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);
46 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
48 static unsigned dm_get_blk_mq_nr_hw_queues(void)
50 return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
53 static unsigned dm_get_blk_mq_queue_depth(void)
55 return __dm_get_module_param(&dm_mq_queue_depth,
56 DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
59 int dm_request_based(struct mapped_device *md)
61 return queue_is_mq(md->queue);
64 void dm_start_queue(struct request_queue *q)
66 blk_mq_unquiesce_queue(q);
67 blk_mq_kick_requeue_list(q);
70 void dm_stop_queue(struct request_queue *q)
72 blk_mq_quiesce_queue(q);
76 * Partial completion handling for request-based dm
78 static void end_clone_bio(struct bio *clone)
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;
91 * An error has already been detected on the request.
92 * Once error occurred, just let clone->end_io() handle
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.
107 * I/O for the bio successfully completed.
108 * Notice the data completion to the upper layer.
110 tio->completed += nr_bytes;
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.
119 blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
122 static struct dm_rq_target_io *tio_from_request(struct request *rq)
124 return blk_mq_rq_to_pdu(rq);
127 static void rq_end_stats(struct mapped_device *md, struct request *orig)
129 if (unlikely(dm_stats_used(&md->stats))) {
130 struct dm_rq_target_io *tio = tio_from_request(orig);
131 tio->duration_jiffies = jiffies - tio->duration_jiffies;
132 dm_stats_account_io(&md->stats, rq_data_dir(orig),
133 blk_rq_pos(orig), tio->n_sectors, true,
134 tio->duration_jiffies, &tio->stats_aux);
139 * Don't touch any member of the md after calling this function because
140 * the md may be freed in dm_put() at the end of this function.
141 * Or do dm_get() before calling this function and dm_put() later.
143 static void rq_completed(struct mapped_device *md)
146 * dm_put() must be at the end of this function. See the comment above
152 * Complete the clone and the original request.
153 * Must be called without clone's queue lock held,
154 * see end_clone_request() for more details.
156 static void dm_end_request(struct request *clone, blk_status_t error)
158 struct dm_rq_target_io *tio = clone->end_io_data;
159 struct mapped_device *md = tio->md;
160 struct request *rq = tio->orig;
162 blk_rq_unprep_clone(clone);
163 tio->ti->type->release_clone_rq(clone, NULL);
165 rq_end_stats(md, rq);
166 blk_mq_end_request(rq, error);
170 static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
172 blk_mq_delay_kick_requeue_list(q, msecs);
175 void dm_mq_kick_requeue_list(struct mapped_device *md)
177 __dm_mq_kick_requeue_list(md->queue, 0);
179 EXPORT_SYMBOL(dm_mq_kick_requeue_list);
181 static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
183 blk_mq_requeue_request(rq, false);
184 __dm_mq_kick_requeue_list(rq->q, msecs);
187 static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
189 struct mapped_device *md = tio->md;
190 struct request *rq = tio->orig;
191 unsigned long delay_ms = delay_requeue ? 100 : 0;
193 rq_end_stats(md, rq);
195 blk_rq_unprep_clone(tio->clone);
196 tio->ti->type->release_clone_rq(tio->clone, NULL);
199 dm_mq_delay_requeue_request(rq, delay_ms);
203 static void dm_done(struct request *clone, blk_status_t error, bool mapped)
205 int r = DM_ENDIO_DONE;
206 struct dm_rq_target_io *tio = clone->end_io_data;
207 dm_request_endio_fn rq_end_io = NULL;
210 rq_end_io = tio->ti->type->rq_end_io;
212 if (mapped && rq_end_io)
213 r = rq_end_io(tio->ti, clone, error, &tio->info);
216 if (unlikely(error == BLK_STS_TARGET)) {
217 if (req_op(clone) == REQ_OP_DISCARD &&
218 !clone->q->limits.max_discard_sectors)
219 disable_discard(tio->md);
220 else if (req_op(clone) == REQ_OP_WRITE_SAME &&
221 !clone->q->limits.max_write_same_sectors)
222 disable_write_same(tio->md);
223 else if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
224 !clone->q->limits.max_write_zeroes_sectors)
225 disable_write_zeroes(tio->md);
230 /* The target wants to complete the I/O */
231 dm_end_request(clone, error);
233 case DM_ENDIO_INCOMPLETE:
234 /* The target will handle the I/O */
236 case DM_ENDIO_REQUEUE:
237 /* The target wants to requeue the I/O */
238 dm_requeue_original_request(tio, false);
240 case DM_ENDIO_DELAY_REQUEUE:
241 /* The target wants to requeue the I/O after a delay */
242 dm_requeue_original_request(tio, true);
245 DMWARN("unimplemented target endio return value: %d", r);
251 * Request completion handler for request-based dm
253 static void dm_softirq_done(struct request *rq)
256 struct dm_rq_target_io *tio = tio_from_request(rq);
257 struct request *clone = tio->clone;
260 struct mapped_device *md = tio->md;
262 rq_end_stats(md, rq);
263 blk_mq_end_request(rq, tio->error);
268 if (rq->rq_flags & RQF_FAILED)
271 dm_done(clone, tio->error, mapped);
275 * Complete the clone and the original request with the error status
276 * through softirq context.
