Linux Audio

Check our new training course

Loading...
v6.8
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Copyright (C) 2016 Red Hat, Inc. All rights reserved.
  4 *
  5 * This file is released under the GPL.
  6 */
  7
  8#include "dm-core.h"
  9#include "dm-rq.h"
 10
 11#include <linux/blk-mq.h>
 12
 13#define DM_MSG_PREFIX "core-rq"
 14
 15/*
 16 * One of these is allocated per request.
 17 */
 18struct dm_rq_target_io {
 19	struct mapped_device *md;
 20	struct dm_target *ti;
 21	struct request *orig, *clone;
 22	struct kthread_work work;
 23	blk_status_t error;
 24	union map_info info;
 25	struct dm_stats_aux stats_aux;
 26	unsigned long duration_jiffies;
 27	unsigned int n_sectors;
 28	unsigned int completed;
 29};
 30
 31#define DM_MQ_NR_HW_QUEUES 1
 32#define DM_MQ_QUEUE_DEPTH 2048
 33static unsigned int dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
 34static unsigned int dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
 35
 36/*
 37 * Request-based DM's mempools' reserved IOs set by the user.
 38 */
 39#define RESERVED_REQUEST_BASED_IOS	256
 40static unsigned int reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
 41
 42unsigned int dm_get_reserved_rq_based_ios(void)
 43{
 44	return __dm_get_module_param(&reserved_rq_based_ios,
 45				     RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
 46}
 47
 48static unsigned int dm_get_blk_mq_nr_hw_queues(void)
 49{
 50	return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
 51}
 52
 53static unsigned int dm_get_blk_mq_queue_depth(void)
 54{
 55	return __dm_get_module_param(&dm_mq_queue_depth,
 56				     DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
 57}
 58
 59int dm_request_based(struct mapped_device *md)
 60{
 61	return queue_is_mq(md->queue);
 62}
 63
 64void dm_start_queue(struct request_queue *q)
 65{
 66	blk_mq_unquiesce_queue(q);
 67	blk_mq_kick_requeue_list(q);
 68}
 69
 70void dm_stop_queue(struct request_queue *q)
 71{
 72	blk_mq_quiesce_queue(q);
 73}
 74
 75/*
 76 * Partial completion handling for request-based dm
 77 */
 78static void end_clone_bio(struct bio *clone)
 79{
 80	struct dm_rq_clone_bio_info *info =
 81		container_of(clone, struct dm_rq_clone_bio_info, clone);
 82	struct dm_rq_target_io *tio = info->tio;
 83	unsigned int nr_bytes = info->orig->bi_iter.bi_size;
 84	blk_status_t error = clone->bi_status;
 85	bool is_last = !clone->bi_next;
 86
 87	bio_put(clone);
 88
 89	if (tio->error)
 90		/*
 91		 * An error has already been detected on the request.
 92		 * Once error occurred, just let clone->end_io() handle
 93		 * the remainder.
 94		 */
 95		return;
 96	else if (error) {
 97		/*
 98		 * Don't notice the error to the upper layer yet.
 99		 * The error handling decision is made by the target driver,
100		 * when the request is completed.
101		 */
102		tio->error = error;
103		goto exit;
104	}
105
106	/*
107	 * I/O for the bio successfully completed.
108	 * Notice the data completion to the upper layer.
109	 */
110	tio->completed += nr_bytes;
111
112	/*
113	 * Update the original request.
114	 * Do not use blk_mq_end_request() here, because it may complete
115	 * the original request before the clone, and break the ordering.
116	 */
117	if (is_last)
118 exit:
119		blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
120}
121
122static struct dm_rq_target_io *tio_from_request(struct request *rq)
123{
124	return blk_mq_rq_to_pdu(rq);
125}
126
127static void rq_end_stats(struct mapped_device *md, struct request *orig)
128{
129	if (unlikely(dm_stats_used(&md->stats))) {
130		struct dm_rq_target_io *tio = tio_from_request(orig);
131
132		tio->duration_jiffies = jiffies - tio->duration_jiffies;
133		dm_stats_account_io(&md->stats, rq_data_dir(orig),
134				    blk_rq_pos(orig), tio->n_sectors, true,
135				    tio->duration_jiffies, &tio->stats_aux);
136	}
137}
138
139/*
140 * Don't touch any member of the md after calling this function because
141 * the md may be freed in dm_put() at the end of this function.
142 * Or do dm_get() before calling this function and dm_put() later.
143 */
144static void rq_completed(struct mapped_device *md)
145{
146	/*
147	 * dm_put() must be at the end of this function. See the comment above
148	 */
149	dm_put(md);
150}
151
152/*
153 * Complete the clone and the original request.
