2 * Copyright (C) 2016 Red Hat, Inc. All rights reserved.
4 * This file is released under the GPL.
10 #include <linux/blk-mq.h>
12 #define DM_MSG_PREFIX "core-rq"
15 * One of these is allocated per request.
17 struct dm_rq_target_io {
18 struct mapped_device *md;
20 struct request *orig, *clone;
21 struct kthread_work work;
24 struct dm_stats_aux stats_aux;
25 unsigned long duration_jiffies;
30 #define DM_MQ_NR_HW_QUEUES 1
31 #define DM_MQ_QUEUE_DEPTH 2048
32 static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
33 static unsigned dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
36 * Request-based DM's mempools' reserved IOs set by the user.
38 #define RESERVED_REQUEST_BASED_IOS 256
39 static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
41 unsigned dm_get_reserved_rq_based_ios(void)
43 return __dm_get_module_param(&reserved_rq_based_ios,
44 RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
47 static unsigned dm_get_blk_mq_nr_hw_queues(void)
49 return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
52 static unsigned dm_get_blk_mq_queue_depth(void)
54 return __dm_get_module_param(&dm_mq_queue_depth,
55 DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
58 int dm_request_based(struct mapped_device *md)
60 return queue_is_mq(md->queue);
63 void dm_start_queue(struct request_queue *q)
65 blk_mq_unquiesce_queue(q);
66 blk_mq_kick_requeue_list(q);
69 void dm_stop_queue(struct request_queue *q)
71 blk_mq_quiesce_queue(q);
75 * Partial completion handling for request-based dm
77 static void end_clone_bio(struct bio *clone)
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;
90 * An error has already been detected on the request.
91 * Once error occurred, just let clone->end_io() handle
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.
106 * I/O for the bio successfully completed.
107 * Notice the data completion to the upper layer.
109 tio->completed += nr_bytes;
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.
118 blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
121 static struct dm_rq_target_io *tio_from_request(struct request *rq)
123 return blk_mq_rq_to_pdu(rq);
126 static void rq_end_stats(struct mapped_device *md, struct request *orig)
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);
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.
142 static void rq_completed(struct mapped_device *md)
145 * dm_put() must be at the end of this function. See the comment above
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.
155 static void dm_end_request(struct request *clone, blk_status_t error)
157 struct dm_rq_target_io *tio = clone->end_io_data;
158 struct mapped_device *md = tio->md;
159 struct request *rq = tio->orig;
161 blk_rq_unprep_clone(clone);
162 tio->ti->type->release_clone_rq(clone, NULL);
164 rq_end_stats(md, rq);
165 blk_mq_end_request(rq, error);
169 static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
171 blk_mq_delay_kick_requeue_list(q, msecs);
174 void dm_mq_kick_requeue_list(struct mapped_device *md)
176 __dm_mq_kick_requeue_list(md->queue, 0);
178 EXPORT_SYMBOL(dm_mq_kick_requeue_list);
180 static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
182 blk_mq_requeue_request(rq, false);
183 __dm_mq_kick_requeue_list(rq->q, msecs);
186 static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
188 struct mapped_device *md = tio->md;
189 struct request *rq = tio->orig;
190 unsigned long delay_ms = delay_requeue ? 100 : 0;
192 rq_end_stats(md, rq);
194 blk_rq_unprep_clone(tio->clone);
195 tio->ti->type->release_clone_rq(tio->clone, NULL);
198 dm_mq_delay_requeue_request(rq, delay_ms);
202 static void dm_done(struct request *clone, blk_status_t error, bool mapped)
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;
209 rq_end_io = tio->ti->type->rq_end_io;
211 if (mapped && rq_end_io)
212 r = rq_end_io(tio->ti, clone, error, &tio->info);
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);
226 /* The target wants to complete the I/O */
227 dm_end_request(clone, error);
229 case DM_ENDIO_INCOMPLETE:
230 /* The target will handle the I/O */
232 case DM_ENDIO_REQUEUE:
233 /* The target wants to requeue the I/O */
234 dm_requeue_original_request(tio, false);
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);
241 DMWARN("unimplemented target endio return value: %d", r);
247 * Request completion handler for request-based dm
249 static void dm_softirq_done(struct request *rq)
252 struct dm_rq_target_io *tio = tio_from_request(rq);
253 struct request *clone = tio->clone;
256 struct mapped_device *md = tio->md;
258 rq_end_stats(md, rq);
259 blk_mq_end_request(rq, tio->error);
264 if (rq->rq_flags & RQF_FAILED)
267 dm_done(clone, tio->error, mapped);
271 * Complete the clone and the original request with the error status
272 * through softirq context.
