dm-rq.c

  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	if (!is_last)
113		return;
114	/*
115	 * At this moment we know this is the last bio of the cloned request,
116	 * and all cloned bios have been released, so reset the clone request's
117	 * bio pointer to avoid double free.
118	 */
119	tio->clone->bio = NULL;
120 exit:
121	/*
122	 * Update the original request.
123	 * Do not use blk_mq_end_request() here, because it may complete
124	 * the original request before the clone, and break the ordering.
125	 */
126	blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
127}
128
129static struct dm_rq_target_io *tio_from_request(struct request *rq)
130{
131	return blk_mq_rq_to_pdu(rq);
132}
133
134static void rq_end_stats(struct mapped_device *md, struct request *orig)
135{
136	if (unlikely(dm_stats_used(&md->stats))) {
137		struct dm_rq_target_io *tio = tio_from_request(orig);
138
139		tio->duration_jiffies = jiffies - tio->duration_jiffies;
140		dm_stats_account_io(&md->stats, rq_data_dir(orig),
141				    blk_rq_pos(orig), tio->n_sectors, true,
142				    tio->duration_jiffies, &tio->stats_aux);
143	}
144}
145
146/*
147 * Don't touch any member of the md after calling this function because
148 * the md may be freed in dm_put() at the end of this function.
149 * Or do dm_get() before calling this function and dm_put() later.
150 */
151static void rq_completed(struct mapped_device *md)
152{
153	/*
154	 * dm_put() must be at the end of this function. See the comment above
155	 */
156	dm_put(md);
157}
158
159/*
160 * Complete the clone and the original request.
161 * Must be called without clone's queue lock held,
162 * see end_clone_request() for more details.
163 */
164static void dm_end_request(struct request *clone, blk_status_t error)
165{
166	struct dm_rq_target_io *tio = clone->end_io_data;
167	struct mapped_device *md = tio->md;
168	struct request *rq = tio->orig;
169
170	blk_rq_unprep_clone(clone);
171	tio->ti->type->release_clone_rq(clone, NULL);
172
173	rq_end_stats(md, rq);
174	blk_mq_end_request(rq, error);
175	rq_completed(md);
176}
177
178static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
179{
180	blk_mq_delay_kick_requeue_list(q, msecs);
181}
182
183void dm_mq_kick_requeue_list(struct mapped_device *md)
184{
185	__dm_mq_kick_requeue_list(md->queue, 0);
186}
187EXPORT_SYMBOL(dm_mq_kick_requeue_list);
188
189static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
190{
191	blk_mq_requeue_request(rq, false);
192	__dm_mq_kick_requeue_list(rq->q, msecs);
193}
194
195static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
196{
197	struct mapped_device *md = tio->md;
198	struct request *rq = tio->orig;
199	unsigned long delay_ms = delay_requeue ? 100 : 0;
200
201	rq_end_stats(md, rq);
202	if (tio->clone) {
203		blk_rq_unprep_clone(tio->clone);
204		tio->ti->type->release_clone_rq(tio->clone, NULL);
205	}
206
207	dm_mq_delay_requeue_request(rq, delay_ms);
208	rq_completed(md);
209}
210
211static void dm_done(struct request *clone, blk_status_t error, bool mapped)
212{
213	int r = DM_ENDIO_DONE;
214	struct dm_rq_target_io *tio = clone->end_io_data;
215	dm_request_endio_fn rq_end_io = NULL;
216
217	if (tio->ti) {
218		rq_end_io = tio->ti->type->rq_end_io;
219
220		if (mapped && rq_end_io)
221			r = rq_end_io(tio->ti, clone, error, &tio->info);
222	}
223
224	if (unlikely(error == BLK_STS_TARGET)) {
225		if (req_op(clone) == REQ_OP_DISCARD &&
226		    !clone->q->limits.max_discard_sectors)
