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"
15 #define DM_MQ_NR_HW_QUEUES 1
16 #define DM_MQ_QUEUE_DEPTH 2048
17 static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
18 static unsigned dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
21 * Request-based DM's mempools' reserved IOs set by the user.
23 #define RESERVED_REQUEST_BASED_IOS 256
24 static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
26 unsigned dm_get_reserved_rq_based_ios(void)
28 return __dm_get_module_param(&reserved_rq_based_ios,
29 RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
31 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
33 static unsigned dm_get_blk_mq_nr_hw_queues(void)
35 return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
38 static unsigned dm_get_blk_mq_queue_depth(void)
40 return __dm_get_module_param(&dm_mq_queue_depth,
41 DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
44 int dm_request_based(struct mapped_device *md)
46 return queue_is_rq_based(md->queue);
49 void dm_start_queue(struct request_queue *q)
51 blk_mq_unquiesce_queue(q);
52 blk_mq_kick_requeue_list(q);
55 void dm_stop_queue(struct request_queue *q)
57 if (blk_mq_queue_stopped(q))
60 blk_mq_quiesce_queue(q);
64 * Partial completion handling for request-based dm
66 static void end_clone_bio(struct bio *clone)
68 struct dm_rq_clone_bio_info *info =
69 container_of(clone, struct dm_rq_clone_bio_info, clone);
70 struct dm_rq_target_io *tio = info->tio;
71 unsigned int nr_bytes = info->orig->bi_iter.bi_size;
72 blk_status_t error = clone->bi_status;
73 bool is_last = !clone->bi_next;
79 * An error has already been detected on the request.
80 * Once error occurred, just let clone->end_io() handle
86 * Don't notice the error to the upper layer yet.
87 * The error handling decision is made by the target driver,
88 * when the request is completed.
95 * I/O for the bio successfully completed.
96 * Notice the data completion to the upper layer.
98 tio->completed += nr_bytes;
101 * Update the original request.
102 * Do not use blk_end_request() here, because it may complete
103 * the original request before the clone, and break the ordering.
107 blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
110 static struct dm_rq_target_io *tio_from_request(struct request *rq)
112 return blk_mq_rq_to_pdu(rq);
115 static void rq_end_stats(struct mapped_device *md, struct request *orig)
117 if (unlikely(dm_stats_used(&md->stats))) {
118 struct dm_rq_target_io *tio = tio_from_request(orig);
119 tio->duration_jiffies = jiffies - tio->duration_jiffies;
120 dm_stats_account_io(&md->stats, rq_data_dir(orig),
121 blk_rq_pos(orig), tio->n_sectors, true,
122 tio->duration_jiffies, &tio->stats_aux);
127 * Don't touch any member of the md after calling this function because
128 * the md may be freed in dm_put() at the end of this function.
129 * Or do dm_get() before calling this function and dm_put() later.
131 static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
133 atomic_dec(&md->pending[rw]);
135 /* nudge anyone waiting on suspend queue */
136 if (!md_in_flight(md))
140 * dm_put() must be at the end of this function. See the comment above
146 * Complete the clone and the original request.
147 * Must be called without clone's queue lock held,
148 * see end_clone_request() for more details.
150 static void dm_end_request(struct request *clone, blk_status_t error)
152 int rw = rq_data_dir(clone);
153 struct dm_rq_target_io *tio = clone->end_io_data;
154 struct mapped_device *md = tio->md;
155 struct request *rq = tio->orig;
157 blk_rq_unprep_clone(clone);
158 tio->ti->type->release_clone_rq(clone);
160 rq_end_stats(md, rq);
161 blk_mq_end_request(rq, error);
162 rq_completed(md, rw, true);
165 static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
167 blk_mq_delay_kick_requeue_list(q, msecs);
170 void dm_mq_kick_requeue_list(struct mapped_device *md)
172 __dm_mq_kick_requeue_list(dm_get_md_queue(md), 0);
174 EXPORT_SYMBOL(dm_mq_kick_requeue_list);
176 static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
178 blk_mq_requeue_request(rq, false);
179 __dm_mq_kick_requeue_list(rq->q, msecs);
182 static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
184 struct mapped_device *md = tio->md;
185 struct request *rq = tio->orig;
186 int rw = rq_data_dir(rq);
187 unsigned long delay_ms = delay_requeue ? 100 : 0;
189 rq_end_stats(md, rq);
191 blk_rq_unprep_clone(tio->clone);
192 tio->ti->type->release_clone_rq(tio->clone);
195 dm_mq_delay_requeue_request(rq, delay_ms);
196 rq_completed(md, rw, false);