154 * Must be called without clone's queue lock held,
155 * see end_clone_request() for more details.
156 */
157static void dm_end_request(struct request *clone, blk_status_t error)
158{
159	struct dm_rq_target_io *tio = clone->end_io_data;
160	struct mapped_device *md = tio->md;
161	struct request *rq = tio->orig;
162
163	blk_rq_unprep_clone(clone);
164	tio->ti->type->release_clone_rq(clone, NULL);
165
166	rq_end_stats(md, rq);
167	blk_mq_end_request(rq, error);
168	rq_completed(md);
169}
170
171static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
172{
173	blk_mq_delay_kick_requeue_list(q, msecs);
174}
175
176void dm_mq_kick_requeue_list(struct mapped_device *md)
177{
178	__dm_mq_kick_requeue_list(md->queue, 0);
179}
180EXPORT_SYMBOL(dm_mq_kick_requeue_list);
181
182static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
183{
184	blk_mq_requeue_request(rq, false);
185	__dm_mq_kick_requeue_list(rq->q, msecs);
186}
187
188static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
189{
190	struct mapped_device *md = tio->md;
191	struct request *rq = tio->orig;
192	unsigned long delay_ms = delay_requeue ? 100 : 0;
193
194	rq_end_stats(md, rq);
195	if (tio->clone) {
196		blk_rq_unprep_clone(tio->clone);
197		tio->ti->type->release_clone_rq(tio->clone, NULL);
198	}
199
200	dm_mq_delay_requeue_request(rq, delay_ms);
201	rq_completed(md);
202}
203
204static void dm_done(struct request *clone, blk_status_t error, bool mapped)
205{
206	int r = DM_ENDIO_DONE;
207	struct dm_rq_target_io *tio = clone->end_io_data;
208	dm_request_endio_fn rq_end_io = NULL;
209
210	if (tio->ti) {
211		rq_end_io = tio->ti->type->rq_end_io;
212
213		if (mapped && rq_end_io)
214			r = rq_end_io(tio->ti, clone, error, &tio->info);
215	}
216
217	if (unlikely(error == BLK_STS_TARGET)) {
218		if (req_op(clone) == REQ_OP_DISCARD &&
219		    !clone->q->limits.max_discard_sectors)
220			disable_discard(tio->md);
221		else if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
222			 !clone->q->limits.max_write_zeroes_sectors)
223			disable_write_zeroes(tio->md);
224	}
225
226	switch (r) {
227	case DM_ENDIO_DONE:
228		/* The target wants to complete the I/O */
229		dm_end_request(clone, error);
230		break;
231	case DM_ENDIO_INCOMPLETE:
232		/* The target will handle the I/O */
233		return;
234	case DM_ENDIO_REQUEUE:
235		/* The target wants to requeue the I/O */
236		dm_requeue_original_request(tio, false);
237		break;
238	case DM_ENDIO_DELAY_REQUEUE:
239		/* The target wants to requeue the I/O after a delay */
240		dm_requeue_original_request(tio, true);
241		break;
242	default:
243		DMCRIT("unimplemented target endio return value: %d", r);
244		BUG();
245	}
246}
247
248/*
249 * Request completion handler for request-based dm
250 */
251static void dm_softirq_done(struct request *rq)
252{
253	bool mapped = true;
254	struct dm_rq_target_io *tio = tio_from_request(rq);
255	struct request *clone = tio->clone;
256
257	if (!clone) {
258		struct mapped_device *md = tio->md;
259
260		rq_end_stats(md, rq);
261		blk_mq_end_request(rq, tio->error);
262		rq_completed(md);
263		return;
264	}
265
266	if (rq->rq_flags & RQF_FAILED)
267		mapped = false;
268
269	dm_done(clone, tio->error, mapped);
270}
271
272/*
273 * Complete the clone and the original request with the error status
274 * through softirq context.
275 */
276static void dm_complete_request(struct request *rq, blk_status_t error)
277{
278	struct dm_rq_target_io *tio = tio_from_request(rq);
279
280	tio->error = error;
281	if (likely(!blk_should_fake_timeout(rq->q)))
282		blk_mq_complete_request(rq);
283}
284
285/*
286 * Complete the not-mapped clone and the original request with the error status
287 * through softirq context.
288 * Target's rq_end_io() function isn't called.
289 * This may be used when the target's clone_and_map_rq() function fails.