227			blk_queue_disable_discard(tio->md->queue);
228		else if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
229			 !clone->q->limits.max_write_zeroes_sectors)
230			blk_queue_disable_write_zeroes(tio->md->queue);
231	}
232
233	switch (r) {
234	case DM_ENDIO_DONE:
235		/* The target wants to complete the I/O */
236		dm_end_request(clone, error);
237		break;
238	case DM_ENDIO_INCOMPLETE:
239		/* The target will handle the I/O */
240		return;
241	case DM_ENDIO_REQUEUE:
242		/* The target wants to requeue the I/O */
243		dm_requeue_original_request(tio, false);
244		break;
245	case DM_ENDIO_DELAY_REQUEUE:
246		/* The target wants to requeue the I/O after a delay */
247		dm_requeue_original_request(tio, true);
248		break;
249	default:
250		DMCRIT("unimplemented target endio return value: %d", r);
251		BUG();
252	}
253}
254
255/*
256 * Request completion handler for request-based dm
257 */
258static void dm_softirq_done(struct request *rq)
259{
260	bool mapped = true;
261	struct dm_rq_target_io *tio = tio_from_request(rq);
262	struct request *clone = tio->clone;
263
264	if (!clone) {
265		struct mapped_device *md = tio->md;
266
267		rq_end_stats(md, rq);
268		blk_mq_end_request(rq, tio->error);
269		rq_completed(md);
270		return;
271	}
272
273	if (rq->rq_flags & RQF_FAILED)
274		mapped = false;
275
276	dm_done(clone, tio->error, mapped);
277}
278
279/*
280 * Complete the clone and the original request with the error status
281 * through softirq context.
282 */
283static void dm_complete_request(struct request *rq, blk_status_t error)
284{
285	struct dm_rq_target_io *tio = tio_from_request(rq);
286
287	tio->error = error;
288	blk_mq_complete_request(rq);
289}
290
291/*
292 * Complete the not-mapped clone and the original request with the error status
293 * through softirq context.
294 * Target's rq_end_io() function isn't called.
295 * This may be used when the target's clone_and_map_rq() function fails.
296 */
297static void dm_kill_unmapped_request(struct request *rq, blk_status_t error)
298{
299	rq->rq_flags |= RQF_FAILED;
300	dm_complete_request(rq, error);
301}
302
303static enum rq_end_io_ret end_clone_request(struct request *clone,
304					    blk_status_t error)
305{
306	struct dm_rq_target_io *tio = clone->end_io_data;
307
308	dm_complete_request(tio->orig, error);
309	return RQ_END_IO_NONE;
310}
311
312static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
313				 void *data)
314{
315	struct dm_rq_target_io *tio = data;
316	struct dm_rq_clone_bio_info *info =
317		container_of(bio, struct dm_rq_clone_bio_info, clone);
318
319	info->orig = bio_orig;
320	info->tio = tio;
321	bio->bi_end_io = end_clone_bio;
322
323	return 0;
324}
325
326static int setup_clone(struct request *clone, struct request *rq,
327		       struct dm_rq_target_io *tio, gfp_t gfp_mask)
328{
329	int r;
330
331	r = blk_rq_prep_clone(clone, rq, &tio->md->mempools->bs, gfp_mask,
332			      dm_rq_bio_constructor, tio);
333	if (r)
334		return r;
335
336	clone->end_io = end_clone_request;
337	clone->end_io_data = tio;
338
339	tio->clone = clone;
340
341	return 0;
342}
343
344static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
345		     struct mapped_device *md)
346{
347	tio->md = md;
348	tio->ti = NULL;
349	tio->clone = NULL;
350	tio->orig = rq;
351	tio->error = 0;
352	tio->completed = 0;
353	/*
354	 * Avoid initializing info for blk-mq; it passes
355	 * target-specific data through info.ptr
356	 * (see: dm_mq_init_request)
357	 */