278 static void dm_complete_request(struct request *rq, blk_status_t error)
280 struct dm_rq_target_io *tio = tio_from_request(rq);
283 if (likely(!blk_should_fake_timeout(rq->q)))
284 blk_mq_complete_request(rq);
288 * Complete the not-mapped clone and the original request with the error status
289 * through softirq context.
290 * Target's rq_end_io() function isn't called.
291 * This may be used when the target's clone_and_map_rq() function fails.
293 static void dm_kill_unmapped_request(struct request *rq, blk_status_t error)
295 rq->rq_flags |= RQF_FAILED;
296 dm_complete_request(rq, error);
299 static void end_clone_request(struct request *clone, blk_status_t error)
301 struct dm_rq_target_io *tio = clone->end_io_data;
303 dm_complete_request(tio->orig, error);
306 static blk_status_t dm_dispatch_clone_request(struct request *clone, struct request *rq)
310 if (blk_queue_io_stat(clone->q))
311 clone->rq_flags |= RQF_IO_STAT;
313 clone->start_time_ns = ktime_get_ns();
314 r = blk_insert_cloned_request(clone->q, clone);
315 if (r != BLK_STS_OK && r != BLK_STS_RESOURCE && r != BLK_STS_DEV_RESOURCE)
316 /* must complete clone in terms of original request */
317 dm_complete_request(rq, r);
321 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
324 struct dm_rq_target_io *tio = data;
325 struct dm_rq_clone_bio_info *info =
326 container_of(bio, struct dm_rq_clone_bio_info, clone);
328 info->orig = bio_orig;
330 bio->bi_end_io = end_clone_bio;
335 static int setup_clone(struct request *clone, struct request *rq,
336 struct dm_rq_target_io *tio, gfp_t gfp_mask)
340 r = blk_rq_prep_clone(clone, rq, &tio->md->bs, gfp_mask,
341 dm_rq_bio_constructor, tio);
345 clone->end_io = end_clone_request;
346 clone->end_io_data = tio;
353 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
354 struct mapped_device *md)
363 * Avoid initializing info for blk-mq; it passes
364 * target-specific data through info.ptr
365 * (see: dm_mq_init_request)
367 if (!md->init_tio_pdu)
368 memset(&tio->info, 0, sizeof(tio->info));
373 * DM_MAPIO_* : the request has been processed as indicated
374 * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
375 * < 0 : the request was completed due to failure
377 static int map_request(struct dm_rq_target_io *tio)
380 struct dm_target *ti = tio->ti;
381 struct mapped_device *md = tio->md;
382 struct request *rq = tio->orig;
383 struct request *clone = NULL;
386 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
388 case DM_MAPIO_SUBMITTED:
389 /* The target has taken the I/O to submit by itself later */
391 case DM_MAPIO_REMAPPED:
392 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
394 ti->type->release_clone_rq(clone, &tio->info);
395 return DM_MAPIO_REQUEUE;
398 /* The target has remapped the I/O so dispatch it */
399 trace_block_rq_remap(clone, disk_devt(dm_disk(md)),
401 ret = dm_dispatch_clone_request(clone, rq);
402 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
403 blk_rq_unprep_clone(clone);
404 blk_mq_cleanup_rq(clone);
405 tio->ti->type->release_clone_rq(clone, &tio->info);
407 return DM_MAPIO_REQUEUE;
410 case DM_MAPIO_REQUEUE:
411 /* The target wants to requeue the I/O */
413 case DM_MAPIO_DELAY_REQUEUE:
414 /* The target wants to requeue the I/O after a delay */
415 dm_requeue_original_request(tio, true);
418 /* The target wants to complete the I/O */
419 dm_kill_unmapped_request(rq, BLK_STS_IOERR);
422 DMWARN("unimplemented target map return value: %d", r);
429 /* DEPRECATED: previously used for request-based merge heuristic in dm_request_fn() */
430 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
432 return sprintf(buf, "%u\n", 0);
435 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
436 const char *buf, size_t count)
441 static void dm_start_request(struct mapped_device *md, struct request *orig)
443 blk_mq_start_request(orig);
445 if (unlikely(dm_stats_used(&md->stats))) {
446 struct dm_rq_target_io *tio = tio_from_request(orig);
447 tio->duration_jiffies = jiffies;
448 tio->n_sectors = blk_rq_sectors(orig);
449 dm_stats_account_io(&md->stats, rq_data_dir(orig),
450 blk_rq_pos(orig), tio->n_sectors, false, 0,
455 * Hold the md reference here for the in-flight I/O.