199 static void dm_done(struct request *clone, blk_status_t error, bool mapped)
201 int r = DM_ENDIO_DONE;
202 struct dm_rq_target_io *tio = clone->end_io_data;
203 dm_request_endio_fn rq_end_io = NULL;
206 rq_end_io = tio->ti->type->rq_end_io;
208 if (mapped && rq_end_io)
209 r = rq_end_io(tio->ti, clone, error, &tio->info);
212 if (unlikely(error == BLK_STS_TARGET)) {
213 if (req_op(clone) == REQ_OP_WRITE_SAME &&
214 !clone->q->limits.max_write_same_sectors)
215 disable_write_same(tio->md);
216 if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
217 !clone->q->limits.max_write_zeroes_sectors)
218 disable_write_zeroes(tio->md);
223 /* The target wants to complete the I/O */
224 dm_end_request(clone, error);
226 case DM_ENDIO_INCOMPLETE:
227 /* The target will handle the I/O */
229 case DM_ENDIO_REQUEUE:
230 /* The target wants to requeue the I/O */
231 dm_requeue_original_request(tio, false);
233 case DM_ENDIO_DELAY_REQUEUE:
234 /* The target wants to requeue the I/O after a delay */
235 dm_requeue_original_request(tio, true);
238 DMWARN("unimplemented target endio return value: %d", r);
244 * Request completion handler for request-based dm
246 static void dm_softirq_done(struct request *rq)
249 struct dm_rq_target_io *tio = tio_from_request(rq);
250 struct request *clone = tio->clone;
254 struct mapped_device *md = tio->md;
256 rq_end_stats(md, rq);
257 rw = rq_data_dir(rq);
258 blk_mq_end_request(rq, tio->error);
259 rq_completed(md, rw, false);
263 if (rq->rq_flags & RQF_FAILED)
266 dm_done(clone, tio->error, mapped);
270 * Complete the clone and the original request with the error status
271 * through softirq context.
273 static void dm_complete_request(struct request *rq, blk_status_t error)
275 struct dm_rq_target_io *tio = tio_from_request(rq);
278 blk_mq_complete_request(rq);
282 * Complete the not-mapped clone and the original request with the error status
283 * through softirq context.
284 * Target's rq_end_io() function isn't called.
285 * This may be used when the target's clone_and_map_rq() function fails.
287 static void dm_kill_unmapped_request(struct request *rq, blk_status_t error)
289 rq->rq_flags |= RQF_FAILED;
290 dm_complete_request(rq, error);
293 static void end_clone_request(struct request *clone, blk_status_t error)
295 struct dm_rq_target_io *tio = clone->end_io_data;
297 dm_complete_request(tio->orig, error);
300 static blk_status_t dm_dispatch_clone_request(struct request *clone, struct request *rq)
304 if (blk_queue_io_stat(clone->q))
305 clone->rq_flags |= RQF_IO_STAT;
307 clone->start_time_ns = ktime_get_ns();
308 r = blk_insert_cloned_request(clone->q, clone);
309 if (r != BLK_STS_OK && r != BLK_STS_RESOURCE && r != BLK_STS_DEV_RESOURCE)
310 /* must complete clone in terms of original request */
311 dm_complete_request(rq, r);
315 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
318 struct dm_rq_target_io *tio = data;
319 struct dm_rq_clone_bio_info *info =
320 container_of(bio, struct dm_rq_clone_bio_info, clone);
322 info->orig = bio_orig;
324 bio->bi_end_io = end_clone_bio;
329 static int setup_clone(struct request *clone, struct request *rq,
330 struct dm_rq_target_io *tio, gfp_t gfp_mask)
334 r = blk_rq_prep_clone(clone, rq, &tio->md->bs, gfp_mask,
335 dm_rq_bio_constructor, tio);
339 clone->end_io = end_clone_request;
340 clone->end_io_data = tio;
347 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
348 struct mapped_device *md)
357 * Avoid initializing info for blk-mq; it passes
358 * target-specific data through info.ptr
359 * (see: dm_mq_init_request)
361 if (!md->init_tio_pdu)
362 memset(&tio->info, 0, sizeof(tio->info));
367 * DM_MAPIO_* : the request has been processed as indicated
368 * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
369 * < 0 : the request was completed due to failure
371 static int map_request(struct dm_rq_target_io *tio)
374 struct dm_target *ti = tio->ti;
375 struct mapped_device *md = tio->md;
376 struct request *rq = tio->orig;
377 struct request *clone = NULL;
380 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
383 case DM_MAPIO_SUBMITTED:
384 /* The target has taken the I/O to submit by itself later */
386 case DM_MAPIO_REMAPPED:
387 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
389 ti->type->release_clone_rq(clone);
390 return DM_MAPIO_REQUEUE;
393 /* The target has remapped the I/O so dispatch it */
394 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
396 ret = dm_dispatch_clone_request(clone, rq);
397 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
398 blk_rq_unprep_clone(clone);
399 tio->ti->type->release_clone_rq(clone);
