1017EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
1018
1019static void blk_mq_hctx_notify(void *data, unsigned long action,
1020 unsigned int cpu)
1021{
1022 struct blk_mq_hw_ctx *hctx = data;
1023 struct blk_mq_ctx *ctx;
1024 LIST_HEAD(tmp);
1025
1026 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1027 return;
1028
1029 /*
1030 * Move ctx entries to new CPU, if this one is going away.
1031 */
1032 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1033
1034 spin_lock(&ctx->lock);
1035 if (!list_empty(&ctx->rq_list)) {
1036 list_splice_init(&ctx->rq_list, &tmp);
1037 clear_bit(ctx->index_hw, hctx->ctx_map);
1038 }
1039 spin_unlock(&ctx->lock);
1040
1041 if (list_empty(&tmp))
1042 return;
1043
1044 ctx = blk_mq_get_ctx(hctx->queue);
1045 spin_lock(&ctx->lock);
1046
1047 while (!list_empty(&tmp)) {
1048 struct request *rq;
1049
1050 rq = list_first_entry(&tmp, struct request, queuelist);
1051 rq->mq_ctx = ctx;
1052 list_move_tail(&rq->queuelist, &ctx->rq_list);
1053 }
1054
1055 blk_mq_hctx_mark_pending(hctx, ctx);
1056
1057 spin_unlock(&ctx->lock);
1058 blk_mq_put_ctx(ctx);
1059}
1060
1061static void blk_mq_init_hw_commands(struct blk_mq_hw_ctx *hctx,
1062 void (*init)(void *, struct blk_mq_hw_ctx *,
1063 struct request *, unsigned int),
1064 void *data)
1065{
1066 unsigned int i;
1067
1068 for (i = 0; i < hctx->queue_depth; i++) {
1069 struct request *rq = hctx->rqs[i];
1070
1071 init(data, hctx, rq, i);
1072 }
1073}
1074
1075void blk_mq_init_commands(struct request_queue *q,
1076 void (*init)(void *, struct blk_mq_hw_ctx *,
1077 struct request *, unsigned int),
1078 void *data)
1079{
1080 struct blk_mq_hw_ctx *hctx;
1081 unsigned int i;
1082
1083 queue_for_each_hw_ctx(q, hctx, i)
1084 blk_mq_init_hw_commands(hctx, init, data);
1085}
1086EXPORT_SYMBOL(blk_mq_init_commands);
1087
1088static void blk_mq_free_rq_map(struct blk_mq_hw_ctx *hctx)
1089{
1090 struct page *page;
1091
1092 while (!list_empty(&hctx->page_list)) {
Dave Hansen67534712014-01-08 20:17:46 -07001093 page = list_first_entry(&hctx->page_list, struct page, lru);
1094 list_del_init(&page->lru);
Jens Axboe320ae512013-10-24 09:20:05 +01001095 __free_pages(page, page->private);
1096 }
1097
1098 kfree(hctx->rqs);
1099
1100 if (hctx->tags)
1101 blk_mq_free_tags(hctx->tags);
1102}
1103
1104static size_t order_to_size(unsigned int order)
1105{
1106 size_t ret = PAGE_SIZE;
1107
1108 while (order--)
1109 ret *= 2;
1110
1111 return ret;
1112}
1113
1114static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
1115 unsigned int reserved_tags, int node)
1116{
1117 unsigned int i, j, entries_per_page, max_order = 4;
1118 size_t rq_size, left;
1119
1120 INIT_LIST_HEAD(&hctx->page_list);
1121
1122 hctx->rqs = kmalloc_node(hctx->queue_depth * sizeof(struct request *),
1123 GFP_KERNEL, node);
1124 if (!hctx->rqs)
1125 return -ENOMEM;
1126
1127 /*
1128 * rq_size is the size of the request plus driver payload, rounded
1129 * to the cacheline size
1130 */
1131 rq_size = round_up(sizeof(struct request) + hctx->cmd_size,
1132 cache_line_size());
1133 left = rq_size * hctx->queue_depth;
1134
1135 for (i = 0; i < hctx->queue_depth;) {
1136 int this_order = max_order;
1137 struct page *page;
1138 int to_do;
1139 void *p;
1140
1141 while (left < order_to_size(this_order - 1) && this_order)
1142 this_order--;
1143
1144 do {
1145 page = alloc_pages_node(node, GFP_KERNEL, this_order);
1146 if (page)
1147 break;
1148 if (!this_order--)
