Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
75bb4625 JA |
2 | /* |
3 | * Block multiqueue core code | |
4 | * | |
5 | * Copyright (C) 2013-2014 Jens Axboe | |
6 | * Copyright (C) 2013-2014 Christoph Hellwig | |
7 | */ | |
320ae51f JA |
8 | #include <linux/kernel.h> |
9 | #include <linux/module.h> | |
10 | #include <linux/backing-dev.h> | |
11 | #include <linux/bio.h> | |
12 | #include <linux/blkdev.h> | |
fe45e630 | 13 | #include <linux/blk-integrity.h> |
f75782e4 | 14 | #include <linux/kmemleak.h> |
320ae51f JA |
15 | #include <linux/mm.h> |
16 | #include <linux/init.h> | |
17 | #include <linux/slab.h> | |
18 | #include <linux/workqueue.h> | |
19 | #include <linux/smp.h> | |
e41d12f5 | 20 | #include <linux/interrupt.h> |
320ae51f | 21 | #include <linux/llist.h> |
320ae51f JA |
22 | #include <linux/cpu.h> |
23 | #include <linux/cache.h> | |
105ab3d8 | 24 | #include <linux/sched/topology.h> |
174cd4b1 | 25 | #include <linux/sched/signal.h> |
320ae51f | 26 | #include <linux/delay.h> |
aedcd72f | 27 | #include <linux/crash_dump.h> |
88c7b2b7 | 28 | #include <linux/prefetch.h> |
a892c8d5 | 29 | #include <linux/blk-crypto.h> |
82d981d4 | 30 | #include <linux/part_stat.h> |
320ae51f JA |
31 | |
32 | #include <trace/events/block.h> | |
33 | ||
54d4e6ab | 34 | #include <linux/t10-pi.h> |
320ae51f JA |
35 | #include "blk.h" |
36 | #include "blk-mq.h" | |
9c1051aa | 37 | #include "blk-mq-debugfs.h" |
986d413b | 38 | #include "blk-pm.h" |
cf43e6be | 39 | #include "blk-stat.h" |
bd166ef1 | 40 | #include "blk-mq-sched.h" |
c1c80384 | 41 | #include "blk-rq-qos.h" |
320ae51f | 42 | |
f9ab4918 | 43 | static DEFINE_PER_CPU(struct llist_head, blk_cpu_done); |
660e802c | 44 | static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd); |
c3077b5d | 45 | |
710fa378 | 46 | static void blk_mq_insert_request(struct request *rq, blk_insert_t flags); |
360f2648 CH |
47 | static void blk_mq_request_bypass_insert(struct request *rq, |
48 | blk_insert_t flags); | |
94aa228c CH |
49 | static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
50 | struct list_head *list); | |
f6c80cff KB |
51 | static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx, |
52 | struct io_comp_batch *iob, unsigned int flags); | |
3e08773c | 53 | |
320ae51f | 54 | /* |
85fae294 YY |
55 | * Check if any of the ctx, dispatch list or elevator |
56 | * have pending work in this hardware queue. | |
320ae51f | 57 | */ |
79f720a7 | 58 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 59 | { |
79f720a7 JA |
60 | return !list_empty_careful(&hctx->dispatch) || |
61 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 62 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
63 | } |
64 | ||
320ae51f JA |
65 | /* |
66 | * Mark this ctx as having pending work in this hardware queue | |
67 | */ | |
68 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
69 | struct blk_mq_ctx *ctx) | |
70 | { | |
f31967f0 JA |
71 | const int bit = ctx->index_hw[hctx->type]; |
72 | ||
73 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
74 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
75 | } |
76 | ||
77 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
78 | struct blk_mq_ctx *ctx) | |
79 | { | |
f31967f0 JA |
80 | const int bit = ctx->index_hw[hctx->type]; |
81 | ||
82 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
83 | } |
84 | ||
f299b7c7 | 85 | struct mq_inflight { |
8446fe92 | 86 | struct block_device *part; |
a2e80f6f | 87 | unsigned int inflight[2]; |
f299b7c7 JA |
88 | }; |
89 | ||
2dd6532e | 90 | static bool blk_mq_check_inflight(struct request *rq, void *priv) |
f299b7c7 JA |
91 | { |
92 | struct mq_inflight *mi = priv; | |
93 | ||
b81c14ca HW |
94 | if (rq->part && blk_do_io_stat(rq) && |
95 | (!mi->part->bd_partno || rq->part == mi->part) && | |
b0d97557 | 96 | blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) |
bb4e6b14 | 97 | mi->inflight[rq_data_dir(rq)]++; |
7baa8572 JA |
98 | |
99 | return true; | |
f299b7c7 JA |
100 | } |
101 | ||
8446fe92 CH |
102 | unsigned int blk_mq_in_flight(struct request_queue *q, |
103 | struct block_device *part) | |
f299b7c7 | 104 | { |
a2e80f6f | 105 | struct mq_inflight mi = { .part = part }; |
f299b7c7 | 106 | |
f299b7c7 | 107 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 | 108 | |
a2e80f6f | 109 | return mi.inflight[0] + mi.inflight[1]; |
bf0ddaba OS |
110 | } |
111 | ||
8446fe92 CH |
112 | void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, |
113 | unsigned int inflight[2]) | |
bf0ddaba | 114 | { |
a2e80f6f | 115 | struct mq_inflight mi = { .part = part }; |
bf0ddaba | 116 | |
bb4e6b14 | 117 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
a2e80f6f PB |
118 | inflight[0] = mi.inflight[0]; |
119 | inflight[1] = mi.inflight[1]; | |
bf0ddaba OS |
120 | } |
121 | ||
1671d522 | 122 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 123 | { |
7996a8b5 BL |
124 | mutex_lock(&q->mq_freeze_lock); |
125 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 126 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 127 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 128 | if (queue_is_mq(q)) |
055f6e18 | 129 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
130 | } else { |
131 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 132 | } |
f3af020b | 133 | } |
1671d522 | 134 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 135 | |
6bae363e | 136 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 137 | { |
3ef28e83 | 138 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 139 | } |
6bae363e | 140 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 141 | |
f91328c4 KB |
142 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
143 | unsigned long timeout) | |
144 | { | |
145 | return wait_event_timeout(q->mq_freeze_wq, | |
146 | percpu_ref_is_zero(&q->q_usage_counter), | |
147 | timeout); | |
148 | } | |
149 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 150 | |
f3af020b TH |
151 | /* |
152 | * Guarantee no request is in use, so we can change any data structure of | |
153 | * the queue afterward. | |
154 | */ | |
3ef28e83 | 155 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 156 | { |
3ef28e83 DW |
157 | /* |
158 | * In the !blk_mq case we are only calling this to kill the | |
159 | * q_usage_counter, otherwise this increases the freeze depth | |
160 | * and waits for it to return to zero. For this reason there is | |
161 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
162 | * exported to drivers as the only user for unfreeze is blk_mq. | |
163 | */ | |
1671d522 | 164 | blk_freeze_queue_start(q); |
f3af020b TH |
165 | blk_mq_freeze_queue_wait(q); |
166 | } | |
3ef28e83 DW |
167 | |
168 | void blk_mq_freeze_queue(struct request_queue *q) | |
169 | { | |
170 | /* | |
171 | * ...just an alias to keep freeze and unfreeze actions balanced | |
172 | * in the blk_mq_* namespace | |
173 | */ | |
174 | blk_freeze_queue(q); | |
175 | } | |
c761d96b | 176 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 177 | |
aec89dc5 | 178 | void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic) |
320ae51f | 179 | { |
7996a8b5 | 180 | mutex_lock(&q->mq_freeze_lock); |
aec89dc5 CH |
181 | if (force_atomic) |
182 | q->q_usage_counter.data->force_atomic = true; | |
7996a8b5 BL |
183 | q->mq_freeze_depth--; |
184 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
185 | if (!q->mq_freeze_depth) { | |
bdd63160 | 186 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 187 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 188 | } |
7996a8b5 | 189 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 190 | } |
aec89dc5 CH |
191 | |
192 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
193 | { | |
194 | __blk_mq_unfreeze_queue(q, false); | |
195 | } | |
b4c6a028 | 196 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 197 | |
852ec809 BVA |
198 | /* |
199 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
200 | * mpt3sas driver such that this function can be removed. | |
201 | */ | |
202 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
203 | { | |
e70feb8b ML |
204 | unsigned long flags; |
205 | ||
206 | spin_lock_irqsave(&q->queue_lock, flags); | |
207 | if (!q->quiesce_depth++) | |
208 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
209 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
852ec809 BVA |
210 | } |
211 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
212 | ||
6a83e74d | 213 | /** |
9ef4d020 | 214 | * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done |
483239c7 | 215 | * @set: tag_set to wait on |
6a83e74d | 216 | * |
9ef4d020 | 217 | * Note: it is driver's responsibility for making sure that quiesce has |
483239c7 CH |
218 | * been started on or more of the request_queues of the tag_set. This |
219 | * function only waits for the quiesce on those request_queues that had | |
220 | * the quiesce flag set using blk_mq_quiesce_queue_nowait. | |
6a83e74d | 221 | */ |
483239c7 | 222 | void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set) |
6a83e74d | 223 | { |
483239c7 CH |
224 | if (set->flags & BLK_MQ_F_BLOCKING) |
225 | synchronize_srcu(set->srcu); | |
704b914f | 226 | else |
6a83e74d BVA |
227 | synchronize_rcu(); |
228 | } | |
9ef4d020 ML |
229 | EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); |
230 | ||
231 | /** | |
232 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished | |
233 | * @q: request queue. | |
234 | * | |
235 | * Note: this function does not prevent that the struct request end_io() | |
236 | * callback function is invoked. Once this function is returned, we make | |
237 | * sure no dispatch can happen until the queue is unquiesced via | |
238 | * blk_mq_unquiesce_queue(). | |
239 | */ | |
240 | void blk_mq_quiesce_queue(struct request_queue *q) | |
241 | { | |
242 | blk_mq_quiesce_queue_nowait(q); | |
8537380b CH |
243 | /* nothing to wait for non-mq queues */ |
244 | if (queue_is_mq(q)) | |
483239c7 | 245 | blk_mq_wait_quiesce_done(q->tag_set); |
9ef4d020 | 246 | } |
6a83e74d BVA |
247 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); |
248 | ||
e4e73913 ML |
249 | /* |
250 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
251 | * @q: request queue. | |
252 | * | |
253 | * This function recovers queue into the state before quiescing | |
254 | * which is done by blk_mq_quiesce_queue. | |
255 | */ | |
256 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
257 | { | |
e70feb8b ML |
258 | unsigned long flags; |
259 | bool run_queue = false; | |
260 | ||
261 | spin_lock_irqsave(&q->queue_lock, flags); | |
262 | if (WARN_ON_ONCE(q->quiesce_depth <= 0)) { | |
263 | ; | |
264 | } else if (!--q->quiesce_depth) { | |
265 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); | |
266 | run_queue = true; | |
267 | } | |
268 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
f4560ffe | 269 | |
1d9e9bc6 | 270 | /* dispatch requests which are inserted during quiescing */ |
e70feb8b ML |
271 | if (run_queue) |
272 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
273 | } |
274 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
275 | ||
414dd48e CL |
276 | void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set) |
277 | { | |
278 | struct request_queue *q; | |
279 | ||
280 | mutex_lock(&set->tag_list_lock); | |
281 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
282 | if (!blk_queue_skip_tagset_quiesce(q)) | |
283 | blk_mq_quiesce_queue_nowait(q); | |
284 | } | |
285 | blk_mq_wait_quiesce_done(set); | |
286 | mutex_unlock(&set->tag_list_lock); | |
287 | } | |
288 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset); | |
289 | ||
290 | void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set) | |
291 | { | |
292 | struct request_queue *q; | |
293 | ||
294 | mutex_lock(&set->tag_list_lock); | |
295 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
296 | if (!blk_queue_skip_tagset_quiesce(q)) | |
297 | blk_mq_unquiesce_queue(q); | |
298 | } | |
299 | mutex_unlock(&set->tag_list_lock); | |
300 | } | |
301 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset); | |
302 | ||
aed3ea94 JA |
303 | void blk_mq_wake_waiters(struct request_queue *q) |
304 | { | |
305 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 306 | unsigned long i; |
aed3ea94 JA |
307 | |
308 | queue_for_each_hw_ctx(q, hctx, i) | |
309 | if (blk_mq_hw_queue_mapped(hctx)) | |
310 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
311 | } | |
312 | ||
52fdbbcc CH |
313 | void blk_rq_init(struct request_queue *q, struct request *rq) |
314 | { | |
315 | memset(rq, 0, sizeof(*rq)); | |
316 | ||
317 | INIT_LIST_HEAD(&rq->queuelist); | |
318 | rq->q = q; | |
319 | rq->__sector = (sector_t) -1; | |
320 | INIT_HLIST_NODE(&rq->hash); | |
321 | RB_CLEAR_NODE(&rq->rb_node); | |
322 | rq->tag = BLK_MQ_NO_TAG; | |
323 | rq->internal_tag = BLK_MQ_NO_TAG; | |
08420cf7 | 324 | rq->start_time_ns = blk_time_get_ns(); |
52fdbbcc CH |
325 | rq->part = NULL; |
326 | blk_crypto_rq_set_defaults(rq); | |
327 | } | |
328 | EXPORT_SYMBOL(blk_rq_init); | |
329 | ||
5c17f45e CZ |
330 | /* Set start and alloc time when the allocated request is actually used */ |
331 | static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns) | |
332 | { | |
333 | if (blk_mq_need_time_stamp(rq)) | |
08420cf7 | 334 | rq->start_time_ns = blk_time_get_ns(); |
5c17f45e CZ |
335 | else |
336 | rq->start_time_ns = 0; | |
337 | ||
338 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME | |
339 | if (blk_queue_rq_alloc_time(rq->q)) | |
340 | rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns; | |
341 | else | |
342 | rq->alloc_time_ns = 0; | |
343 | #endif | |
344 | } | |
345 | ||
e4cdf1a1 | 346 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
5c17f45e | 347 | struct blk_mq_tags *tags, unsigned int tag) |
320ae51f | 348 | { |
605f784e PB |
349 | struct blk_mq_ctx *ctx = data->ctx; |
350 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
351 | struct request_queue *q = data->q; | |
e4cdf1a1 | 352 | struct request *rq = tags->static_rqs[tag]; |
c3a148d2 | 353 | |
c7b84d42 JA |
354 | rq->q = q; |
355 | rq->mq_ctx = ctx; | |
356 | rq->mq_hctx = hctx; | |
357 | rq->cmd_flags = data->cmd_flags; | |
358 | ||
359 | if (data->flags & BLK_MQ_REQ_PM) | |
360 | data->rq_flags |= RQF_PM; | |
361 | if (blk_queue_io_stat(q)) | |
362 | data->rq_flags |= RQF_IO_STAT; | |
363 | rq->rq_flags = data->rq_flags; | |
364 | ||
dd6216bb | 365 | if (data->rq_flags & RQF_SCHED_TAGS) { |
56f8da64 JA |
366 | rq->tag = BLK_MQ_NO_TAG; |
367 | rq->internal_tag = tag; | |
dd6216bb CH |
368 | } else { |
369 | rq->tag = tag; | |
370 | rq->internal_tag = BLK_MQ_NO_TAG; | |
e4cdf1a1 | 371 | } |
c7b84d42 | 372 | rq->timeout = 0; |
e4cdf1a1 | 373 | |
af76e555 | 374 | rq->part = NULL; |
544ccc8d | 375 | rq->io_start_time_ns = 0; |
3d244306 | 376 | rq->stats_sectors = 0; |
af76e555 CH |
377 | rq->nr_phys_segments = 0; |
378 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
379 | rq->nr_integrity_segments = 0; | |
380 | #endif | |
af76e555 CH |
381 | rq->end_io = NULL; |
382 | rq->end_io_data = NULL; | |
af76e555 | 383 | |
4f266f2b PB |
384 | blk_crypto_rq_set_defaults(rq); |
385 | INIT_LIST_HEAD(&rq->queuelist); | |
386 | /* tag was already set */ | |
387 | WRITE_ONCE(rq->deadline, 0); | |
0a467d0f | 388 | req_ref_set(rq, 1); |
7ea4d8a4 | 389 | |
dd6216bb | 390 | if (rq->rq_flags & RQF_USE_SCHED) { |
7ea4d8a4 CH |
391 | struct elevator_queue *e = data->q->elevator; |
392 | ||
4f266f2b PB |
393 | INIT_HLIST_NODE(&rq->hash); |
394 | RB_CLEAR_NODE(&rq->rb_node); | |
395 | ||
dd6216bb | 396 | if (e->type->ops.prepare_request) |
7ea4d8a4 | 397 | e->type->ops.prepare_request(rq); |
7ea4d8a4 CH |
398 | } |
399 | ||
e4cdf1a1 | 400 | return rq; |
5dee8577 CH |
401 | } |
402 | ||
349302da | 403 | static inline struct request * |
5c17f45e | 404 | __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data) |
349302da JA |
405 | { |
406 | unsigned int tag, tag_offset; | |
fe6134f6 | 407 | struct blk_mq_tags *tags; |
349302da | 408 | struct request *rq; |
fe6134f6 | 409 | unsigned long tag_mask; |
349302da JA |
410 | int i, nr = 0; |
411 | ||
fe6134f6 JA |
412 | tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset); |
413 | if (unlikely(!tag_mask)) | |
349302da JA |
414 | return NULL; |
415 | ||
fe6134f6 JA |
416 | tags = blk_mq_tags_from_data(data); |
417 | for (i = 0; tag_mask; i++) { | |
418 | if (!(tag_mask & (1UL << i))) | |
349302da JA |
419 | continue; |
420 | tag = tag_offset + i; | |
a22c00be | 421 | prefetch(tags->static_rqs[tag]); |
fe6134f6 | 422 | tag_mask &= ~(1UL << i); |
5c17f45e | 423 | rq = blk_mq_rq_ctx_init(data, tags, tag); |
013a7f95 | 424 | rq_list_add(data->cached_rq, rq); |
c5fc7b93 | 425 | nr++; |
349302da | 426 | } |
b8643d68 CZ |
427 | if (!(data->rq_flags & RQF_SCHED_TAGS)) |
428 | blk_mq_add_active_requests(data->hctx, nr); | |
c5fc7b93 JA |
429 | /* caller already holds a reference, add for remainder */ |
430 | percpu_ref_get_many(&data->q->q_usage_counter, nr - 1); | |
349302da JA |
431 | data->nr_tags -= nr; |
432 | ||
013a7f95 | 433 | return rq_list_pop(data->cached_rq); |
349302da JA |
434 | } |
435 | ||
b90cfaed | 436 | static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) |
d2c0d383 | 437 | { |
e6e7abff | 438 | struct request_queue *q = data->q; |
6f816b4b | 439 | u64 alloc_time_ns = 0; |
47c122e3 | 440 | struct request *rq; |
600c3b0c | 441 | unsigned int tag; |
d2c0d383 | 442 | |
6f816b4b TH |
443 | /* alloc_time includes depth and tag waits */ |
444 | if (blk_queue_rq_alloc_time(q)) | |
08420cf7 | 445 | alloc_time_ns = blk_time_get_ns(); |
6f816b4b | 446 | |
f9afca4d | 447 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 448 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 | 449 | |
781dd830 | 450 | if (q->elevator) { |
dd6216bb CH |
451 | /* |
452 | * All requests use scheduler tags when an I/O scheduler is | |
453 | * enabled for the queue. | |
454 | */ | |
455 | data->rq_flags |= RQF_SCHED_TAGS; | |
781dd830 | 456 | |
d2c0d383 | 457 | /* |
8d663f34 | 458 | * Flush/passthrough requests are special and go directly to the |
dd6216bb | 459 | * dispatch list. |
d2c0d383 | 460 | */ |
be4c4278 | 461 | if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH && |
dd6216bb CH |
462 | !blk_op_is_passthrough(data->cmd_flags)) { |
463 | struct elevator_mq_ops *ops = &q->elevator->type->ops; | |
464 | ||
465 | WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED); | |
466 | ||
467 | data->rq_flags |= RQF_USE_SCHED; | |
468 | if (ops->limit_depth) | |
469 | ops->limit_depth(data->cmd_flags, data); | |
470 | } | |
d2c0d383 CH |
471 | } |
472 | ||
bf0beec0 | 473 | retry: |
600c3b0c CH |
474 | data->ctx = blk_mq_get_ctx(q); |
475 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
dd6216bb | 476 | if (!(data->rq_flags & RQF_SCHED_TAGS)) |
600c3b0c CH |
477 | blk_mq_tag_busy(data->hctx); |
478 | ||
99e48cd6 JG |
479 | if (data->flags & BLK_MQ_REQ_RESERVED) |
480 | data->rq_flags |= RQF_RESV; | |
481 | ||
349302da JA |
482 | /* |
483 | * Try batched alloc if we want more than 1 tag. | |
484 | */ | |
485 | if (data->nr_tags > 1) { | |
5c17f45e CZ |
486 | rq = __blk_mq_alloc_requests_batch(data); |
487 | if (rq) { | |
488 | blk_mq_rq_time_init(rq, alloc_time_ns); | |
349302da | 489 | return rq; |
5c17f45e | 490 | } |
349302da JA |
491 | data->nr_tags = 1; |
492 | } | |
493 | ||
bf0beec0 ML |
494 | /* |
495 | * Waiting allocations only fail because of an inactive hctx. In that | |
496 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
497 | * should have migrated us to an online CPU by now. | |
498 | */ | |
e4cdf1a1 | 499 | tag = blk_mq_get_tag(data); |
bf0beec0 ML |
500 | if (tag == BLK_MQ_NO_TAG) { |
501 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
502 | return NULL; | |
bf0beec0 | 503 | /* |
349302da JA |
504 | * Give up the CPU and sleep for a random short time to |
505 | * ensure that thread using a realtime scheduling class | |
506 | * are migrated off the CPU, and thus off the hctx that | |
507 | * is going away. | |
bf0beec0 ML |
508 | */ |
509 | msleep(3); | |
510 | goto retry; | |
511 | } | |
47c122e3 | 512 | |
b8643d68 CZ |
513 | if (!(data->rq_flags & RQF_SCHED_TAGS)) |
514 | blk_mq_inc_active_requests(data->hctx); | |
5c17f45e CZ |
515 | rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag); |
516 | blk_mq_rq_time_init(rq, alloc_time_ns); | |
517 | return rq; | |
d2c0d383 CH |
518 | } |
519 | ||
4b6a5d9c JA |
520 | static struct request *blk_mq_rq_cache_fill(struct request_queue *q, |
521 | struct blk_plug *plug, | |
522 | blk_opf_t opf, | |
523 | blk_mq_req_flags_t flags) | |
320ae51f | 524 | { |
e6e7abff CH |
525 | struct blk_mq_alloc_data data = { |
526 | .q = q, | |
527 | .flags = flags, | |
16458cf3 | 528 | .cmd_flags = opf, |
4b6a5d9c JA |
529 | .nr_tags = plug->nr_ios, |
530 | .cached_rq = &plug->cached_rq, | |
e6e7abff | 531 | }; |
bd166ef1 | 532 | struct request *rq; |
320ae51f | 533 | |
4b6a5d9c JA |
534 | if (blk_queue_enter(q, flags)) |
535 | return NULL; | |
536 | ||
537 | plug->nr_ios = 1; | |
320ae51f | 538 | |
b90cfaed | 539 | rq = __blk_mq_alloc_requests(&data); |
4b6a5d9c JA |
540 | if (unlikely(!rq)) |
541 | blk_queue_exit(q); | |
542 | return rq; | |
543 | } | |
544 | ||
545 | static struct request *blk_mq_alloc_cached_request(struct request_queue *q, | |
546 | blk_opf_t opf, | |
547 | blk_mq_req_flags_t flags) | |
548 | { | |
549 | struct blk_plug *plug = current->plug; | |
550 | struct request *rq; | |
551 | ||
552 | if (!plug) | |
553 | return NULL; | |
40467282 | 554 | |
4b6a5d9c JA |
555 | if (rq_list_empty(plug->cached_rq)) { |
556 | if (plug->nr_ios == 1) | |
557 | return NULL; | |
558 | rq = blk_mq_rq_cache_fill(q, plug, opf, flags); | |
40467282 JC |
559 | if (!rq) |
560 | return NULL; | |
561 | } else { | |
562 | rq = rq_list_peek(&plug->cached_rq); | |
563 | if (!rq || rq->q != q) | |
564 | return NULL; | |
4b6a5d9c | 565 | |
40467282 JC |
566 | if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type) |
567 | return NULL; | |
568 | if (op_is_flush(rq->cmd_flags) != op_is_flush(opf)) | |
569 | return NULL; | |
570 | ||
571 | plug->cached_rq = rq_list_next(rq); | |
5c17f45e | 572 | blk_mq_rq_time_init(rq, 0); |
40467282 | 573 | } |
4b6a5d9c | 574 | |
4b6a5d9c JA |
575 | rq->cmd_flags = opf; |
576 | INIT_LIST_HEAD(&rq->queuelist); | |
577 | return rq; | |
578 | } | |
579 | ||
580 | struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf, | |
581 | blk_mq_req_flags_t flags) | |
582 | { | |
583 | struct request *rq; | |
584 | ||
585 | rq = blk_mq_alloc_cached_request(q, opf, flags); | |
586 | if (!rq) { | |
587 | struct blk_mq_alloc_data data = { | |
588 | .q = q, | |
589 | .flags = flags, | |
590 | .cmd_flags = opf, | |
591 | .nr_tags = 1, | |
592 | }; | |
593 | int ret; | |
594 | ||
595 | ret = blk_queue_enter(q, flags); | |
596 | if (ret) | |
597 | return ERR_PTR(ret); | |
598 | ||
599 | rq = __blk_mq_alloc_requests(&data); | |
600 | if (!rq) | |
601 | goto out_queue_exit; | |
602 | } | |
0c4de0f3 CH |
603 | rq->__data_len = 0; |
604 | rq->__sector = (sector_t) -1; | |
605 | rq->bio = rq->biotail = NULL; | |
320ae51f | 606 | return rq; |
a5ea5811 CH |
607 | out_queue_exit: |
608 | blk_queue_exit(q); | |
609 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 610 | } |
4bb659b1 | 611 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 612 | |
cd6ce148 | 613 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
16458cf3 | 614 | blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 615 | { |
e6e7abff CH |
616 | struct blk_mq_alloc_data data = { |
617 | .q = q, | |
