1 | // SPDX-License-Identifier: GPL-2.0-only |
2 | /* |
3 | * Copyright (C) 2016 Red Hat, Inc. All rights reserved. |
4 | * |
5 | * This file is released under the GPL. |
6 | */ |
7 | |
8 | #include "dm-core.h" |
9 | #include "dm-rq.h" |
10 | |
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 | */ |
18 | struct 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 int n_sectors; |
28 | unsigned int completed; |
29 | }; |
30 | |
31 | #define DM_MQ_NR_HW_QUEUES 1 |
32 | #define DM_MQ_QUEUE_DEPTH 2048 |
33 | static unsigned int dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES; |
34 | static unsigned int 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 |
40 | static unsigned int reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS; |
41 | |
42 | unsigned int dm_get_reserved_rq_based_ios(void) |
43 | { |
44 | return __dm_get_module_param(module_param: &reserved_rq_based_ios, |
45 | RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS); |
46 | } |
47 | |
48 | static unsigned int dm_get_blk_mq_nr_hw_queues(void) |
49 | { |
50 | return __dm_get_module_param(module_param: &dm_mq_nr_hw_queues, def: 1, max: 32); |
51 | } |
52 | |
53 | static unsigned int dm_get_blk_mq_queue_depth(void) |
54 | { |
55 | return __dm_get_module_param(module_param: &dm_mq_queue_depth, |
56 | DM_MQ_QUEUE_DEPTH, max: BLK_MQ_MAX_DEPTH); |
57 | } |
58 | |
59 | int dm_request_based(struct mapped_device *md) |
60 | { |
61 | return queue_is_mq(q: md->queue); |
62 | } |
63 | |
64 | void dm_start_queue(struct request_queue *q) |
65 | { |
66 | blk_mq_unquiesce_queue(q); |
67 | blk_mq_kick_requeue_list(q); |
68 | } |
69 | |
70 | void dm_stop_queue(struct request_queue *q) |
71 | { |
72 | blk_mq_quiesce_queue(q); |
73 | } |
74 | |
75 | /* |
76 | * Partial completion handling for request-based dm |
77 | */ |
78 | static void end_clone_bio(struct bio *clone) |
79 | { |
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; |
86 | |
87 | bio_put(clone); |
88 | |
89 | if (tio->error) |
90 | /* |
91 | * An error has already been detected on the request. |
92 | * Once error occurred, just let clone->end_io() handle |
93 | * the remainder. |
94 | */ |
95 | return; |
96 | else if (error) { |
97 | /* |
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. |
101 | */ |
102 | tio->error = error; |
103 | goto exit; |
104 | } |
105 | |
106 | /* |
107 | * I/O for the bio successfully completed. |
108 | * Notice the data completion to the upper layer. |
109 | */ |
110 | tio->completed += nr_bytes; |
111 | |
112 | /* |
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. |
116 | */ |
117 | if (is_last) |
118 | exit: |
119 | blk_update_request(rq: tio->orig, BLK_STS_OK, nr_bytes: tio->completed); |
120 | } |
121 | |
122 | static struct dm_rq_target_io *tio_from_request(struct request *rq) |
123 | { |
124 | return blk_mq_rq_to_pdu(rq); |
125 | } |
126 | |
127 | static void rq_end_stats(struct mapped_device *md, struct request *orig) |
128 | { |
129 | if (unlikely(dm_stats_used(&md->stats))) { |
130 | struct dm_rq_target_io *tio = tio_from_request(rq: orig); |
131 | |
132 | tio->duration_jiffies = jiffies - tio->duration_jiffies; |
133 | dm_stats_account_io(stats: &md->stats, rq_data_dir(orig), |
134 | bi_sector: blk_rq_pos(rq: orig), bi_sectors: tio->n_sectors, end: true, |
135 | start_time: tio->duration_jiffies, aux: &tio->stats_aux); |
136 | } |
137 | } |
138 | |
139 | /* |
140 | * Don't touch any member of the md after calling this function because |
141 | * the md may be freed in dm_put() at the end of this function. |
142 | * Or do dm_get() before calling this function and dm_put() later. |
143 | */ |
144 | static void rq_completed(struct mapped_device *md) |
145 | { |
146 | /* |