1149 break;
1150 if (order_to_size(this_order) < rq_size)
1151 break;
1152 } while (1);
1153
1154 if (!page)
1155 break;
1156
1157 page->private = this_order;
Dave Hansen67534712014-01-08 20:17:46 -07001158 list_add_tail(&page->lru, &hctx->page_list);
Jens Axboe320ae512013-10-24 09:20:05 +01001159
1160 p = page_address(page);
1161 entries_per_page = order_to_size(this_order) / rq_size;
1162 to_do = min(entries_per_page, hctx->queue_depth - i);
1163 left -= to_do * rq_size;
1164 for (j = 0; j < to_do; j++) {
1165 hctx->rqs[i] = p;
1166 blk_mq_rq_init(hctx, hctx->rqs[i]);
1167 p += rq_size;
1168 i++;
1169 }
1170 }
1171
1172 if (i < (reserved_tags + BLK_MQ_TAG_MIN))
1173 goto err_rq_map;
1174 else if (i != hctx->queue_depth) {
1175 hctx->queue_depth = i;
1176 pr_warn("%s: queue depth set to %u because of low memory\n",
1177 __func__, i);
1178 }
1179
1180 hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node);
1181 if (!hctx->tags) {
1182err_rq_map:
1183 blk_mq_free_rq_map(hctx);
1184 return -ENOMEM;
1185 }
1186
1187 return 0;
1188}
1189
1190static int blk_mq_init_hw_queues(struct request_queue *q,
1191 struct blk_mq_reg *reg, void *driver_data)
1192{
1193 struct blk_mq_hw_ctx *hctx;
1194 unsigned int i, j;
1195
1196 /*
1197 * Initialize hardware queues
1198 */
1199 queue_for_each_hw_ctx(q, hctx, i) {
1200 unsigned int num_maps;
1201 int node;
1202
1203 node = hctx->numa_node;
1204 if (node == NUMA_NO_NODE)
1205 node = hctx->numa_node = reg->numa_node;
1206
1207 INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn);
1208 spin_lock_init(&hctx->lock);
1209 INIT_LIST_HEAD(&hctx->dispatch);
1210 hctx->queue = q;
1211 hctx->queue_num = i;
1212 hctx->flags = reg->flags;
1213 hctx->queue_depth = reg->queue_depth;
1214 hctx->cmd_size = reg->cmd_size;
1215
1216 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1217 blk_mq_hctx_notify, hctx);
1218 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1219
1220 if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node))
1221 break;
1222
1223 /*
1224 * Allocate space for all possible cpus to avoid allocation in
1225 * runtime
1226 */
1227 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1228 GFP_KERNEL, node);
1229 if (!hctx->ctxs)
1230 break;
1231
1232 num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1233 hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1234 GFP_KERNEL, node);
1235 if (!hctx->ctx_map)
1236 break;
1237
1238 hctx->nr_ctx_map = num_maps;
1239 hctx->nr_ctx = 0;
1240
1241 if (reg->ops->init_hctx &&
1242 reg->ops->init_hctx(hctx, driver_data, i))
1243 break;
1244 }
1245
1246 if (i == q->nr_hw_queues)
1247 return 0;
1248
1249 /*
1250 * Init failed
1251 */
1252 queue_for_each_hw_ctx(q, hctx, j) {
1253 if (i == j)
1254 break;
1255
1256 if (reg->ops->exit_hctx)
1257 reg->ops->exit_hctx(hctx, j);
1258
1259 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1260 blk_mq_free_rq_map(hctx);
1261 kfree(hctx->ctxs);
1262 }
1263
1264 return 1;
1265}
1266
1267static void blk_mq_init_cpu_queues(struct request_queue *q,
1268 unsigned int nr_hw_queues)
1269{
1270 unsigned int i;
1271
1272 for_each_possible_cpu(i) {
1273 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1274 struct blk_mq_hw_ctx *hctx;
1275
1276 memset(__ctx, 0, sizeof(*__ctx));
1277 __ctx->cpu = i;
1278 spin_lock_init(&__ctx->lock);
1279 INIT_LIST_HEAD(&__ctx->rq_list);
1280 __ctx->queue = q;
1281
1282 /* If the cpu isn't online, the cpu is mapped to first hctx */