618 | .flags = flags, | |
16458cf3 | 619 | .cmd_flags = opf, |
47c122e3 | 620 | .nr_tags = 1, |
e6e7abff | 621 | }; |
600c3b0c | 622 | u64 alloc_time_ns = 0; |
e3c5a78c | 623 | struct request *rq; |
6d2809d5 | 624 | unsigned int cpu; |
600c3b0c | 625 | unsigned int tag; |
1f5bd336 ML |
626 | int ret; |
627 | ||
600c3b0c CH |
628 | /* alloc_time includes depth and tag waits */ |
629 | if (blk_queue_rq_alloc_time(q)) | |
08420cf7 | 630 | alloc_time_ns = blk_time_get_ns(); |
600c3b0c | 631 | |
1f5bd336 ML |
632 | /* |
633 | * If the tag allocator sleeps we could get an allocation for a | |
634 | * different hardware context. No need to complicate the low level | |
635 | * allocator for this for the rare use case of a command tied to | |
636 | * a specific queue. | |
637 | */ | |
6ee858a3 KS |
638 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) || |
639 | WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED))) | |
1f5bd336 ML |
640 | return ERR_PTR(-EINVAL); |
641 | ||
642 | if (hctx_idx >= q->nr_hw_queues) | |
643 | return ERR_PTR(-EIO); | |
644 | ||
3a0a5299 | 645 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
646 | if (ret) |
647 | return ERR_PTR(ret); | |
648 | ||
c8712c6a CH |
649 | /* |
650 | * Check if the hardware context is actually mapped to anything. | |
651 | * If not tell the caller that it should skip this queue. | |
652 | */ | |
a5ea5811 | 653 | ret = -EXDEV; |
4e5cc99e | 654 | data.hctx = xa_load(&q->hctx_table, hctx_idx); |
e6e7abff | 655 | if (!blk_mq_hw_queue_mapped(data.hctx)) |
a5ea5811 | 656 | goto out_queue_exit; |
e6e7abff | 657 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
14dc7a18 BVA |
658 | if (cpu >= nr_cpu_ids) |
659 | goto out_queue_exit; | |
e6e7abff | 660 | data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 661 | |
dd6216bb CH |
662 | if (q->elevator) |
663 | data.rq_flags |= RQF_SCHED_TAGS; | |
781dd830 | 664 | else |
dd6216bb | 665 | blk_mq_tag_busy(data.hctx); |
600c3b0c | 666 | |
99e48cd6 JG |
667 | if (flags & BLK_MQ_REQ_RESERVED) |
668 | data.rq_flags |= RQF_RESV; | |
669 | ||
a5ea5811 | 670 | ret = -EWOULDBLOCK; |
600c3b0c CH |
671 | tag = blk_mq_get_tag(&data); |
672 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 673 | goto out_queue_exit; |
b8643d68 CZ |
674 | if (!(data.rq_flags & RQF_SCHED_TAGS)) |
675 | blk_mq_inc_active_requests(data.hctx); | |
5c17f45e CZ |
676 | rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag); |
677 | blk_mq_rq_time_init(rq, alloc_time_ns); | |
e3c5a78c JG |
678 | rq->__data_len = 0; |
679 | rq->__sector = (sector_t) -1; | |
680 | rq->bio = rq->biotail = NULL; | |
681 | return rq; | |
600c3b0c | 682 | |
a5ea5811 CH |
683 | out_queue_exit: |
684 | blk_queue_exit(q); | |
685 | return ERR_PTR(ret); | |
1f5bd336 ML |
686 | } |
687 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
688 | ||
e5c0ca13 CZ |
689 | static void blk_mq_finish_request(struct request *rq) |
690 | { | |
691 | struct request_queue *q = rq->q; | |
692 | ||
693 | if (rq->rq_flags & RQF_USE_SCHED) { | |
694 | q->elevator->type->ops.finish_request(rq); | |
695 | /* | |
696 | * For postflush request that may need to be | |
697 | * completed twice, we should clear this flag | |
698 | * to avoid double finish_request() on the rq. | |
699 | */ | |
700 | rq->rq_flags &= ~RQF_USE_SCHED; | |
701 | } | |
702 | } | |
703 | ||
12f5b931 KB |
704 | static void __blk_mq_free_request(struct request *rq) |
705 | { | |
706 | struct request_queue *q = rq->q; | |
707 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 708 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
709 | const int sched_tag = rq->internal_tag; |
710 | ||
a892c8d5 | 711 | blk_crypto_free_request(rq); |
986d413b | 712 | blk_pm_mark_last_busy(rq); |
ea4f995e | 713 | rq->mq_hctx = NULL; |
ddad5933 | 714 | |
b8643d68 CZ |
715 | if (rq->tag != BLK_MQ_NO_TAG) { |
716 | blk_mq_dec_active_requests(hctx); | |
cae740a0 | 717 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
b8643d68 | 718 | } |
76647368 | 719 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 720 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
721 | blk_mq_sched_restart(hctx); |
722 | blk_queue_exit(q); | |
723 | } | |
724 | ||
6af54051 | 725 | void blk_mq_free_request(struct request *rq) |
320ae51f | 726 | { |
320ae51f | 727 | struct request_queue *q = rq->q; |
6af54051 | 728 | |
e5c0ca13 | 729 | blk_mq_finish_request(rq); |
320ae51f | 730 | |
7beb2f84 | 731 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 732 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 733 | |
a7905043 | 734 | rq_qos_done(q, rq); |
0d2602ca | 735 | |
12f5b931 | 736 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 737 | if (req_ref_put_and_test(rq)) |
12f5b931 | 738 | __blk_mq_free_request(rq); |
320ae51f | 739 | } |
1a3b595a | 740 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 741 | |
47c122e3 | 742 | void blk_mq_free_plug_rqs(struct blk_plug *plug) |
320ae51f | 743 | { |
013a7f95 | 744 | struct request *rq; |
fe1f4526 | 745 | |
c5fc7b93 | 746 | while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) |
47c122e3 | 747 | blk_mq_free_request(rq); |
47c122e3 | 748 | } |
522a7775 | 749 | |
22350ad7 CH |
750 | void blk_dump_rq_flags(struct request *rq, char *msg) |
751 | { | |
752 | printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, | |
f3fa33ac | 753 | rq->q->disk ? rq->q->disk->disk_name : "?", |
16458cf3 | 754 | (__force unsigned long long) rq->cmd_flags); |
22350ad7 CH |
755 | |
756 | printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", | |
757 | (unsigned long long)blk_rq_pos(rq), | |
758 | blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); | |
759 | printk(KERN_INFO " bio %p, biotail %p, len %u\n", | |
760 | rq->bio, rq->biotail, blk_rq_bytes(rq)); | |
761 | } | |
762 | EXPORT_SYMBOL(blk_dump_rq_flags); | |
763 | ||
9be3e06f JA |
764 | static void req_bio_endio(struct request *rq, struct bio *bio, |
765 | unsigned int nbytes, blk_status_t error) | |
766 | { | |
478eb72b | 767 | if (unlikely(error)) { |
9be3e06f | 768 | bio->bi_status = error; |
478eb72b | 769 | } else if (req_op(rq) == REQ_OP_ZONE_APPEND) { |
9be3e06f JA |
770 | /* |
771 | * Partial zone append completions cannot be supported as the | |
772 | * BIO fragments may end up not being written sequentially. | |
748dc0b6 DLM |
773 | * For such case, force the completed nbytes to be equal to |
774 | * the BIO size so that bio_advance() sets the BIO remaining | |
775 | * size to 0 and we end up calling bio_endio() before returning. | |
9be3e06f | 776 | */ |
748dc0b6 | 777 | if (bio->bi_iter.bi_size != nbytes) { |
9be3e06f | 778 | bio->bi_status = BLK_STS_IOERR; |
748dc0b6 DLM |
779 | nbytes = bio->bi_iter.bi_size; |
780 | } else { | |
9be3e06f | 781 | bio->bi_iter.bi_sector = rq->__sector; |
748dc0b6 | 782 | } |
9be3e06f JA |
783 | } |
784 | ||
478eb72b PB |
785 | bio_advance(bio, nbytes); |
786 | ||
787 | if (unlikely(rq->rq_flags & RQF_QUIET)) | |
788 | bio_set_flag(bio, BIO_QUIET); | |
9be3e06f JA |
789 | /* don't actually finish bio if it's part of flush sequence */ |
790 | if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ)) | |
791 | bio_endio(bio); | |
792 | } | |
793 | ||
794 | static void blk_account_io_completion(struct request *req, unsigned int bytes) | |
795 | { | |
796 | if (req->part && blk_do_io_stat(req)) { | |
797 | const int sgrp = op_stat_group(req_op(req)); | |
798 | ||
799 | part_stat_lock(); | |
800 | part_stat_add(req->part, sectors[sgrp], bytes >> 9); | |
801 | part_stat_unlock(); | |
802 | } | |
803 | } | |
804 | ||
0d7a29a2 CH |
805 | static void blk_print_req_error(struct request *req, blk_status_t status) |
806 | { | |
807 | printk_ratelimited(KERN_ERR | |
808 | "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x " | |
809 | "phys_seg %u prio class %u\n", | |
810 | blk_status_to_str(status), | |
f3fa33ac | 811 | req->q->disk ? req->q->disk->disk_name : "?", |
16458cf3 BVA |
812 | blk_rq_pos(req), (__force u32)req_op(req), |
813 | blk_op_str(req_op(req)), | |
814 | (__force u32)(req->cmd_flags & ~REQ_OP_MASK), | |
0d7a29a2 CH |
815 | req->nr_phys_segments, |
816 | IOPRIO_PRIO_CLASS(req->ioprio)); | |
817 | } | |
818 | ||
5581a5dd JA |
819 | /* |
820 | * Fully end IO on a request. Does not support partial completions, or | |
821 | * errors. | |
822 | */ | |
823 | static void blk_complete_request(struct request *req) | |
824 | { | |
825 | const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0; | |
826 | int total_bytes = blk_rq_bytes(req); | |
827 | struct bio *bio = req->bio; | |
828 | ||
829 | trace_block_rq_complete(req, BLK_STS_OK, total_bytes); | |
830 | ||
831 | if (!bio) | |
832 | return; | |
833 | ||
834 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
835 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ) | |
836 | req->q->integrity.profile->complete_fn(req, total_bytes); | |
837 | #endif | |
838 | ||
9cd1e566 EB |
839 | /* |
840 | * Upper layers may call blk_crypto_evict_key() anytime after the last | |
841 | * bio_endio(). Therefore, the keyslot must be released before that. | |
842 | */ | |
843 | blk_crypto_rq_put_keyslot(req); | |
844 | ||
5581a5dd JA |
845 | blk_account_io_completion(req, total_bytes); |
846 | ||
847 | do { | |
848 | struct bio *next = bio->bi_next; | |
849 | ||
850 | /* Completion has already been traced */ | |
851 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
a12821d5 PR |
852 | |
853 | if (req_op(req) == REQ_OP_ZONE_APPEND) | |
854 | bio->bi_iter.bi_sector = req->__sector; | |
855 | ||
5581a5dd JA |
856 | if (!is_flush) |
857 | bio_endio(bio); | |
858 | bio = next; | |
859 | } while (bio); | |
860 | ||
861 | /* | |
862 | * Reset counters so that the request stacking driver | |
863 | * can find how many bytes remain in the request | |
864 | * later. | |
865 | */ | |
ab3e1d3b JA |
866 | if (!req->end_io) { |
867 | req->bio = NULL; | |
868 | req->__data_len = 0; | |
869 | } | |
5581a5dd JA |
870 | } |
871 | ||
9be3e06f JA |
872 | /** |
873 | * blk_update_request - Complete multiple bytes without completing the request | |
874 | * @req: the request being processed | |
875 | * @error: block status code | |
876 | * @nr_bytes: number of bytes to complete for @req | |
877 | * | |
878 | * Description: | |
879 | * Ends I/O on a number of bytes attached to @req, but doesn't complete | |
880 | * the request structure even if @req doesn't have leftover. | |
881 | * If @req has leftover, sets it up for the next range of segments. | |
882 | * | |
883 | * Passing the result of blk_rq_bytes() as @nr_bytes guarantees | |
884 | * %false return from this function. | |
885 | * | |
886 | * Note: | |
887 | * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function | |
888 | * except in the consistency check at the end of this function. | |
889 | * | |
890 | * Return: | |
891 | * %false - this request doesn't have any more data | |
892 | * %true - this request has more data | |
893 | **/ | |
894 | bool blk_update_request(struct request *req, blk_status_t error, | |
895 | unsigned int nr_bytes) | |
896 | { | |
897 | int total_bytes; | |
898 | ||
8a7d267b | 899 | trace_block_rq_complete(req, error, nr_bytes); |
9be3e06f JA |
900 | |
901 | if (!req->bio) | |
902 | return false; | |
903 | ||
904 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
905 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && | |
906 | error == BLK_STS_OK) | |
907 | req->q->integrity.profile->complete_fn(req, nr_bytes); | |
908 | #endif | |
909 | ||
9cd1e566 EB |
910 | /* |
911 | * Upper layers may call blk_crypto_evict_key() anytime after the last | |
912 | * bio_endio(). Therefore, the keyslot must be released before that. | |
913 | */ | |
914 | if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req)) | |
915 | __blk_crypto_rq_put_keyslot(req); | |
916 | ||
9be3e06f | 917 | if (unlikely(error && !blk_rq_is_passthrough(req) && |
3d973a76 CH |
918 | !(req->rq_flags & RQF_QUIET)) && |
919 | !test_bit(GD_DEAD, &req->q->disk->state)) { | |
9be3e06f | 920 | blk_print_req_error(req, error); |
d5869fdc YS |
921 | trace_block_rq_error(req, error, nr_bytes); |
922 | } | |
9be3e06f JA |
923 | |
924 | blk_account_io_completion(req, nr_bytes); | |
925 | ||
926 | total_bytes = 0; | |
927 | while (req->bio) { | |
928 | struct bio *bio = req->bio; | |
929 | unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); | |
930 | ||
931 | if (bio_bytes == bio->bi_iter.bi_size) | |
932 | req->bio = bio->bi_next; | |
933 | ||
934 | /* Completion has already been traced */ | |
935 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
936 | req_bio_endio(req, bio, bio_bytes, error); | |
937 | ||
938 | total_bytes += bio_bytes; | |
939 | nr_bytes -= bio_bytes; | |
940 | ||
941 | if (!nr_bytes) | |
942 | break; | |
943 | } | |
944 | ||
945 | /* | |
946 | * completely done | |
947 | */ | |
948 | if (!req->bio) { | |
949 | /* | |
950 | * Reset counters so that the request stacking driver | |
951 | * can find how many bytes remain in the request | |
952 | * later. | |
953 | */ | |
954 | req->__data_len = 0; | |
955 | return false; | |
956 | } | |
957 | ||
958 | req->__data_len -= total_bytes; | |
959 | ||
960 | /* update sector only for requests with clear definition of sector */ | |
961 | if (!blk_rq_is_passthrough(req)) | |
962 | req->__sector += total_bytes >> 9; | |
963 | ||
964 | /* mixed attributes always follow the first bio */ | |
965 | if (req->rq_flags & RQF_MIXED_MERGE) { | |
966 | req->cmd_flags &= ~REQ_FAILFAST_MASK; | |
967 | req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; | |
968 | } | |
969 | ||
970 | if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { | |
971 | /* | |
972 | * If total number of sectors is less than the first segment | |
973 | * size, something has gone terribly wrong. | |
974 | */ | |
975 | if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { | |
976 | blk_dump_rq_flags(req, "request botched"); | |
977 | req->__data_len = blk_rq_cur_bytes(req); | |
978 | } | |
979 | ||
980 | /* recalculate the number of segments */ | |
981 | req->nr_phys_segments = blk_recalc_rq_segments(req); | |
982 | } | |
983 | ||
984 | return true; | |
985 | } | |
986 | EXPORT_SYMBOL_GPL(blk_update_request); | |
987 | ||
450b7879 CH |
988 | static inline void blk_account_io_done(struct request *req, u64 now) |
989 | { | |
5a80bd07 HC |
990 | trace_block_io_done(req); |
991 | ||
450b7879 CH |
992 | /* |
993 | * Account IO completion. flush_rq isn't accounted as a | |
994 | * normal IO on queueing nor completion. Accounting the | |
995 | * containing request is enough. | |
996 | */ | |
997 | if (blk_do_io_stat(req) && req->part && | |
06965037 CK |
998 | !(req->rq_flags & RQF_FLUSH_SEQ)) { |
999 | const int sgrp = op_stat_group(req_op(req)); | |
450b7879 | 1000 | |
06965037 CK |
1001 | part_stat_lock(); |
1002 | update_io_ticks(req->part, jiffies, true); | |
1003 | part_stat_inc(req->part, ios[sgrp]); | |
1004 | part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns); | |
1005 | part_stat_unlock(); | |
1006 | } | |
450b7879 CH |
1007 | } |
1008 | ||
1009 | static inline void blk_account_io_start(struct request *req) | |
1010 | { | |
5a80bd07 HC |
1011 | trace_block_io_start(req); |
1012 | ||
e165fb4d CK |
1013 | if (blk_do_io_stat(req)) { |
1014 | /* | |
1015 | * All non-passthrough requests are created from a bio with one | |
1016 | * exception: when a flush command that is part of a flush sequence | |
1017 | * generated by the state machine in blk-flush.c is cloned onto the | |
1018 | * lower device by dm-multipath we can get here without a bio. | |
1019 | */ | |
1020 | if (req->bio) | |
1021 | req->part = req->bio->bi_bdev; | |
1022 | else | |
1023 | req->part = req->q->disk->part0; | |
1024 | ||
1025 | part_stat_lock(); | |
1026 | update_io_ticks(req->part, jiffies, false); | |
1027 | part_stat_unlock(); | |
1028 | } | |
450b7879 CH |
1029 | } |
1030 | ||
f794f335 | 1031 | static inline void __blk_mq_end_request_acct(struct request *rq, u64 now) |
320ae51f | 1032 | { |
54bdd67d | 1033 | if (rq->rq_flags & RQF_STATS) |
522a7775 | 1034 | blk_stat_add(rq, now); |
4bc6339a | 1035 | |
87890092 | 1036 | blk_mq_sched_completed_request(rq, now); |
522a7775 | 1037 | blk_account_io_done(rq, now); |
f794f335 | 1038 | } |
522a7775 | 1039 | |
f794f335 JA |
1040 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
1041 | { | |
1042 | if (blk_mq_need_time_stamp(rq)) | |
08420cf7 | 1043 | __blk_mq_end_request_acct(rq, blk_time_get_ns()); |
0d11e6ac | 1044 | |
e5c0ca13 CZ |
1045 | blk_mq_finish_request(rq); |
1046 | ||
91b63639 | 1047 | if (rq->end_io) { |
a7905043 | 1048 | rq_qos_done(rq->q, rq); |
de671d61 JA |
1049 | if (rq->end_io(rq, error) == RQ_END_IO_FREE) |
1050 | blk_mq_free_request(rq); | |
91b63639 | 1051 | } else { |
320ae51f | 1052 | blk_mq_free_request(rq); |
91b63639 | 1053 | } |
320ae51f | 1054 | } |
c8a446ad | 1055 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 1056 | |
2a842aca | 1057 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
1058 | { |
1059 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
1060 | BUG(); | |
c8a446ad | 1061 | __blk_mq_end_request(rq, error); |
63151a44 | 1062 | } |
c8a446ad | 1063 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 1064 | |
f794f335 JA |
1065 | #define TAG_COMP_BATCH 32 |
1066 | ||
1067 | static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx, | |
1068 | int *tag_array, int nr_tags) | |
1069 | { | |
1070 | struct request_queue *q = hctx->queue; | |
1071 | ||
b8643d68 | 1072 | blk_mq_sub_active_requests(hctx, nr_tags); |
3b87c6ea | 1073 | |
f794f335 JA |
1074 | blk_mq_put_tags(hctx->tags, tag_array, nr_tags); |
1075 | percpu_ref_put_many(&q->q_usage_counter, nr_tags); | |
1076 | } | |
1077 | ||
1078 | void blk_mq_end_request_batch(struct io_comp_batch *iob) | |
1079 | { | |
1080 | int tags[TAG_COMP_BATCH], nr_tags = 0; | |
02f7eab0 | 1081 | struct blk_mq_hw_ctx *cur_hctx = NULL; |
f794f335 JA |
1082 | struct request *rq; |
1083 | u64 now = 0; | |
1084 | ||
1085 | if (iob->need_ts) | |
08420cf7 | 1086 | now = blk_time_get_ns(); |
f794f335 JA |
1087 | |
1088 | while ((rq = rq_list_pop(&iob->req_list)) != NULL) { | |
1089 | prefetch(rq->bio); | |
1090 | prefetch(rq->rq_next); | |
1091 | ||
5581a5dd | 1092 | blk_complete_request(rq); |
f794f335 JA |
1093 | if (iob->need_ts) |
1094 | __blk_mq_end_request_acct(rq, now); | |
1095 | ||
e5c0ca13 CZ |
1096 | blk_mq_finish_request(rq); |
1097 | ||
98b26a0e JA |
1098 | rq_qos_done(rq->q, rq); |
1099 | ||
ab3e1d3b JA |
1100 | /* |
1101 | * If end_io handler returns NONE, then it still has | |
1102 | * ownership of the request. | |
1103 | */ | |
1104 | if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE) | |
1105 | continue; | |
1106 | ||
f794f335 | 1107 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 1108 | if (!req_ref_put_and_test(rq)) |
f794f335 JA |
1109 | continue; |
1110 | ||
1111 | blk_crypto_free_request(rq); | |
1112 | blk_pm_mark_last_busy(rq); | |
f794f335 | 1113 | |
02f7eab0 JA |
1114 | if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) { |
1115 | if (cur_hctx) | |
1116 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); | |
f794f335 | 1117 | nr_tags = 0; |
02f7eab0 | 1118 | cur_hctx = rq->mq_hctx; |
f794f335 JA |
1119 | } |
1120 | tags[nr_tags++] = rq->tag; | |
f794f335 JA |
1121 | } |
1122 | ||
1123 | if (nr_tags) | |
02f7eab0 | 1124 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); |
f794f335 JA |
1125 | } |
1126 | EXPORT_SYMBOL_GPL(blk_mq_end_request_batch); | |
1127 | ||
f9ab4918 | 1128 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 1129 | { |
f9ab4918 SAS |
1130 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
1131 | struct request *rq, *next; | |
c3077b5d | 1132 | |
f9ab4918 | 1133 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 1134 | rq->q->mq_ops->complete(rq); |
320ae51f | 1135 | } |
320ae51f | 1136 | |
f9ab4918 | 1137 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 1138 | { |
f9ab4918 | 1139 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
1140 | } |
1141 | ||
c3077b5d CH |
1142 | static int blk_softirq_cpu_dead(unsigned int cpu) |
1143 | { | |
f9ab4918 | 1144 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
1145 | return 0; |
1146 | } | |
1147 | ||
40d09b53 | 1148 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 1149 | { |
f9ab4918 | 1150 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
1151 | } |
1152 | ||
96339526 CH |
1153 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
1154 | { | |
1155 | int cpu = raw_smp_processor_id(); | |
1156 | ||
1157 | if (!IS_ENABLED(CONFIG_SMP) || | |
1158 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
1159 | return false; | |
71425189 SAS |
1160 | /* |
1161 | * With force threaded interrupts enabled, raising softirq from an SMP | |
1162 | * function call will always result in waking the ksoftirqd thread. | |
1163 | * This is probably worse than completing the request on a different | |
1164 | * cache domain. | |
1165 | */ | |
91cc470e | 1166 | if (force_irqthreads()) |
71425189 | 1167 | return false; |
96339526 | 1168 | |
af550e4c | 1169 | /* same CPU or cache domain and capacity? Complete locally */ |
96339526 CH |
1170 | if (cpu == rq->mq_ctx->cpu || |
1171 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
af550e4c QY |
1172 | cpus_share_cache(cpu, rq->mq_ctx->cpu) && |
1173 | cpus_equal_capacity(cpu, rq->mq_ctx->cpu))) | |
96339526 CH |
1174 | return false; |
1175 | ||
1176 | /* don't try to IPI to an offline CPU */ | |
1177 | return cpu_online(rq->mq_ctx->cpu); | |
1178 | } | |
1179 | ||
f9ab4918 SAS |
1180 | static void blk_mq_complete_send_ipi(struct request *rq) |
1181 | { | |
f9ab4918 SAS |
1182 | unsigned int cpu; |
1183 | ||
1184 | cpu = rq->mq_ctx->cpu; | |
660e802c CZ |
1185 | if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu))) |
1186 | smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu)); | |
f9ab4918 SAS |
1187 | } |
1188 | ||
1189 | static void blk_mq_raise_softirq(struct request *rq) | |
1190 | { | |
1191 | struct llist_head *list; | |
1192 | ||
1193 | preempt_disable(); | |
1194 | list = this_cpu_ptr(&blk_cpu_done); | |
1195 | if (llist_add(&rq->ipi_list, list)) | |
1196 | raise_softirq(BLOCK_SOFTIRQ); | |
1197 | preempt_enable(); | |
1198 | } | |
1199 | ||
40d09b53 | 1200 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 1201 | { |
af78ff7c | 1202 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 1203 | |
4ab32bf3 | 1204 | /* |
f168420c LS |
1205 | * For request which hctx has only one ctx mapping, |
1206 | * or a polled request, always complete locally, | |
1207 | * it's pointless to redirect the completion. | |
4ab32bf3 | 1208 | */ |
30654614 ET |
1209 | if ((rq->mq_hctx->nr_ctx == 1 && |
1210 | rq->mq_ctx->cpu == raw_smp_processor_id()) || | |
1211 | rq->cmd_flags & REQ_POLLED) | |
40d09b53 | 1212 | return false; |
38535201 | 1213 | |
96339526 | 1214 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
1215 | blk_mq_complete_send_ipi(rq); |
1216 | return true; | |
3d6efbf6 | 1217 | } |
40d09b53 | 1218 | |
f9ab4918 SAS |
1219 | if (rq->q->nr_hw_queues == 1) { |
1220 | blk_mq_raise_softirq(rq); | |
1221 | return true; | |
1222 | } | |
1223 | return false; | |
40d09b53 CH |
1224 | } |
1225 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
1226 | ||
1227 | /** | |
1228 | * blk_mq_complete_request - end I/O on a request | |
1229 | * @rq: the request being processed | |
1230 | * | |
1231 | * Description: | |
1232 | * Complete a request by scheduling the ->complete_rq operation. | |
1233 | **/ | |