147 | * dm_put() must be at the end of this function. See the comment above |
148 | */ |
149 | dm_put(md); |
150 | } |
151 | |
152 | /* |
153 | * Complete the clone and the original request. |
154 | * Must be called without clone's queue lock held, |
155 | * see end_clone_request() for more details. |
156 | */ |
157 | static void dm_end_request(struct request *clone, blk_status_t error) |
158 | { |
159 | struct dm_rq_target_io *tio = clone->end_io_data; |
160 | struct mapped_device *md = tio->md; |
161 | struct request *rq = tio->orig; |
162 | |
163 | blk_rq_unprep_clone(rq: clone); |
164 | tio->ti->type->release_clone_rq(clone, NULL); |
165 | |
166 | rq_end_stats(md, orig: rq); |
167 | blk_mq_end_request(rq, error); |
168 | rq_completed(md); |
169 | } |
170 | |
171 | static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs) |
172 | { |
173 | blk_mq_delay_kick_requeue_list(q, msecs); |
174 | } |
175 | |
176 | void dm_mq_kick_requeue_list(struct mapped_device *md) |
177 | { |
178 | __dm_mq_kick_requeue_list(q: md->queue, msecs: 0); |
179 | } |
180 | EXPORT_SYMBOL(dm_mq_kick_requeue_list); |
181 | |
182 | static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs) |
183 | { |
184 | blk_mq_requeue_request(rq, kick_requeue_list: false); |
185 | __dm_mq_kick_requeue_list(q: rq->q, msecs); |
186 | } |
187 | |
188 | static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue) |
189 | { |
190 | struct mapped_device *md = tio->md; |
191 | struct request *rq = tio->orig; |
192 | unsigned long delay_ms = delay_requeue ? 100 : 0; |
193 | |
194 | rq_end_stats(md, orig: rq); |
195 | if (tio->clone) { |
196 | blk_rq_unprep_clone(rq: tio->clone); |
197 | tio->ti->type->release_clone_rq(tio->clone, NULL); |
198 | } |
199 | |
200 | dm_mq_delay_requeue_request(rq, msecs: delay_ms); |
201 | rq_completed(md); |
202 | } |
203 | |
204 | static void dm_done(struct request *clone, blk_status_t error, bool mapped) |
205 | { |
206 | int r = DM_ENDIO_DONE; |
207 | struct dm_rq_target_io *tio = clone->end_io_data; |
208 | dm_request_endio_fn rq_end_io = NULL; |
209 | |
210 | if (tio->ti) { |
211 | rq_end_io = tio->ti->type->rq_end_io; |
212 | |
213 | if (mapped && rq_end_io) |
214 | r = rq_end_io(tio->ti, clone, error, &tio->info); |
215 | } |
216 | |
217 | if (unlikely(error == BLK_STS_TARGET)) { |
218 | if (req_op(req: clone) == REQ_OP_DISCARD && |
219 | !clone->q->limits.max_discard_sectors) |
220 | disable_discard(md: tio->md); |
221 | else if (req_op(req: clone) == REQ_OP_WRITE_ZEROES && |
222 | !clone->q->limits.max_write_zeroes_sectors) |
223 | disable_write_zeroes(md: tio->md); |
224 | } |
225 | |
226 | switch (r) { |
227 | case DM_ENDIO_DONE: |
228 | /* The target wants to complete the I/O */ |
229 | dm_end_request(clone, error); |
230 | break; |
231 | case DM_ENDIO_INCOMPLETE: |
232 | /* The target will handle the I/O */ |
233 | return; |
234 | case DM_ENDIO_REQUEUE: |
235 | /* The target wants to requeue the I/O */ |
236 | dm_requeue_original_request(tio, delay_requeue: false); |
237 | break; |
238 | case DM_ENDIO_DELAY_REQUEUE: |
239 | /* The target wants to requeue the I/O after a delay */ |
240 | dm_requeue_original_request(tio, delay_requeue: true); |
241 | break; |
242 | default: |
243 | DMCRIT("unimplemented target endio return value: %d" , r); |
244 | BUG(); |
245 | } |
246 | } |
247 | |
248 | /* |
249 | * Request completion handler for request-based dm |
250 | */ |
251 | static void dm_softirq_done(struct request *rq) |
252 | { |
253 | bool mapped = true; |
254 | struct dm_rq_target_io *tio = tio_from_request(rq); |
255 | struct request *clone = tio->clone; |
256 | |
257 | if (!clone) { |
258 | struct mapped_device *md = tio->md; |
259 | |
260 | rq_end_stats(md, orig: rq); |