274 static void dm_complete_request(struct request *rq, blk_status_t error)
276 struct dm_rq_target_io *tio = tio_from_request(rq);
279 if (likely(!blk_should_fake_timeout(rq->q)))
280 blk_mq_complete_request(rq);
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.
289 static void dm_kill_unmapped_request(struct request *rq, blk_status_t error)
291 rq->rq_flags |= RQF_FAILED;
292 dm_complete_request(rq, error);
295 static void end_clone_request(struct request *clone, blk_status_t error)
297 struct dm_rq_target_io *tio = clone->end_io_data;
299 dm_complete_request(tio->orig, error);
302 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
305 struct dm_rq_target_io *tio = data;
306 struct dm_rq_clone_bio_info *info =
307 container_of(bio, struct dm_rq_clone_bio_info, clone);
309 info->orig = bio_orig;
311 bio->bi_end_io = end_clone_bio;
316 static int setup_clone(struct request *clone, struct request *rq,
317 struct dm_rq_target_io *tio, gfp_t gfp_mask)
321 r = blk_rq_prep_clone(clone, rq, &tio->md->mempools->bs, gfp_mask,
322 dm_rq_bio_constructor, tio);
326 clone->end_io = end_clone_request;
327 clone->end_io_data = tio;
334 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
335 struct mapped_device *md)
344 * Avoid initializing info for blk-mq; it passes
345 * target-specific data through info.ptr
346 * (see: dm_mq_init_request)
348 if (!md->init_tio_pdu)
349 memset(&tio->info, 0, sizeof(tio->info));
354 * DM_MAPIO_* : the request has been processed as indicated
355 * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
356 * < 0 : the request was completed due to failure
358 static int map_request(struct dm_rq_target_io *tio)
361 struct dm_target *ti = tio->ti;
362 struct mapped_device *md = tio->md;
363 struct request *rq = tio->orig;
364 struct request *clone = NULL;
367 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
369 case DM_MAPIO_SUBMITTED:
370 /* The target has taken the I/O to submit by itself later */
372 case DM_MAPIO_REMAPPED:
373 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
375 ti->type->release_clone_rq(clone, &tio->info);
376 return DM_MAPIO_REQUEUE;
379 /* The target has remapped the I/O so dispatch it */
380 trace_block_rq_remap(clone, disk_devt(dm_disk(md)),
382 ret = blk_insert_cloned_request(clone);
386 case BLK_STS_RESOURCE:
387 case BLK_STS_DEV_RESOURCE:
388 blk_rq_unprep_clone(clone);
389 blk_mq_cleanup_rq(clone);
390 tio->ti->type->release_clone_rq(clone, &tio->info);
392 return DM_MAPIO_REQUEUE;
394 /* must complete clone in terms of original request */
395 dm_complete_request(rq, ret);
398 case DM_MAPIO_REQUEUE:
399 /* The target wants to requeue the I/O */
401 case DM_MAPIO_DELAY_REQUEUE:
402 /* The target wants to requeue the I/O after a delay */
403 dm_requeue_original_request(tio, true);
406 /* The target wants to complete the I/O */
407 dm_kill_unmapped_request(rq, BLK_STS_IOERR);
410 DMWARN("unimplemented target map return value: %d", r);
417 /* DEPRECATED: previously used for request-based merge heuristic in dm_request_fn() */
418 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
420 return sprintf(buf, "%u\n", 0);
423 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
424 const char *buf, size_t count)
429 static void dm_start_request(struct mapped_device *md, struct request *orig)
431 blk_mq_start_request(orig);
433 if (unlikely(dm_stats_used(&md->stats))) {
434 struct dm_rq_target_io *tio = tio_from_request(orig);
435 tio->duration_jiffies = jiffies;
436 tio->n_sectors = blk_rq_sectors(orig);
437 dm_stats_account_io(&md->stats, rq_data_dir(orig),
438 blk_rq_pos(orig), tio->n_sectors, false, 0,
443 * Hold the md reference here for the in-flight I/O.