1234 | void blk_mq_complete_request(struct request *rq) | |
1235 | { | |
1236 | if (!blk_mq_complete_request_remote(rq)) | |
1237 | rq->q->mq_ops->complete(rq); | |
320ae51f | 1238 | } |
15f73f5b | 1239 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 1240 | |
105663f7 AA |
1241 | /** |
1242 | * blk_mq_start_request - Start processing a request | |
1243 | * @rq: Pointer to request to be started | |
1244 | * | |
1245 | * Function used by device drivers to notify the block layer that a request | |
1246 | * is going to be processed now, so blk layer can do proper initializations | |
1247 | * such as starting the timeout timer. | |
1248 | */ | |
e2490073 | 1249 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
1250 | { |
1251 | struct request_queue *q = rq->q; | |
1252 | ||
a54895fa | 1253 | trace_block_rq_issue(rq); |
320ae51f | 1254 | |
847c5bcd KK |
1255 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) && |
1256 | !blk_rq_is_passthrough(rq)) { | |
08420cf7 | 1257 | rq->io_start_time_ns = blk_time_get_ns(); |
3d244306 | 1258 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 1259 | rq->rq_flags |= RQF_STATS; |
a7905043 | 1260 | rq_qos_issue(q, rq); |
cf43e6be JA |
1261 | } |
1262 | ||
1d9bd516 | 1263 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 1264 | |
1d9bd516 | 1265 | blk_add_timer(rq); |
12f5b931 | 1266 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
217b613a | 1267 | rq->mq_hctx->tags->rqs[rq->tag] = rq; |
49f5baa5 | 1268 | |
54d4e6ab MG |
1269 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
1270 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
1271 | q->integrity.profile->prepare_fn(rq); | |
1272 | #endif | |
3e08773c | 1273 | if (rq->bio && rq->bio->bi_opf & REQ_POLLED) |
f6c80cff | 1274 | WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num); |
320ae51f | 1275 | } |
e2490073 | 1276 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 1277 | |
a327c341 ML |
1278 | /* |
1279 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple | |
1280 | * queues. This is important for md arrays to benefit from merging | |
1281 | * requests. | |
1282 | */ | |
1283 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
1284 | { | |
1285 | if (plug->multiple_queues) | |
1286 | return BLK_MAX_REQUEST_COUNT * 2; | |
1287 | return BLK_MAX_REQUEST_COUNT; | |
1288 | } | |
1289 | ||
1290 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) | |
1291 | { | |
1292 | struct request *last = rq_list_peek(&plug->mq_list); | |
1293 | ||
1294 | if (!plug->rq_count) { | |
1295 | trace_block_plug(rq->q); | |
1296 | } else if (plug->rq_count >= blk_plug_max_rq_count(plug) || | |
1297 | (!blk_queue_nomerges(rq->q) && | |
1298 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1299 | blk_mq_flush_plug_list(plug, false); | |
878eb6e4 | 1300 | last = NULL; |
a327c341 ML |
1301 | trace_block_plug(rq->q); |
1302 | } | |
1303 | ||
1304 | if (!plug->multiple_queues && last && last->q != rq->q) | |
1305 | plug->multiple_queues = true; | |
c6b7a3a2 ML |
1306 | /* |
1307 | * Any request allocated from sched tags can't be issued to | |
1308 | * ->queue_rqs() directly | |
1309 | */ | |
1310 | if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS)) | |
a327c341 ML |
1311 | plug->has_elevator = true; |
1312 | rq->rq_next = NULL; | |
1313 | rq_list_add(&plug->mq_list, rq); | |
1314 | plug->rq_count++; | |
1315 | } | |
1316 | ||
4054cff9 CH |
1317 | /** |
1318 | * blk_execute_rq_nowait - insert a request to I/O scheduler for execution | |
4054cff9 CH |
1319 | * @rq: request to insert |
1320 | * @at_head: insert request at head or tail of queue | |
4054cff9 CH |
1321 | * |
1322 | * Description: | |
1323 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1324 | * for execution. Don't wait for completion. | |
1325 | * | |
1326 | * Note: | |
1327 | * This function will invoke @done directly if the queue is dead. | |
1328 | */ | |
e2e53086 | 1329 | void blk_execute_rq_nowait(struct request *rq, bool at_head) |
4054cff9 | 1330 | { |
f0dbe6e8 CH |
1331 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
1332 | ||
ae948fd6 CH |
1333 | WARN_ON(irqs_disabled()); |
1334 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 | 1335 | |
ae948fd6 | 1336 | blk_account_io_start(rq); |
110fdb44 PR |
1337 | |
1338 | /* | |
1339 | * As plugging can be enabled for passthrough requests on a zoned | |
1340 | * device, directly accessing the plug instead of using blk_mq_plug() | |
1341 | * should not have any consequences. | |
1342 | */ | |
f0dbe6e8 | 1343 | if (current->plug && !at_head) { |
ae948fd6 | 1344 | blk_add_rq_to_plug(current->plug, rq); |
f0dbe6e8 CH |
1345 | return; |
1346 | } | |
1347 | ||
710fa378 | 1348 | blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); |
65a558f6 | 1349 | blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING); |
4054cff9 CH |
1350 | } |
1351 | EXPORT_SYMBOL_GPL(blk_execute_rq_nowait); | |
1352 | ||
32ac5a9b CH |
1353 | struct blk_rq_wait { |
1354 | struct completion done; | |
1355 | blk_status_t ret; | |
1356 | }; | |
1357 | ||
de671d61 | 1358 | static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret) |
32ac5a9b CH |
1359 | { |
1360 | struct blk_rq_wait *wait = rq->end_io_data; | |
1361 | ||
1362 | wait->ret = ret; | |
1363 | complete(&wait->done); | |
de671d61 | 1364 | return RQ_END_IO_NONE; |
32ac5a9b CH |
1365 | } |
1366 | ||
c6e99ea4 | 1367 | bool blk_rq_is_poll(struct request *rq) |
4054cff9 CH |
1368 | { |
1369 | if (!rq->mq_hctx) | |
1370 | return false; | |
1371 | if (rq->mq_hctx->type != HCTX_TYPE_POLL) | |
1372 | return false; | |
4054cff9 CH |
1373 | return true; |
1374 | } | |
c6e99ea4 | 1375 | EXPORT_SYMBOL_GPL(blk_rq_is_poll); |
4054cff9 CH |
1376 | |
1377 | static void blk_rq_poll_completion(struct request *rq, struct completion *wait) | |
1378 | { | |
1379 | do { | |
f6c80cff | 1380 | blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0); |
4054cff9 CH |
1381 | cond_resched(); |
1382 | } while (!completion_done(wait)); | |
1383 | } | |
1384 | ||
1385 | /** | |
1386 | * blk_execute_rq - insert a request into queue for execution | |
4054cff9 CH |
1387 | * @rq: request to insert |
1388 | * @at_head: insert request at head or tail of queue | |
1389 | * | |
1390 | * Description: | |
1391 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1392 | * for execution and wait for completion. | |
1393 | * Return: The blk_status_t result provided to blk_mq_end_request(). | |
1394 | */ | |
b84ba30b | 1395 | blk_status_t blk_execute_rq(struct request *rq, bool at_head) |
4054cff9 | 1396 | { |
f0dbe6e8 | 1397 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
32ac5a9b CH |
1398 | struct blk_rq_wait wait = { |
1399 | .done = COMPLETION_INITIALIZER_ONSTACK(wait.done), | |
1400 | }; | |
4054cff9 | 1401 | |
ae948fd6 CH |
1402 | WARN_ON(irqs_disabled()); |
1403 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 CH |
1404 | |
1405 | rq->end_io_data = &wait; | |
ae948fd6 | 1406 | rq->end_io = blk_end_sync_rq; |
4054cff9 | 1407 | |
ae948fd6 | 1408 | blk_account_io_start(rq); |
710fa378 | 1409 | blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); |
f0dbe6e8 | 1410 | blk_mq_run_hw_queue(hctx, false); |
4054cff9 | 1411 | |
0eb4db47 | 1412 | if (blk_rq_is_poll(rq)) |
32ac5a9b | 1413 | blk_rq_poll_completion(rq, &wait.done); |
0eb4db47 KB |
1414 | else |
1415 | blk_wait_io(&wait.done); | |
4054cff9 | 1416 | |
32ac5a9b | 1417 | return wait.ret; |
4054cff9 CH |
1418 | } |
1419 | EXPORT_SYMBOL(blk_execute_rq); | |
1420 | ||
ed0791b2 | 1421 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
1422 | { |
1423 | struct request_queue *q = rq->q; | |
1424 | ||
923218f6 ML |
1425 | blk_mq_put_driver_tag(rq); |
1426 | ||
a54895fa | 1427 | trace_block_rq_requeue(rq); |
a7905043 | 1428 | rq_qos_requeue(q, rq); |
49f5baa5 | 1429 | |
12f5b931 KB |
1430 | if (blk_mq_request_started(rq)) { |
1431 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 1432 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 1433 | } |
320ae51f JA |
1434 | } |
1435 | ||
2b053aca | 1436 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 1437 | { |
214a4418 | 1438 | struct request_queue *q = rq->q; |
9a67aa52 | 1439 | unsigned long flags; |
214a4418 | 1440 | |
ed0791b2 | 1441 | __blk_mq_requeue_request(rq); |
ed0791b2 | 1442 | |
105976f5 ML |
1443 | /* this request will be re-inserted to io scheduler queue */ |
1444 | blk_mq_sched_requeue_request(rq); | |
1445 | ||
9a67aa52 CH |
1446 | spin_lock_irqsave(&q->requeue_lock, flags); |
1447 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
1448 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
214a4418 CH |
1449 | |
1450 | if (kick_requeue_list) | |
1451 | blk_mq_kick_requeue_list(q); | |
ed0791b2 CH |
1452 | } |
1453 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
1454 | ||
6fca6a61 CH |
1455 | static void blk_mq_requeue_work(struct work_struct *work) |
1456 | { | |
1457 | struct request_queue *q = | |
2849450a | 1458 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 | 1459 | LIST_HEAD(rq_list); |
9a67aa52 CH |
1460 | LIST_HEAD(flush_list); |
1461 | struct request *rq; | |
6fca6a61 | 1462 | |
18e9781d | 1463 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 1464 | list_splice_init(&q->requeue_list, &rq_list); |
9a67aa52 | 1465 | list_splice_init(&q->flush_list, &flush_list); |
18e9781d | 1466 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 | 1467 | |
9a67aa52 CH |
1468 | while (!list_empty(&rq_list)) { |
1469 | rq = list_entry(rq_list.next, struct request, queuelist); | |
aef1897c | 1470 | /* |
a1e948b8 CH |
1471 | * If RQF_DONTPREP ist set, the request has been started by the |
1472 | * driver already and might have driver-specific data allocated | |
1473 | * already. Insert it into the hctx dispatch list to avoid | |
1474 | * block layer merges for the request. | |
aef1897c | 1475 | */ |
a1e948b8 | 1476 | if (rq->rq_flags & RQF_DONTPREP) { |
a1e948b8 | 1477 | list_del_init(&rq->queuelist); |
2b597613 | 1478 | blk_mq_request_bypass_insert(rq, 0); |
9a67aa52 | 1479 | } else { |
a1e948b8 | 1480 | list_del_init(&rq->queuelist); |
710fa378 | 1481 | blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD); |
a1e948b8 | 1482 | } |
6fca6a61 CH |
1483 | } |
1484 | ||
9a67aa52 CH |
1485 | while (!list_empty(&flush_list)) { |
1486 | rq = list_entry(flush_list.next, struct request, queuelist); | |
6fca6a61 | 1487 | list_del_init(&rq->queuelist); |
710fa378 | 1488 | blk_mq_insert_request(rq, 0); |
6fca6a61 CH |
1489 | } |
1490 | ||
52d7f1b5 | 1491 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
1492 | } |
1493 | ||
6fca6a61 CH |
1494 | void blk_mq_kick_requeue_list(struct request_queue *q) |
1495 | { | |
ae943d20 | 1496 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
1497 | } |
1498 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
1499 | ||
2849450a MS |
1500 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
1501 | unsigned long msecs) | |
1502 | { | |
d4acf365 BVA |
1503 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
1504 | msecs_to_jiffies(msecs)); | |
2849450a MS |
1505 | } |
1506 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
1507 | ||
0e4237ae ML |
1508 | static bool blk_is_flush_data_rq(struct request *rq) |
1509 | { | |
1510 | return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq); | |
1511 | } | |
1512 | ||
2dd6532e | 1513 | static bool blk_mq_rq_inflight(struct request *rq, void *priv) |
ae879912 JA |
1514 | { |
1515 | /* | |
8ab30a33 JG |
1516 | * If we find a request that isn't idle we know the queue is busy |
1517 | * as it's checked in the iter. | |
1518 | * Return false to stop the iteration. | |
0e4237ae ML |
1519 | * |
1520 | * In case of queue quiesce, if one flush data request is completed, | |
1521 | * don't count it as inflight given the flush sequence is suspended, | |
1522 | * and the original flush data request is invisible to driver, just | |
1523 | * like other pending requests because of quiesce | |
ae879912 | 1524 | */ |
0e4237ae ML |
1525 | if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) && |
1526 | blk_is_flush_data_rq(rq) && | |
1527 | blk_mq_request_completed(rq))) { | |
ae879912 JA |
1528 | bool *busy = priv; |
1529 | ||
1530 | *busy = true; | |
1531 | return false; | |
1532 | } | |
1533 | ||
1534 | return true; | |
1535 | } | |
1536 | ||
3c94d83c | 1537 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
1538 | { |
1539 | bool busy = false; | |
1540 | ||
3c94d83c | 1541 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
1542 | return busy; |
1543 | } | |
3c94d83c | 1544 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 1545 | |
9bdb4833 | 1546 | static void blk_mq_rq_timed_out(struct request *req) |
320ae51f | 1547 | { |
da661267 | 1548 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
1549 | if (req->q->mq_ops->timeout) { |
1550 | enum blk_eh_timer_return ret; | |
1551 | ||
9bdb4833 | 1552 | ret = req->q->mq_ops->timeout(req); |
d1210d5a CH |
1553 | if (ret == BLK_EH_DONE) |
1554 | return; | |
1555 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 1556 | } |
d1210d5a CH |
1557 | |
1558 | blk_add_timer(req); | |
87ee7b11 | 1559 | } |
5b3f25fc | 1560 | |
82c22947 DJ |
1561 | struct blk_expired_data { |
1562 | bool has_timedout_rq; | |
1563 | unsigned long next; | |
1564 | unsigned long timeout_start; | |
1565 | }; | |
1566 | ||
1567 | static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired) | |
81481eb4 | 1568 | { |
12f5b931 | 1569 | unsigned long deadline; |
87ee7b11 | 1570 | |
12f5b931 KB |
1571 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
1572 | return false; | |
da661267 CH |
1573 | if (rq->rq_flags & RQF_TIMED_OUT) |
1574 | return false; | |
a7af0af3 | 1575 | |
079076b3 | 1576 | deadline = READ_ONCE(rq->deadline); |
82c22947 | 1577 | if (time_after_eq(expired->timeout_start, deadline)) |
12f5b931 | 1578 | return true; |
a7af0af3 | 1579 | |
82c22947 DJ |
1580 | if (expired->next == 0) |
1581 | expired->next = deadline; | |
1582 | else if (time_after(expired->next, deadline)) | |
1583 | expired->next = deadline; | |
12f5b931 | 1584 | return false; |
87ee7b11 JA |
1585 | } |
1586 | ||
2e315dc0 ML |
1587 | void blk_mq_put_rq_ref(struct request *rq) |
1588 | { | |
de671d61 JA |
1589 | if (is_flush_rq(rq)) { |
1590 | if (rq->end_io(rq, 0) == RQ_END_IO_FREE) | |
1591 | blk_mq_free_request(rq); | |
1592 | } else if (req_ref_put_and_test(rq)) { | |
2e315dc0 | 1593 | __blk_mq_free_request(rq); |
de671d61 | 1594 | } |
2e315dc0 ML |
1595 | } |
1596 | ||
2dd6532e | 1597 | static bool blk_mq_check_expired(struct request *rq, void *priv) |
1d9bd516 | 1598 | { |
82c22947 | 1599 | struct blk_expired_data *expired = priv; |
12f5b931 KB |
1600 | |
1601 | /* | |
c797b40c ML |
1602 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
1603 | * be reallocated underneath the timeout handler's processing, then | |
1604 | * the expire check is reliable. If the request is not expired, then | |
1605 | * it was completed and reallocated as a new request after returning | |
1606 | * from blk_mq_check_expired(). | |
1d9bd516 | 1607 | */ |
82c22947 DJ |
1608 | if (blk_mq_req_expired(rq, expired)) { |
1609 | expired->has_timedout_rq = true; | |
1610 | return false; | |
1611 | } | |
1612 | return true; | |
1613 | } | |
1614 | ||
1615 | static bool blk_mq_handle_expired(struct request *rq, void *priv) | |
1616 | { | |
1617 | struct blk_expired_data *expired = priv; | |
1618 | ||
1619 | if (blk_mq_req_expired(rq, expired)) | |
9bdb4833 | 1620 | blk_mq_rq_timed_out(rq); |
7baa8572 | 1621 | return true; |
1d9bd516 TH |
1622 | } |
1623 | ||
287922eb | 1624 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 1625 | { |
287922eb CH |
1626 | struct request_queue *q = |
1627 | container_of(work, struct request_queue, timeout_work); | |
82c22947 DJ |
1628 | struct blk_expired_data expired = { |
1629 | .timeout_start = jiffies, | |
1630 | }; | |
1d9bd516 | 1631 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 1632 | unsigned long i; |
320ae51f | 1633 | |
71f79fb3 GKB |
1634 | /* A deadlock might occur if a request is stuck requiring a |
1635 | * timeout at the same time a queue freeze is waiting | |
1636 | * completion, since the timeout code would not be able to | |
1637 | * acquire the queue reference here. | |
1638 | * | |
1639 | * That's why we don't use blk_queue_enter here; instead, we use | |
1640 | * percpu_ref_tryget directly, because we need to be able to | |
1641 | * obtain a reference even in the short window between the queue | |
1642 | * starting to freeze, by dropping the first reference in | |
1671d522 | 1643 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
1644 | * consumed, marked by the instant q_usage_counter reaches |
1645 | * zero. | |
1646 | */ | |
1647 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
1648 | return; |
1649 | ||
82c22947 DJ |
1650 | /* check if there is any timed-out request */ |
1651 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired); | |
1652 | if (expired.has_timedout_rq) { | |
1653 | /* | |
1654 | * Before walking tags, we must ensure any submit started | |
1655 | * before the current time has finished. Since the submit | |
1656 | * uses srcu or rcu, wait for a synchronization point to | |
1657 | * ensure all running submits have finished | |
1658 | */ | |
483239c7 | 1659 | blk_mq_wait_quiesce_done(q->tag_set); |
82c22947 DJ |
1660 | |
1661 | expired.next = 0; | |
1662 | blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired); | |
1663 | } | |
320ae51f | 1664 | |
82c22947 DJ |
1665 | if (expired.next != 0) { |
1666 | mod_timer(&q->timeout, expired.next); | |
0d2602ca | 1667 | } else { |
fcd36c36 BVA |
1668 | /* |
1669 | * Request timeouts are handled as a forward rolling timer. If | |
1670 | * we end up here it means that no requests are pending and | |
1671 | * also that no request has been pending for a while. Mark | |
1672 | * each hctx as idle. | |
1673 | */ | |
f054b56c ML |
1674 | queue_for_each_hw_ctx(q, hctx, i) { |
1675 | /* the hctx may be unmapped, so check it here */ | |
1676 | if (blk_mq_hw_queue_mapped(hctx)) | |
1677 | blk_mq_tag_idle(hctx); | |
1678 | } | |
0d2602ca | 1679 | } |
287922eb | 1680 | blk_queue_exit(q); |
320ae51f JA |
1681 | } |
1682 | ||
88459642 OS |
1683 | struct flush_busy_ctx_data { |
1684 | struct blk_mq_hw_ctx *hctx; | |
1685 | struct list_head *list; | |
1686 | }; | |
1687 | ||
1688 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1689 | { | |
1690 | struct flush_busy_ctx_data *flush_data = data; | |
1691 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1692 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1693 | enum hctx_type type = hctx->type; |
88459642 | 1694 | |
88459642 | 1695 | spin_lock(&ctx->lock); |
c16d6b5a | 1696 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1697 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1698 | spin_unlock(&ctx->lock); |
1699 | return true; | |
1700 | } | |
1701 | ||
1429d7c9 JA |
1702 | /* |
1703 | * Process software queues that have been marked busy, splicing them | |
1704 | * to the for-dispatch | |
1705 | */ | |
2c3ad667 | 1706 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1707 | { |
88459642 OS |
1708 | struct flush_busy_ctx_data data = { |
1709 | .hctx = hctx, | |
1710 | .list = list, | |
1711 | }; | |
1429d7c9 | 1712 | |
88459642 | 1713 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1714 | } |
2c3ad667 | 1715 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1716 | |
b347689f ML |
1717 | struct dispatch_rq_data { |
1718 | struct blk_mq_hw_ctx *hctx; | |
1719 | struct request *rq; | |
1720 | }; | |
1721 | ||
1722 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1723 | void *data) | |
1724 | { | |
1725 | struct dispatch_rq_data *dispatch_data = data; | |
1726 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1727 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1728 | enum hctx_type type = hctx->type; |
b347689f ML |
1729 | |
1730 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1731 | if (!list_empty(&ctx->rq_lists[type])) { |
1732 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1733 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1734 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1735 | sbitmap_clear_bit(sb, bitnr); |
1736 | } | |
1737 | spin_unlock(&ctx->lock); | |
1738 | ||
1739 | return !dispatch_data->rq; | |
1740 | } | |
1741 | ||
1742 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1743 | struct blk_mq_ctx *start) | |
1744 | { | |
f31967f0 | 1745 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1746 | struct dispatch_rq_data data = { |
1747 | .hctx = hctx, | |
1748 | .rq = NULL, | |
1749 | }; | |
1750 | ||
1751 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1752 | dispatch_rq_from_ctx, &data); | |
1753 | ||
1754 | return data.rq; | |
1755 | } | |
1756 | ||
b8643d68 | 1757 | bool __blk_mq_alloc_driver_tag(struct request *rq) |
570e9b73 | 1758 | { |
ae0f1a73 | 1759 | struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1760 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1761 | int tag; |
1762 | ||
568f2700 ML |
1763 | blk_mq_tag_busy(rq->mq_hctx); |
1764 | ||
570e9b73 | 1765 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
ae0f1a73 | 1766 | bt = &rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1767 | tag_offset = 0; |
28500850 ML |
1768 | } else { |
1769 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1770 | return false; | |
570e9b73 ML |
1771 | } |
1772 | ||
570e9b73 ML |
1773 | tag = __sbitmap_queue_get(bt); |
1774 | if (tag == BLK_MQ_NO_TAG) | |
1775 | return false; | |
1776 | ||
1777 | rq->tag = tag + tag_offset; | |
b8643d68 | 1778 | blk_mq_inc_active_requests(rq->mq_hctx); |
568f2700 | 1779 | return true; |
570e9b73 ML |
1780 | } |
1781 | ||
eb619fdb JA |
1782 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1783 | int flags, void *key) | |
da55f2cc OS |
1784 | { |
1785 | struct blk_mq_hw_ctx *hctx; | |
1786 | ||
1787 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1788 | ||
5815839b | 1789 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1790 | if (!list_empty(&wait->entry)) { |
1791 | struct sbitmap_queue *sbq; | |
1792 | ||
1793 | list_del_init(&wait->entry); | |
ae0f1a73 | 1794 | sbq = &hctx->tags->bitmap_tags; |
e8618575 JA |
1795 | atomic_dec(&sbq->ws_active); |
1796 | } | |
5815839b ML |
1797 | spin_unlock(&hctx->dispatch_wait_lock); |
1798 | ||
da55f2cc OS |
1799 | blk_mq_run_hw_queue(hctx, true); |
1800 | return 1; | |
1801 | } | |
1802 | ||
f906a6a0 JA |
1803 | /* |
1804 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1805 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1806 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1807 | * marking us as waiting. |
1808 | */ | |
2278d69f | 1809 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1810 | struct request *rq) |
da55f2cc | 1811 | { |
98b99e94 | 1812 | struct sbitmap_queue *sbq; |
5815839b | 1813 | struct wait_queue_head *wq; |
f906a6a0 JA |
1814 | wait_queue_entry_t *wait; |
1815 | bool ret; | |
da55f2cc | 1816 | |
47df9ce9 KS |
1817 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
1818 | !(blk_mq_is_shared_tags(hctx->flags))) { | |
684b7324 | 1819 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1820 | |
c27d53fb BVA |
1821 | /* |
1822 | * It's possible that a tag was freed in the window between the | |
1823 | * allocation failure and adding the hardware queue to the wait | |
1824 | * queue. | |
1825 | * | |
1826 | * Don't clear RESTART here, someone else could have set it. | |
1827 | * At most this will cost an extra queue run. | |