290 */
291static void dm_kill_unmapped_request(struct request *rq, blk_status_t error)
292{
293	rq->rq_flags |= RQF_FAILED;
294	dm_complete_request(rq, error);
295}
296
297static enum rq_end_io_ret end_clone_request(struct request *clone,
298					    blk_status_t error)
299{
300	struct dm_rq_target_io *tio = clone->end_io_data;
301
302	dm_complete_request(tio->orig, error);
303	return RQ_END_IO_NONE;
304}
305
306static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
307				 void *data)
308{
309	struct dm_rq_target_io *tio = data;
310	struct dm_rq_clone_bio_info *info =
311		container_of(bio, struct dm_rq_clone_bio_info, clone);
312
313	info->orig = bio_orig;
314	info->tio = tio;
315	bio->bi_end_io = end_clone_bio;
316
317	return 0;
318}
319
320static int setup_clone(struct request *clone, struct request *rq,
321		       struct dm_rq_target_io *tio, gfp_t gfp_mask)
322{
323	int r;
324
325	r = blk_rq_prep_clone(clone, rq, &tio->md->mempools->bs, gfp_mask,
326			      dm_rq_bio_constructor, tio);
327	if (r)
328		return r;
329
330	clone->end_io = end_clone_request;
331	clone->end_io_data = tio;
332
333	tio->clone = clone;
334
335	return 0;
336}
337
338static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
339		     struct mapped_device *md)
340{
341	tio->md = md;
342	tio->ti = NULL;
343	tio->clone = NULL;
344	tio->orig = rq;
345	tio->error = 0;
346	tio->completed = 0;
347	/*
348	 * Avoid initializing info for blk-mq; it passes
349	 * target-specific data through info.ptr
350	 * (see: dm_mq_init_request)
351	 */
352	if (!md->init_tio_pdu)
353		memset(&tio->info, 0, sizeof(tio->info));
354}
355
356/*
357 * Returns:
358 * DM_MAPIO_*       : the request has been processed as indicated
359 * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
360 * < 0              : the request was completed due to failure
361 */
362static int map_request(struct dm_rq_target_io *tio)
363{
364	int r;
365	struct dm_target *ti = tio->ti;
366	struct mapped_device *md = tio->md;
367	struct request *rq = tio->orig;
368	struct request *clone = NULL;
369	blk_status_t ret;
370
371	r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
372	switch (r) {
373	case DM_MAPIO_SUBMITTED:
374		/* The target has taken the I/O to submit by itself later */
375		break;
376	case DM_MAPIO_REMAPPED:
377		if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
378			/* -ENOMEM */
379			ti->type->release_clone_rq(clone, &tio->info);
380			return DM_MAPIO_REQUEUE;
381		}
382
383		/* The target has remapped the I/O so dispatch it */
384		trace_block_rq_remap(clone, disk_devt(dm_disk(md)),
385				     blk_rq_pos(rq));
386		ret = blk_insert_cloned_request(clone);
387		switch (ret) {
388		case BLK_STS_OK:
389			break;
390		case BLK_STS_RESOURCE:
391		case BLK_STS_DEV_RESOURCE:
392			blk_rq_unprep_clone(clone);
393			blk_mq_cleanup_rq(clone);
394			tio->ti->type->release_clone_rq(clone, &tio->info);
395			tio->clone = NULL;
396			return DM_MAPIO_REQUEUE;
397		default:
398			/* must complete clone in terms of original request */
399			dm_complete_request(rq, ret);
400		}
401		break;
402	case DM_MAPIO_REQUEUE:
403		/* The target wants to requeue the I/O */
404		break;
405	case DM_MAPIO_DELAY_REQUEUE:
406		/* The target wants to requeue the I/O after a delay */
407		dm_requeue_original_request(tio, true);
408		break;
409	case DM_MAPIO_KILL:
410		/* The target wants to complete the I/O */
411		dm_kill_unmapped_request(rq, BLK_STS_IOERR);
412		break;
413	default:
414		DMCRIT("unimplemented target map return value: %d", r);
415		BUG();
416	}
417
418	return r;
419}
420
421/* DEPRECATED: previously used for request-based merge heuristic in dm_request_fn() */
422ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
423{
424	return sprintf(buf, "%u\n", 0);
425}
426
427ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
428						     const char *buf, size_t count)
429{
430	return count;
431}
432
433static void dm_start_request(struct mapped_device *md, struct request *orig)
434{
435	blk_mq_start_request(orig);
436
437	if (unlikely(dm_stats_used(&md->stats))) {
438		struct dm_rq_target_io *tio = tio_from_request(orig);
439
440		tio->duration_jiffies = jiffies;
441		tio->n_sectors = blk_rq_sectors(orig);
442		dm_stats_account_io(&md->stats, rq_data_dir(orig),
443				    blk_rq_pos(orig), tio->n_sectors, false, 0,
444				    &tio->stats_aux);
445	}
446
447	/*
448	 * Hold the md reference here for the in-flight I/O.