261 | blk_mq_end_request(rq, error: tio->error); |
262 | rq_completed(md); |
263 | return; |
264 | } |
265 | |
266 | if (rq->rq_flags & RQF_FAILED) |
267 | mapped = false; |
268 | |
269 | dm_done(clone, error: tio->error, mapped); |
270 | } |
271 | |
272 | /* |
273 | * Complete the clone and the original request with the error status |
274 | * through softirq context. |
275 | */ |
276 | static void dm_complete_request(struct request *rq, blk_status_t error) |
277 | { |
278 | struct dm_rq_target_io *tio = tio_from_request(rq); |
279 | |
280 | tio->error = error; |
281 | if (likely(!blk_should_fake_timeout(rq->q))) |
282 | blk_mq_complete_request(rq); |
283 | } |
284 | |
285 | /* |
286 | * Complete the not-mapped clone and the original request with the error status |
287 | * through softirq context. |
288 | * Target's rq_end_io() function isn't called. |
289 | * This may be used when the target's clone_and_map_rq() function fails. |
290 | */ |
291 | static void dm_kill_unmapped_request(struct request *rq, blk_status_t error) |
292 | { |
293 | rq->rq_flags |= RQF_FAILED; |
294 | dm_complete_request(rq, error); |
295 | } |
296 | |
297 | static enum rq_end_io_ret end_clone_request(struct request *clone, |
298 | blk_status_t error) |
299 | { |
300 | struct dm_rq_target_io *tio = clone->end_io_data; |
301 | |
302 | dm_complete_request(rq: tio->orig, error); |
303 | return RQ_END_IO_NONE; |
304 | } |
305 | |
306 | static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig, |
307 | void *data) |
308 | { |
309 | struct dm_rq_target_io *tio = data; |
310 | struct dm_rq_clone_bio_info *info = |
311 | container_of(bio, struct dm_rq_clone_bio_info, clone); |
312 | |
313 | info->orig = bio_orig; |
314 | info->tio = tio; |
315 | bio->bi_end_io = end_clone_bio; |
316 | |
317 | return 0; |
318 | } |
319 | |
320 | static int setup_clone(struct request *clone, struct request *rq, |
321 | struct dm_rq_target_io *tio, gfp_t gfp_mask) |
322 | { |
323 | int r; |
324 | |
325 | r = blk_rq_prep_clone(rq: clone, rq_src: rq, bs: &tio->md->mempools->bs, gfp_mask, |
326 | bio_ctr: dm_rq_bio_constructor, data: tio); |
327 | if (r) |
328 | return r; |
329 | |
330 | clone->end_io = end_clone_request; |
331 | clone->end_io_data = tio; |
332 | |
333 | tio->clone = clone; |
334 | |
335 | return 0; |
336 | } |
337 | |
338 | static void init_tio(struct dm_rq_target_io *tio, struct request *rq, |
339 | struct mapped_device *md) |
340 | { |
341 | tio->md = md; |
342 | tio->ti = NULL; |
343 | tio->clone = NULL; |
344 | tio->orig = rq; |
345 | tio->error = 0; |
346 | tio->completed = 0; |
347 | /* |
348 | * Avoid initializing info for blk-mq; it passes |
349 | * target-specific data through info.ptr |
350 | * (see: dm_mq_init_request) |
351 | */ |
352 | if (!md->init_tio_pdu) |
353 | memset(&tio->info, 0, sizeof(tio->info)); |
354 | } |
355 | |
356 | /* |
357 | * Returns: |
358 | * DM_MAPIO_* : the request has been processed as indicated |
359 | * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued |
360 | * < 0 : the request was completed due to failure |
361 | */ |
362 | static int map_request(struct dm_rq_target_io *tio) |
363 | { |
364 | int r; |
365 | struct dm_target *ti = tio->ti; |
366 | struct mapped_device *md = tio->md; |
367 | struct request *rq = tio->orig; |
368 | struct request *clone = NULL; |
369 | blk_status_t ret; |
370 | |
371 | r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone); |
372 | switch (r) { |
373 | case DM_MAPIO_SUBMITTED: |
374 | /* The target has taken the I/O to submit by itself later */ |
375 | break; |
376 | case DM_MAPIO_REMAPPED: |
377 | if (setup_clone(clone, rq, tio, GFP_ATOMIC)) { |
378 | /* -ENOMEM */ |
379 | ti->type->release_clone_rq(clone, &tio->info); |
380 | return DM_MAPIO_REQUEUE; |