1828 | */ | |
8ab6bb9e | 1829 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1830 | } |
eb619fdb | 1831 | |
2278d69f | 1832 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1833 | if (!list_empty_careful(&wait->entry)) |
1834 | return false; | |
1835 | ||
98b99e94 KS |
1836 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) |
1837 | sbq = &hctx->tags->breserved_tags; | |
1838 | else | |
1839 | sbq = &hctx->tags->bitmap_tags; | |
e8618575 | 1840 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1841 | |
1842 | spin_lock_irq(&wq->lock); | |
1843 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1844 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1845 | spin_unlock(&hctx->dispatch_wait_lock); |
1846 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1847 | return false; |
eb619fdb JA |
1848 | } |
1849 | ||
e8618575 | 1850 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1851 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1852 | __add_wait_queue(wq, wait); | |
c27d53fb | 1853 | |
5266caaf ML |
1854 | /* |
1855 | * Add one explicit barrier since blk_mq_get_driver_tag() may | |
1856 | * not imply barrier in case of failure. | |
1857 | * | |
1858 | * Order adding us to wait queue and allocating driver tag. | |
1859 | * | |
1860 | * The pair is the one implied in sbitmap_queue_wake_up() which | |
1861 | * orders clearing sbitmap tag bits and waitqueue_active() in | |
1862 | * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless | |
1863 | * | |
1864 | * Otherwise, re-order of adding wait queue and getting driver tag | |
1865 | * may cause __sbitmap_queue_wake_up() to wake up nothing because | |
1866 | * the waitqueue_active() may not observe us in wait queue. | |
1867 | */ | |
1868 | smp_mb(); | |
1869 | ||
da55f2cc | 1870 | /* |
eb619fdb JA |
1871 | * It's possible that a tag was freed in the window between the |
1872 | * allocation failure and adding the hardware queue to the wait | |
1873 | * queue. | |
da55f2cc | 1874 | */ |
8ab6bb9e | 1875 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1876 | if (!ret) { |
5815839b ML |
1877 | spin_unlock(&hctx->dispatch_wait_lock); |
1878 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1879 | return false; |
eb619fdb | 1880 | } |
c27d53fb BVA |
1881 | |
1882 | /* | |
1883 | * We got a tag, remove ourselves from the wait queue to ensure | |
1884 | * someone else gets the wakeup. | |
1885 | */ | |
c27d53fb | 1886 | list_del_init(&wait->entry); |
e8618575 | 1887 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1888 | spin_unlock(&hctx->dispatch_wait_lock); |
1889 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1890 | |
1891 | return true; | |
da55f2cc OS |
1892 | } |
1893 | ||
6e768717 ML |
1894 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1895 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1896 | /* | |
1897 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1898 | * - EWMA is one simple way to compute running average value | |
1899 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1900 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1901 | * factor doesn't matter because EWMA decreases exponentially | |
1902 | */ | |
1903 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1904 | { | |
1905 | unsigned int ewma; | |
1906 | ||
6e768717 ML |
1907 | ewma = hctx->dispatch_busy; |
1908 | ||
1909 | if (!ewma && !busy) | |
1910 | return; | |
1911 | ||
1912 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1913 | if (busy) | |
1914 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1915 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1916 | ||
1917 | hctx->dispatch_busy = ewma; | |
1918 | } | |
1919 | ||
86ff7c2a ML |
1920 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1921 | ||
c92a4103 JT |
1922 | static void blk_mq_handle_dev_resource(struct request *rq, |
1923 | struct list_head *list) | |
1924 | { | |
c92a4103 JT |
1925 | list_add(&rq->queuelist, list); |
1926 | __blk_mq_requeue_request(rq); | |
1927 | } | |
1928 | ||
0512a75b KB |
1929 | static void blk_mq_handle_zone_resource(struct request *rq, |
1930 | struct list_head *zone_list) | |
1931 | { | |
1932 | /* | |
1933 | * If we end up here it is because we cannot dispatch a request to a | |
1934 | * specific zone due to LLD level zone-write locking or other zone | |
1935 | * related resource not being available. In this case, set the request | |
1936 | * aside in zone_list for retrying it later. | |
1937 | */ | |
1938 | list_add(&rq->queuelist, zone_list); | |
1939 | __blk_mq_requeue_request(rq); | |
1940 | } | |
1941 | ||
75383524 ML |
1942 | enum prep_dispatch { |
1943 | PREP_DISPATCH_OK, | |
1944 | PREP_DISPATCH_NO_TAG, | |
1945 | PREP_DISPATCH_NO_BUDGET, | |
1946 | }; | |
1947 | ||
1948 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1949 | bool need_budget) | |
1950 | { | |
1951 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1952 | int budget_token = -1; |
75383524 | 1953 | |
2a5a24aa ML |
1954 | if (need_budget) { |
1955 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1956 | if (budget_token < 0) { | |
1957 | blk_mq_put_driver_tag(rq); | |
1958 | return PREP_DISPATCH_NO_BUDGET; | |
1959 | } | |
1960 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1961 | } |
1962 | ||
1963 | if (!blk_mq_get_driver_tag(rq)) { | |
1964 | /* | |
1965 | * The initial allocation attempt failed, so we need to | |
1966 | * rerun the hardware queue when a tag is freed. The | |
1967 | * waitqueue takes care of that. If the queue is run | |
1968 | * before we add this entry back on the dispatch list, | |
1969 | * we'll re-run it below. | |
1970 | */ | |
1971 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1972 | /* |
1973 | * All budgets not got from this function will be put | |
1974 | * together during handling partial dispatch | |
1975 | */ | |
1976 | if (need_budget) | |
2a5a24aa | 1977 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1978 | return PREP_DISPATCH_NO_TAG; |
1979 | } | |
1980 | } | |
1981 | ||
1982 | return PREP_DISPATCH_OK; | |
1983 | } | |
1984 | ||
1fd40b5e ML |
1985 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1986 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1987 | struct list_head *list) |
1fd40b5e | 1988 | { |
2a5a24aa | 1989 | struct request *rq; |
1fd40b5e | 1990 | |
2a5a24aa ML |
1991 | list_for_each_entry(rq, list, queuelist) { |
1992 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1993 | |
2a5a24aa ML |
1994 | if (budget_token >= 0) |
1995 | blk_mq_put_dispatch_budget(q, budget_token); | |
1996 | } | |
1fd40b5e ML |
1997 | } |
1998 | ||
34c9f547 KS |
1999 | /* |
2000 | * blk_mq_commit_rqs will notify driver using bd->last that there is no | |
2001 | * more requests. (See comment in struct blk_mq_ops for commit_rqs for | |
2002 | * details) | |
2003 | * Attention, we should explicitly call this in unusual cases: | |
2004 | * 1) did not queue everything initially scheduled to queue | |
2005 | * 2) the last attempt to queue a request failed | |
2006 | */ | |
2007 | static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued, | |
2008 | bool from_schedule) | |
2009 | { | |
2010 | if (hctx->queue->mq_ops->commit_rqs && queued) { | |
2011 | trace_block_unplug(hctx->queue, queued, !from_schedule); | |
2012 | hctx->queue->mq_ops->commit_rqs(hctx); | |
2013 | } | |
2014 | } | |
2015 | ||
1f57f8d4 JA |
2016 | /* |
2017 | * Returns true if we did some work AND can potentially do more. | |
2018 | */ | |
445874e8 | 2019 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 2020 | unsigned int nr_budgets) |
320ae51f | 2021 | { |
75383524 | 2022 | enum prep_dispatch prep; |
445874e8 | 2023 | struct request_queue *q = hctx->queue; |
f1ce99f7 | 2024 | struct request *rq; |
4ea58fe4 | 2025 | int queued; |
86ff7c2a | 2026 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 2027 | LIST_HEAD(zone_list); |
9586e67b | 2028 | bool needs_resource = false; |
320ae51f | 2029 | |
81380ca1 OS |
2030 | if (list_empty(list)) |
2031 | return false; | |
2032 | ||
320ae51f JA |
2033 | /* |
2034 | * Now process all the entries, sending them to the driver. | |
2035 | */ | |
4ea58fe4 | 2036 | queued = 0; |
81380ca1 | 2037 | do { |
74c45052 | 2038 | struct blk_mq_queue_data bd; |
320ae51f | 2039 | |
f04c3df3 | 2040 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 2041 | |
445874e8 | 2042 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 2043 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 2044 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 2045 | break; |
de148297 | 2046 | |
320ae51f | 2047 | list_del_init(&rq->queuelist); |
320ae51f | 2048 | |
74c45052 | 2049 | bd.rq = rq; |
f1ce99f7 | 2050 | bd.last = list_empty(list); |
74c45052 | 2051 | |
1fd40b5e ML |
2052 | /* |
2053 | * once the request is queued to lld, no need to cover the | |
2054 | * budget any more | |
2055 | */ | |
2056 | if (nr_budgets) | |
2057 | nr_budgets--; | |
74c45052 | 2058 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
2059 | switch (ret) { |
2060 | case BLK_STS_OK: | |
2061 | queued++; | |
320ae51f | 2062 | break; |
7bf13729 | 2063 | case BLK_STS_RESOURCE: |
9586e67b NA |
2064 | needs_resource = true; |
2065 | fallthrough; | |
7bf13729 ML |
2066 | case BLK_STS_DEV_RESOURCE: |
2067 | blk_mq_handle_dev_resource(rq, list); | |
2068 | goto out; | |
2069 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
2070 | /* |
2071 | * Move the request to zone_list and keep going through | |
2072 | * the dispatch list to find more requests the drive can | |
2073 | * accept. | |
2074 | */ | |
2075 | blk_mq_handle_zone_resource(rq, &zone_list); | |
9586e67b | 2076 | needs_resource = true; |
7bf13729 ML |
2077 | break; |
2078 | default: | |
e21ee5a6 | 2079 | blk_mq_end_request(rq, ret); |
320ae51f | 2080 | } |
81380ca1 | 2081 | } while (!list_empty(list)); |
7bf13729 | 2082 | out: |
0512a75b KB |
2083 | if (!list_empty(&zone_list)) |
2084 | list_splice_tail_init(&zone_list, list); | |
2085 | ||
632bfb63 | 2086 | /* If we didn't flush the entire list, we could have told the driver |
2087 | * there was more coming, but that turned out to be a lie. | |
2088 | */ | |
e4ef2e05 KS |
2089 | if (!list_empty(list) || ret != BLK_STS_OK) |
2090 | blk_mq_commit_rqs(hctx, queued, false); | |
2091 | ||
320ae51f JA |
2092 | /* |
2093 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
2094 | * that is where we will continue on next queue run. | |
2095 | */ | |
f04c3df3 | 2096 | if (!list_empty(list)) { |
86ff7c2a | 2097 | bool needs_restart; |
75383524 ML |
2098 | /* For non-shared tags, the RESTART check will suffice */ |
2099 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
47df9ce9 KS |
2100 | ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) || |
2101 | blk_mq_is_shared_tags(hctx->flags)); | |
86ff7c2a | 2102 | |
2a5a24aa ML |
2103 | if (nr_budgets) |
2104 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 2105 | |
320ae51f | 2106 | spin_lock(&hctx->lock); |
01e99aec | 2107 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 2108 | spin_unlock(&hctx->lock); |
f04c3df3 | 2109 | |
d7d8535f ML |
2110 | /* |
2111 | * Order adding requests to hctx->dispatch and checking | |
2112 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
2113 | * in blk_mq_sched_restart(). Avoid restart code path to | |
2114 | * miss the new added requests to hctx->dispatch, meantime | |
2115 | * SCHED_RESTART is observed here. | |
2116 | */ | |
2117 | smp_mb(); | |
2118 | ||
9ba52e58 | 2119 | /* |
710c785f BVA |
2120 | * If SCHED_RESTART was set by the caller of this function and |
2121 | * it is no longer set that means that it was cleared by another | |
2122 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 2123 | * |
eb619fdb JA |
2124 | * If 'no_tag' is set, that means that we failed getting |
2125 | * a driver tag with an I/O scheduler attached. If our dispatch | |
2126 | * waitqueue is no longer active, ensure that we run the queue | |
2127 | * AFTER adding our entries back to the list. | |
bd166ef1 | 2128 | * |
710c785f BVA |
2129 | * If no I/O scheduler has been configured it is possible that |
2130 | * the hardware queue got stopped and restarted before requests | |
2131 | * were pushed back onto the dispatch list. Rerun the queue to | |
2132 | * avoid starvation. Notes: | |
2133 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
2134 | * been stopped before rerunning a queue. | |
2135 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 2136 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 2137 | * and dm-rq. |
86ff7c2a ML |
2138 | * |
2139 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
2140 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 | 2141 | * that could otherwise occur if the queue is idle. We'll do |
9586e67b NA |
2142 | * similar if we couldn't get budget or couldn't lock a zone |
2143 | * and SCHED_RESTART is set. | |
bd166ef1 | 2144 | */ |
86ff7c2a | 2145 | needs_restart = blk_mq_sched_needs_restart(hctx); |
9586e67b NA |
2146 | if (prep == PREP_DISPATCH_NO_BUDGET) |
2147 | needs_resource = true; | |
86ff7c2a | 2148 | if (!needs_restart || |
eb619fdb | 2149 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 2150 | blk_mq_run_hw_queue(hctx, true); |
6d5e8d21 | 2151 | else if (needs_resource) |
86ff7c2a | 2152 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 2153 | |
6e768717 | 2154 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 2155 | return false; |
4ea58fe4 | 2156 | } |
f04c3df3 | 2157 | |
4ea58fe4 KS |
2158 | blk_mq_update_dispatch_busy(hctx, false); |
2159 | return true; | |
f04c3df3 JA |
2160 | } |
2161 | ||
f82ddf19 ML |
2162 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
2163 | { | |
2164 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
2165 | ||
2166 | if (cpu >= nr_cpu_ids) | |
2167 | cpu = cpumask_first(hctx->cpumask); | |
2168 | return cpu; | |
2169 | } | |
2170 | ||
506e931f JA |
2171 | /* |
2172 | * It'd be great if the workqueue API had a way to pass | |
2173 | * in a mask and had some smarts for more clever placement. | |
2174 | * For now we just round-robin here, switching for every | |
2175 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
2176 | */ | |
2177 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
2178 | { | |
7bed4595 | 2179 | bool tried = false; |
476f8c98 | 2180 | int next_cpu = hctx->next_cpu; |
7bed4595 | 2181 | |
b657d7e6 CH |
2182 | if (hctx->queue->nr_hw_queues == 1) |
2183 | return WORK_CPU_UNBOUND; | |
506e931f JA |
2184 | |
2185 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 2186 | select_cpu: |
476f8c98 | 2187 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 2188 | cpu_online_mask); |
506e931f | 2189 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 2190 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2191 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2192 | } | |
2193 | ||
7bed4595 ML |
2194 | /* |
2195 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
2196 | * and it should only happen in the path of handling CPU DEAD. | |
2197 | */ | |
476f8c98 | 2198 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
2199 | if (!tried) { |
2200 | tried = true; | |
2201 | goto select_cpu; | |
2202 | } | |
2203 | ||
2204 | /* | |
2205 | * Make sure to re-select CPU next time once after CPUs | |
2206 | * in hctx->cpumask become online again. | |
2207 | */ | |
476f8c98 | 2208 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
2209 | hctx->next_cpu_batch = 1; |
2210 | return WORK_CPU_UNBOUND; | |
2211 | } | |
476f8c98 ML |
2212 | |
2213 | hctx->next_cpu = next_cpu; | |
2214 | return next_cpu; | |
506e931f JA |
2215 | } |
2216 | ||
105663f7 | 2217 | /** |
1aa8d875 | 2218 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. |
105663f7 | 2219 | * @hctx: Pointer to the hardware queue to run. |
fa94ba8a | 2220 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 | 2221 | * |
1aa8d875 | 2222 | * Run a hardware queue asynchronously with a delay of @msecs. |
105663f7 | 2223 | */ |
1aa8d875 | 2224 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
320ae51f | 2225 | { |
5435c023 | 2226 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f | 2227 | return; |
ae943d20 BVA |
2228 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
2229 | msecs_to_jiffies(msecs)); | |
7587a5ae | 2230 | } |
7587a5ae BVA |
2231 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); |
2232 | ||
105663f7 AA |
2233 | /** |
2234 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
2235 | * @hctx: Pointer to the hardware queue to run. | |
2236 | * @async: If we want to run the queue asynchronously. | |
2237 | * | |
2238 | * Check if the request queue is not in a quiesced state and if there are | |
2239 | * pending requests to be sent. If this is true, run the queue to send requests | |
2240 | * to hardware. | |
2241 | */ | |
626fb735 | 2242 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 2243 | { |
24f5a90f ML |
2244 | bool need_run; |
2245 | ||
4d5bba5b CH |
2246 | /* |
2247 | * We can't run the queue inline with interrupts disabled. | |
2248 | */ | |
2249 | WARN_ON_ONCE(!async && in_interrupt()); | |
2250 | ||
65a558f6 BVA |
2251 | might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING); |
2252 | ||
24f5a90f ML |
2253 | /* |
2254 | * When queue is quiesced, we may be switching io scheduler, or | |
2255 | * updating nr_hw_queues, or other things, and we can't run queue | |
2256 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
2257 | * | |
2258 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
2259 | * quiesced. | |
2260 | */ | |
41adf531 | 2261 | __blk_mq_run_dispatch_ops(hctx->queue, false, |
2a904d00 ML |
2262 | need_run = !blk_queue_quiesced(hctx->queue) && |
2263 | blk_mq_hctx_has_pending(hctx)); | |
24f5a90f | 2264 | |
1aa8d875 CH |
2265 | if (!need_run) |
2266 | return; | |
2267 | ||
65a558f6 | 2268 | if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) { |
1aa8d875 CH |
2269 | blk_mq_delay_run_hw_queue(hctx, 0); |
2270 | return; | |
2271 | } | |
2272 | ||
4d5bba5b CH |
2273 | blk_mq_run_dispatch_ops(hctx->queue, |
2274 | blk_mq_sched_dispatch_requests(hctx)); | |
320ae51f | 2275 | } |
5b727272 | 2276 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 2277 | |
b6e68ee8 JK |
2278 | /* |
2279 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
2280 | * scheduler. | |
2281 | */ | |
2282 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
2283 | { | |
5d05426e | 2284 | struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); |
b6e68ee8 JK |
2285 | /* |
2286 | * If the IO scheduler does not respect hardware queues when | |
2287 | * dispatching, we just don't bother with multiple HW queues and | |
2288 | * dispatch from hctx for the current CPU since running multiple queues | |
2289 | * just causes lock contention inside the scheduler and pointless cache | |
2290 | * bouncing. | |
2291 | */ | |
51ab80f0 | 2292 | struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT]; |
5d05426e | 2293 | |
b6e68ee8 JK |
2294 | if (!blk_mq_hctx_stopped(hctx)) |
2295 | return hctx; | |
2296 | return NULL; | |
2297 | } | |
2298 | ||
105663f7 | 2299 | /** |
24f7bb88 | 2300 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
2301 | * @q: Pointer to the request queue to run. |
2302 | * @async: If we want to run the queue asynchronously. | |
2303 | */ | |
b94ec296 | 2304 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 2305 | { |
b6e68ee8 | 2306 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2307 | unsigned long i; |
320ae51f | 2308 | |
b6e68ee8 | 2309 | sq_hctx = NULL; |
4d337ceb | 2310 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2311 | sq_hctx = blk_mq_get_sq_hctx(q); |
320ae51f | 2312 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 2313 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 2314 | continue; |
b6e68ee8 JK |
2315 | /* |
2316 | * Dispatch from this hctx either if there's no hctx preferred | |
2317 | * by IO scheduler or if it has requests that bypass the | |
2318 | * scheduler. | |
2319 | */ | |
2320 | if (!sq_hctx || sq_hctx == hctx || | |
2321 | !list_empty_careful(&hctx->dispatch)) | |
2322 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
2323 | } |
2324 | } | |
b94ec296 | 2325 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 2326 | |
b9151e7b DA |
2327 | /** |
2328 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
2329 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 2330 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
2331 | */ |
2332 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
2333 | { | |
b6e68ee8 | 2334 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2335 | unsigned long i; |
b9151e7b | 2336 | |
b6e68ee8 | 2337 | sq_hctx = NULL; |
4d337ceb | 2338 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2339 | sq_hctx = blk_mq_get_sq_hctx(q); |
b9151e7b DA |
2340 | queue_for_each_hw_ctx(q, hctx, i) { |
2341 | if (blk_mq_hctx_stopped(hctx)) | |
2342 | continue; | |
8f5fea65 DJ |
2343 | /* |
2344 | * If there is already a run_work pending, leave the | |
2345 | * pending delay untouched. Otherwise, a hctx can stall | |
2346 | * if another hctx is re-delaying the other's work | |
2347 | * before the work executes. | |
2348 | */ | |
2349 | if (delayed_work_pending(&hctx->run_work)) | |
2350 | continue; | |
b6e68ee8 JK |
2351 | /* |
2352 | * Dispatch from this hctx either if there's no hctx preferred | |
2353 | * by IO scheduler or if it has requests that bypass the | |
2354 | * scheduler. | |
2355 | */ | |
2356 | if (!sq_hctx || sq_hctx == hctx || | |
2357 | !list_empty_careful(&hctx->dispatch)) | |
2358 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
2359 | } |
2360 | } | |
2361 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
2362 | ||
39a70c76 ML |
2363 | /* |
2364 | * This function is often used for pausing .queue_rq() by driver when | |
2365 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2366 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2367 | * |
2368 | * We do not guarantee that dispatch can be drained or blocked | |
2369 | * after blk_mq_stop_hw_queue() returns. Please use | |
2370 | * blk_mq_quiesce_queue() for that requirement. | |
2371 | */ | |
2719aa21 JA |
2372 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
2373 | { | |
641a9ed6 | 2374 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 2375 | |
641a9ed6 | 2376 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 2377 | } |
641a9ed6 | 2378 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 2379 | |
39a70c76 ML |
2380 | /* |
2381 | * This function is often used for pausing .queue_rq() by driver when | |
2382 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2383 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2384 | * |
2385 | * We do not guarantee that dispatch can be drained or blocked | |
2386 | * after blk_mq_stop_hw_queues() returns. Please use | |
2387 | * blk_mq_quiesce_queue() for that requirement. | |
2388 | */ | |
2719aa21 JA |
2389 | void blk_mq_stop_hw_queues(struct request_queue *q) |
2390 | { | |
641a9ed6 | 2391 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 2392 | unsigned long i; |
641a9ed6 ML |
2393 | |
2394 | queue_for_each_hw_ctx(q, hctx, i) | |
2395 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
2396 | } |
2397 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
2398 | ||
320ae51f JA |
2399 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
2400 | { | |
2401 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 2402 | |
65a558f6 | 2403 | blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING); |
320ae51f JA |
2404 | } |