358	if (!md->init_tio_pdu)
359		memset(&tio->info, 0, sizeof(tio->info));
360}
361
362/*
363 * Returns:
364 * DM_MAPIO_*       : the request has been processed as indicated
365 * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
366 * < 0              : the request was completed due to failure
367 */
368static int map_request(struct dm_rq_target_io *tio)
369{
370	int r;
371	struct dm_target *ti = tio->ti;
372	struct mapped_device *md = tio->md;
373	struct request *rq = tio->orig;
374	struct request *clone = NULL;
375	blk_status_t ret;
376
377	r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
378	switch (r) {
379	case DM_MAPIO_SUBMITTED:
380		/* The target has taken the I/O to submit by itself later */
381		break;
382	case DM_MAPIO_REMAPPED:
383		if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
384			/* -ENOMEM */
385			ti->type->release_clone_rq(clone, &tio->info);
386			return DM_MAPIO_REQUEUE;
387		}
388
389		/* The target has remapped the I/O so dispatch it */
390		trace_block_rq_remap(clone, disk_devt(dm_disk(md)),
391				     blk_rq_pos(rq));
392		ret = blk_insert_cloned_request(clone);
393		switch (ret) {
394		case BLK_STS_OK:
395			break;
396		case BLK_STS_RESOURCE:
397		case BLK_STS_DEV_RESOURCE:
398			blk_rq_unprep_clone(clone);
399			blk_mq_cleanup_rq(clone);
400			tio->ti->type->release_clone_rq(clone, &tio->info);
401			tio->clone = NULL;
402			return DM_MAPIO_REQUEUE;
403		default:
404			/* must complete clone in terms of original request */
405			dm_complete_request(rq, ret);
406		}
407		break;
408	case DM_MAPIO_REQUEUE:
409		/* The target wants to requeue the I/O */
410		break;
411	case DM_MAPIO_DELAY_REQUEUE:
412		/* The target wants to requeue the I/O after a delay */
413		dm_requeue_original_request(tio, true);
414		break;
415	case DM_MAPIO_KILL:
416		/* The target wants to complete the I/O */
417		dm_kill_unmapped_request(rq, BLK_STS_IOERR);
418		break;
419	default:
420		DMCRIT("unimplemented target map return value: %d", r);
421		BUG();
422	}
423
424	return r;
425}
426
427/* DEPRECATED: previously used for request-based merge heuristic in dm_request_fn() */
428ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
429{
430	return sprintf(buf, "%u\n", 0);
431}
432
433ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
434						     const char *buf, size_t count)
435{
436	return count;
437}
438
439static void dm_start_request(struct mapped_device *md, struct request *orig)
440{
441	blk_mq_start_request(orig);
442
443	if (unlikely(dm_stats_used(&md->stats))) {
444		struct dm_rq_target_io *tio = tio_from_request(orig);
445
446		tio->duration_jiffies = jiffies;
447		tio->n_sectors = blk_rq_sectors(orig);
448		dm_stats_account_io(&md->stats, rq_data_dir(orig),
449				    blk_rq_pos(orig), tio->n_sectors, false, 0,
450				    &tio->stats_aux);
451	}
452
453	/*
454	 * Hold the md reference here for the in-flight I/O.
455	 * We can't rely on the reference count by device opener,
456	 * because the device may be closed during the request completion
457	 * when all bios are completed.
458	 * See the comment in rq_completed() too.
459	 */
460	dm_get(md);
461}
462
463static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
464			      unsigned int hctx_idx, unsigned int numa_node)
465{
466	struct mapped_device *md = set->driver_data;
467	struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
468
469	/*
470	 * Must initialize md member of tio, otherwise it won't
471	 * be available in dm_mq_queue_rq.