1283 hctx = q->mq_ops->map_queue(q, i);
1284 hctx->nr_ctx++;
1285
1286 if (!cpu_online(i))
1287 continue;
1288
1289 /*
1290 * Set local node, IFF we have more than one hw queue. If
1291 * not, we remain on the home node of the device
1292 */
1293 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1294 hctx->numa_node = cpu_to_node(i);
1295 }
1296}
1297
1298static void blk_mq_map_swqueue(struct request_queue *q)
1299{
1300 unsigned int i;
1301 struct blk_mq_hw_ctx *hctx;
1302 struct blk_mq_ctx *ctx;
1303
1304 queue_for_each_hw_ctx(q, hctx, i) {
1305 hctx->nr_ctx = 0;
1306 }
1307
1308 /*
1309 * Map software to hardware queues
1310 */
1311 queue_for_each_ctx(q, ctx, i) {
1312 /* If the cpu isn't online, the cpu is mapped to first hctx */
1313 hctx = q->mq_ops->map_queue(q, i);
1314 ctx->index_hw = hctx->nr_ctx;
1315 hctx->ctxs[hctx->nr_ctx++] = ctx;
1316 }
1317}
1318
1319struct request_queue *blk_mq_init_queue(struct blk_mq_reg *reg,
1320 void *driver_data)
1321{
1322 struct blk_mq_hw_ctx **hctxs;
1323 struct blk_mq_ctx *ctx;
1324 struct request_queue *q;
1325 int i;
1326
1327 if (!reg->nr_hw_queues ||
1328 !reg->ops->queue_rq || !reg->ops->map_queue ||
1329 !reg->ops->alloc_hctx || !reg->ops->free_hctx)
1330 return ERR_PTR(-EINVAL);
1331
1332 if (!reg->queue_depth)
1333 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1334 else if (reg->queue_depth > BLK_MQ_MAX_DEPTH) {
1335 pr_err("blk-mq: queuedepth too large (%u)\n", reg->queue_depth);
1336 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1337 }
1338
Jens Axboe320ae512013-10-24 09:20:05 +01001339 if (reg->queue_depth < (reg->reserved_tags + BLK_MQ_TAG_MIN))
1340 return ERR_PTR(-EINVAL);
1341
1342 ctx = alloc_percpu(struct blk_mq_ctx);
1343 if (!ctx)
1344 return ERR_PTR(-ENOMEM);
1345
1346 hctxs = kmalloc_node(reg->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1347 reg->numa_node);
1348
1349 if (!hctxs)
1350 goto err_percpu;
1351
1352 for (i = 0; i < reg->nr_hw_queues; i++) {
1353 hctxs[i] = reg->ops->alloc_hctx(reg, i);
1354 if (!hctxs[i])
1355 goto err_hctxs;
1356
1357 hctxs[i]->numa_node = NUMA_NO_NODE;
1358 hctxs[i]->queue_num = i;
1359 }
1360
1361 q = blk_alloc_queue_node(GFP_KERNEL, reg->numa_node);
1362 if (!q)
1363 goto err_hctxs;
1364
1365 q->mq_map = blk_mq_make_queue_map(reg);
1366 if (!q->mq_map)
1367 goto err_map;
1368
1369 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1370 blk_queue_rq_timeout(q, 30000);
1371
1372 q->nr_queues = nr_cpu_ids;
1373 q->nr_hw_queues = reg->nr_hw_queues;
1374
1375 q->queue_ctx = ctx;
1376 q->queue_hw_ctx = hctxs;
1377
1378 q->mq_ops = reg->ops;
Jens Axboe94eddfb2013-11-19 09:25:07 -07001379 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
Jens Axboe320ae512013-10-24 09:20:05 +01001380
Christoph Hellwig1be036e2014-02-07 10:22:39 -08001381 q->sg_reserved_size = INT_MAX;
1382
Jens Axboe320ae512013-10-24 09:20:05 +01001383 blk_queue_make_request(q, blk_mq_make_request);
1384 blk_queue_rq_timed_out(q, reg->ops->timeout);
1385 if (reg->timeout)
1386 blk_queue_rq_timeout(q, reg->timeout);
1387
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08001388 if (reg->ops->complete)
1389 blk_queue_softirq_done(q, reg->ops->complete);
1390
Jens Axboe320ae512013-10-24 09:20:05 +01001391 blk_mq_init_flush(q);
1392 blk_mq_init_cpu_queues(q, reg->nr_hw_queues);
1393
Christoph Hellwig18741982014-02-10 09:29:00 -07001394 q->flush_rq = kzalloc(round_up(sizeof(struct request) + reg->cmd_size,