2405 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
2406 | ||
2f268556 CH |
2407 | void blk_mq_start_hw_queues(struct request_queue *q) |
2408 | { | |
2409 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2410 | unsigned long i; |
2f268556 CH |
2411 | |
2412 | queue_for_each_hw_ctx(q, hctx, i) | |
2413 | blk_mq_start_hw_queue(hctx); | |
2414 | } | |
2415 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
2416 | ||
ae911c5e JA |
2417 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
2418 | { | |
2419 | if (!blk_mq_hctx_stopped(hctx)) | |
2420 | return; | |
2421 | ||
2422 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
2423 | blk_mq_run_hw_queue(hctx, async); | |
2424 | } | |
2425 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
2426 | ||
1b4a3258 | 2427 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
2428 | { |
2429 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2430 | unsigned long i; |
320ae51f | 2431 | |
ae911c5e | 2432 | queue_for_each_hw_ctx(q, hctx, i) |
65a558f6 BVA |
2433 | blk_mq_start_stopped_hw_queue(hctx, async || |
2434 | (hctx->flags & BLK_MQ_F_BLOCKING)); | |
320ae51f JA |
2435 | } |
2436 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
2437 | ||
70f4db63 | 2438 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f | 2439 | { |
c20a1a2c CH |
2440 | struct blk_mq_hw_ctx *hctx = |
2441 | container_of(work, struct blk_mq_hw_ctx, run_work.work); | |
7b607814 | 2442 | |
4d5bba5b CH |
2443 | blk_mq_run_dispatch_ops(hctx->queue, |
2444 | blk_mq_sched_dispatch_requests(hctx)); | |
320ae51f JA |
2445 | } |
2446 | ||
105663f7 AA |
2447 | /** |
2448 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
2449 | * @rq: Pointer to request to be inserted. | |
2b597613 | 2450 | * @flags: BLK_MQ_INSERT_* |
105663f7 | 2451 | * |
157f377b JA |
2452 | * Should only be used carefully, when the caller knows we want to |
2453 | * bypass a potential IO scheduler on the target device. | |
2454 | */ | |
360f2648 | 2455 | static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags) |
157f377b | 2456 | { |
ea4f995e | 2457 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
2458 | |
2459 | spin_lock(&hctx->lock); | |
2b597613 | 2460 | if (flags & BLK_MQ_INSERT_AT_HEAD) |
01e99aec ML |
2461 | list_add(&rq->queuelist, &hctx->dispatch); |
2462 | else | |
2463 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b | 2464 | spin_unlock(&hctx->lock); |
157f377b JA |
2465 | } |
2466 | ||
05a93117 CH |
2467 | static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, |
2468 | struct blk_mq_ctx *ctx, struct list_head *list, | |
2469 | bool run_queue_async) | |
320ae51f | 2470 | { |
3f0cedc7 | 2471 | struct request *rq; |
c16d6b5a | 2472 | enum hctx_type type = hctx->type; |
3f0cedc7 | 2473 | |
94aa228c CH |
2474 | /* |
2475 | * Try to issue requests directly if the hw queue isn't busy to save an | |
2476 | * extra enqueue & dequeue to the sw queue. | |
2477 | */ | |
2478 | if (!hctx->dispatch_busy && !run_queue_async) { | |
2479 | blk_mq_run_dispatch_ops(hctx->queue, | |
2480 | blk_mq_try_issue_list_directly(hctx, list)); | |
2481 | if (list_empty(list)) | |
2482 | goto out; | |
2483 | } | |
2484 | ||
320ae51f JA |
2485 | /* |
2486 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
2487 | * offline now | |
2488 | */ | |
3f0cedc7 | 2489 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 2490 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 2491 | trace_block_rq_insert(rq); |
65a558f6 BVA |
2492 | if (rq->cmd_flags & REQ_NOWAIT) |
2493 | run_queue_async = true; | |
320ae51f | 2494 | } |
3f0cedc7 ML |
2495 | |
2496 | spin_lock(&ctx->lock); | |
c16d6b5a | 2497 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 2498 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 2499 | spin_unlock(&ctx->lock); |
94aa228c CH |
2500 | out: |
2501 | blk_mq_run_hw_queue(hctx, run_queue_async); | |
320ae51f JA |
2502 | } |
2503 | ||
710fa378 | 2504 | static void blk_mq_insert_request(struct request *rq, blk_insert_t flags) |
2bd215df CH |
2505 | { |
2506 | struct request_queue *q = rq->q; | |
2bd215df CH |
2507 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
2508 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2509 | ||
53548d2a CH |
2510 | if (blk_rq_is_passthrough(rq)) { |
2511 | /* | |
2512 | * Passthrough request have to be added to hctx->dispatch | |
2513 | * directly. The device may be in a situation where it can't | |
2514 | * handle FS request, and always returns BLK_STS_RESOURCE for | |
2515 | * them, which gets them added to hctx->dispatch. | |
2516 | * | |
2517 | * If a passthrough request is required to unblock the queues, | |
2518 | * and it is added to the scheduler queue, there is no chance to | |
2519 | * dispatch it given we prioritize requests in hctx->dispatch. | |
2520 | */ | |
2b597613 | 2521 | blk_mq_request_bypass_insert(rq, flags); |
be4c4278 | 2522 | } else if (req_op(rq) == REQ_OP_FLUSH) { |
2bd215df CH |
2523 | /* |
2524 | * Firstly normal IO request is inserted to scheduler queue or | |
2525 | * sw queue, meantime we add flush request to dispatch queue( | |
2526 | * hctx->dispatch) directly and there is at most one in-flight | |
2527 | * flush request for each hw queue, so it doesn't matter to add | |
2528 | * flush request to tail or front of the dispatch queue. | |
2529 | * | |
2530 | * Secondly in case of NCQ, flush request belongs to non-NCQ | |
2531 | * command, and queueing it will fail when there is any | |
2532 | * in-flight normal IO request(NCQ command). When adding flush | |
2533 | * rq to the front of hctx->dispatch, it is easier to introduce | |
2534 | * extra time to flush rq's latency because of S_SCHED_RESTART | |
2535 | * compared with adding to the tail of dispatch queue, then | |
2536 | * chance of flush merge is increased, and less flush requests | |
2537 | * will be issued to controller. It is observed that ~10% time | |
2538 | * is saved in blktests block/004 on disk attached to AHCI/NCQ | |
2539 | * drive when adding flush rq to the front of hctx->dispatch. | |
2540 | * | |
2541 | * Simply queue flush rq to the front of hctx->dispatch so that | |
2542 | * intensive flush workloads can benefit in case of NCQ HW. | |
2543 | */ | |
2b597613 | 2544 | blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD); |
53548d2a | 2545 | } else if (q->elevator) { |
2bd215df CH |
2546 | LIST_HEAD(list); |
2547 | ||
53548d2a CH |
2548 | WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG); |
2549 | ||
2bd215df | 2550 | list_add(&rq->queuelist, &list); |
93fffe16 | 2551 | q->elevator->type->ops.insert_requests(hctx, &list, flags); |
2bd215df | 2552 | } else { |
4ec5c055 CH |
2553 | trace_block_rq_insert(rq); |
2554 | ||
2bd215df | 2555 | spin_lock(&ctx->lock); |
710fa378 | 2556 | if (flags & BLK_MQ_INSERT_AT_HEAD) |
4ec5c055 CH |
2557 | list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]); |
2558 | else | |
2559 | list_add_tail(&rq->queuelist, | |
2560 | &ctx->rq_lists[hctx->type]); | |
a88db1e0 | 2561 | blk_mq_hctx_mark_pending(hctx, ctx); |
2bd215df CH |
2562 | spin_unlock(&ctx->lock); |
2563 | } | |
320ae51f JA |
2564 | } |
2565 | ||
14ccb66b CH |
2566 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
2567 | unsigned int nr_segs) | |
320ae51f | 2568 | { |
93f221ae EB |
2569 | int err; |
2570 | ||
f924cdde CH |
2571 | if (bio->bi_opf & REQ_RAHEAD) |
2572 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
2573 | ||
2574 | rq->__sector = bio->bi_iter.bi_sector; | |
44981351 | 2575 | rq->write_hint = bio->bi_write_hint; |
14ccb66b | 2576 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
2577 | |
2578 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
2579 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
2580 | WARN_ON_ONCE(err); | |
4b570521 | 2581 | |
b5af37ab | 2582 | blk_account_io_start(rq); |
320ae51f JA |
2583 | } |
2584 | ||
0f95549c | 2585 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2586 | struct request *rq, bool last) |
f984df1f | 2587 | { |
f984df1f | 2588 | struct request_queue *q = rq->q; |
f984df1f SL |
2589 | struct blk_mq_queue_data bd = { |
2590 | .rq = rq, | |
be94f058 | 2591 | .last = last, |
f984df1f | 2592 | }; |
f06345ad | 2593 | blk_status_t ret; |
0f95549c | 2594 | |
0f95549c MS |
2595 | /* |
2596 | * For OK queue, we are done. For error, caller may kill it. | |
2597 | * Any other error (busy), just add it to our list as we | |
2598 | * previously would have done. | |
2599 | */ | |
2600 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
2601 | switch (ret) { | |
2602 | case BLK_STS_OK: | |
6ce3dd6e | 2603 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2604 | break; |
2605 | case BLK_STS_RESOURCE: | |
86ff7c2a | 2606 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 2607 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
2608 | __blk_mq_requeue_request(rq); |
2609 | break; | |
2610 | default: | |
6ce3dd6e | 2611 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2612 | break; |
2613 | } | |
2614 | ||
2615 | return ret; | |
2616 | } | |
2617 | ||
2b71b877 | 2618 | static bool blk_mq_get_budget_and_tag(struct request *rq) |
0f95549c | 2619 | { |
2a5a24aa | 2620 | int budget_token; |
d964f04a | 2621 | |
2b71b877 | 2622 | budget_token = blk_mq_get_dispatch_budget(rq->q); |
2a5a24aa | 2623 | if (budget_token < 0) |
2b71b877 | 2624 | return false; |
2a5a24aa | 2625 | blk_mq_set_rq_budget_token(rq, budget_token); |
8ab6bb9e | 2626 | if (!blk_mq_get_driver_tag(rq)) { |
2b71b877 CH |
2627 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
2628 | return false; | |
88022d72 | 2629 | } |
2b71b877 | 2630 | return true; |
fd9c40f6 BVA |
2631 | } |
2632 | ||
105663f7 AA |
2633 | /** |
2634 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2635 | * @hctx: Pointer of the associated hardware queue. | |
2636 | * @rq: Pointer to request to be sent. | |
105663f7 AA |
2637 | * |
2638 | * If the device has enough resources to accept a new request now, send the | |
2639 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2640 | * we can try send it another time in the future. Requests inserted at this | |
2641 | * queue have higher priority. | |
2642 | */ | |
fd9c40f6 | 2643 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2644 | struct request *rq) |
fd9c40f6 | 2645 | { |
e1f44ac0 CH |
2646 | blk_status_t ret; |
2647 | ||
2648 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { | |
710fa378 | 2649 | blk_mq_insert_request(rq, 0); |
e1f44ac0 CH |
2650 | return; |
2651 | } | |
2652 | ||
dd6216bb | 2653 | if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) { |
710fa378 | 2654 | blk_mq_insert_request(rq, 0); |
65a558f6 | 2655 | blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT); |
e1f44ac0 CH |
2656 | return; |
2657 | } | |
fd9c40f6 | 2658 | |
e1f44ac0 CH |
2659 | ret = __blk_mq_issue_directly(hctx, rq, true); |
2660 | switch (ret) { | |
2661 | case BLK_STS_OK: | |
2662 | break; | |
2663 | case BLK_STS_RESOURCE: | |
2664 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2665 | blk_mq_request_bypass_insert(rq, 0); |
2394395c | 2666 | blk_mq_run_hw_queue(hctx, false); |
e1f44ac0 CH |
2667 | break; |
2668 | default: | |
fd9c40f6 | 2669 | blk_mq_end_request(rq, ret); |
e1f44ac0 CH |
2670 | break; |
2671 | } | |
fd9c40f6 BVA |
2672 | } |
2673 | ||
06c8c691 | 2674 | static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) |
fd9c40f6 | 2675 | { |
e1f44ac0 CH |
2676 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
2677 | ||
2678 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { | |
710fa378 | 2679 | blk_mq_insert_request(rq, 0); |
e1f44ac0 CH |
2680 | return BLK_STS_OK; |
2681 | } | |
2682 | ||
2683 | if (!blk_mq_get_budget_and_tag(rq)) | |
2684 | return BLK_STS_RESOURCE; | |
2685 | return __blk_mq_issue_directly(hctx, rq, last); | |
5eb6126e CH |
2686 | } |
2687 | ||
3e368fb0 | 2688 | static void blk_mq_plug_issue_direct(struct blk_plug *plug) |
b84c5b50 CH |
2689 | { |
2690 | struct blk_mq_hw_ctx *hctx = NULL; | |
2691 | struct request *rq; | |
2692 | int queued = 0; | |
0d617a83 | 2693 | blk_status_t ret = BLK_STS_OK; |
b84c5b50 CH |
2694 | |
2695 | while ((rq = rq_list_pop(&plug->mq_list))) { | |
2696 | bool last = rq_list_empty(plug->mq_list); | |
b84c5b50 CH |
2697 | |
2698 | if (hctx != rq->mq_hctx) { | |
34c9f547 KS |
2699 | if (hctx) { |
2700 | blk_mq_commit_rqs(hctx, queued, false); | |
2701 | queued = 0; | |
2702 | } | |
b84c5b50 CH |
2703 | hctx = rq->mq_hctx; |
2704 | } | |
2705 | ||
2706 | ret = blk_mq_request_issue_directly(rq, last); | |
2707 | switch (ret) { | |
2708 | case BLK_STS_OK: | |
2709 | queued++; | |
2710 | break; | |
2711 | case BLK_STS_RESOURCE: | |
2712 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2713 | blk_mq_request_bypass_insert(rq, 0); |
2394395c | 2714 | blk_mq_run_hw_queue(hctx, false); |
0d617a83 | 2715 | goto out; |
b84c5b50 CH |
2716 | default: |
2717 | blk_mq_end_request(rq, ret); | |
b84c5b50 CH |
2718 | break; |
2719 | } | |
2720 | } | |
2721 | ||
0d617a83 KS |
2722 | out: |
2723 | if (ret != BLK_STS_OK) | |
34c9f547 | 2724 | blk_mq_commit_rqs(hctx, queued, false); |
b84c5b50 CH |
2725 | } |
2726 | ||
518579a9 KB |
2727 | static void __blk_mq_flush_plug_list(struct request_queue *q, |
2728 | struct blk_plug *plug) | |
2729 | { | |
2730 | if (blk_queue_quiesced(q)) | |
2731 | return; | |
2732 | q->mq_ops->queue_rqs(&plug->mq_list); | |
2733 | } | |
2734 | ||
26fed4ac JA |
2735 | static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched) |
2736 | { | |
2737 | struct blk_mq_hw_ctx *this_hctx = NULL; | |
2738 | struct blk_mq_ctx *this_ctx = NULL; | |
2739 | struct request *requeue_list = NULL; | |
34e0a279 | 2740 | struct request **requeue_lastp = &requeue_list; |
26fed4ac | 2741 | unsigned int depth = 0; |
d97217e7 | 2742 | bool is_passthrough = false; |
26fed4ac JA |
2743 | LIST_HEAD(list); |
2744 | ||
2745 | do { | |
2746 | struct request *rq = rq_list_pop(&plug->mq_list); | |
2747 | ||
2748 | if (!this_hctx) { | |
2749 | this_hctx = rq->mq_hctx; | |
2750 | this_ctx = rq->mq_ctx; | |
d97217e7 ML |
2751 | is_passthrough = blk_rq_is_passthrough(rq); |
2752 | } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx || | |
2753 | is_passthrough != blk_rq_is_passthrough(rq)) { | |
34e0a279 | 2754 | rq_list_add_tail(&requeue_lastp, rq); |
26fed4ac JA |
2755 | continue; |
2756 | } | |
34e0a279 | 2757 | list_add(&rq->queuelist, &list); |
26fed4ac JA |
2758 | depth++; |
2759 | } while (!rq_list_empty(plug->mq_list)); | |
2760 | ||
2761 | plug->mq_list = requeue_list; | |
2762 | trace_block_unplug(this_hctx->queue, depth, !from_sched); | |
05a93117 CH |
2763 | |
2764 | percpu_ref_get(&this_hctx->queue->q_usage_counter); | |
d97217e7 | 2765 | /* passthrough requests should never be issued to the I/O scheduler */ |
2293cae7 ML |
2766 | if (is_passthrough) { |
2767 | spin_lock(&this_hctx->lock); | |
2768 | list_splice_tail_init(&list, &this_hctx->dispatch); | |
2769 | spin_unlock(&this_hctx->lock); | |
2770 | blk_mq_run_hw_queue(this_hctx, from_sched); | |
2771 | } else if (this_hctx->queue->elevator) { | |
05a93117 | 2772 | this_hctx->queue->elevator->type->ops.insert_requests(this_hctx, |
93fffe16 | 2773 | &list, 0); |
05a93117 CH |
2774 | blk_mq_run_hw_queue(this_hctx, from_sched); |
2775 | } else { | |
2776 | blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched); | |
2777 | } | |
2778 | percpu_ref_put(&this_hctx->queue->q_usage_counter); | |
26fed4ac JA |
2779 | } |
2780 | ||
b84c5b50 CH |
2781 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
2782 | { | |
3c67d44d | 2783 | struct request *rq; |
b84c5b50 | 2784 | |
70904263 RL |
2785 | /* |
2786 | * We may have been called recursively midway through handling | |
2787 | * plug->mq_list via a schedule() in the driver's queue_rq() callback. | |
2788 | * To avoid mq_list changing under our feet, clear rq_count early and | |
2789 | * bail out specifically if rq_count is 0 rather than checking | |
2790 | * whether the mq_list is empty. | |
2791 | */ | |
2792 | if (plug->rq_count == 0) | |
b84c5b50 CH |
2793 | return; |
2794 | plug->rq_count = 0; | |
2795 | ||
2796 | if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) { | |
3c67d44d JA |
2797 | struct request_queue *q; |
2798 | ||
2799 | rq = rq_list_peek(&plug->mq_list); | |
2800 | q = rq->q; | |
2801 | ||
2802 | /* | |
2803 | * Peek first request and see if we have a ->queue_rqs() hook. | |
2804 | * If we do, we can dispatch the whole plug list in one go. We | |
2805 | * already know at this point that all requests belong to the | |
2806 | * same queue, caller must ensure that's the case. | |
3c67d44d | 2807 | */ |
434097ee | 2808 | if (q->mq_ops->queue_rqs) { |
3c67d44d | 2809 | blk_mq_run_dispatch_ops(q, |
518579a9 | 2810 | __blk_mq_flush_plug_list(q, plug)); |
3c67d44d JA |
2811 | if (rq_list_empty(plug->mq_list)) |
2812 | return; | |
2813 | } | |
73f3760e ML |
2814 | |
2815 | blk_mq_run_dispatch_ops(q, | |
3e368fb0 | 2816 | blk_mq_plug_issue_direct(plug)); |
b84c5b50 CH |
2817 | if (rq_list_empty(plug->mq_list)) |
2818 | return; | |
2819 | } | |
2820 | ||
b84c5b50 | 2821 | do { |
26fed4ac | 2822 | blk_mq_dispatch_plug_list(plug, from_schedule); |
b84c5b50 | 2823 | } while (!rq_list_empty(plug->mq_list)); |
b84c5b50 CH |
2824 | } |
2825 | ||
94aa228c | 2826 | static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
6ce3dd6e ML |
2827 | struct list_head *list) |
2828 | { | |
536167d4 | 2829 | int queued = 0; |
984ce0a7 | 2830 | blk_status_t ret = BLK_STS_OK; |
536167d4 | 2831 | |
6ce3dd6e | 2832 | while (!list_empty(list)) { |
6ce3dd6e ML |
2833 | struct request *rq = list_first_entry(list, struct request, |
2834 | queuelist); | |
2835 | ||
2836 | list_del_init(&rq->queuelist); | |
fd9c40f6 | 2837 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
27e8b2bb KS |
2838 | switch (ret) { |
2839 | case BLK_STS_OK: | |
536167d4 | 2840 | queued++; |
27e8b2bb KS |
2841 | break; |
2842 | case BLK_STS_RESOURCE: | |
2843 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2844 | blk_mq_request_bypass_insert(rq, 0); |
2394395c CH |
2845 | if (list_empty(list)) |
2846 | blk_mq_run_hw_queue(hctx, false); | |
27e8b2bb KS |
2847 | goto out; |
2848 | default: | |
2849 | blk_mq_end_request(rq, ret); | |
2850 | break; | |
2851 | } | |
6ce3dd6e | 2852 | } |
d666ba98 | 2853 | |
27e8b2bb | 2854 | out: |
984ce0a7 KS |
2855 | if (ret != BLK_STS_OK) |
2856 | blk_mq_commit_rqs(hctx, queued, false); | |
6ce3dd6e ML |
2857 | } |
2858 | ||
b131f201 | 2859 | static bool blk_mq_attempt_bio_merge(struct request_queue *q, |
0c5bcc92 | 2860 | struct bio *bio, unsigned int nr_segs) |
900e0807 JA |
2861 | { |
2862 | if (!blk_queue_nomerges(q) && bio_mergeable(bio)) { | |
0c5bcc92 | 2863 | if (blk_attempt_plug_merge(q, bio, nr_segs)) |
900e0807 JA |
2864 | return true; |
2865 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) | |
2866 | return true; | |
2867 | } | |
2868 | return false; | |
2869 | } | |
2870 | ||
71539717 JA |
2871 | static struct request *blk_mq_get_new_requests(struct request_queue *q, |
2872 | struct blk_plug *plug, | |
0a5aa8d1 SK |
2873 | struct bio *bio, |
2874 | unsigned int nsegs) | |
71539717 JA |
2875 | { |
2876 | struct blk_mq_alloc_data data = { | |
2877 | .q = q, | |
2878 | .nr_tags = 1, | |
9d497e29 | 2879 | .cmd_flags = bio->bi_opf, |
71539717 JA |
2880 | }; |
2881 | struct request *rq; | |
2882 | ||
0a5aa8d1 SK |
2883 | rq_qos_throttle(q, bio); |
2884 | ||
71539717 JA |
2885 | if (plug) { |
2886 | data.nr_tags = plug->nr_ios; | |
2887 | plug->nr_ios = 1; | |
2888 | data.cached_rq = &plug->cached_rq; | |
2889 | } | |
2890 | ||
2891 | rq = __blk_mq_alloc_requests(&data); | |
373b5416 JA |
2892 | if (rq) |
2893 | return rq; | |
71539717 JA |
2894 | rq_qos_cleanup(q, bio); |
2895 | if (bio->bi_opf & REQ_NOWAIT) | |
2896 | bio_wouldblock_error(bio); | |
2897 | return NULL; | |
2898 | } | |
2899 | ||
309ce674 | 2900 | /* |
337e89fe | 2901 | * Check if there is a suitable cached request and return it. |
309ce674 | 2902 | */ |
337e89fe CH |
2903 | static struct request *blk_mq_peek_cached_request(struct blk_plug *plug, |
2904 | struct request_queue *q, blk_opf_t opf) | |
71539717 | 2905 | { |
337e89fe CH |
2906 | enum hctx_type type = blk_mq_get_hctx_type(opf); |
2907 | struct request *rq; | |
71539717 | 2908 | |
337e89fe CH |
2909 | if (!plug) |
2910 | return NULL; | |
2911 | rq = rq_list_peek(&plug->cached_rq); | |
2912 | if (!rq || rq->q != q) | |
2913 | return NULL; | |
2914 | if (type != rq->mq_hctx->type && | |
2915 | (type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT)) | |
2916 | return NULL; | |
2917 | if (op_is_flush(rq->cmd_flags) != op_is_flush(opf)) | |
2918 | return NULL; | |
2919 | return rq; | |
2920 | } | |
0a5aa8d1 | 2921 | |
337e89fe CH |
2922 | static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug, |
2923 | struct bio *bio) | |
2924 | { | |
2925 | WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq); | |
5b13bc8a | 2926 | |
2645672f JA |
2927 | /* |
2928 | * If any qos ->throttle() end up blocking, we will have flushed the | |
2929 | * plug and hence killed the cached_rq list as well. Pop this entry | |
2930 | * before we throttle. | |
2931 | */ | |
5b13bc8a | 2932 | plug->cached_rq = rq_list_next(rq); |
b0077e26 | 2933 | rq_qos_throttle(rq->q, bio); |
2645672f | 2934 | |
5c17f45e | 2935 | blk_mq_rq_time_init(rq, 0); |
b0077e26 | 2936 | rq->cmd_flags = bio->bi_opf; |
5b13bc8a | 2937 | INIT_LIST_HEAD(&rq->queuelist); |
71539717 JA |
2938 | } |
2939 | ||
105663f7 | 2940 | /** |
c62b37d9 | 2941 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2942 | * @bio: Bio pointer. |
2943 | * | |
2944 | * Builds up a request structure from @q and @bio and send to the device. The | |
2945 | * request may not be queued directly to hardware if: | |
2946 | * * This request can be merged with another one | |
2947 | * * We want to place request at plug queue for possible future merging | |
2948 | * * There is an IO scheduler active at this queue | |
2949 | * | |
2950 | * It will not queue the request if there is an error with the bio, or at the | |
2951 | * request creation. | |
105663f7 | 2952 | */ |
3e08773c | 2953 | void blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2954 | { |
ed6cddef | 2955 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
6deacb3b | 2956 | struct blk_plug *plug = blk_mq_plug(bio); |
ef295ecf | 2957 | const int is_sync = op_is_sync(bio->bi_opf); |
f0dbe6e8 | 2958 | struct blk_mq_hw_ctx *hctx; |
abd45c15 | 2959 | unsigned int nr_segs = 1; |
72e84e90 | 2960 | struct request *rq; |
a892c8d5 | 2961 | blk_status_t ret; |
07068d5b | 2962 | |
51d798cd | 2963 | bio = blk_queue_bounce(bio, q); |
9c6227e0 | 2964 | |
72e84e90 CH |
2965 | /* |
2966 | * If the plug has a cached request for this queue, try use it. | |
2967 | * | |
2968 | * The cached request already holds a q_usage_counter reference and we | |
2969 | * don't have to acquire a new one if we use it. | |
2970 | */ | |
337e89fe | 2971 | rq = blk_mq_peek_cached_request(plug, q, bio->bi_opf); |
72e84e90 | 2972 | if (!rq) { |
b0077e26 | 2973 | if (unlikely(bio_queue_enter(bio))) |
5b13bc8a | 2974 | return; |
b0077e26 CH |
2975 | } |
2976 | ||
337e89fe CH |
2977 | if (unlikely(bio_may_exceed_limits(bio, &q->limits))) { |
2978 | bio = __bio_split_to_limits(bio, &q->limits, &nr_segs); | |
2979 | if (!bio) | |
0f299da5 | 2980 | goto queue_exit; |
b0077e26 | 2981 | } |
337e89fe CH |
2982 | if (!bio_integrity_prep(bio)) |
2983 | goto queue_exit; | |
b0077e26 | 2984 | |
0f299da5 CH |
2985 | if (blk_mq_attempt_bio_merge(q, bio, nr_segs)) |