381 | } |
382 | |
383 | /* The target has remapped the I/O so dispatch it */ |
384 | trace_block_rq_remap(rq: clone, dev: disk_devt(disk: dm_disk(md)), |
385 | from: blk_rq_pos(rq)); |
386 | ret = blk_insert_cloned_request(rq: clone); |
387 | switch (ret) { |
388 | case BLK_STS_OK: |
389 | break; |
390 | case BLK_STS_RESOURCE: |
391 | case BLK_STS_DEV_RESOURCE: |
392 | blk_rq_unprep_clone(rq: clone); |
393 | blk_mq_cleanup_rq(rq: clone); |
394 | tio->ti->type->release_clone_rq(clone, &tio->info); |
395 | tio->clone = NULL; |
396 | return DM_MAPIO_REQUEUE; |
397 | default: |
398 | /* must complete clone in terms of original request */ |
399 | dm_complete_request(rq, error: ret); |
400 | } |
401 | break; |
402 | case DM_MAPIO_REQUEUE: |
403 | /* The target wants to requeue the I/O */ |
404 | break; |
405 | case DM_MAPIO_DELAY_REQUEUE: |
406 | /* The target wants to requeue the I/O after a delay */ |
407 | dm_requeue_original_request(tio, delay_requeue: true); |
408 | break; |
409 | case DM_MAPIO_KILL: |
410 | /* The target wants to complete the I/O */ |
411 | dm_kill_unmapped_request(rq, BLK_STS_IOERR); |
412 | break; |
413 | default: |
414 | DMCRIT("unimplemented target map return value: %d" , r); |
415 | BUG(); |
416 | } |
417 | |
418 | return r; |
419 | } |
420 | |
421 | /* DEPRECATED: previously used for request-based merge heuristic in dm_request_fn() */ |
422 | ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf) |
423 | { |
424 | return sprintf(buf, fmt: "%u\n" , 0); |
425 | } |
426 | |
427 | ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md, |
428 | const char *buf, size_t count) |
429 | { |
430 | return count; |
431 | } |
432 | |
433 | static void dm_start_request(struct mapped_device *md, struct request *orig) |
434 | { |
435 | blk_mq_start_request(rq: orig); |
436 | |
437 | if (unlikely(dm_stats_used(&md->stats))) { |
438 | struct dm_rq_target_io *tio = tio_from_request(rq: orig); |
439 | |
440 | tio->duration_jiffies = jiffies; |
441 | tio->n_sectors = blk_rq_sectors(rq: orig); |
442 | dm_stats_account_io(stats: &md->stats, rq_data_dir(orig), |
443 | bi_sector: blk_rq_pos(rq: orig), bi_sectors: tio->n_sectors, end: false, start_time: 0, |
444 | aux: &tio->stats_aux); |
445 | } |
446 | |
447 | /* |
448 | * Hold the md reference here for the in-flight I/O. |
449 | * We can't rely on the reference count by device opener, |
450 | * because the device may be closed during the request completion |
451 | * when all bios are completed. |
452 | * See the comment in rq_completed() too. |
453 | */ |
454 | dm_get(md); |
455 | } |
456 | |
457 | static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
458 | unsigned int hctx_idx, unsigned int numa_node) |
459 | { |
460 | struct mapped_device *md = set->driver_data; |
461 | struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq); |
462 | |
463 | /* |
464 | * Must initialize md member of tio, otherwise it won't |
465 | * be available in dm_mq_queue_rq. |
466 | */ |
467 | tio->md = md; |
468 | |
469 | if (md->init_tio_pdu) { |
470 | /* target-specific per-io data is immediately after the tio */ |
471 | tio->info.ptr = tio + 1; |
472 | } |
473 | |
474 | return 0; |
475 | } |
476 | |
477 | static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx, |
478 | const struct blk_mq_queue_data *bd) |
479 | { |
480 | struct request *rq = bd->rq; |
481 | struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq); |
482 | struct mapped_device *md = tio->md; |
483 | struct dm_target *ti = md->immutable_target; |
484 | |
485 | /* |
486 | * blk-mq's unquiesce may come from outside events, such as |
487 | * elevator switch, updating nr_requests or others, and request may |
488 | * come during suspend, so simply ask for blk-mq to requeue it. |