472	 */
473	tio->md = md;
474
475	if (md->init_tio_pdu) {
476		/* target-specific per-io data is immediately after the tio */
477		tio->info.ptr = tio + 1;
478	}
479
480	return 0;
481}
482
483static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
484			  const struct blk_mq_queue_data *bd)
485{
486	struct request *rq = bd->rq;
487	struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
488	struct mapped_device *md = tio->md;
489	struct dm_target *ti = md->immutable_target;
490
491	/*
492	 * blk-mq's unquiesce may come from outside events, such as
493	 * elevator switch, updating nr_requests or others, and request may
494	 * come during suspend, so simply ask for blk-mq to requeue it.
495	 */
496	if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)))
497		return BLK_STS_RESOURCE;
498
499	if (unlikely(!ti)) {
500		int srcu_idx;
501		struct dm_table *map;
502
503		map = dm_get_live_table(md, &srcu_idx);
504		if (unlikely(!map)) {
505			DMERR_LIMIT("%s: mapping table unavailable, erroring io",
506				    dm_device_name(md));
507			dm_put_live_table(md, srcu_idx);
508			return BLK_STS_IOERR;
509		}
510		ti = dm_table_find_target(map, 0);
511		dm_put_live_table(md, srcu_idx);
512	}
513
514	if (ti->type->busy && ti->type->busy(ti))
515		return BLK_STS_RESOURCE;
516
517	dm_start_request(md, rq);
518
519	/* Init tio using md established in .init_request */
520	init_tio(tio, rq, md);
521
522	/*
523	 * Establish tio->ti before calling map_request().
524	 */
525	tio->ti = ti;
526
527	/* Direct call is fine since .queue_rq allows allocations */
528	if (map_request(tio) == DM_MAPIO_REQUEUE) {
529		/* Undo dm_start_request() before requeuing */
530		rq_end_stats(md, rq);
531		rq_completed(md);
532		return BLK_STS_RESOURCE;
533	}
534
535	return BLK_STS_OK;
536}
537
538static const struct blk_mq_ops dm_mq_ops = {
539	.queue_rq = dm_mq_queue_rq,
540	.complete = dm_softirq_done,
541	.init_request = dm_mq_init_request,
542};
543
544int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
545{
546	struct dm_target *immutable_tgt;
547	int err;
548
549	md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
550	if (!md->tag_set)
551		return -ENOMEM;
552
553	md->tag_set->ops = &dm_mq_ops;
554	md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
555	md->tag_set->numa_node = md->numa_node_id;
556	md->tag_set->flags = BLK_MQ_F_STACKING;
557	md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
558	md->tag_set->driver_data = md;
559
560	md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
561	immutable_tgt = dm_table_get_immutable_target(t);
562	if (immutable_tgt && immutable_tgt->per_io_data_size) {
563		/* any target-specific per-io data is immediately after the tio */
564		md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
565		md->init_tio_pdu = true;
566	}
567
568	err = blk_mq_alloc_tag_set(md->tag_set);
569	if (err)
570		goto out_kfree_tag_set;
571
572	err = blk_mq_init_allocated_queue(md->tag_set, md->queue);
573	if (err)
574		goto out_tag_set;
575	return 0;
576
577out_tag_set:
578	blk_mq_free_tag_set(md->tag_set);
579out_kfree_tag_set:
580	kfree(md->tag_set);
581	md->tag_set = NULL;
582
583	return err;
584}
585
586void dm_mq_cleanup_mapped_device(struct mapped_device *md)
587{
588	if (md->tag_set) {
589		blk_mq_free_tag_set(md->tag_set);
590		kfree(md->tag_set);
591		md->tag_set = NULL;
592	}
593}
594
595module_param(reserved_rq_based_ios, uint, 0644);
596MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
597
598/* Unused, but preserved for userspace compatibility */
599static bool use_blk_mq = true;
600module_param(use_blk_mq, bool, 0644);
601MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
602
603module_param(dm_mq_nr_hw_queues, uint, 0644);
604MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
605
606module_param(dm_mq_queue_depth, uint, 0644);
607MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");