449	 * We can't rely on the reference count by device opener,
450	 * because the device may be closed during the request completion
451	 * when all bios are completed.
452	 * See the comment in rq_completed() too.
453	 */
454	dm_get(md);
455}
456
457static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
458			      unsigned int hctx_idx, unsigned int numa_node)
459{
460	struct mapped_device *md = set->driver_data;
461	struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
462
463	/*
464	 * Must initialize md member of tio, otherwise it won't
465	 * be available in dm_mq_queue_rq.
466	 */
467	tio->md = md;
468
469	if (md->init_tio_pdu) {
470		/* target-specific per-io data is immediately after the tio */
471		tio->info.ptr = tio + 1;
472	}
473
474	return 0;
475}
476
477static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
478			  const struct blk_mq_queue_data *bd)
479{
480	struct request *rq = bd->rq;
481	struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
482	struct mapped_device *md = tio->md;
483	struct dm_target *ti = md->immutable_target;
484
485	/*
486	 * blk-mq's unquiesce may come from outside events, such as
487	 * elevator switch, updating nr_requests or others, and request may
488	 * come during suspend, so simply ask for blk-mq to requeue it.
489	 */
490	if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)))
491		return BLK_STS_RESOURCE;
492
493	if (unlikely(!ti)) {
494		int srcu_idx;
495		struct dm_table *map;
496
497		map = dm_get_live_table(md, &srcu_idx);
498		if (unlikely(!map)) {
499			dm_put_live_table(md, srcu_idx);
500			return BLK_STS_RESOURCE;
501		}
502		ti = dm_table_find_target(map, 0);
503		dm_put_live_table(md, srcu_idx);
504	}
505
506	if (ti->type->busy && ti->type->busy(ti))
507		return BLK_STS_RESOURCE;
508
509	dm_start_request(md, rq);
510
511	/* Init tio using md established in .init_request */
512	init_tio(tio, rq, md);
513
514	/*
515	 * Establish tio->ti before calling map_request().
516	 */
517	tio->ti = ti;
518
519	/* Direct call is fine since .queue_rq allows allocations */
520	if (map_request(tio) == DM_MAPIO_REQUEUE) {
521		/* Undo dm_start_request() before requeuing */
522		rq_end_stats(md, rq);
523		rq_completed(md);
524		return BLK_STS_RESOURCE;
525	}
526
527	return BLK_STS_OK;
528}
529
530static const struct blk_mq_ops dm_mq_ops = {
531	.queue_rq = dm_mq_queue_rq,
532	.complete = dm_softirq_done,
533	.init_request = dm_mq_init_request,
534};
535
536int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
537{
538	struct dm_target *immutable_tgt;
539	int err;
540
541	md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
542	if (!md->tag_set)
543		return -ENOMEM;
544
545	md->tag_set->ops = &dm_mq_ops;
546	md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
547	md->tag_set->numa_node = md->numa_node_id;
548	md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_STACKING;
549	md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
550	md->tag_set->driver_data = md;
551
552	md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
553	immutable_tgt = dm_table_get_immutable_target(t);
554	if (immutable_tgt && immutable_tgt->per_io_data_size) {
555		/* any target-specific per-io data is immediately after the tio */
556		md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
557		md->init_tio_pdu = true;
558	}
559
560	err = blk_mq_alloc_tag_set(md->tag_set);
561	if (err)
562		goto out_kfree_tag_set;
563
564	err = blk_mq_init_allocated_queue(md->tag_set, md->queue);
565	if (err)
566		goto out_tag_set;
567	return 0;
568
569out_tag_set:
570	blk_mq_free_tag_set(md->tag_set);
571out_kfree_tag_set:
572	kfree(md->tag_set);
573	md->tag_set = NULL;
574
575	return err;
576}
577
578void dm_mq_cleanup_mapped_device(struct mapped_device *md)
579{
580	if (md->tag_set) {
581		blk_mq_free_tag_set(md->tag_set);
582		kfree(md->tag_set);
583		md->tag_set = NULL;
584	}
585}
586
587module_param(reserved_rq_based_ios, uint, 0644);
588MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
589
590/* Unused, but preserved for userspace compatibility */
591static bool use_blk_mq = true;
592module_param(use_blk_mq, bool, 0644);
593MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
594
595module_param(dm_mq_nr_hw_queues, uint, 0644);
596MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
597
598module_param(dm_mq_queue_depth, uint, 0644);
599MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");
v6.2
 
  1/*
  2 * Copyright (C) 2016 Red Hat, Inc. All rights reserved.