2986 | goto queue_exit; | |
2987 | ||
72e84e90 CH |
2988 | if (!rq) { |
2989 | rq = blk_mq_get_new_requests(q, plug, bio, nr_segs); | |
2990 | if (unlikely(!rq)) | |
2991 | goto queue_exit; | |
2992 | } else { | |
2993 | blk_mq_use_cached_rq(rq, plug, bio); | |
5b13bc8a | 2994 | } |
87760e5e | 2995 | |
e8a676d6 | 2996 | trace_block_getrq(bio); |
d6f1dda2 | 2997 | |
c1c80384 | 2998 | rq_qos_track(q, rq, bio); |
07068d5b | 2999 | |
970d168d BVA |
3000 | blk_mq_bio_to_request(rq, bio, nr_segs); |
3001 | ||
9cd1e566 | 3002 | ret = blk_crypto_rq_get_keyslot(rq); |
a892c8d5 ST |
3003 | if (ret != BLK_STS_OK) { |
3004 | bio->bi_status = ret; | |
3005 | bio_endio(bio); | |
3006 | blk_mq_free_request(rq); | |
3e08773c | 3007 | return; |
a892c8d5 ST |
3008 | } |
3009 | ||
360f2648 | 3010 | if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq)) |
d92ca9d8 CH |
3011 | return; |
3012 | ||
f0dbe6e8 | 3013 | if (plug) { |
ce5b009c | 3014 | blk_add_rq_to_plug(plug, rq); |
f0dbe6e8 CH |
3015 | return; |
3016 | } | |
3017 | ||
3018 | hctx = rq->mq_hctx; | |
dd6216bb | 3019 | if ((rq->rq_flags & RQF_USE_SCHED) || |
f0dbe6e8 | 3020 | (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) { |
710fa378 | 3021 | blk_mq_insert_request(rq, 0); |
f0dbe6e8 CH |
3022 | blk_mq_run_hw_queue(hctx, true); |
3023 | } else { | |
3024 | blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq)); | |
3025 | } | |
0f299da5 CH |
3026 | return; |
3027 | ||
3028 | queue_exit: | |
72e84e90 CH |
3029 | /* |
3030 | * Don't drop the queue reference if we were trying to use a cached | |
3031 | * request and thus didn't acquire one. | |
3032 | */ | |
3033 | if (!rq) | |
3034 | blk_queue_exit(q); | |
320ae51f JA |
3035 | } |
3036 | ||
248c7933 | 3037 | #ifdef CONFIG_BLK_MQ_STACKING |
06c8c691 | 3038 | /** |
a5efda3c | 3039 | * blk_insert_cloned_request - Helper for stacking drivers to submit a request |
a5efda3c | 3040 | * @rq: the request being queued |
06c8c691 | 3041 | */ |
28db4711 | 3042 | blk_status_t blk_insert_cloned_request(struct request *rq) |
06c8c691 | 3043 | { |
28db4711 | 3044 | struct request_queue *q = rq->q; |
06c8c691 | 3045 | unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); |
49d24398 | 3046 | unsigned int max_segments = blk_rq_get_max_segments(rq); |
a5efda3c | 3047 | blk_status_t ret; |
06c8c691 CH |
3048 | |
3049 | if (blk_rq_sectors(rq) > max_sectors) { | |
3050 | /* | |
3051 | * SCSI device does not have a good way to return if | |
3052 | * Write Same/Zero is actually supported. If a device rejects | |
3053 | * a non-read/write command (discard, write same,etc.) the | |
3054 | * low-level device driver will set the relevant queue limit to | |
3055 | * 0 to prevent blk-lib from issuing more of the offending | |
3056 | * operations. Commands queued prior to the queue limit being | |
3057 | * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O | |
3058 | * errors being propagated to upper layers. | |
3059 | */ | |
3060 | if (max_sectors == 0) | |
3061 | return BLK_STS_NOTSUPP; | |
3062 | ||
3063 | printk(KERN_ERR "%s: over max size limit. (%u > %u)\n", | |
3064 | __func__, blk_rq_sectors(rq), max_sectors); | |
3065 | return BLK_STS_IOERR; | |
3066 | } | |
3067 | ||
3068 | /* | |
3069 | * The queue settings related to segment counting may differ from the | |
3070 | * original queue. | |
3071 | */ | |
3072 | rq->nr_phys_segments = blk_recalc_rq_segments(rq); | |
49d24398 US |
3073 | if (rq->nr_phys_segments > max_segments) { |
3074 | printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n", | |
3075 | __func__, rq->nr_phys_segments, max_segments); | |
06c8c691 CH |
3076 | return BLK_STS_IOERR; |
3077 | } | |
3078 | ||
28db4711 | 3079 | if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq))) |
06c8c691 CH |
3080 | return BLK_STS_IOERR; |
3081 | ||
5b8562f0 EB |
3082 | ret = blk_crypto_rq_get_keyslot(rq); |
3083 | if (ret != BLK_STS_OK) | |
3084 | return ret; | |
06c8c691 CH |
3085 | |
3086 | blk_account_io_start(rq); | |
3087 | ||
3088 | /* | |
3089 | * Since we have a scheduler attached on the top device, | |
3090 | * bypass a potential scheduler on the bottom device for | |
3091 | * insert. | |
3092 | */ | |
28db4711 | 3093 | blk_mq_run_dispatch_ops(q, |
4cafe86c | 3094 | ret = blk_mq_request_issue_directly(rq, true)); |
592ee119 | 3095 | if (ret) |
08420cf7 | 3096 | blk_account_io_done(rq, blk_time_get_ns()); |
4cafe86c | 3097 | return ret; |
06c8c691 CH |
3098 | } |
3099 | EXPORT_SYMBOL_GPL(blk_insert_cloned_request); | |
3100 | ||
3101 | /** | |
3102 | * blk_rq_unprep_clone - Helper function to free all bios in a cloned request | |
3103 | * @rq: the clone request to be cleaned up | |
3104 | * | |
3105 | * Description: | |
3106 | * Free all bios in @rq for a cloned request. | |
3107 | */ | |
3108 | void blk_rq_unprep_clone(struct request *rq) | |
3109 | { | |
3110 | struct bio *bio; | |
3111 | ||
3112 | while ((bio = rq->bio) != NULL) { | |
3113 | rq->bio = bio->bi_next; | |
3114 | ||
3115 | bio_put(bio); | |
3116 | } | |
3117 | } | |
3118 | EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); | |
3119 | ||
3120 | /** | |
3121 | * blk_rq_prep_clone - Helper function to setup clone request | |
3122 | * @rq: the request to be setup | |
3123 | * @rq_src: original request to be cloned | |
3124 | * @bs: bio_set that bios for clone are allocated from | |
3125 | * @gfp_mask: memory allocation mask for bio | |
3126 | * @bio_ctr: setup function to be called for each clone bio. | |
3127 | * Returns %0 for success, non %0 for failure. | |
3128 | * @data: private data to be passed to @bio_ctr | |
3129 | * | |
3130 | * Description: | |
3131 | * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. | |
3132 | * Also, pages which the original bios are pointing to are not copied | |
3133 | * and the cloned bios just point same pages. | |
3134 | * So cloned bios must be completed before original bios, which means | |
3135 | * the caller must complete @rq before @rq_src. | |
3136 | */ | |
3137 | int blk_rq_prep_clone(struct request *rq, struct request *rq_src, | |
3138 | struct bio_set *bs, gfp_t gfp_mask, | |
3139 | int (*bio_ctr)(struct bio *, struct bio *, void *), | |
3140 | void *data) | |
3141 | { | |
3142 | struct bio *bio, *bio_src; | |
3143 | ||
3144 | if (!bs) | |
3145 | bs = &fs_bio_set; | |
3146 | ||
3147 | __rq_for_each_bio(bio_src, rq_src) { | |
abfc426d CH |
3148 | bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask, |
3149 | bs); | |
06c8c691 CH |
3150 | if (!bio) |
3151 | goto free_and_out; | |
3152 | ||
3153 | if (bio_ctr && bio_ctr(bio, bio_src, data)) | |
3154 | goto free_and_out; | |
3155 | ||
3156 | if (rq->bio) { | |
3157 | rq->biotail->bi_next = bio; | |
3158 | rq->biotail = bio; | |
3159 | } else { | |
3160 | rq->bio = rq->biotail = bio; | |
3161 | } | |
3162 | bio = NULL; | |
3163 | } | |
3164 | ||
3165 | /* Copy attributes of the original request to the clone request. */ | |
3166 | rq->__sector = blk_rq_pos(rq_src); | |
3167 | rq->__data_len = blk_rq_bytes(rq_src); | |
3168 | if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) { | |
3169 | rq->rq_flags |= RQF_SPECIAL_PAYLOAD; | |
3170 | rq->special_vec = rq_src->special_vec; | |
3171 | } | |
3172 | rq->nr_phys_segments = rq_src->nr_phys_segments; | |
3173 | rq->ioprio = rq_src->ioprio; | |
44981351 | 3174 | rq->write_hint = rq_src->write_hint; |
06c8c691 CH |
3175 | |
3176 | if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0) | |
3177 | goto free_and_out; | |
3178 | ||
3179 | return 0; | |
3180 | ||
3181 | free_and_out: | |
3182 | if (bio) | |
3183 | bio_put(bio); | |
3184 | blk_rq_unprep_clone(rq); | |
3185 | ||
3186 | return -ENOMEM; | |
3187 | } | |
3188 | EXPORT_SYMBOL_GPL(blk_rq_prep_clone); | |
248c7933 | 3189 | #endif /* CONFIG_BLK_MQ_STACKING */ |
06c8c691 | 3190 | |
f2b8f3ce CH |
3191 | /* |
3192 | * Steal bios from a request and add them to a bio list. | |
3193 | * The request must not have been partially completed before. | |
3194 | */ | |
3195 | void blk_steal_bios(struct bio_list *list, struct request *rq) | |
3196 | { | |
3197 | if (rq->bio) { | |
3198 | if (list->tail) | |
3199 | list->tail->bi_next = rq->bio; | |
3200 | else | |
3201 | list->head = rq->bio; | |
3202 | list->tail = rq->biotail; | |
3203 | ||
3204 | rq->bio = NULL; | |
3205 | rq->biotail = NULL; | |
3206 | } | |
3207 | ||
3208 | rq->__data_len = 0; | |
3209 | } | |
3210 | EXPORT_SYMBOL_GPL(blk_steal_bios); | |
3211 | ||
bd63141d ML |
3212 | static size_t order_to_size(unsigned int order) |
3213 | { | |
3214 | return (size_t)PAGE_SIZE << order; | |
3215 | } | |
3216 | ||
3217 | /* called before freeing request pool in @tags */ | |
f32e4eaf JG |
3218 | static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, |
3219 | struct blk_mq_tags *tags) | |
bd63141d | 3220 | { |
bd63141d ML |
3221 | struct page *page; |
3222 | unsigned long flags; | |
3223 | ||
76dd2980 YK |
3224 | /* |
3225 | * There is no need to clear mapping if driver tags is not initialized | |
3226 | * or the mapping belongs to the driver tags. | |
3227 | */ | |
3228 | if (!drv_tags || drv_tags == tags) | |
4f245d5b JG |
3229 | return; |
3230 | ||
bd63141d ML |
3231 | list_for_each_entry(page, &tags->page_list, lru) { |
3232 | unsigned long start = (unsigned long)page_address(page); | |
3233 | unsigned long end = start + order_to_size(page->private); | |
3234 | int i; | |
3235 | ||
f32e4eaf | 3236 | for (i = 0; i < drv_tags->nr_tags; i++) { |
bd63141d ML |
3237 | struct request *rq = drv_tags->rqs[i]; |
3238 | unsigned long rq_addr = (unsigned long)rq; | |
3239 | ||
3240 | if (rq_addr >= start && rq_addr < end) { | |
0a467d0f | 3241 | WARN_ON_ONCE(req_ref_read(rq) != 0); |
bd63141d ML |
3242 | cmpxchg(&drv_tags->rqs[i], rq, NULL); |
3243 | } | |
3244 | } | |
3245 | } | |
3246 | ||
3247 | /* | |
3248 | * Wait until all pending iteration is done. | |
3249 | * | |
3250 | * Request reference is cleared and it is guaranteed to be observed | |
3251 | * after the ->lock is released. | |
3252 | */ | |
3253 | spin_lock_irqsave(&drv_tags->lock, flags); | |
3254 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
3255 | } | |
3256 | ||
cc71a6f4 JA |
3257 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
3258 | unsigned int hctx_idx) | |
95363efd | 3259 | { |
f32e4eaf | 3260 | struct blk_mq_tags *drv_tags; |
e9b267d9 | 3261 | struct page *page; |
320ae51f | 3262 | |
e02657ea ML |
3263 | if (list_empty(&tags->page_list)) |
3264 | return; | |
3265 | ||
079a2e3e JG |
3266 | if (blk_mq_is_shared_tags(set->flags)) |
3267 | drv_tags = set->shared_tags; | |
e155b0c2 JG |
3268 | else |
3269 | drv_tags = set->tags[hctx_idx]; | |
f32e4eaf | 3270 | |
65de57bb | 3271 | if (tags->static_rqs && set->ops->exit_request) { |
e9b267d9 | 3272 | int i; |
320ae51f | 3273 | |
24d2f903 | 3274 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
3275 | struct request *rq = tags->static_rqs[i]; |
3276 | ||
3277 | if (!rq) | |
e9b267d9 | 3278 | continue; |
d6296d39 | 3279 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 3280 | tags->static_rqs[i] = NULL; |
e9b267d9 | 3281 | } |
320ae51f | 3282 | } |
320ae51f | 3283 | |
f32e4eaf | 3284 | blk_mq_clear_rq_mapping(drv_tags, tags); |
bd63141d | 3285 | |
24d2f903 CH |
3286 | while (!list_empty(&tags->page_list)) { |
3287 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 3288 | list_del_init(&page->lru); |
f75782e4 CM |
3289 | /* |
3290 | * Remove kmemleak object previously allocated in | |
273938bf | 3291 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
3292 | */ |
3293 | kmemleak_free(page_address(page)); | |
320ae51f JA |
3294 | __free_pages(page, page->private); |
3295 | } | |
cc71a6f4 | 3296 | } |
320ae51f | 3297 | |
e155b0c2 | 3298 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
cc71a6f4 | 3299 | { |
24d2f903 | 3300 | kfree(tags->rqs); |
cc71a6f4 | 3301 | tags->rqs = NULL; |
2af8cbe3 JA |
3302 | kfree(tags->static_rqs); |
3303 | tags->static_rqs = NULL; | |
320ae51f | 3304 | |
e155b0c2 | 3305 | blk_mq_free_tags(tags); |
320ae51f JA |
3306 | } |
3307 | ||
4d805131 ML |
3308 | static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set, |
3309 | unsigned int hctx_idx) | |
3310 | { | |
3311 | int i; | |
3312 | ||
3313 | for (i = 0; i < set->nr_maps; i++) { | |
3314 | unsigned int start = set->map[i].queue_offset; | |
3315 | unsigned int end = start + set->map[i].nr_queues; | |
3316 | ||
3317 | if (hctx_idx >= start && hctx_idx < end) | |
3318 | break; | |
3319 | } | |
3320 | ||
3321 | if (i >= set->nr_maps) | |
3322 | i = HCTX_TYPE_DEFAULT; | |
3323 | ||
3324 | return i; | |
3325 | } | |
3326 | ||
3327 | static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set, | |
3328 | unsigned int hctx_idx) | |
3329 | { | |
3330 | enum hctx_type type = hctx_idx_to_type(set, hctx_idx); | |
3331 | ||
3332 | return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx); | |
3333 | } | |
3334 | ||
63064be1 JG |
3335 | static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
3336 | unsigned int hctx_idx, | |
3337 | unsigned int nr_tags, | |
e155b0c2 | 3338 | unsigned int reserved_tags) |
320ae51f | 3339 | { |
4d805131 | 3340 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
24d2f903 | 3341 | struct blk_mq_tags *tags; |
320ae51f | 3342 | |
59f082e4 SL |
3343 | if (node == NUMA_NO_NODE) |
3344 | node = set->numa_node; | |
3345 | ||
e155b0c2 JG |
3346 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
3347 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
3348 | if (!tags) |
3349 | return NULL; | |
320ae51f | 3350 | |
590b5b7d | 3351 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 3352 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 3353 | node); |
7edfd681 JC |
3354 | if (!tags->rqs) |
3355 | goto err_free_tags; | |
320ae51f | 3356 | |
590b5b7d KC |
3357 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
3358 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
3359 | node); | |
7edfd681 JC |
3360 | if (!tags->static_rqs) |
3361 | goto err_free_rqs; | |
2af8cbe3 | 3362 | |
cc71a6f4 | 3363 | return tags; |
7edfd681 JC |
3364 | |
3365 | err_free_rqs: | |
3366 | kfree(tags->rqs); | |
3367 | err_free_tags: | |
3368 | blk_mq_free_tags(tags); | |
3369 | return NULL; | |
cc71a6f4 JA |
3370 | } |
3371 | ||
1d9bd516 TH |
3372 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
3373 | unsigned int hctx_idx, int node) | |
3374 | { | |
3375 | int ret; | |
3376 | ||
3377 | if (set->ops->init_request) { | |
3378 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
3379 | if (ret) | |
3380 | return ret; | |
3381 | } | |
3382 | ||
12f5b931 | 3383 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
3384 | return 0; |
3385 | } | |
3386 | ||
63064be1 JG |
3387 | static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, |
3388 | struct blk_mq_tags *tags, | |
3389 | unsigned int hctx_idx, unsigned int depth) | |
cc71a6f4 JA |
3390 | { |
3391 | unsigned int i, j, entries_per_page, max_order = 4; | |
4d805131 | 3392 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
cc71a6f4 | 3393 | size_t rq_size, left; |
59f082e4 | 3394 | |
59f082e4 SL |
3395 | if (node == NUMA_NO_NODE) |
3396 | node = set->numa_node; | |
cc71a6f4 JA |
3397 | |
3398 | INIT_LIST_HEAD(&tags->page_list); | |
3399 | ||
320ae51f JA |
3400 | /* |
3401 | * rq_size is the size of the request plus driver payload, rounded | |
3402 | * to the cacheline size | |
3403 | */ | |
24d2f903 | 3404 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 3405 | cache_line_size()); |
cc71a6f4 | 3406 | left = rq_size * depth; |
320ae51f | 3407 | |
cc71a6f4 | 3408 | for (i = 0; i < depth; ) { |
320ae51f JA |
3409 | int this_order = max_order; |
3410 | struct page *page; | |
3411 | int to_do; | |
3412 | void *p; | |
3413 | ||
b3a834b1 | 3414 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
3415 | this_order--; |
3416 | ||
3417 | do { | |
59f082e4 | 3418 | page = alloc_pages_node(node, |
36e1f3d1 | 3419 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 3420 | this_order); |
320ae51f JA |
3421 | if (page) |
3422 | break; | |
3423 | if (!this_order--) | |
3424 | break; | |
3425 | if (order_to_size(this_order) < rq_size) | |
3426 | break; | |
3427 | } while (1); | |
3428 | ||
3429 | if (!page) | |
24d2f903 | 3430 | goto fail; |
320ae51f JA |
3431 | |
3432 | page->private = this_order; | |
24d2f903 | 3433 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
3434 | |
3435 | p = page_address(page); | |
f75782e4 CM |
3436 | /* |
3437 | * Allow kmemleak to scan these pages as they contain pointers | |
3438 | * to additional allocations like via ops->init_request(). | |
3439 | */ | |
36e1f3d1 | 3440 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 3441 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 3442 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
3443 | left -= to_do * rq_size; |
3444 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
3445 | struct request *rq = p; |
3446 | ||
3447 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
3448 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
3449 | tags->static_rqs[i] = NULL; | |
3450 | goto fail; | |
e9b267d9 CH |
3451 | } |
3452 | ||
320ae51f JA |
3453 | p += rq_size; |
3454 | i++; | |
3455 | } | |
3456 | } | |
cc71a6f4 | 3457 | return 0; |
320ae51f | 3458 | |
24d2f903 | 3459 | fail: |
cc71a6f4 JA |
3460 | blk_mq_free_rqs(set, tags, hctx_idx); |
3461 | return -ENOMEM; | |
320ae51f JA |
3462 | } |
3463 | ||
bf0beec0 ML |
3464 | struct rq_iter_data { |
3465 | struct blk_mq_hw_ctx *hctx; | |
3466 | bool has_rq; | |
3467 | }; | |
3468 | ||
2dd6532e | 3469 | static bool blk_mq_has_request(struct request *rq, void *data) |
bf0beec0 ML |
3470 | { |
3471 | struct rq_iter_data *iter_data = data; | |
3472 | ||
3473 | if (rq->mq_hctx != iter_data->hctx) | |
3474 | return true; | |
3475 | iter_data->has_rq = true; | |
3476 | return false; | |
3477 | } | |
3478 | ||
3479 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
3480 | { | |
3481 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
3482 | hctx->sched_tags : hctx->tags; | |
3483 | struct rq_iter_data data = { | |
3484 | .hctx = hctx, | |
3485 | }; | |
3486 | ||
3487 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
3488 | return data.has_rq; | |
3489 | } | |
3490 | ||
3491 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
3492 | struct blk_mq_hw_ctx *hctx) | |
3493 | { | |
9b51d9d8 | 3494 | if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu) |
bf0beec0 ML |
3495 | return false; |
3496 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
3497 | return false; | |
3498 | return true; | |
3499 | } | |
3500 | ||
3501 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
3502 | { | |
3503 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3504 | struct blk_mq_hw_ctx, cpuhp_online); | |
3505 | ||
3506 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
3507 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
3508 | return 0; | |
3509 | ||
3510 | /* | |
3511 | * Prevent new request from being allocated on the current hctx. | |
3512 | * | |
3513 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
3514 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
3515 | * seen once we return from the tag allocator. | |
3516 | */ | |
3517 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3518 | smp_mb__after_atomic(); | |
3519 | ||
3520 | /* | |
3521 | * Try to grab a reference to the queue and wait for any outstanding | |
3522 | * requests. If we could not grab a reference the queue has been | |
3523 | * frozen and there are no requests. | |
3524 | */ | |
3525 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
3526 | while (blk_mq_hctx_has_requests(hctx)) | |
3527 | msleep(5); | |
3528 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
3529 | } | |
3530 | ||
3531 | return 0; | |
3532 | } | |
3533 | ||
3534 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
3535 | { | |
3536 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3537 | struct blk_mq_hw_ctx, cpuhp_online); | |
3538 | ||
3539 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
3540 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3541 | return 0; | |
3542 | } | |
3543 | ||
e57690fe JA |
3544 | /* |
3545 | * 'cpu' is going away. splice any existing rq_list entries from this | |
3546 | * software queue to the hw queue dispatch list, and ensure that it | |
3547 | * gets run. | |
3548 | */ | |
9467f859 | 3549 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 3550 | { |
9467f859 | 3551 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
3552 | struct blk_mq_ctx *ctx; |
3553 | LIST_HEAD(tmp); | |
c16d6b5a | 3554 | enum hctx_type type; |
484b4061 | 3555 | |
9467f859 | 3556 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
3557 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
3558 | return 0; | |
3559 | ||
e57690fe | 3560 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 3561 | type = hctx->type; |
484b4061 JA |
3562 | |
3563 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
3564 | if (!list_empty(&ctx->rq_lists[type])) { |
3565 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
3566 | blk_mq_hctx_clear_pending(hctx, ctx); |
3567 | } | |
3568 | spin_unlock(&ctx->lock); | |
3569 | ||
3570 | if (list_empty(&tmp)) | |
9467f859 | 3571 | return 0; |
484b4061 | 3572 | |
e57690fe JA |
3573 | spin_lock(&hctx->lock); |
3574 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
3575 | spin_unlock(&hctx->lock); | |
484b4061 JA |
3576 | |
3577 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 3578 | return 0; |
484b4061 JA |
3579 | } |
3580 | ||
9467f859 | 3581 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 3582 | { |
bf0beec0 ML |
3583 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3584 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3585 | &hctx->cpuhp_online); | |
9467f859 TG |
3586 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
3587 | &hctx->cpuhp_dead); | |
484b4061 JA |
3588 | } |
3589 | ||
364b6181 ML |
3590 | /* |
3591 | * Before freeing hw queue, clearing the flush request reference in | |
3592 | * tags->rqs[] for avoiding potential UAF. | |
3593 | */ | |
3594 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
3595 | unsigned int queue_depth, struct request *flush_rq) | |
3596 | { | |
3597 | int i; | |
3598 | unsigned long flags; | |
3599 | ||
3600 | /* The hw queue may not be mapped yet */ | |
3601 | if (!tags) | |
3602 | return; | |
3603 | ||
0a467d0f | 3604 | WARN_ON_ONCE(req_ref_read(flush_rq) != 0); |
364b6181 ML |
3605 | |
3606 | for (i = 0; i < queue_depth; i++) | |
3607 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
3608 | ||
3609 | /* | |
3610 | * Wait until all pending iteration is done. | |
3611 | * | |
3612 | * Request reference is cleared and it is guaranteed to be observed | |
3613 | * after the ->lock is released. | |
3614 | */ | |
3615 | spin_lock_irqsave(&tags->lock, flags); | |
3616 | spin_unlock_irqrestore(&tags->lock, flags); | |
3617 | } | |
3618 | ||
c3b4afca | 3619 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
3620 | static void blk_mq_exit_hctx(struct request_queue *q, |
3621 | struct blk_mq_tag_set *set, | |
3622 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
3623 | { | |
364b6181 ML |
3624 | struct request *flush_rq = hctx->fq->flush_rq; |
3625 | ||
8ab0b7dc ML |
3626 | if (blk_mq_hw_queue_mapped(hctx)) |
3627 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 3628 | |
6cfeadbf ML |
3629 | if (blk_queue_init_done(q)) |
3630 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], | |
3631 | set->queue_depth, flush_rq); | |