1395 cache_line_size()), GFP_KERNEL);
1396 if (!q->flush_rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001397 goto err_hw;
1398
Christoph Hellwig18741982014-02-10 09:29:00 -07001399 if (blk_mq_init_hw_queues(q, reg, driver_data))
1400 goto err_flush_rq;
1401
Jens Axboe320ae512013-10-24 09:20:05 +01001402 blk_mq_map_swqueue(q);
1403
1404 mutex_lock(&all_q_mutex);
1405 list_add_tail(&q->all_q_node, &all_q_list);
1406 mutex_unlock(&all_q_mutex);
1407
1408 return q;
Christoph Hellwig18741982014-02-10 09:29:00 -07001409
1410err_flush_rq:
1411 kfree(q->flush_rq);
Jens Axboe320ae512013-10-24 09:20:05 +01001412err_hw:
1413 kfree(q->mq_map);
1414err_map:
1415 blk_cleanup_queue(q);
1416err_hctxs:
1417 for (i = 0; i < reg->nr_hw_queues; i++) {
1418 if (!hctxs[i])
1419 break;
1420 reg->ops->free_hctx(hctxs[i], i);
1421 }
1422 kfree(hctxs);
1423err_percpu:
1424 free_percpu(ctx);
1425 return ERR_PTR(-ENOMEM);
1426}
1427EXPORT_SYMBOL(blk_mq_init_queue);
1428
1429void blk_mq_free_queue(struct request_queue *q)
1430{
1431 struct blk_mq_hw_ctx *hctx;
1432 int i;
1433
1434 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboe320ae512013-10-24 09:20:05 +01001435 kfree(hctx->ctx_map);
1436 kfree(hctx->ctxs);
1437 blk_mq_free_rq_map(hctx);
1438 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1439 if (q->mq_ops->exit_hctx)
1440 q->mq_ops->exit_hctx(hctx, i);
1441 q->mq_ops->free_hctx(hctx, i);
1442 }
1443
1444 free_percpu(q->queue_ctx);
1445 kfree(q->queue_hw_ctx);
1446 kfree(q->mq_map);
1447
1448 q->queue_ctx = NULL;
1449 q->queue_hw_ctx = NULL;
1450 q->mq_map = NULL;
1451
1452 mutex_lock(&all_q_mutex);
1453 list_del_init(&q->all_q_node);
1454 mutex_unlock(&all_q_mutex);
1455}
Jens Axboe320ae512013-10-24 09:20:05 +01001456
1457/* Basically redo blk_mq_init_queue with queue frozen */
Paul Gortmakerf618ef72013-11-14 08:26:02 -07001458static void blk_mq_queue_reinit(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01001459{
1460 blk_mq_freeze_queue(q);
1461
1462 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1463
1464 /*
1465 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1466 * we should change hctx numa_node according to new topology (this
1467 * involves free and re-allocate memory, worthy doing?)
1468 */
1469
1470 blk_mq_map_swqueue(q);
1471
1472 blk_mq_unfreeze_queue(q);
1473}
1474
Paul Gortmakerf618ef72013-11-14 08:26:02 -07001475static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1476 unsigned long action, void *hcpu)
Jens Axboe320ae512013-10-24 09:20:05 +01001477{
1478 struct request_queue *q;
1479
1480 /*
1481 * Before new mapping is established, hotadded cpu might already start
1482 * handling requests. This doesn't break anything as we map offline
1483 * CPUs to first hardware queue. We will re-init queue below to get
1484 * optimal settings.
1485 */
1486 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1487 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1488 return NOTIFY_OK;
1489
1490 mutex_lock(&all_q_mutex);
1491 list_for_each_entry(q, &all_q_list, all_q_node)
1492 blk_mq_queue_reinit(q);
1493 mutex_unlock(&all_q_mutex);
1494 return NOTIFY_OK;
1495}
1496
1497static int __init blk_mq_init(void)
1498{
Jens Axboe320ae512013-10-24 09:20:05 +01001499 blk_mq_cpu_init();
1500
1501 /* Must be called after percpu_counter_hotcpu_callback() */
1502 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1503
1504 return 0;
1505}
1506subsys_initcall(blk_mq_init);