  3 *
  4 * This file is released under the GPL.
  5 */
  6
  7#include "dm-core.h"
  8#include "dm-rq.h"
  9
 10#include <linux/blk-mq.h>
 11
 12#define DM_MSG_PREFIX "core-rq"
 13
 14/*
 15 * One of these is allocated per request.
 16 */
 17struct dm_rq_target_io {
 18	struct mapped_device *md;
 19	struct dm_target *ti;
 20	struct request *orig, *clone;
 21	struct kthread_work work;
 22	blk_status_t error;
 23	union map_info info;
 24	struct dm_stats_aux stats_aux;
 25	unsigned long duration_jiffies;
 26	unsigned n_sectors;
 27	unsigned completed;
 28};
 29
 30#define DM_MQ_NR_HW_QUEUES 1
 31#define DM_MQ_QUEUE_DEPTH 2048
 32static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
 33static unsigned dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
 34
 35/*
 36 * Request-based DM's mempools' reserved IOs set by the user.
 37 */
 38#define RESERVED_REQUEST_BASED_IOS	256
 39static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
 40
 41unsigned dm_get_reserved_rq_based_ios(void)
 42{
 43	return __dm_get_module_param(&reserved_rq_based_ios,
 44				     RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
 45}
 46
 47static unsigned dm_get_blk_mq_nr_hw_queues(void)
 48{
 49	return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
 50}
 51
 52static unsigned dm_get_blk_mq_queue_depth(void)
 53{
 54	return __dm_get_module_param(&dm_mq_queue_depth,
 55				     DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
 56}
 57
 58int dm_request_based(struct mapped_device *md)
 59{
 60	return queue_is_mq(md->queue);
 61}
 62
 63void dm_start_queue(struct request_queue *q)
 64{
 65	blk_mq_unquiesce_queue(q);
 66	blk_mq_kick_requeue_list(q);
 67}
 68
 69void dm_stop_queue(struct request_queue *q)
 70{
 71	blk_mq_quiesce_queue(q);
 72}
 73
 74/*
 75 * Partial completion handling for request-based dm
 76 */
 77static void end_clone_bio(struct bio *clone)
 78{
 79	struct dm_rq_clone_bio_info *info =
 80		container_of(clone, struct dm_rq_clone_bio_info, clone);
 81	struct dm_rq_target_io *tio = info->tio;
 82	unsigned int nr_bytes = info->orig->bi_iter.bi_size;
 83	blk_status_t error = clone->bi_status;
 84	bool is_last = !clone->bi_next;
 85
 86	bio_put(clone);
 87
 88	if (tio->error)
 89		/*
 90		 * An error has already been detected on the request.
 91		 * Once error occurred, just let clone->end_io() handle
 92		 * the remainder.
 93		 */
 94		return;
 95	else if (error) {
 96		/*
 97		 * Don't notice the error to the upper layer yet.
 98		 * The error handling decision is made by the target driver,
 99		 * when the request is completed.
100		 */
101		tio->error = error;
102		goto exit;
103	}
104
105	/*
106	 * I/O for the bio successfully completed.
107	 * Notice the data completion to the upper layer.
108	 */
109	tio->completed += nr_bytes;
110
111	/*
112	 * Update the original request.
113	 * Do not use blk_mq_end_request() here, because it may complete
114	 * the original request before the clone, and break the ordering.
115	 */
116	if (is_last)
117 exit:
118		blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
119}
120
121static struct dm_rq_target_io *tio_from_request(struct request *rq)
122{
123	return blk_mq_rq_to_pdu(rq);
124}
125
126static void rq_end_stats(struct mapped_device *md, struct request *orig)
127{
128	if (unlikely(dm_stats_used(&md->stats))) {
129		struct dm_rq_target_io *tio = tio_from_request(orig);
 
130		tio->duration_jiffies = jiffies - tio->duration_jiffies;
131		dm_stats_account_io(&md->stats, rq_data_dir(orig),
132				    blk_rq_pos(orig), tio->n_sectors, true,
133				    tio->duration_jiffies, &tio->stats_aux);
134	}
135}
136
137/*
138 * Don't touch any member of the md after calling this function because
139 * the md may be freed in dm_put() at the end of this function.
140 * Or do dm_get() before calling this function and dm_put() later.
