2 * Copyright (C) 2016 Red Hat, Inc. All rights reserved.
4 * This file is released under the GPL.
10 #include <linux/elevator.h> /* for rq_end_sector() */
11 #include <linux/blk-mq.h>
13 #define DM_MSG_PREFIX "core-rq"
16 * One of these is allocated per request.
18 struct dm_rq_target_io {
19 struct mapped_device *md;
21 struct request *orig, *clone;
22 struct kthread_work work;
25 struct dm_stats_aux stats_aux;
26 unsigned long duration_jiffies;
31 #define DM_MQ_NR_HW_QUEUES 1
32 #define DM_MQ_QUEUE_DEPTH 2048
33 static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
34 static unsigned dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
37 * Request-based DM's mempools' reserved IOs set by the user.
39 #define RESERVED_REQUEST_BASED_IOS 256
40 static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
42 unsigned dm_get_reserved_rq_based_ios(void)
44 return __dm_get_module_param(&reserved_rq_based_ios,
45 RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
47 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
49 static unsigned dm_get_blk_mq_nr_hw_queues(void)
51 return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
54 static unsigned dm_get_blk_mq_queue_depth(void)
56 return __dm_get_module_param(&dm_mq_queue_depth,
57 DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
60 int dm_request_based(struct mapped_device *md)
62 return queue_is_mq(md->queue);
65 void dm_start_queue(struct request_queue *q)
67 blk_mq_unquiesce_queue(q);
68 blk_mq_kick_requeue_list(q);
71 void dm_stop_queue(struct request_queue *q)
73 if (blk_mq_queue_stopped(q))
76 blk_mq_quiesce_queue(q);
80 * Partial completion handling for request-based dm
82 static void end_clone_bio(struct bio *clone)
84 struct dm_rq_clone_bio_info *info =
85 container_of(clone, struct dm_rq_clone_bio_info, clone);
86 struct dm_rq_target_io *tio = info->tio;
87 unsigned int nr_bytes = info->orig->bi_iter.bi_size;
88 blk_status_t error = clone->bi_status;
89 bool is_last = !clone->bi_next;
95 * An error has already been detected on the request.
96 * Once error occurred, just let clone->end_io() handle
102 * Don't notice the error to the upper layer yet.
103 * The error handling decision is made by the target driver,
104 * when the request is completed.
111 * I/O for the bio successfully completed.
112 * Notice the data completion to the upper layer.
114 tio->completed += nr_bytes;
117 * Update the original request.
118 * Do not use blk_mq_end_request() here, because it may complete
119 * the original request before the clone, and break the ordering.
123 blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
126 static struct dm_rq_target_io *tio_from_request(struct request *rq)
128 return blk_mq_rq_to_pdu(rq);
131 static void rq_end_stats(struct mapped_device *md, struct request *orig)
133 if (unlikely(dm_stats_used(&md->stats))) {
134 struct dm_rq_target_io *tio = tio_from_request(orig);
135 tio->duration_jiffies = jiffies - tio->duration_jiffies;
136 dm_stats_account_io(&md->stats, rq_data_dir(orig),
137 blk_rq_pos(orig), tio->n_sectors, true,
138 tio->duration_jiffies, &tio->stats_aux);
143 * Don't touch any member of the md after calling this function because
144 * the md may be freed in dm_put() at the end of this function.
145 * Or do dm_get() before calling this function and dm_put() later.
147 static void rq_completed(struct mapped_device *md)
150 * dm_put() must be at the end of this function. See the comment above
156 * Complete the clone and the original request.
157 * Must be called without clone's queue lock held,
158 * see end_clone_request() for more details.