f70ced09 | 3632 | if (set->ops->exit_request) |
364b6181 | 3633 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 3634 | |
08e98fc6 ML |
3635 | if (set->ops->exit_hctx) |
3636 | set->ops->exit_hctx(hctx, hctx_idx); | |
3637 | ||
9467f859 | 3638 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 | 3639 | |
4e5cc99e ML |
3640 | xa_erase(&q->hctx_table, hctx_idx); |
3641 | ||
2f8f1336 ML |
3642 | spin_lock(&q->unused_hctx_lock); |
3643 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
3644 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
3645 | } |
3646 | ||
624dbe47 ML |
3647 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
3648 | struct blk_mq_tag_set *set, int nr_queue) | |
3649 | { | |
3650 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3651 | unsigned long i; |
624dbe47 ML |
3652 | |
3653 | queue_for_each_hw_ctx(q, hctx, i) { | |
3654 | if (i == nr_queue) | |
3655 | break; | |
08e98fc6 | 3656 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 3657 | } |
624dbe47 ML |
3658 | } |
3659 | ||
08e98fc6 ML |
3660 | static int blk_mq_init_hctx(struct request_queue *q, |
3661 | struct blk_mq_tag_set *set, | |
3662 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 3663 | { |
7c6c5b7c ML |
3664 | hctx->queue_num = hctx_idx; |
3665 | ||
bf0beec0 ML |
3666 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3667 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3668 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
3669 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
3670 | ||
3671 | hctx->tags = set->tags[hctx_idx]; | |
3672 | ||
3673 | if (set->ops->init_hctx && | |
3674 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
3675 | goto unregister_cpu_notifier; | |
08e98fc6 | 3676 | |
7c6c5b7c ML |
3677 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
3678 | hctx->numa_node)) | |
3679 | goto exit_hctx; | |
4e5cc99e ML |
3680 | |
3681 | if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL)) | |
3682 | goto exit_flush_rq; | |
3683 | ||
7c6c5b7c ML |
3684 | return 0; |
3685 | ||
4e5cc99e ML |
3686 | exit_flush_rq: |
3687 | if (set->ops->exit_request) | |
3688 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); | |
7c6c5b7c ML |
3689 | exit_hctx: |
3690 | if (set->ops->exit_hctx) | |
3691 | set->ops->exit_hctx(hctx, hctx_idx); | |
3692 | unregister_cpu_notifier: | |
3693 | blk_mq_remove_cpuhp(hctx); | |
3694 | return -1; | |
3695 | } | |
3696 | ||
3697 | static struct blk_mq_hw_ctx * | |
3698 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
3699 | int node) | |
3700 | { | |
3701 | struct blk_mq_hw_ctx *hctx; | |
3702 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
3703 | ||
704b914f | 3704 | hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node); |
7c6c5b7c ML |
3705 | if (!hctx) |
3706 | goto fail_alloc_hctx; | |
3707 | ||
3708 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
3709 | goto free_hctx; | |
3710 | ||
3711 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 3712 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
3713 | node = set->numa_node; |
3714 | hctx->numa_node = node; | |
08e98fc6 | 3715 | |
9f993737 | 3716 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
3717 | spin_lock_init(&hctx->lock); |
3718 | INIT_LIST_HEAD(&hctx->dispatch); | |
3719 | hctx->queue = q; | |
51db1c37 | 3720 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 3721 | |
2f8f1336 ML |
3722 | INIT_LIST_HEAD(&hctx->hctx_list); |
3723 | ||
320ae51f | 3724 | /* |
08e98fc6 ML |
3725 | * Allocate space for all possible cpus to avoid allocation at |
3726 | * runtime | |
320ae51f | 3727 | */ |
d904bfa7 | 3728 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 3729 | gfp, node); |
08e98fc6 | 3730 | if (!hctx->ctxs) |
7c6c5b7c | 3731 | goto free_cpumask; |
320ae51f | 3732 | |
5b202853 | 3733 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 3734 | gfp, node, false, false)) |
08e98fc6 | 3735 | goto free_ctxs; |
08e98fc6 | 3736 | hctx->nr_ctx = 0; |
320ae51f | 3737 | |
5815839b | 3738 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
3739 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
3740 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
3741 | ||
754a1572 | 3742 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 3743 | if (!hctx->fq) |
7c6c5b7c | 3744 | goto free_bitmap; |
320ae51f | 3745 | |
7c6c5b7c | 3746 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 3747 | |
7c6c5b7c | 3748 | return hctx; |
320ae51f | 3749 | |
08e98fc6 | 3750 | free_bitmap: |
88459642 | 3751 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
3752 | free_ctxs: |
3753 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
3754 | free_cpumask: |
3755 | free_cpumask_var(hctx->cpumask); | |
3756 | free_hctx: | |
3757 | kfree(hctx); | |
3758 | fail_alloc_hctx: | |
3759 | return NULL; | |
08e98fc6 | 3760 | } |
320ae51f | 3761 | |
320ae51f JA |
3762 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
3763 | unsigned int nr_hw_queues) | |
3764 | { | |
b3c661b1 JA |
3765 | struct blk_mq_tag_set *set = q->tag_set; |
3766 | unsigned int i, j; | |
320ae51f JA |
3767 | |
3768 | for_each_possible_cpu(i) { | |
3769 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
3770 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 3771 | int k; |
320ae51f | 3772 | |
320ae51f JA |
3773 | __ctx->cpu = i; |
3774 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
3775 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
3776 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
3777 | ||
320ae51f JA |
3778 | __ctx->queue = q; |
3779 | ||
320ae51f JA |
3780 | /* |
3781 | * Set local node, IFF we have more than one hw queue. If | |
3782 | * not, we remain on the home node of the device | |
3783 | */ | |
b3c661b1 JA |
3784 | for (j = 0; j < set->nr_maps; j++) { |
3785 | hctx = blk_mq_map_queue_type(q, j, i); | |
3786 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 3787 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 3788 | } |
320ae51f JA |
3789 | } |
3790 | } | |
3791 | ||
63064be1 JG |
3792 | struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3793 | unsigned int hctx_idx, | |
3794 | unsigned int depth) | |
cc71a6f4 | 3795 | { |
63064be1 JG |
3796 | struct blk_mq_tags *tags; |
3797 | int ret; | |
cc71a6f4 | 3798 | |
e155b0c2 | 3799 | tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); |
63064be1 JG |
3800 | if (!tags) |
3801 | return NULL; | |
cc71a6f4 | 3802 | |
63064be1 JG |
3803 | ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); |
3804 | if (ret) { | |
e155b0c2 | 3805 | blk_mq_free_rq_map(tags); |
63064be1 JG |
3806 | return NULL; |
3807 | } | |
cc71a6f4 | 3808 | |
63064be1 | 3809 | return tags; |
cc71a6f4 JA |
3810 | } |
3811 | ||
63064be1 JG |
3812 | static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3813 | int hctx_idx) | |
cc71a6f4 | 3814 | { |
079a2e3e JG |
3815 | if (blk_mq_is_shared_tags(set->flags)) { |
3816 | set->tags[hctx_idx] = set->shared_tags; | |
1c0706a7 | 3817 | |
e155b0c2 | 3818 | return true; |
bd166ef1 | 3819 | } |
e155b0c2 | 3820 | |
63064be1 JG |
3821 | set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, |
3822 | set->queue_depth); | |
3823 | ||
3824 | return set->tags[hctx_idx]; | |
cc71a6f4 JA |
3825 | } |
3826 | ||
645db34e JG |
3827 | void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3828 | struct blk_mq_tags *tags, | |
3829 | unsigned int hctx_idx) | |
cc71a6f4 | 3830 | { |
645db34e JG |
3831 | if (tags) { |
3832 | blk_mq_free_rqs(set, tags, hctx_idx); | |
e155b0c2 | 3833 | blk_mq_free_rq_map(tags); |
bd166ef1 | 3834 | } |
cc71a6f4 JA |
3835 | } |
3836 | ||
e155b0c2 JG |
3837 | static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3838 | unsigned int hctx_idx) | |
3839 | { | |
079a2e3e | 3840 | if (!blk_mq_is_shared_tags(set->flags)) |
e155b0c2 JG |
3841 | blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); |
3842 | ||
3843 | set->tags[hctx_idx] = NULL; | |
cc71a6f4 JA |
3844 | } |
3845 | ||
4b855ad3 | 3846 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 3847 | { |
4f481208 ML |
3848 | unsigned int j, hctx_idx; |
3849 | unsigned long i; | |
320ae51f JA |
3850 | struct blk_mq_hw_ctx *hctx; |
3851 | struct blk_mq_ctx *ctx; | |
2a34c087 | 3852 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
3853 | |
3854 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 3855 | cpumask_clear(hctx->cpumask); |
320ae51f | 3856 | hctx->nr_ctx = 0; |
d416c92c | 3857 | hctx->dispatch_from = NULL; |
320ae51f JA |
3858 | } |
3859 | ||
3860 | /* | |
4b855ad3 | 3861 | * Map software to hardware queues. |
4412efec ML |
3862 | * |
3863 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 3864 | */ |
20e4d813 | 3865 | for_each_possible_cpu(i) { |
4412efec | 3866 | |
897bb0c7 | 3867 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 3868 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
3869 | if (!set->map[j].nr_queues) { |
3870 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3871 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 3872 | continue; |
bb94aea1 | 3873 | } |
fd689871 ML |
3874 | hctx_idx = set->map[j].mq_map[i]; |
3875 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
3876 | if (!set->tags[hctx_idx] && | |
63064be1 | 3877 | !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { |
fd689871 ML |
3878 | /* |
3879 | * If tags initialization fail for some hctx, | |
3880 | * that hctx won't be brought online. In this | |
3881 | * case, remap the current ctx to hctx[0] which | |
3882 | * is guaranteed to always have tags allocated | |
3883 | */ | |
3884 | set->map[j].mq_map[i] = 0; | |
3885 | } | |
e5edd5f2 | 3886 | |
b3c661b1 | 3887 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 3888 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
3889 | /* |
3890 | * If the CPU is already set in the mask, then we've | |
3891 | * mapped this one already. This can happen if | |
3892 | * devices share queues across queue maps. | |
3893 | */ | |
3894 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
3895 | continue; | |
3896 | ||
3897 | cpumask_set_cpu(i, hctx->cpumask); | |
3898 | hctx->type = j; | |
3899 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
3900 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
3901 | ||
3902 | /* | |
3903 | * If the nr_ctx type overflows, we have exceeded the | |
3904 | * amount of sw queues we can support. | |
3905 | */ | |
3906 | BUG_ON(!hctx->nr_ctx); | |
3907 | } | |
bb94aea1 JW |
3908 | |
3909 | for (; j < HCTX_MAX_TYPES; j++) | |
3910 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3911 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 3912 | } |
506e931f JA |
3913 | |
3914 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
3915 | /* |
3916 | * If no software queues are mapped to this hardware queue, | |
3917 | * disable it and free the request entries. | |
3918 | */ | |
3919 | if (!hctx->nr_ctx) { | |
3920 | /* Never unmap queue 0. We need it as a | |
3921 | * fallback in case of a new remap fails | |
3922 | * allocation | |
3923 | */ | |
e155b0c2 JG |
3924 | if (i) |
3925 | __blk_mq_free_map_and_rqs(set, i); | |
4412efec ML |
3926 | |
3927 | hctx->tags = NULL; | |
3928 | continue; | |
3929 | } | |
484b4061 | 3930 | |
2a34c087 ML |
3931 | hctx->tags = set->tags[i]; |
3932 | WARN_ON(!hctx->tags); | |
3933 | ||
889fa31f CY |
3934 | /* |
3935 | * Set the map size to the number of mapped software queues. | |
3936 | * This is more accurate and more efficient than looping | |
3937 | * over all possibly mapped software queues. | |
3938 | */ | |
88459642 | 3939 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 3940 | |
484b4061 JA |
3941 | /* |
3942 | * Initialize batch roundrobin counts | |
3943 | */ | |
f82ddf19 | 3944 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
3945 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
3946 | } | |
320ae51f JA |
3947 | } |
3948 | ||
8e8320c9 JA |
3949 | /* |
3950 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
3951 | * the queue isn't live yet. | |
3952 | */ | |
2404e607 | 3953 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
3954 | { |
3955 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3956 | unsigned long i; |
0d2602ca | 3957 | |
2404e607 | 3958 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 3959 | if (shared) { |
51db1c37 | 3960 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
3961 | } else { |
3962 | blk_mq_tag_idle(hctx); | |
51db1c37 | 3963 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 3964 | } |
2404e607 JM |
3965 | } |
3966 | } | |
3967 | ||
655ac300 HR |
3968 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
3969 | bool shared) | |
2404e607 JM |
3970 | { |
3971 | struct request_queue *q; | |
0d2602ca | 3972 | |
705cda97 BVA |
3973 | lockdep_assert_held(&set->tag_list_lock); |
3974 | ||
0d2602ca JA |
3975 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3976 | blk_mq_freeze_queue(q); | |
2404e607 | 3977 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
3978 | blk_mq_unfreeze_queue(q); |
3979 | } | |
3980 | } | |
3981 | ||
3982 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
3983 | { | |
3984 | struct blk_mq_tag_set *set = q->tag_set; | |
3985 | ||
0d2602ca | 3986 | mutex_lock(&set->tag_list_lock); |
08c875cb | 3987 | list_del(&q->tag_set_list); |
2404e607 JM |
3988 | if (list_is_singular(&set->tag_list)) { |
3989 | /* just transitioned to unshared */ | |
51db1c37 | 3990 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 3991 | /* update existing queue */ |
655ac300 | 3992 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 3993 | } |
0d2602ca | 3994 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 3995 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
3996 | } |
3997 | ||
3998 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
3999 | struct request_queue *q) | |
4000 | { | |
0d2602ca | 4001 | mutex_lock(&set->tag_list_lock); |
2404e607 | 4002 | |
ff821d27 JA |
4003 | /* |
4004 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
4005 | */ | |
4006 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
4007 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
4008 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 4009 | /* update existing queue */ |
655ac300 | 4010 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 4011 | } |
51db1c37 | 4012 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 4013 | queue_set_hctx_shared(q, true); |
08c875cb | 4014 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 4015 | |
0d2602ca JA |
4016 | mutex_unlock(&set->tag_list_lock); |
4017 | } | |
4018 | ||
1db4909e ML |
4019 | /* All allocations will be freed in release handler of q->mq_kobj */ |
4020 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
4021 | { | |
4022 | struct blk_mq_ctxs *ctxs; | |
4023 | int cpu; | |
4024 | ||
4025 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
4026 | if (!ctxs) | |
4027 | return -ENOMEM; | |
4028 | ||
4029 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
4030 | if (!ctxs->queue_ctx) | |
4031 | goto fail; | |
4032 | ||
4033 | for_each_possible_cpu(cpu) { | |
4034 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
4035 | ctx->ctxs = ctxs; | |
4036 | } | |
4037 | ||
4038 | q->mq_kobj = &ctxs->kobj; | |
4039 | q->queue_ctx = ctxs->queue_ctx; | |
4040 | ||
4041 | return 0; | |
4042 | fail: | |
4043 | kfree(ctxs); | |
4044 | return -ENOMEM; | |
4045 | } | |
4046 | ||
e09aae7e ML |
4047 | /* |
4048 | * It is the actual release handler for mq, but we do it from | |
4049 | * request queue's release handler for avoiding use-after-free | |
4050 | * and headache because q->mq_kobj shouldn't have been introduced, | |
4051 | * but we can't group ctx/kctx kobj without it. | |
4052 | */ | |
4053 | void blk_mq_release(struct request_queue *q) | |
4054 | { | |
2f8f1336 | 4055 | struct blk_mq_hw_ctx *hctx, *next; |
4f481208 | 4056 | unsigned long i; |
e09aae7e | 4057 | |
2f8f1336 ML |
4058 | queue_for_each_hw_ctx(q, hctx, i) |
4059 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
4060 | ||
4061 | /* all hctx are in .unused_hctx_list now */ | |
4062 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
4063 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 4064 | kobject_put(&hctx->kobj); |
c3b4afca | 4065 | } |
e09aae7e | 4066 | |
4e5cc99e | 4067 | xa_destroy(&q->hctx_table); |
e09aae7e | 4068 | |
7ea5fe31 ML |
4069 | /* |
4070 | * release .mq_kobj and sw queue's kobject now because | |
4071 | * both share lifetime with request queue. | |
4072 | */ | |
4073 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
4074 | } |
4075 | ||
9ac4dd8c CH |
4076 | struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set, |
4077 | struct queue_limits *lim, void *queuedata) | |
b62c21b7 | 4078 | { |
9ac4dd8c | 4079 | struct queue_limits default_lim = { }; |
26a9750a CH |
4080 | struct request_queue *q; |
4081 | int ret; | |
b62c21b7 | 4082 | |
9ac4dd8c | 4083 | q = blk_alloc_queue(lim ? lim : &default_lim, set->numa_node); |
ad751ba1 CH |
4084 | if (IS_ERR(q)) |
4085 | return q; | |
26a9750a CH |
4086 | q->queuedata = queuedata; |
4087 | ret = blk_mq_init_allocated_queue(set, q); | |
4088 | if (ret) { | |
6f8191fd | 4089 | blk_put_queue(q); |
26a9750a CH |
4090 | return ERR_PTR(ret); |
4091 | } | |
b62c21b7 MS |
4092 | return q; |
4093 | } | |
9ac4dd8c | 4094 | EXPORT_SYMBOL(blk_mq_alloc_queue); |
b62c21b7 | 4095 | |
6f8191fd CH |
4096 | /** |
4097 | * blk_mq_destroy_queue - shutdown a request queue | |
4098 | * @q: request queue to shutdown | |
4099 | * | |
9ac4dd8c | 4100 | * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future |
81ea42b9 | 4101 | * requests will be failed with -ENODEV. The caller is responsible for dropping |
9ac4dd8c | 4102 | * the reference from blk_mq_alloc_queue() by calling blk_put_queue(). |
6f8191fd CH |
4103 | * |
4104 | * Context: can sleep | |
4105 | */ | |
4106 | void blk_mq_destroy_queue(struct request_queue *q) | |
4107 | { | |
4108 | WARN_ON_ONCE(!queue_is_mq(q)); | |
4109 | WARN_ON_ONCE(blk_queue_registered(q)); | |
4110 | ||
4111 | might_sleep(); | |
4112 | ||
4113 | blk_queue_flag_set(QUEUE_FLAG_DYING, q); | |
4114 | blk_queue_start_drain(q); | |
56c1ee92 | 4115 | blk_mq_freeze_queue_wait(q); |
6f8191fd CH |
4116 | |
4117 | blk_sync_queue(q); | |
4118 | blk_mq_cancel_work_sync(q); | |
4119 | blk_mq_exit_queue(q); | |
6f8191fd CH |
4120 | } |
4121 | EXPORT_SYMBOL(blk_mq_destroy_queue); | |
4122 | ||
27e32cd2 CH |
4123 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, |
4124 | struct queue_limits *lim, void *queuedata, | |
4dcc4874 | 4125 | struct lock_class_key *lkclass) |
9316a9ed JA |
4126 | { |
4127 | struct request_queue *q; | |
b461dfc4 | 4128 | struct gendisk *disk; |
9316a9ed | 4129 | |
27e32cd2 | 4130 | q = blk_mq_alloc_queue(set, lim, queuedata); |
b461dfc4 CH |
4131 | if (IS_ERR(q)) |
4132 | return ERR_CAST(q); | |
9316a9ed | 4133 | |
4a1fa41d | 4134 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 | 4135 | if (!disk) { |
0a3e5cc7 | 4136 | blk_mq_destroy_queue(q); |
2b3f056f | 4137 | blk_put_queue(q); |
b461dfc4 | 4138 | return ERR_PTR(-ENOMEM); |
9316a9ed | 4139 | } |
6f8191fd | 4140 | set_bit(GD_OWNS_QUEUE, &disk->state); |
b461dfc4 | 4141 | return disk; |
9316a9ed | 4142 | } |
b461dfc4 | 4143 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 4144 | |
6f8191fd CH |
4145 | struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q, |
4146 | struct lock_class_key *lkclass) | |
4147 | { | |
22c17e27 CH |
4148 | struct gendisk *disk; |
4149 | ||
6f8191fd CH |
4150 | if (!blk_get_queue(q)) |
4151 | return NULL; | |
22c17e27 CH |
4152 | disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass); |
4153 | if (!disk) | |
4154 | blk_put_queue(q); | |
4155 | return disk; | |
6f8191fd CH |
4156 | } |
4157 | EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue); | |
4158 | ||
34d11ffa JW |
4159 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
4160 | struct blk_mq_tag_set *set, struct request_queue *q, | |
4161 | int hctx_idx, int node) | |
4162 | { | |
2f8f1336 | 4163 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 4164 | |
2f8f1336 ML |
4165 | /* reuse dead hctx first */ |
4166 | spin_lock(&q->unused_hctx_lock); | |
4167 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
4168 | if (tmp->numa_node == node) { | |
4169 | hctx = tmp; | |
4170 | break; | |
4171 | } | |
4172 | } | |
4173 | if (hctx) | |
4174 | list_del_init(&hctx->hctx_list); | |
4175 | spin_unlock(&q->unused_hctx_lock); | |
4176 | ||
4177 | if (!hctx) | |
4178 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 4179 | if (!hctx) |
7c6c5b7c | 4180 | goto fail; |
34d11ffa | 4181 | |
7c6c5b7c ML |
4182 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
4183 | goto free_hctx; | |
34d11ffa JW |
4184 | |
4185 | return hctx; | |
7c6c5b7c ML |
4186 | |
4187 | free_hctx: | |
4188 | kobject_put(&hctx->kobj); | |
4189 | fail: | |
4190 | return NULL; | |
34d11ffa JW |
4191 | } |
4192 | ||
868f2f0b KB |
4193 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
4194 | struct request_queue *q) | |
320ae51f | 4195 | { |
4e5cc99e ML |
4196 | struct blk_mq_hw_ctx *hctx; |
4197 | unsigned long i, j; | |
ac0d6b92 | 4198 | |
fb350e0a ML |
4199 | /* protect against switching io scheduler */ |
4200 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 4201 | for (i = 0; i < set->nr_hw_queues; i++) { |
306f13ee | 4202 | int old_node; |
4d805131 | 4203 | int node = blk_mq_get_hctx_node(set, i); |
4e5cc99e | 4204 | struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i); |
868f2f0b | 4205 | |
306f13ee ML |
4206 | if (old_hctx) { |
4207 | old_node = old_hctx->numa_node; | |
4208 | blk_mq_exit_hctx(q, set, old_hctx, i); | |
4209 | } | |
868f2f0b | 4210 | |
4e5cc99e | 4211 | if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) { |
306f13ee | 4212 | if (!old_hctx) |
34d11ffa | 4213 | break; |
306f13ee ML |
4214 | pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n", |
4215 | node, old_node); | |
4e5cc99e ML |
4216 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node); |
4217 | WARN_ON_ONCE(!hctx); | |
868f2f0b | 4218 | } |
320ae51f | 4219 | } |
e01ad46d JW |
4220 | /* |
4221 | * Increasing nr_hw_queues fails. Free the newly allocated | |
4222 | * hctxs and keep the previous q->nr_hw_queues. | |
4223 | */ | |
4224 | if (i != set->nr_hw_queues) { | |
4225 | j = q->nr_hw_queues; | |
e01ad46d JW |
4226 | } else { |
4227 | j = i; | |
e01ad46d JW |
4228 | q->nr_hw_queues = set->nr_hw_queues; |
4229 | } | |
34d11ffa | 4230 | |
4e5cc99e ML |
4231 | xa_for_each_start(&q->hctx_table, j, hctx, j) |
4232 | blk_mq_exit_hctx(q, set, hctx, j); | |
fb350e0a | 4233 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
4234 | } |
4235 | ||
42ee3061 ML |
4236 | static void blk_mq_update_poll_flag(struct request_queue *q) |
4237 | { | |
4238 | struct blk_mq_tag_set *set = q->tag_set; | |
4239 | ||
4240 | if (set->nr_maps > HCTX_TYPE_POLL && | |
4241 | set->map[HCTX_TYPE_POLL].nr_queues) | |
4242 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); | |
4243 | else | |
4244 | blk_queue_flag_clear(QUEUE_FLAG_POLL, q); | |
4245 | } | |
4246 | ||
26a9750a CH |
4247 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
4248 | struct request_queue *q) | |
868f2f0b | 4249 | { |
66841672 ML |
4250 | /* mark the queue as mq asap */ |
4251 | q->mq_ops = set->ops; | |
4252 | ||
1db4909e | 4253 | if (blk_mq_alloc_ctxs(q)) |
54bdd67d | 4254 | goto err_exit; |
868f2f0b | 4255 | |
737f98cf ML |
4256 | /* init q->mq_kobj and sw queues' kobjects */ |
4257 | blk_mq_sysfs_init(q); | |
4258 | ||
2f8f1336 ML |
