1/*
2 * Copyright (C) 2016 Red Hat, Inc. All rights reserved.
3 *
4 * This file is released under the GPL.
5 */
6
7#include "dm-core.h"
8#include "dm-rq.h"
9
10#include <linux/elevator.h> /* for rq_end_sector() */
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 n_sectors;
28 unsigned completed;
29};
30
31#define DM_MQ_NR_HW_QUEUES 1
32#define DM_MQ_QUEUE_DEPTH 2048
33static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
34static unsigned 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 reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
41
42unsigned 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}
47EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
48
49static unsigned dm_get_blk_mq_nr_hw_queues(void)
50{
51 return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
52}
53
54static unsigned dm_get_blk_mq_queue_depth(void)
55{
56 return __dm_get_module_param(&dm_mq_queue_depth,
57 DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
58}
59
60int dm_request_based(struct mapped_device *md)
61{
62 return queue_is_mq(md->queue);
63}
64
65void dm_start_queue(struct request_queue *q)
66{
67 blk_mq_unquiesce_queue(q);
68 blk_mq_kick_requeue_list(q);
69}
70
71void dm_stop_queue(struct request_queue *q)
72{
73 if (blk_mq_queue_stopped(q))
74 return;
75
76 blk_mq_quiesce_queue(q);
77}
78
79/*
80 * Partial completion handling for request-based dm
81 */
82static void end_clone_bio(struct bio *clone)
83{
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;
90
91 bio_put(clone);
92
93 if (tio->error)
94 /*
95 * An error has already been detected on the request.
96 * Once error occurred, just let clone->end_io() handle
97 * the remainder.
98 */
99 return;
100 else if (error) {
101 /*
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.
105 */
106 tio->error = error;
107 goto exit;
108 }
109
110 /*
111 * I/O for the bio successfully completed.
112 * Notice the data completion to the upper layer.
113 */
114 tio->completed += nr_bytes;
115
116 /*
117 * Update the original request.
118 * Do not use blk_end_request() here, because it may complete
119 * the original request before the clone, and break the ordering.
120 */
121 if (is_last)
122 exit:
123 blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
124}
125
126static struct dm_rq_target_io *tio_from_request(struct request *rq)
127{
128 return blk_mq_rq_to_pdu(rq);
129}
130
131static void rq_end_stats(struct mapped_device *md, struct request *orig)
132{
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);
139 }
140}
141
142/*
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.
146 */
147static void rq_completed(struct mapped_device *md)
148{
149 /* nudge anyone waiting on suspend queue */
150 if (unlikely(wq_has_sleeper(&md->wait)))
151 wake_up(&md->wait);
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);
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(dm_get_md_queue(md), 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);
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_WRITE_SAME &&
226 !clone->q->limits.max_write_same_sectors)
227 disable_write_same(tio->md);
228 if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
229 !clone->q->limits.max_write_zeroes_sectors)
230 disable_write_zeroes(tio->md);
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 DMWARN("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 void end_clone_request(struct request *clone, blk_status_t error)
304{
305 struct dm_rq_target_io *tio = clone->end_io_data;
306
307 dm_complete_request(tio->orig, error);
308}
309
310static blk_status_t dm_dispatch_clone_request(struct request *clone, struct request *rq)
311{
312 blk_status_t r;
313
314 if (blk_queue_io_stat(clone->q))
315 clone->rq_flags |= RQF_IO_STAT;
316
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);
322 return r;
323}
324
325static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
326 void *data)
327{
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);
331
332 info->orig = bio_orig;
333 info->tio = tio;
334 bio->bi_end_io = end_clone_bio;
335
336 return 0;
337}
338
339static int setup_clone(struct request *clone, struct request *rq,
340 struct dm_rq_target_io *tio, gfp_t gfp_mask)
341{
342 int r;
343
344 r = blk_rq_prep_clone(clone, rq, &tio->md->bs, gfp_mask,
345 dm_rq_bio_constructor, tio);
346 if (r)
347 return r;
348
349 clone->end_io = end_clone_request;
350 clone->end_io_data = tio;
351
352 tio->clone = clone;
353
354 return 0;
355}
356
357static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
358 struct mapped_device *md)
359{
360 tio->md = md;
361 tio->ti = NULL;
362 tio->clone = NULL;
363 tio->orig = rq;
364 tio->error = 0;
365 tio->completed = 0;
366 /*
367 * Avoid initializing info for blk-mq; it passes
368 * target-specific data through info.ptr
369 * (see: dm_mq_init_request)
370 */
371 if (!md->init_tio_pdu)
372 memset(&tio->info, 0, sizeof(tio->info));
373}
374
375/*
376 * Returns:
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
380 */
381static int map_request(struct dm_rq_target_io *tio)
382{
383 int r;
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;
388 blk_status_t ret;
389
390 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
391 switch (r) {
392 case DM_MAPIO_SUBMITTED:
393 /* The target has taken the I/O to submit by itself later */
394 break;
395 case DM_MAPIO_REMAPPED:
396 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
397 /* -ENOMEM */
398 ti->type->release_clone_rq(clone);
399 return DM_MAPIO_REQUEUE;
400 }
401
402 /* The target has remapped the I/O so dispatch it */