160 static void dm_end_request(struct request *clone, blk_status_t error)
162 struct dm_rq_target_io *tio = clone->end_io_data;
163 struct mapped_device *md = tio->md;
164 struct request *rq = tio->orig;
166 blk_rq_unprep_clone(clone);
167 tio->ti->type->release_clone_rq(clone, NULL);
169 rq_end_stats(md, rq);
170 blk_mq_end_request(rq, error);
174 static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
176 blk_mq_delay_kick_requeue_list(q, msecs);
179 void dm_mq_kick_requeue_list(struct mapped_device *md)
181 __dm_mq_kick_requeue_list(dm_get_md_queue(md), 0);
183 EXPORT_SYMBOL(dm_mq_kick_requeue_list);
185 static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
187 blk_mq_requeue_request(rq, false);
188 __dm_mq_kick_requeue_list(rq->q, msecs);
191 static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
193 struct mapped_device *md = tio->md;
194 struct request *rq = tio->orig;
195 unsigned long delay_ms = delay_requeue ? 100 : 0;
197 rq_end_stats(md, rq);
199 blk_rq_unprep_clone(tio->clone);
200 tio->ti->type->release_clone_rq(tio->clone, NULL);
203 dm_mq_delay_requeue_request(rq, delay_ms);
207 static void dm_done(struct request *clone, blk_status_t error, bool mapped)
209 int r = DM_ENDIO_DONE;
210 struct dm_rq_target_io *tio = clone->end_io_data;
211 dm_request_endio_fn rq_end_io = NULL;
214 rq_end_io = tio->ti->type->rq_end_io;
216 if (mapped && rq_end_io)
217 r = rq_end_io(tio->ti, clone, error, &tio->info);
220 if (unlikely(error == BLK_STS_TARGET)) {
221 if (req_op(clone) == REQ_OP_DISCARD &&
222 !clone->q->limits.max_discard_sectors)
223 disable_discard(tio->md);
224 else if (req_op(clone) == REQ_OP_WRITE_SAME &&
225 !clone->q->limits.max_write_same_sectors)
226 disable_write_same(tio->md);
227 else if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
228 !clone->q->limits.max_write_zeroes_sectors)
229 disable_write_zeroes(tio->md);
234 /* The target wants to complete the I/O */
235 dm_end_request(clone, error);
237 case DM_ENDIO_INCOMPLETE:
238 /* The target will handle the I/O */
240 case DM_ENDIO_REQUEUE:
241 /* The target wants to requeue the I/O */
242 dm_requeue_original_request(tio, false);
244 case DM_ENDIO_DELAY_REQUEUE:
245 /* The target wants to requeue the I/O after a delay */
246 dm_requeue_original_request(tio, true);
249 DMWARN("unimplemented target endio return value: %d", r);
255 * Request completion handler for request-based dm
257 static void dm_softirq_done(struct request *rq)
260 struct dm_rq_target_io *tio = tio_from_request(rq);
261 struct request *clone = tio->clone;
264 struct mapped_device *md = tio->md;
266 rq_end_stats(md, rq);
267 blk_mq_end_request(rq, tio->error);
272 if (rq->rq_flags & RQF_FAILED)
275 dm_done(clone, tio->error, mapped);
279 * Complete the clone and the original request with the error status
280 * through softirq context.
282 static void dm_complete_request(struct request *rq, blk_status_t error)
284 struct dm_rq_target_io *tio = tio_from_request(rq);
287 if (likely(!blk_should_fake_timeout(rq->q)))
288 blk_mq_complete_request(rq);
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.
297 static void dm_kill_unmapped_request(struct request *rq, blk_status_t error)
299 rq->rq_flags |= RQF_FAILED;
300 dm_complete_request(rq, error);
303 static void end_clone_request(struct request *clone, blk_status_t error)
305 struct dm_rq_target_io *tio = clone->end_io_data;
307 dm_complete_request(tio->orig, error);
310 static blk_status_t dm_dispatch_clone_request(struct request *clone, struct request *rq)
314 if (blk_queue_io_stat(clone->q))
315 clone->rq_flags |= RQF_IO_STAT;
317 clone->start_time_ns = ktime_get_ns();
318 r = blk_insert_cloned_request(clone->q, clone);
319 if (r != BLK_STS_OK && r != BLK_STS_RESOURCE && r != BLK_STS_DEV_RESOURCE)
320 /* must complete clone in terms of original request */
321 dm_complete_request(rq, r);
325 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
328 struct dm_rq_target_io *tio = data;
329 struct dm_rq_clone_bio_info *info =
330 container_of(bio, struct dm_rq_clone_bio_info, clone);
332 info->orig = bio_orig;
334 bio->bi_end_io = end_clone_bio;
339 static int setup_clone(struct request *clone, struct request *rq,
340 struct dm_rq_target_io *tio, gfp_t gfp_mask)
344 r = blk_rq_prep_clone(clone, rq, &tio->md->bs, gfp_mask,
345 dm_rq_bio_constructor, tio);
349 clone->end_io = end_clone_request;
350 clone->end_io_data = tio;
357 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
358 struct mapped_device *md)
367 * Avoid initializing info for blk-mq; it passes
368 * target-specific data through info.ptr
369 * (see: dm_mq_init_request)
371 if (!md->init_tio_pdu)
372 memset(&tio->info, 0, sizeof(tio->info));
377 * DM_MAPIO_* : the request has been processed as indicated
378 * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
379 * < 0 : the request was completed due to failure
381 static int map_request(struct dm_rq_target_io *tio)
384 struct dm_target *ti = tio->ti;
385 struct mapped_device *md = tio->md;
386 struct request *rq = tio->orig;
387 struct request *clone = NULL;
390 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