4259 | INIT_LIST_HEAD(&q->unused_hctx_list); |
4260 | spin_lock_init(&q->unused_hctx_lock); | |
4261 | ||
4e5cc99e ML |
4262 | xa_init(&q->hctx_table); |
4263 | ||
868f2f0b KB |
4264 | blk_mq_realloc_hw_ctxs(set, q); |
4265 | if (!q->nr_hw_queues) | |
4266 | goto err_hctxs; | |
320ae51f | 4267 | |
287922eb | 4268 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 4269 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 4270 | |
a8908939 | 4271 | q->tag_set = set; |
320ae51f | 4272 | |
94eddfbe | 4273 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
42ee3061 | 4274 | blk_mq_update_poll_flag(q); |
320ae51f | 4275 | |
2849450a | 4276 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
9a67aa52 | 4277 | INIT_LIST_HEAD(&q->flush_list); |
6fca6a61 CH |
4278 | INIT_LIST_HEAD(&q->requeue_list); |
4279 | spin_lock_init(&q->requeue_lock); | |
4280 | ||
eba71768 JA |
4281 | q->nr_requests = set->queue_depth; |
4282 | ||
24d2f903 | 4283 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 4284 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 4285 | blk_mq_map_swqueue(q); |
26a9750a | 4286 | return 0; |
18741986 | 4287 | |
320ae51f | 4288 | err_hctxs: |
943f45b9 | 4289 | blk_mq_release(q); |
c7de5726 ML |
4290 | err_exit: |
4291 | q->mq_ops = NULL; | |
26a9750a | 4292 | return -ENOMEM; |
320ae51f | 4293 | } |
b62c21b7 | 4294 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 4295 | |
c7e2d94b ML |
4296 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
4297 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 4298 | { |
630ef623 | 4299 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 4300 | |
630ef623 | 4301 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 4302 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
4303 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
4304 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 4305 | } |
320ae51f | 4306 | |
a5164405 JA |
4307 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
4308 | { | |
4309 | int i; | |
4310 | ||
079a2e3e JG |
4311 | if (blk_mq_is_shared_tags(set->flags)) { |
4312 | set->shared_tags = blk_mq_alloc_map_and_rqs(set, | |
e155b0c2 JG |
4313 | BLK_MQ_NO_HCTX_IDX, |
4314 | set->queue_depth); | |
079a2e3e | 4315 | if (!set->shared_tags) |
e155b0c2 JG |
4316 | return -ENOMEM; |
4317 | } | |
4318 | ||
8229cca8 | 4319 | for (i = 0; i < set->nr_hw_queues; i++) { |
63064be1 | 4320 | if (!__blk_mq_alloc_map_and_rqs(set, i)) |
a5164405 | 4321 | goto out_unwind; |
8229cca8 XT |
4322 | cond_resched(); |
4323 | } | |
a5164405 JA |
4324 | |
4325 | return 0; | |
4326 | ||
4327 | out_unwind: | |
4328 | while (--i >= 0) | |
e155b0c2 JG |
4329 | __blk_mq_free_map_and_rqs(set, i); |
4330 | ||
079a2e3e JG |
4331 | if (blk_mq_is_shared_tags(set->flags)) { |
4332 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 | 4333 | BLK_MQ_NO_HCTX_IDX); |
645db34e | 4334 | } |
a5164405 | 4335 | |
a5164405 JA |
4336 | return -ENOMEM; |
4337 | } | |
4338 | ||
4339 | /* | |
4340 | * Allocate the request maps associated with this tag_set. Note that this | |
4341 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
4342 | * will be updated to reflect the allocated depth. | |
4343 | */ | |
63064be1 | 4344 | static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) |
a5164405 JA |
4345 | { |
4346 | unsigned int depth; | |
4347 | int err; | |
4348 | ||
4349 | depth = set->queue_depth; | |
4350 | do { | |
4351 | err = __blk_mq_alloc_rq_maps(set); | |
4352 | if (!err) | |
4353 | break; | |
4354 | ||
4355 | set->queue_depth >>= 1; | |
4356 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
4357 | err = -ENOMEM; | |
4358 | break; | |
4359 | } | |
4360 | } while (set->queue_depth); | |
4361 | ||
4362 | if (!set->queue_depth || err) { | |
4363 | pr_err("blk-mq: failed to allocate request map\n"); | |
4364 | return -ENOMEM; | |
4365 | } | |
4366 | ||
4367 | if (depth != set->queue_depth) | |
4368 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
4369 | depth, set->queue_depth); | |
4370 | ||
4371 | return 0; | |
4372 | } | |
4373 | ||
a4e1d0b7 | 4374 | static void blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
ebe8bddb | 4375 | { |
6e66b493 BVA |
4376 | /* |
4377 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
4378 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
4379 | * number of hardware queues. | |
4380 | */ | |
4381 | if (set->nr_maps == 1) | |
4382 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
4383 | ||
ec30b461 | 4384 | if (set->ops->map_queues) { |
b3c661b1 JA |
4385 | int i; |
4386 | ||
7d4901a9 ML |
4387 | /* |
4388 | * transport .map_queues is usually done in the following | |
4389 | * way: | |
4390 | * | |
4391 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
4392 | * mask = get_cpu_mask(queue) | |
4393 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 4394 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
4395 | * } |
4396 | * | |
4397 | * When we need to remap, the table has to be cleared for | |
4398 | * killing stale mapping since one CPU may not be mapped | |
4399 | * to any hw queue. | |
4400 | */ | |
b3c661b1 JA |
4401 | for (i = 0; i < set->nr_maps; i++) |
4402 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 4403 | |
a4e1d0b7 | 4404 | set->ops->map_queues(set); |
b3c661b1 JA |
4405 | } else { |
4406 | BUG_ON(set->nr_maps > 1); | |
a4e1d0b7 | 4407 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 4408 | } |
ebe8bddb OS |
4409 | } |
4410 | ||
f7e76dbc | 4411 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
ee9d5521 | 4412 | int new_nr_hw_queues) |
f7e76dbc BVA |
4413 | { |
4414 | struct blk_mq_tags **new_tags; | |
e1dd7bc9 | 4415 | int i; |
f7e76dbc | 4416 | |
6be6d112 | 4417 | if (set->nr_hw_queues >= new_nr_hw_queues) |
d4b2e0d4 | 4418 | goto done; |
f7e76dbc BVA |
4419 | |
4420 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
4421 | GFP_KERNEL, set->numa_node); | |
4422 | if (!new_tags) | |
4423 | return -ENOMEM; | |
4424 | ||
4425 | if (set->tags) | |
ee9d5521 | 4426 | memcpy(new_tags, set->tags, set->nr_hw_queues * |
f7e76dbc BVA |
4427 | sizeof(*set->tags)); |
4428 | kfree(set->tags); | |
4429 | set->tags = new_tags; | |
7222657e CZ |
4430 | |
4431 | for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) { | |
4432 | if (!__blk_mq_alloc_map_and_rqs(set, i)) { | |
4433 | while (--i >= set->nr_hw_queues) | |
4434 | __blk_mq_free_map_and_rqs(set, i); | |
4435 | return -ENOMEM; | |
4436 | } | |
4437 | cond_resched(); | |
4438 | } | |
4439 | ||
d4b2e0d4 | 4440 | done: |
f7e76dbc | 4441 | set->nr_hw_queues = new_nr_hw_queues; |
f7e76dbc BVA |
4442 | return 0; |
4443 | } | |
4444 | ||
a4391c64 JA |
4445 | /* |
4446 | * Alloc a tag set to be associated with one or more request queues. | |
4447 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 4448 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
4449 | * value will be stored in set->queue_depth. |
4450 | */ | |
24d2f903 CH |
4451 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
4452 | { | |
b3c661b1 | 4453 | int i, ret; |
da695ba2 | 4454 | |
205fb5f5 BVA |
4455 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
4456 | ||
24d2f903 CH |
4457 | if (!set->nr_hw_queues) |
4458 | return -EINVAL; | |
a4391c64 | 4459 | if (!set->queue_depth) |
24d2f903 CH |
4460 | return -EINVAL; |
4461 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
4462 | return -EINVAL; | |
4463 | ||
7d7e0f90 | 4464 | if (!set->ops->queue_rq) |
24d2f903 CH |
4465 | return -EINVAL; |
4466 | ||
de148297 ML |
4467 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
4468 | return -EINVAL; | |
4469 | ||
a4391c64 JA |
4470 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
4471 | pr_info("blk-mq: reduced tag depth to %u\n", | |
4472 | BLK_MQ_MAX_DEPTH); | |
4473 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
4474 | } | |
24d2f903 | 4475 | |
b3c661b1 JA |
4476 | if (!set->nr_maps) |
4477 | set->nr_maps = 1; | |
4478 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
4479 | return -EINVAL; | |
4480 | ||
6637fadf SL |
4481 | /* |
4482 | * If a crashdump is active, then we are potentially in a very | |
ec30b461 ML |
4483 | * memory constrained environment. Limit us to 64 tags to prevent |
4484 | * using too much memory. | |
6637fadf | 4485 | */ |
ec30b461 | 4486 | if (is_kdump_kernel()) |
6637fadf | 4487 | set->queue_depth = min(64U, set->queue_depth); |
ec30b461 | 4488 | |
868f2f0b | 4489 | /* |
392546ae JA |
4490 | * There is no use for more h/w queues than cpus if we just have |
4491 | * a single map | |
868f2f0b | 4492 | */ |
392546ae | 4493 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4494 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 4495 | |
80bd4a7a CH |
4496 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4497 | set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL); | |
4498 | if (!set->srcu) | |
4499 | return -ENOMEM; | |
4500 | ret = init_srcu_struct(set->srcu); | |
4501 | if (ret) | |
4502 | goto out_free_srcu; | |
4503 | } | |
24d2f903 | 4504 | |
da695ba2 | 4505 | ret = -ENOMEM; |
5ee20298 CH |
4506 | set->tags = kcalloc_node(set->nr_hw_queues, |
4507 | sizeof(struct blk_mq_tags *), GFP_KERNEL, | |
4508 | set->numa_node); | |
4509 | if (!set->tags) | |
80bd4a7a | 4510 | goto out_cleanup_srcu; |
24d2f903 | 4511 | |
b3c661b1 JA |
4512 | for (i = 0; i < set->nr_maps; i++) { |
4513 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 4514 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
4515 | GFP_KERNEL, set->numa_node); |
4516 | if (!set->map[i].mq_map) | |
4517 | goto out_free_mq_map; | |
ec30b461 | 4518 | set->map[i].nr_queues = set->nr_hw_queues; |
b3c661b1 | 4519 | } |
bdd17e75 | 4520 | |
a4e1d0b7 | 4521 | blk_mq_update_queue_map(set); |
da695ba2 | 4522 | |
63064be1 | 4523 | ret = blk_mq_alloc_set_map_and_rqs(set); |
da695ba2 | 4524 | if (ret) |
bdd17e75 | 4525 | goto out_free_mq_map; |
24d2f903 | 4526 | |
0d2602ca JA |
4527 | mutex_init(&set->tag_list_lock); |
4528 | INIT_LIST_HEAD(&set->tag_list); | |
4529 | ||
24d2f903 | 4530 | return 0; |
bdd17e75 CH |
4531 | |
4532 | out_free_mq_map: | |
b3c661b1 JA |
4533 | for (i = 0; i < set->nr_maps; i++) { |
4534 | kfree(set->map[i].mq_map); | |
4535 | set->map[i].mq_map = NULL; | |
4536 | } | |
5676e7b6 RE |
4537 | kfree(set->tags); |
4538 | set->tags = NULL; | |
80bd4a7a CH |
4539 | out_cleanup_srcu: |
4540 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4541 | cleanup_srcu_struct(set->srcu); | |
4542 | out_free_srcu: | |
4543 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4544 | kfree(set->srcu); | |
da695ba2 | 4545 | return ret; |
24d2f903 CH |
4546 | } |
4547 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
4548 | ||
cdb14e0f CH |
4549 | /* allocate and initialize a tagset for a simple single-queue device */ |
4550 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
4551 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
4552 | unsigned int set_flags) | |
4553 | { | |
4554 | memset(set, 0, sizeof(*set)); | |
4555 | set->ops = ops; | |
4556 | set->nr_hw_queues = 1; | |
4557 | set->nr_maps = 1; | |
4558 | set->queue_depth = queue_depth; | |
4559 | set->numa_node = NUMA_NO_NODE; | |
4560 | set->flags = set_flags; | |
4561 | return blk_mq_alloc_tag_set(set); | |
4562 | } | |
4563 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
4564 | ||
24d2f903 CH |
4565 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
4566 | { | |
b3c661b1 | 4567 | int i, j; |
24d2f903 | 4568 | |
f7e76dbc | 4569 | for (i = 0; i < set->nr_hw_queues; i++) |
e155b0c2 | 4570 | __blk_mq_free_map_and_rqs(set, i); |
484b4061 | 4571 | |
079a2e3e JG |
4572 | if (blk_mq_is_shared_tags(set->flags)) { |
4573 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 JG |
4574 | BLK_MQ_NO_HCTX_IDX); |
4575 | } | |
32bc15af | 4576 | |
b3c661b1 JA |
4577 | for (j = 0; j < set->nr_maps; j++) { |
4578 | kfree(set->map[j].mq_map); | |
4579 | set->map[j].mq_map = NULL; | |
4580 | } | |
bdd17e75 | 4581 | |
981bd189 | 4582 | kfree(set->tags); |
5676e7b6 | 4583 | set->tags = NULL; |
80bd4a7a CH |
4584 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4585 | cleanup_srcu_struct(set->srcu); | |
4586 | kfree(set->srcu); | |
4587 | } | |
24d2f903 CH |
4588 | } |
4589 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
4590 | ||
e3a2b3f9 JA |
4591 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
4592 | { | |
4593 | struct blk_mq_tag_set *set = q->tag_set; | |
4594 | struct blk_mq_hw_ctx *hctx; | |
4f481208 ML |
4595 | int ret; |
4596 | unsigned long i; | |
e3a2b3f9 | 4597 | |
bd166ef1 | 4598 | if (!set) |
e3a2b3f9 JA |
4599 | return -EINVAL; |
4600 | ||
e5fa8140 AZ |
4601 | if (q->nr_requests == nr) |
4602 | return 0; | |
4603 | ||
70f36b60 | 4604 | blk_mq_freeze_queue(q); |
24f5a90f | 4605 | blk_mq_quiesce_queue(q); |
70f36b60 | 4606 | |
e3a2b3f9 JA |
4607 | ret = 0; |
4608 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
4609 | if (!hctx->tags) |
4610 | continue; | |
bd166ef1 JA |
4611 | /* |
4612 | * If we're using an MQ scheduler, just update the scheduler | |
4613 | * queue depth. This is similar to what the old code would do. | |
4614 | */ | |
f6adcef5 | 4615 | if (hctx->sched_tags) { |
70f36b60 | 4616 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
f6adcef5 | 4617 | nr, true); |
f6adcef5 JG |
4618 | } else { |
4619 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
4620 | false); | |
70f36b60 | 4621 | } |
e3a2b3f9 JA |
4622 | if (ret) |
4623 | break; | |
77f1e0a5 JA |
4624 | if (q->elevator && q->elevator->type->ops.depth_updated) |
4625 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 4626 | } |
d97e594c | 4627 | if (!ret) { |
e3a2b3f9 | 4628 | q->nr_requests = nr; |
079a2e3e | 4629 | if (blk_mq_is_shared_tags(set->flags)) { |
8fa04464 | 4630 | if (q->elevator) |
079a2e3e | 4631 | blk_mq_tag_update_sched_shared_tags(q); |
8fa04464 | 4632 | else |
079a2e3e | 4633 | blk_mq_tag_resize_shared_tags(set, nr); |
8fa04464 | 4634 | } |
d97e594c | 4635 | } |
e3a2b3f9 | 4636 | |
24f5a90f | 4637 | blk_mq_unquiesce_queue(q); |
70f36b60 | 4638 | blk_mq_unfreeze_queue(q); |
70f36b60 | 4639 | |
e3a2b3f9 JA |
4640 | return ret; |
4641 | } | |
4642 | ||
d48ece20 JW |
4643 | /* |
4644 | * request_queue and elevator_type pair. | |
4645 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
4646 | * the elevator_type associated with a request_queue. | |
4647 | */ | |
4648 | struct blk_mq_qe_pair { | |
4649 | struct list_head node; | |
4650 | struct request_queue *q; | |
4651 | struct elevator_type *type; | |
4652 | }; | |
4653 | ||
4654 | /* | |
4655 | * Cache the elevator_type in qe pair list and switch the | |
4656 | * io scheduler to 'none' | |
4657 | */ | |
4658 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
4659 | struct request_queue *q) | |
4660 | { | |
4661 | struct blk_mq_qe_pair *qe; | |
4662 | ||
d48ece20 JW |
4663 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); |
4664 | if (!qe) | |
4665 | return false; | |
4666 | ||
5fd7a84a ML |
4667 | /* q->elevator needs protection from ->sysfs_lock */ |
4668 | mutex_lock(&q->sysfs_lock); | |
4669 | ||
24516565 ML |
4670 | /* the check has to be done with holding sysfs_lock */ |
4671 | if (!q->elevator) { | |
4672 | kfree(qe); | |
4673 | goto unlock; | |
4674 | } | |
4675 | ||
d48ece20 JW |
4676 | INIT_LIST_HEAD(&qe->node); |
4677 | qe->q = q; | |
4678 | qe->type = q->elevator->type; | |
dd6f7f17 CH |
4679 | /* keep a reference to the elevator module as we'll switch back */ |
4680 | __elevator_get(qe->type); | |
d48ece20 | 4681 | list_add(&qe->node, head); |
64b36075 | 4682 | elevator_disable(q); |
24516565 | 4683 | unlock: |
d48ece20 JW |
4684 | mutex_unlock(&q->sysfs_lock); |
4685 | ||
4686 | return true; | |
4687 | } | |
4688 | ||
4a3b666e JK |
4689 | static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head, |
4690 | struct request_queue *q) | |
d48ece20 JW |
4691 | { |
4692 | struct blk_mq_qe_pair *qe; | |
d48ece20 JW |
4693 | |
4694 | list_for_each_entry(qe, head, node) | |
4a3b666e JK |
4695 | if (qe->q == q) |
4696 | return qe; | |
d48ece20 | 4697 | |
4a3b666e JK |
4698 | return NULL; |
4699 | } | |
d48ece20 | 4700 | |
4a3b666e JK |
4701 | static void blk_mq_elv_switch_back(struct list_head *head, |
4702 | struct request_queue *q) | |
4703 | { | |
4704 | struct blk_mq_qe_pair *qe; | |
4705 | struct elevator_type *t; | |
4706 | ||
4707 | qe = blk_lookup_qe_pair(head, q); | |
4708 | if (!qe) | |
4709 | return; | |
4710 | t = qe->type; | |
d48ece20 JW |
4711 | list_del(&qe->node); |
4712 | kfree(qe); | |
4713 | ||
4714 | mutex_lock(&q->sysfs_lock); | |
8237c01f | 4715 | elevator_switch(q, t); |
8ed40ee3 JC |
4716 | /* drop the reference acquired in blk_mq_elv_switch_none */ |
4717 | elevator_put(t); | |
d48ece20 JW |
4718 | mutex_unlock(&q->sysfs_lock); |
4719 | } | |
4720 | ||
e4dc2b32 KB |
4721 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
4722 | int nr_hw_queues) | |
868f2f0b KB |
4723 | { |
4724 | struct request_queue *q; | |
d48ece20 | 4725 | LIST_HEAD(head); |
6be6d112 CZ |
4726 | int prev_nr_hw_queues = set->nr_hw_queues; |
4727 | int i; | |
868f2f0b | 4728 | |
705cda97 BVA |
4729 | lockdep_assert_held(&set->tag_list_lock); |
4730 | ||
392546ae | 4731 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4732 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
4733 | if (nr_hw_queues < 1) |
4734 | return; | |
4735 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
4736 | return; |
4737 | ||
4738 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4739 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
4740 | /* |
4741 | * Switch IO scheduler to 'none', cleaning up the data associated | |
4742 | * with the previous scheduler. We will switch back once we are done | |
4743 | * updating the new sw to hw queue mappings. | |
4744 | */ | |
4745 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4746 | if (!blk_mq_elv_switch_none(&head, q)) | |
4747 | goto switch_back; | |
868f2f0b | 4748 | |
477e19de JW |
4749 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4750 | blk_mq_debugfs_unregister_hctxs(q); | |
eaa870f9 | 4751 | blk_mq_sysfs_unregister_hctxs(q); |
477e19de JW |
4752 | } |
4753 | ||
ee9d5521 | 4754 | if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0) |
f7e76dbc BVA |
4755 | goto reregister; |
4756 | ||
e01ad46d | 4757 | fallback: |
aa880ad6 | 4758 | blk_mq_update_queue_map(set); |
868f2f0b KB |
4759 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4760 | blk_mq_realloc_hw_ctxs(set, q); | |
42ee3061 | 4761 | blk_mq_update_poll_flag(q); |
e01ad46d | 4762 | if (q->nr_hw_queues != set->nr_hw_queues) { |
a846a8e6 YB |
4763 | int i = prev_nr_hw_queues; |
4764 | ||
e01ad46d JW |
4765 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", |
4766 | nr_hw_queues, prev_nr_hw_queues); | |
a846a8e6 YB |
4767 | for (; i < set->nr_hw_queues; i++) |
4768 | __blk_mq_free_map_and_rqs(set, i); | |
4769 | ||
e01ad46d | 4770 | set->nr_hw_queues = prev_nr_hw_queues; |
e01ad46d JW |
4771 | goto fallback; |
4772 | } | |
477e19de JW |
4773 | blk_mq_map_swqueue(q); |
4774 | } | |
4775 | ||
f7e76dbc | 4776 | reregister: |
477e19de | 4777 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
eaa870f9 | 4778 | blk_mq_sysfs_register_hctxs(q); |
477e19de | 4779 | blk_mq_debugfs_register_hctxs(q); |
868f2f0b KB |
4780 | } |
4781 | ||
d48ece20 JW |
4782 | switch_back: |
4783 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4784 | blk_mq_elv_switch_back(&head, q); | |
4785 | ||
868f2f0b KB |
4786 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
4787 | blk_mq_unfreeze_queue(q); | |
6be6d112 CZ |
4788 | |
4789 | /* Free the excess tags when nr_hw_queues shrink. */ | |
4790 | for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++) | |
4791 | __blk_mq_free_map_and_rqs(set, i); | |
868f2f0b | 4792 | } |
e4dc2b32 KB |
4793 | |
4794 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
4795 | { | |
4796 | mutex_lock(&set->tag_list_lock); | |
4797 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
4798 | mutex_unlock(&set->tag_list_lock); | |
4799 | } | |
868f2f0b KB |
4800 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
4801 | ||
f6c80cff KB |
4802 | static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx, |
4803 | struct io_comp_batch *iob, unsigned int flags) | |
bbd7bb70 | 4804 | { |
c6699d6f CH |
4805 | long state = get_current_state(); |
4806 | int ret; | |
bbd7bb70 | 4807 | |
aa61bec3 | 4808 | do { |
5a72e899 | 4809 | ret = q->mq_ops->poll(hctx, iob); |
bbd7bb70 | 4810 | if (ret > 0) { |
849a3700 | 4811 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4812 | return ret; |
bbd7bb70 JA |
4813 | } |
4814 | ||
4815 | if (signal_pending_state(state, current)) | |
849a3700 | 4816 | __set_current_state(TASK_RUNNING); |
b03fbd4f | 4817 | if (task_is_running(current)) |
85f4d4b6 | 4818 | return 1; |
c6699d6f | 4819 | |
ef99b2d3 | 4820 | if (ret < 0 || (flags & BLK_POLL_ONESHOT)) |
bbd7bb70 JA |
4821 | break; |
4822 | cpu_relax(); | |
aa61bec3 | 4823 | } while (!need_resched()); |
bbd7bb70 | 4824 | |
67b4110f | 4825 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4826 | return 0; |
bbd7bb70 | 4827 | } |
1052b8ac | 4828 | |
f6c80cff KB |
4829 | int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, |
4830 | struct io_comp_batch *iob, unsigned int flags) | |
4831 | { | |
4832 | struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie); | |
4833 | ||
4834 | return blk_hctx_poll(q, hctx, iob, flags); | |
4835 | } | |
4836 | ||
4837 | int blk_rq_poll(struct request *rq, struct io_comp_batch *iob, | |
4838 | unsigned int poll_flags) | |
4839 | { | |
4840 | struct request_queue *q = rq->q; | |
4841 | int ret; | |
4842 | ||
4843 | if (!blk_rq_is_poll(rq)) | |
4844 | return 0; | |
4845 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
4846 | return 0; | |
4847 | ||
4848 | ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags); | |
4849 | blk_queue_exit(q); | |
4850 | ||
4851 | return ret; | |
4852 | } | |
4853 | EXPORT_SYMBOL_GPL(blk_rq_poll); | |
4854 | ||
9cf2bab6 JA |
4855 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4856 | { | |
4857 | return rq->mq_ctx->cpu; | |
4858 | } | |
4859 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4860 | ||
2a19b28f ML |
4861 | void blk_mq_cancel_work_sync(struct request_queue *q) |
4862 | { | |
219cf43c JC |
4863 | struct blk_mq_hw_ctx *hctx; |
4864 | unsigned long i; | |
2a19b28f | 4865 | |
219cf43c | 4866 | cancel_delayed_work_sync(&q->requeue_work); |
2a19b28f | 4867 | |
219cf43c JC |
4868 | queue_for_each_hw_ctx(q, hctx, i) |
4869 | cancel_delayed_work_sync(&hctx->run_work); | |
2a19b28f ML |
4870 | } |
4871 | ||
320ae51f JA |
4872 | static int __init blk_mq_init(void) |
4873 | { | |
c3077b5d CH |
4874 | int i; |
4875 | ||
4876 | for_each_possible_cpu(i) | |
f9ab4918 | 4877 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
660e802c CZ |
4878 | for_each_possible_cpu(i) |
4879 | INIT_CSD(&per_cpu(blk_cpu_csd, i), | |
4880 | __blk_mq_complete_request_remote, NULL); | |
c3077b5d CH |
4881 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4882 | ||
4883 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4884 | "block/softirq:dead", NULL, | |
4885 | blk_softirq_cpu_dead); | |
9467f859 TG |
4886 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4887 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4888 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4889 | blk_mq_hctx_notify_online, | |
4890 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4891 | return 0; |
4892 | } | |
4893 | subsys_initcall(blk_mq_init); |