141 */
142static void rq_completed(struct mapped_device *md)
143{
144	/*
145	 * dm_put() must be at the end of this function. See the comment above
146	 */
147	dm_put(md);
148}
149
150/*
151 * Complete the clone and the original request.
152 * Must be called without clone's queue lock held,
153 * see end_clone_request() for more details.
154 */
155static void dm_end_request(struct request *clone, blk_status_t error)
156{
157	struct dm_rq_target_io *tio = clone->end_io_data;
158	struct mapped_device *md = tio->md;
159	struct request *rq = tio->orig;
160
161	blk_rq_unprep_clone(clone);
162	tio->ti->type->release_clone_rq(clone, NULL);
163
164	rq_end_stats(md, rq);
165	blk_mq_end_request(rq, error);
166	rq_completed(md);
167}
168
169static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
170{
171	blk_mq_delay_kick_requeue_list(q, msecs);
172}
173
174void dm_mq_kick_requeue_list(struct mapped_device *md)
175{
176	__dm_mq_kick_requeue_list(md->queue, 0);
177}
178EXPORT_SYMBOL(dm_mq_kick_requeue_list);
179
180static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
181{
182	blk_mq_requeue_request(rq, false);
183	__dm_mq_kick_requeue_list(rq->q, msecs);
184}
185
186static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
187{
188	struct mapped_device *md = tio->md;
189	struct request *rq = tio->orig;
190	unsigned long delay_ms = delay_requeue ? 100 : 0;
191
192	rq_end_stats(md, rq);
193	if (tio->clone) {
194		blk_rq_unprep_clone(tio->clone);
195		tio->ti->type->release_clone_rq(tio->clone, NULL);
196	}
197
198	dm_mq_delay_requeue_request(rq, delay_ms);
199	rq_completed(md);
200}
201
202static void dm_done(struct request *clone, blk_status_t error, bool mapped)
203{
204	int r = DM_ENDIO_DONE;
205	struct dm_rq_target_io *tio = clone->end_io_data;
206	dm_request_endio_fn rq_end_io = NULL;
207
208	if (tio->ti) {
209		rq_end_io = tio->ti->type->rq_end_io;
210
211		if (mapped && rq_end_io)
212			r = rq_end_io(tio->ti, clone, error, &tio->info);
213	}
214
215	if (unlikely(error == BLK_STS_TARGET)) {
216		if (req_op(clone) == REQ_OP_DISCARD &&
217		    !clone->q->limits.max_discard_sectors)
218			disable_discard(tio->md);
219		else if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
220			 !clone->q->limits.max_write_zeroes_sectors)
221			disable_write_zeroes(tio->md);
222	}
223
224	switch (r) {
225	case DM_ENDIO_DONE:
226		/* The target wants to complete the I/O */
227		dm_end_request(clone, error);
228		break;
229	case DM_ENDIO_INCOMPLETE:
230		/* The target will handle the I/O */
231		return;
232	case DM_ENDIO_REQUEUE:
233		/* The target wants to requeue the I/O */
234		dm_requeue_original_request(tio, false);
235		break;
236	case DM_ENDIO_DELAY_REQUEUE:
237		/* The target wants to requeue the I/O after a delay */
238		dm_requeue_original_request(tio, true);
239		break;
240	default:
241		DMCRIT("unimplemented target endio return value: %d", r);
242		BUG();
243	}
244}
245
246/*
247 * Request completion handler for request-based dm
248 */
249static void dm_softirq_done(struct request *rq)
250{
251	bool mapped = true;
252	struct dm_rq_target_io *tio = tio_from_request(rq);
253	struct request *clone = tio->clone;
254
255	if (!clone) {
256		struct mapped_device *md = tio->md;
257
258		rq_end_stats(md, rq);
259		blk_mq_end_request(rq, tio->error);
260		rq_completed(md);
261		return;
262	}
263
264	if (rq->rq_flags & RQF_FAILED)
265		mapped = false;
266
267	dm_done(clone, tio->error, mapped);
268}
269
270/*
271 * Complete the clone and the original request with the error status
272 * through softirq context.
273 */
274static void dm_complete_request(struct request *rq, blk_status_t error)
275{
276	struct dm_rq_target_io *tio = tio_from_request(rq);
277
278	tio->error = error;
279	if (likely(!blk_should_fake_timeout(rq->q)))
280		blk_mq_complete_request(rq);
281}
282
283/*
284 * Complete the not-mapped clone and the original request with the error status
285 * through softirq context.
286 * Target's rq_end_io() function isn't called.
287 * This may be used when the target's clone_and_map_rq() function fails.