489 | */ |
490 | if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) |
491 | return BLK_STS_RESOURCE; |
492 | |
493 | if (unlikely(!ti)) { |
494 | int srcu_idx; |
495 | struct dm_table *map; |
496 | |
497 | map = dm_get_live_table(md, srcu_idx: &srcu_idx); |
498 | if (unlikely(!map)) { |
499 | dm_put_live_table(md, srcu_idx); |
500 | return BLK_STS_RESOURCE; |
501 | } |
502 | ti = dm_table_find_target(t: map, sector: 0); |
503 | dm_put_live_table(md, srcu_idx); |
504 | } |
505 | |
506 | if (ti->type->busy && ti->type->busy(ti)) |
507 | return BLK_STS_RESOURCE; |
508 | |
509 | dm_start_request(md, orig: rq); |
510 | |
511 | /* Init tio using md established in .init_request */ |
512 | init_tio(tio, rq, md); |
513 | |
514 | /* |
515 | * Establish tio->ti before calling map_request(). |
516 | */ |
517 | tio->ti = ti; |
518 | |
519 | /* Direct call is fine since .queue_rq allows allocations */ |
520 | if (map_request(tio) == DM_MAPIO_REQUEUE) { |
521 | /* Undo dm_start_request() before requeuing */ |
522 | rq_end_stats(md, orig: rq); |
523 | rq_completed(md); |
524 | return BLK_STS_RESOURCE; |
525 | } |
526 | |
527 | return BLK_STS_OK; |
528 | } |
529 | |
530 | static const struct blk_mq_ops dm_mq_ops = { |
531 | .queue_rq = dm_mq_queue_rq, |
532 | .complete = dm_softirq_done, |
533 | .init_request = dm_mq_init_request, |
534 | }; |
535 | |
536 | int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t) |
537 | { |
538 | struct dm_target *immutable_tgt; |
539 | int err; |
540 | |
541 | md->tag_set = kzalloc_node(size: sizeof(struct blk_mq_tag_set), GFP_KERNEL, node: md->numa_node_id); |
542 | if (!md->tag_set) |
543 | return -ENOMEM; |
544 | |
545 | md->tag_set->ops = &dm_mq_ops; |
546 | md->tag_set->queue_depth = dm_get_blk_mq_queue_depth(); |
547 | md->tag_set->numa_node = md->numa_node_id; |
548 | md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_STACKING; |
549 | md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues(); |
550 | md->tag_set->driver_data = md; |
551 | |
552 | md->tag_set->cmd_size = sizeof(struct dm_rq_target_io); |
553 | immutable_tgt = dm_table_get_immutable_target(t); |
554 | if (immutable_tgt && immutable_tgt->per_io_data_size) { |
555 | /* any target-specific per-io data is immediately after the tio */ |
556 | md->tag_set->cmd_size += immutable_tgt->per_io_data_size; |
557 | md->init_tio_pdu = true; |
558 | } |
559 | |
560 | err = blk_mq_alloc_tag_set(set: md->tag_set); |
561 | if (err) |
562 | goto out_kfree_tag_set; |
563 | |
564 | err = blk_mq_init_allocated_queue(set: md->tag_set, q: md->queue); |
565 | if (err) |
566 | goto out_tag_set; |
567 | return 0; |
568 | |
569 | out_tag_set: |
570 | blk_mq_free_tag_set(set: md->tag_set); |
571 | out_kfree_tag_set: |
572 | kfree(objp: md->tag_set); |
573 | md->tag_set = NULL; |
574 | |
575 | return err; |
576 | } |
577 | |
578 | void dm_mq_cleanup_mapped_device(struct mapped_device *md) |
579 | { |
580 | if (md->tag_set) { |
581 | blk_mq_free_tag_set(set: md->tag_set); |
582 | kfree(objp: md->tag_set); |
583 | md->tag_set = NULL; |
584 | } |
585 | } |
586 | |
587 | module_param(reserved_rq_based_ios, uint, 0644); |
588 | MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools" ); |
589 | |
590 | /* Unused, but preserved for userspace compatibility */ |
591 | static bool use_blk_mq = true; |
592 | module_param(use_blk_mq, bool, 0644); |
593 | MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices" ); |
594 | |
595 | module_param(dm_mq_nr_hw_queues, uint, 0644); |
596 | MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices" ); |
597 | |
598 | module_param(dm_mq_queue_depth, uint, 0644); |
599 | MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices" ); |
600 | |