392 case DM_MAPIO_SUBMITTED:
393 /* The target has taken the I/O to submit by itself later */
395 case DM_MAPIO_REMAPPED:
396 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
398 ti->type->release_clone_rq(clone, &tio->info);
399 return DM_MAPIO_REQUEUE;
402 /* The target has remapped the I/O so dispatch it */
403 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
405 ret = dm_dispatch_clone_request(clone, rq);
406 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
407 blk_rq_unprep_clone(clone);
408 blk_mq_cleanup_rq(clone);
409 tio->ti->type->release_clone_rq(clone, &tio->info);
411 return DM_MAPIO_REQUEUE;
414 case DM_MAPIO_REQUEUE:
415 /* The target wants to requeue the I/O */
417 case DM_MAPIO_DELAY_REQUEUE:
418 /* The target wants to requeue the I/O after a delay */
419 dm_requeue_original_request(tio, true);
422 /* The target wants to complete the I/O */
423 dm_kill_unmapped_request(rq, BLK_STS_IOERR);
426 DMWARN("unimplemented target map return value: %d", r);
433 /* DEPRECATED: previously used for request-based merge heuristic in dm_request_fn() */
434 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
436 return sprintf(buf, "%u\n", 0);
439 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
440 const char *buf, size_t count)
445 static void dm_start_request(struct mapped_device *md, struct request *orig)
447 blk_mq_start_request(orig);
449 if (unlikely(dm_stats_used(&md->stats))) {
450 struct dm_rq_target_io *tio = tio_from_request(orig);
451 tio->duration_jiffies = jiffies;
452 tio->n_sectors = blk_rq_sectors(orig);
453 dm_stats_account_io(&md->stats, rq_data_dir(orig),
454 blk_rq_pos(orig), tio->n_sectors, false, 0,
459 * Hold the md reference here for the in-flight I/O.
460 * We can't rely on the reference count by device opener,
461 * because the device may be closed during the request completion
462 * when all bios are completed.
463 * See the comment in rq_completed() too.
468 static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
469 unsigned int hctx_idx, unsigned int numa_node)
471 struct mapped_device *md = set->driver_data;
472 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
475 * Must initialize md member of tio, otherwise it won't
476 * be available in dm_mq_queue_rq.
480 if (md->init_tio_pdu) {
481 /* target-specific per-io data is immediately after the tio */
482 tio->info.ptr = tio + 1;
488 static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
489 const struct blk_mq_queue_data *bd)
491 struct request *rq = bd->rq;
492 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
493 struct mapped_device *md = tio->md;
494 struct dm_target *ti = md->immutable_target;
498 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
500 ti = dm_table_find_target(map, 0);
501 dm_put_live_table(md, srcu_idx);
504 if (ti->type->busy && ti->type->busy(ti))
505 return BLK_STS_RESOURCE;
507 dm_start_request(md, rq);
509 /* Init tio using md established in .init_request */
510 init_tio(tio, rq, md);
513 * Establish tio->ti before calling map_request().
517 /* Direct call is fine since .queue_rq allows allocations */
518 if (map_request(tio) == DM_MAPIO_REQUEUE) {
519 /* Undo dm_start_request() before requeuing */
520 rq_end_stats(md, rq);
522 return BLK_STS_RESOURCE;
528 static const struct blk_mq_ops dm_mq_ops = {
529 .queue_rq = dm_mq_queue_rq,
530 .complete = dm_softirq_done,
531 .init_request = dm_mq_init_request,
534 int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
536 struct request_queue *q;
537 struct dm_target *immutable_tgt;
540 md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
544 md->tag_set->ops = &dm_mq_ops;
545 md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
546 md->tag_set->numa_node = md->numa_node_id;
547 md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_STACKING;
548 md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
549 md->tag_set->driver_data = md;
551 md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
552 immutable_tgt = dm_table_get_immutable_target(t);
553 if (immutable_tgt && immutable_tgt->per_io_data_size) {
554 /* any target-specific per-io data is immediately after the tio */
555 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
556 md->init_tio_pdu = true;
559 err = blk_mq_alloc_tag_set(md->tag_set);
561 goto out_kfree_tag_set;
563 q = blk_mq_init_allocated_queue(md->tag_set, md->queue, true);
572 blk_mq_free_tag_set(md->tag_set);
579 void dm_mq_cleanup_mapped_device(struct mapped_device *md)
582 blk_mq_free_tag_set(md->tag_set);
587 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
588 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
590 /* Unused, but preserved for userspace compatibility */
591 static bool use_blk_mq = true;
592 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
593 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
595 module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
596 MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
598 module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
599 MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");