288 */
289static void dm_kill_unmapped_request(struct request *rq, blk_status_t error)
290{
291	rq->rq_flags |= RQF_FAILED;
292	dm_complete_request(rq, error);
293}
294
295static enum rq_end_io_ret end_clone_request(struct request *clone,
296					    blk_status_t error)
297{
298	struct dm_rq_target_io *tio = clone->end_io_data;
299
300	dm_complete_request(tio->orig, error);
301	return RQ_END_IO_NONE;
302}
303
304static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
305				 void *data)
306{
307	struct dm_rq_target_io *tio = data;
308	struct dm_rq_clone_bio_info *info =
309		container_of(bio, struct dm_rq_clone_bio_info, clone);
310
311	info->orig = bio_orig;
312	info->tio = tio;
313	bio->bi_end_io = end_clone_bio;
314
315	return 0;
316}
317
318static int setup_clone(struct request *clone, struct request *rq,
319		       struct dm_rq_target_io *tio, gfp_t gfp_mask)
320{
321	int r;
322
323	r = blk_rq_prep_clone(clone, rq, &tio->md->mempools->bs, gfp_mask,
324			      dm_rq_bio_constructor, tio);
325	if (r)
326		return r;
327
328	clone->end_io = end_clone_request;
329	clone->end_io_data = tio;
330
331	tio->clone = clone;
332
333	return 0;
334}
335
336static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
337		     struct mapped_device *md)
338{
339	tio->md = md;
340	tio->ti = NULL;
341	tio->clone = NULL;
342	tio->orig = rq;
343	tio->error = 0;
344	tio->completed = 0;
345	/*
346	 * Avoid initializing info for blk-mq; it passes
347	 * target-specific data through info.ptr
348	 * (see: dm_mq_init_request)
349	 */
350	if (!md->init_tio_pdu)
351		memset(&tio->info, 0, sizeof(tio->info));
352}
353
354/*
355 * Returns:
356 * DM_MAPIO_*       : the request has been processed as indicated
357 * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
358 * < 0              : the request was completed due to failure
359 */
360static int map_request(struct dm_rq_target_io *tio)
361{
362	int r;
363	struct dm_target *ti = tio->ti;
364	struct mapped_device *md = tio->md;
365	struct request *rq = tio->orig;
366	struct request *clone = NULL;
367	blk_status_t ret;
368
369	r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
370	switch (r) {
371	case DM_MAPIO_SUBMITTED:
372		/* The target has taken the I/O to submit by itself later */
373		break;
374	case DM_MAPIO_REMAPPED:
375		if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
376			/* -ENOMEM */
377			ti->type->release_clone_rq(clone, &tio->info);
378			return DM_MAPIO_REQUEUE;
379		}
380
381		/* The target has remapped the I/O so dispatch it */
382		trace_block_rq_remap(clone, disk_devt(dm_disk(md)),
383				     blk_rq_pos(rq));
384		ret = blk_insert_cloned_request(clone);
385		switch (ret) {
386		case BLK_STS_OK:
387			break;
388		case BLK_STS_RESOURCE:
389		case BLK_STS_DEV_RESOURCE:
390			blk_rq_unprep_clone(clone);
391			blk_mq_cleanup_rq(clone);
392			tio->ti->type->release_clone_rq(clone, &tio->info);
393			tio->clone = NULL;
394			return DM_MAPIO_REQUEUE;
395		default:
396			/* must complete clone in terms of original request */
397			dm_complete_request(rq, ret);
398		}
399		break;
400	case DM_MAPIO_REQUEUE:
401		/* The target wants to requeue the I/O */
402		break;
403	case DM_MAPIO_DELAY_REQUEUE:
404		/* The target wants to requeue the I/O after a delay */
405		dm_requeue_original_request(tio, true);
406		break;
407	case DM_MAPIO_KILL:
408		/* The target wants to complete the I/O */
409		dm_kill_unmapped_request(rq, BLK_STS_IOERR);
410		break;
411	default:
412		DMCRIT("unimplemented target map return value: %d", r);
413		BUG();
414	}
415
416	return r;
417}
418
419/* DEPRECATED: previously used for request-based merge heuristic in dm_request_fn() */
420ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
421{
422	return sprintf(buf, "%u\n", 0);
423}
424
425ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
426						     const char *buf, size_t count)
427{
428	return count;
429}
430
431static void dm_start_request(struct mapped_device *md, struct request *orig)
432{
433	blk_mq_start_request(orig);
434
435	if (unlikely(dm_stats_used(&md->stats))) {
436		struct dm_rq_target_io *tio = tio_from_request(orig);
 
437		tio->duration_jiffies = jiffies;
438		tio->n_sectors = blk_rq_sectors(orig);
439		dm_stats_account_io(&md->stats, rq_data_dir(orig),
440				    blk_rq_pos(orig), tio->n_sectors, false, 0,
441				    &tio->stats_aux);
442	}
443
444	/*
445	 * Hold the md reference here for the in-flight I/O.