456 * We can't rely on the reference count by device opener,
457 * because the device may be closed during the request completion
458 * when all bios are completed.
459 * See the comment in rq_completed() too.
464 static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
465 unsigned int hctx_idx, unsigned int numa_node)
467 struct mapped_device *md = set->driver_data;
468 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
471 * Must initialize md member of tio, otherwise it won't
472 * be available in dm_mq_queue_rq.
476 if (md->init_tio_pdu) {
477 /* target-specific per-io data is immediately after the tio */
478 tio->info.ptr = tio + 1;
484 static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
485 const struct blk_mq_queue_data *bd)
487 struct request *rq = bd->rq;
488 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
489 struct mapped_device *md = tio->md;
490 struct dm_target *ti = md->immutable_target;
493 * blk-mq's unquiesce may come from outside events, such as
494 * elevator switch, updating nr_requests or others, and request may
495 * come during suspend, so simply ask for blk-mq to requeue it.
497 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)))
498 return BLK_STS_RESOURCE;
502 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
504 ti = dm_table_find_target(map, 0);
505 dm_put_live_table(md, srcu_idx);
508 if (ti->type->busy && ti->type->busy(ti))
509 return BLK_STS_RESOURCE;
511 dm_start_request(md, rq);
513 /* Init tio using md established in .init_request */
514 init_tio(tio, rq, md);
517 * Establish tio->ti before calling map_request().
521 /* Direct call is fine since .queue_rq allows allocations */
522 if (map_request(tio) == DM_MAPIO_REQUEUE) {
523 /* Undo dm_start_request() before requeuing */
524 rq_end_stats(md, rq);
526 return BLK_STS_RESOURCE;
532 static const struct blk_mq_ops dm_mq_ops = {
533 .queue_rq = dm_mq_queue_rq,
534 .complete = dm_softirq_done,
535 .init_request = dm_mq_init_request,
538 int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
540 struct dm_target *immutable_tgt;
543 md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
547 md->tag_set->ops = &dm_mq_ops;
548 md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
549 md->tag_set->numa_node = md->numa_node_id;
550 md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_STACKING;
551 md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
552 md->tag_set->driver_data = md;
554 md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
555 immutable_tgt = dm_table_get_immutable_target(t);
556 if (immutable_tgt && immutable_tgt->per_io_data_size) {
557 /* any target-specific per-io data is immediately after the tio */
558 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
559 md->init_tio_pdu = true;
562 err = blk_mq_alloc_tag_set(md->tag_set);
564 goto out_kfree_tag_set;
566 err = blk_mq_init_allocated_queue(md->tag_set, md->queue);
572 blk_mq_free_tag_set(md->tag_set);
580 void dm_mq_cleanup_mapped_device(struct mapped_device *md)
583 blk_mq_free_tag_set(md->tag_set);
589 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
590 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
592 /* Unused, but preserved for userspace compatibility */
593 static bool use_blk_mq = true;
594 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
595 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
597 module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
598 MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
600 module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
601 MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");