401 r = DM_MAPIO_REQUEUE;
405 case DM_MAPIO_REQUEUE:
406 /* The target wants to requeue the I/O */
408 case DM_MAPIO_DELAY_REQUEUE:
409 /* The target wants to requeue the I/O after a delay */
410 dm_requeue_original_request(tio, true);
413 /* The target wants to complete the I/O */
414 dm_kill_unmapped_request(rq, BLK_STS_IOERR);
417 DMWARN("unimplemented target map return value: %d", r);
424 /* DEPRECATED: previously used for request-based merge heuristic in dm_request_fn() */
425 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
427 return sprintf(buf, "%u\n", 0);
430 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
431 const char *buf, size_t count)
436 static void dm_start_request(struct mapped_device *md, struct request *orig)
438 blk_mq_start_request(orig);
439 atomic_inc(&md->pending[rq_data_dir(orig)]);
441 if (unlikely(dm_stats_used(&md->stats))) {
442 struct dm_rq_target_io *tio = tio_from_request(orig);
443 tio->duration_jiffies = jiffies;
444 tio->n_sectors = blk_rq_sectors(orig);
445 dm_stats_account_io(&md->stats, rq_data_dir(orig),
446 blk_rq_pos(orig), tio->n_sectors, false, 0,
451 * Hold the md reference here for the in-flight I/O.
452 * We can't rely on the reference count by device opener,
453 * because the device may be closed during the request completion
454 * when all bios are completed.
455 * See the comment in rq_completed() too.
460 static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
461 unsigned int hctx_idx, unsigned int numa_node)
463 struct mapped_device *md = set->driver_data;
464 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
467 * Must initialize md member of tio, otherwise it won't
468 * be available in dm_mq_queue_rq.
472 if (md->init_tio_pdu) {
473 /* target-specific per-io data is immediately after the tio */
474 tio->info.ptr = tio + 1;
480 static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
481 const struct blk_mq_queue_data *bd)
483 struct request *rq = bd->rq;
484 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
485 struct mapped_device *md = tio->md;
486 struct dm_target *ti = md->immutable_target;
490 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
492 ti = dm_table_find_target(map, 0);
493 dm_put_live_table(md, srcu_idx);
496 if (ti->type->busy && ti->type->busy(ti))
497 return BLK_STS_RESOURCE;
499 dm_start_request(md, rq);
501 /* Init tio using md established in .init_request */
502 init_tio(tio, rq, md);
505 * Establish tio->ti before calling map_request().
509 /* Direct call is fine since .queue_rq allows allocations */
510 if (map_request(tio) == DM_MAPIO_REQUEUE) {
511 /* Undo dm_start_request() before requeuing */
512 rq_end_stats(md, rq);
513 rq_completed(md, rq_data_dir(rq), false);
514 return BLK_STS_RESOURCE;
520 static const struct blk_mq_ops dm_mq_ops = {
521 .queue_rq = dm_mq_queue_rq,
522 .complete = dm_softirq_done,
523 .init_request = dm_mq_init_request,
526 int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
528 struct request_queue *q;
529 struct dm_target *immutable_tgt;
532 md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
536 md->tag_set->ops = &dm_mq_ops;
537 md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
538 md->tag_set->numa_node = md->numa_node_id;
539 md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
540 md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
541 md->tag_set->driver_data = md;
543 md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
544 immutable_tgt = dm_table_get_immutable_target(t);
545 if (immutable_tgt && immutable_tgt->per_io_data_size) {
546 /* any target-specific per-io data is immediately after the tio */
547 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
548 md->init_tio_pdu = true;
551 err = blk_mq_alloc_tag_set(md->tag_set);
553 goto out_kfree_tag_set;
555 q = blk_mq_init_allocated_queue(md->tag_set, md->queue);
564 blk_mq_free_tag_set(md->tag_set);
571 void dm_mq_cleanup_mapped_device(struct mapped_device *md)
574 blk_mq_free_tag_set(md->tag_set);
579 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
580 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
582 /* Unused, but preserved for userspace compatibility */
583 static bool use_blk_mq = true;
584 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
585 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
587 module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
588 MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
590 module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
591 MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");