446	 * We can't rely on the reference count by device opener,
447	 * because the device may be closed during the request completion
448	 * when all bios are completed.
449	 * See the comment in rq_completed() too.
450	 */
451	dm_get(md);
452}
453
454static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
455			      unsigned int hctx_idx, unsigned int numa_node)
456{
457	struct mapped_device *md = set->driver_data;
458	struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
459
460	/*
461	 * Must initialize md member of tio, otherwise it won't
462	 * be available in dm_mq_queue_rq.
463	 */
464	tio->md = md;
465
466	if (md->init_tio_pdu) {
467		/* target-specific per-io data is immediately after the tio */
468		tio->info.ptr = tio + 1;
469	}
470
471	return 0;
472}
473
474static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
475			  const struct blk_mq_queue_data *bd)
476{
477	struct request *rq = bd->rq;
478	struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
479	struct mapped_device *md = tio->md;
480	struct dm_target *ti = md->immutable_target;
481
482	/*
483	 * blk-mq's unquiesce may come from outside events, such as
484	 * elevator switch, updating nr_requests or others, and request may
485	 * come during suspend, so simply ask for blk-mq to requeue it.
486	 */
487	if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)))
488		return BLK_STS_RESOURCE;
489
490	if (unlikely(!ti)) {
491		int srcu_idx;
492		struct dm_table *map;
493
494		map = dm_get_live_table(md, &srcu_idx);
495		if (unlikely(!map)) {
496			dm_put_live_table(md, srcu_idx);
497			return BLK_STS_RESOURCE;
498		}
499		ti = dm_table_find_target(map, 0);
500		dm_put_live_table(md, srcu_idx);
501	}
502
503	if (ti->type->busy && ti->type->busy(ti))
504		return BLK_STS_RESOURCE;
505
506	dm_start_request(md, rq);
507
508	/* Init tio using md established in .init_request */
509	init_tio(tio, rq, md);
510
511	/*
512	 * Establish tio->ti before calling map_request().
513	 */
514	tio->ti = ti;
515
516	/* Direct call is fine since .queue_rq allows allocations */
517	if (map_request(tio) == DM_MAPIO_REQUEUE) {
518		/* Undo dm_start_request() before requeuing */
519		rq_end_stats(md, rq);
520		rq_completed(md);
521		return BLK_STS_RESOURCE;
522	}
523
524	return BLK_STS_OK;
525}
526
527static const struct blk_mq_ops dm_mq_ops = {
528	.queue_rq = dm_mq_queue_rq,
529	.complete = dm_softirq_done,
530	.init_request = dm_mq_init_request,
531};
532
533int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
534{
535	struct dm_target *immutable_tgt;
536	int err;
537
538	md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
539	if (!md->tag_set)
540		return -ENOMEM;
541
542	md->tag_set->ops = &dm_mq_ops;
543	md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
544	md->tag_set->numa_node = md->numa_node_id;
545	md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_STACKING;
546	md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
547	md->tag_set->driver_data = md;
548
549	md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
550	immutable_tgt = dm_table_get_immutable_target(t);
551	if (immutable_tgt && immutable_tgt->per_io_data_size) {
552		/* any target-specific per-io data is immediately after the tio */
553		md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
554		md->init_tio_pdu = true;
555	}
556
557	err = blk_mq_alloc_tag_set(md->tag_set);
558	if (err)
559		goto out_kfree_tag_set;
560
561	err = blk_mq_init_allocated_queue(md->tag_set, md->queue);
562	if (err)
563		goto out_tag_set;
564	return 0;
565
566out_tag_set:
567	blk_mq_free_tag_set(md->tag_set);
568out_kfree_tag_set:
569	kfree(md->tag_set);
570	md->tag_set = NULL;
571
572	return err;
573}
574
575void dm_mq_cleanup_mapped_device(struct mapped_device *md)
576{
577	if (md->tag_set) {
578		blk_mq_free_tag_set(md->tag_set);
579		kfree(md->tag_set);
580		md->tag_set = NULL;
581	}
582}
583
584module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
585MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
586
587/* Unused, but preserved for userspace compatibility */
588static bool use_blk_mq = true;
589module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
590MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
591
592module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
593MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
594
595module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
596MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");