403 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
404 blk_rq_pos(rq));
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 tio->ti->type->release_clone_rq(clone);
409 tio->clone = NULL;
410 return DM_MAPIO_REQUEUE;
411 }
412 break;
413 case DM_MAPIO_REQUEUE:
414 /* The target wants to requeue the I/O */
415 break;
416 case DM_MAPIO_DELAY_REQUEUE:
417 /* The target wants to requeue the I/O after a delay */
418 dm_requeue_original_request(tio, true);
419 break;
420 case DM_MAPIO_KILL:
421 /* The target wants to complete the I/O */
422 dm_kill_unmapped_request(rq, BLK_STS_IOERR);
423 break;
424 default:
425 DMWARN("unimplemented target map return value: %d", r);
426 BUG();
427 }
428
429 return r;
430}
431
432/* DEPRECATED: previously used for request-based merge heuristic in dm_request_fn() */
433ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
434{
435 return sprintf(buf, "%u\n", 0);
436}
437
438ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
439 const char *buf, size_t count)
440{
441 return count;
442}
443
444static void dm_start_request(struct mapped_device *md, struct request *orig)
445{
446 blk_mq_start_request(orig);
447
448 if (unlikely(dm_stats_used(&md->stats))) {
449 struct dm_rq_target_io *tio = tio_from_request(orig);
450 tio->duration_jiffies = jiffies;
451 tio->n_sectors = blk_rq_sectors(orig);
452 dm_stats_account_io(&md->stats, rq_data_dir(orig),
453 blk_rq_pos(orig), tio->n_sectors, false, 0,
454 &tio->stats_aux);
455 }
456
457 /*
458 * Hold the md reference here for the in-flight I/O.
459 * We can't rely on the reference count by device opener,
460 * because the device may be closed during the request completion
461 * when all bios are completed.
462 * See the comment in rq_completed() too.
463 */
464 dm_get(md);
465}
466
467static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
468 unsigned int hctx_idx, unsigned int numa_node)
469{
470 struct mapped_device *md = set->driver_data;
471 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
472
473 /*
474 * Must initialize md member of tio, otherwise it won't
475 * be available in dm_mq_queue_rq.
476 */
477 tio->md = md;
478
479 if (md->init_tio_pdu) {
480 /* target-specific per-io data is immediately after the tio */
481 tio->info.ptr = tio + 1;
482 }
483
484 return 0;
485}
486
487static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
488 const struct blk_mq_queue_data *bd)
489{
490 struct request *rq = bd->rq;
491 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
492 struct mapped_device *md = tio->md;
493 struct dm_target *ti = md->immutable_target;
494
495 if (unlikely(!ti)) {
496 int srcu_idx;
497 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
498
499 ti = dm_table_find_target(map, 0);
500 dm_put_live_table(md, srcu_idx);
501 }
502
503 if (ti->type->busy && ti->type->busy(ti))
504 return BLK_STS_RESOURCE;
505
506 dm_start_request(md, rq);
507
508 /* Init tio using md established in .init_request */
509 init_tio(tio, rq, md);
510
511 /*
512 * Establish tio->ti before calling map_request().
513 */
514 tio->ti = ti;
515
516 /* Direct call is fine since .queue_rq allows allocations */
517 if (map_request(tio) == DM_MAPIO_REQUEUE) {
518 /* Undo dm_start_request() before requeuing */
519 rq_end_stats(md, rq);
520 rq_completed(md);
521 return BLK_STS_RESOURCE;
522 }
523
524 return BLK_STS_OK;
525}
526
527static const struct blk_mq_ops dm_mq_ops = {
528 .queue_rq = dm_mq_queue_rq,
529 .complete = dm_softirq_done,
530 .init_request = dm_mq_init_request,
531};
532
533int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
534{
535 struct request_queue *q;
536 struct dm_target *immutable_tgt;
537 int err;
538
539 md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
540 if (!md->tag_set)
541 return -ENOMEM;
542
543 md->tag_set->ops = &dm_mq_ops;
544 md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
545 md->tag_set->numa_node = md->numa_node_id;
546 md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE;
547 md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
548 md->tag_set->driver_data = md;
549
550 md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
551 immutable_tgt = dm_table_get_immutable_target(t);
552 if (immutable_tgt && immutable_tgt->per_io_data_size) {
553 /* any target-specific per-io data is immediately after the tio */
554 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
555 md->init_tio_pdu = true;
556 }
557
558 err = blk_mq_alloc_tag_set(md->tag_set);
559 if (err)
560 goto out_kfree_tag_set;
561
562 q = blk_mq_init_allocated_queue(md->tag_set, md->queue);
563 if (IS_ERR(q)) {
564 err = PTR_ERR(q);
565 goto out_tag_set;
566 }
567
568 return 0;
569
570out_tag_set:
571 blk_mq_free_tag_set(md->tag_set);
572out_kfree_tag_set:
573 kfree(md->tag_set);
574
575 return err;
576}
577
578void dm_mq_cleanup_mapped_device(struct mapped_device *md)
579{
580 if (md->tag_set) {
581 blk_mq_free_tag_set(md->tag_set);
582 kfree(md->tag_set);
583 }
584}
585
586module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
587MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
588
589/* Unused, but preserved for userspace compatibility */
590static bool use_blk_mq = true;
591module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
592MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
593
594module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
595MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
596
597module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
598MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");
599