444 * We can't rely on the reference count by device opener,
445 * because the device may be closed during the request completion
446 * when all bios are completed.
447 * See the comment in rq_completed() too.
452 static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
453 unsigned int hctx_idx, unsigned int numa_node)
455 struct mapped_device *md = set->driver_data;
456 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
459 * Must initialize md member of tio, otherwise it won't
460 * be available in dm_mq_queue_rq.
464 if (md->init_tio_pdu) {
465 /* target-specific per-io data is immediately after the tio */
466 tio->info.ptr = tio + 1;
472 static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
473 const struct blk_mq_queue_data *bd)
475 struct request *rq = bd->rq;
476 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
477 struct mapped_device *md = tio->md;
478 struct dm_target *ti = md->immutable_target;
481 * blk-mq's unquiesce may come from outside events, such as
482 * elevator switch, updating nr_requests or others, and request may
483 * come during suspend, so simply ask for blk-mq to requeue it.
485 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)))
486 return BLK_STS_RESOURCE;
490 struct dm_table *map;
492 map = dm_get_live_table(md, &srcu_idx);
493 if (unlikely(!map)) {
494 dm_put_live_table(md, srcu_idx);
495 return BLK_STS_RESOURCE;
497 ti = dm_table_find_target(map, 0);
498 dm_put_live_table(md, srcu_idx);
501 if (ti->type->busy && ti->type->busy(ti))
502 return BLK_STS_RESOURCE;
504 dm_start_request(md, rq);
506 /* Init tio using md established in .init_request */
507 init_tio(tio, rq, md);
510 * Establish tio->ti before calling map_request().
514 /* Direct call is fine since .queue_rq allows allocations */
515 if (map_request(tio) == DM_MAPIO_REQUEUE) {
516 /* Undo dm_start_request() before requeuing */
517 rq_end_stats(md, rq);
519 return BLK_STS_RESOURCE;
525 static const struct blk_mq_ops dm_mq_ops = {
526 .queue_rq = dm_mq_queue_rq,
527 .complete = dm_softirq_done,
528 .init_request = dm_mq_init_request,
531 int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
533 struct dm_target *immutable_tgt;
536 md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
540 md->tag_set->ops = &dm_mq_ops;
541 md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
542 md->tag_set->numa_node = md->numa_node_id;
543 md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_STACKING;
544 md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
545 md->tag_set->driver_data = md;
547 md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
548 immutable_tgt = dm_table_get_immutable_target(t);
549 if (immutable_tgt && immutable_tgt->per_io_data_size) {
550 /* any target-specific per-io data is immediately after the tio */
551 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
552 md->init_tio_pdu = true;
555 err = blk_mq_alloc_tag_set(md->tag_set);
557 goto out_kfree_tag_set;
559 err = blk_mq_init_allocated_queue(md->tag_set, md->queue);
565 blk_mq_free_tag_set(md->tag_set);
573 void dm_mq_cleanup_mapped_device(struct mapped_device *md)
576 blk_mq_free_tag_set(md->tag_set);
582 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
583 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
585 /* Unused, but preserved for userspace compatibility */
586 static bool use_blk_mq = true;
587 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
588 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
590 module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
591 MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
593 module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
594 MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");