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> | |
24 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 25 | #include <linux/sched/topology.h> |
174cd4b1 | 26 | #include <linux/sched/signal.h> |
320ae51f | 27 | #include <linux/delay.h> |
aedcd72f | 28 | #include <linux/crash_dump.h> |
88c7b2b7 | 29 | #include <linux/prefetch.h> |
a892c8d5 | 30 | #include <linux/blk-crypto.h> |
82d981d4 | 31 | #include <linux/part_stat.h> |
320ae51f JA |
32 | |
33 | #include <trace/events/block.h> | |
34 | ||
54d4e6ab | 35 | #include <linux/t10-pi.h> |
320ae51f JA |
36 | #include "blk.h" |
37 | #include "blk-mq.h" | |
9c1051aa | 38 | #include "blk-mq-debugfs.h" |
986d413b | 39 | #include "blk-pm.h" |
cf43e6be | 40 | #include "blk-stat.h" |
bd166ef1 | 41 | #include "blk-mq-sched.h" |
c1c80384 | 42 | #include "blk-rq-qos.h" |
82b74cac | 43 | #include "blk-ioprio.h" |
320ae51f | 44 | |
f9ab4918 | 45 | static DEFINE_PER_CPU(struct llist_head, blk_cpu_done); |
660e802c | 46 | static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd); |
c3077b5d | 47 | |
710fa378 | 48 | static void blk_mq_insert_request(struct request *rq, blk_insert_t flags); |
360f2648 CH |
49 | static void blk_mq_request_bypass_insert(struct request *rq, |
50 | blk_insert_t flags); | |
94aa228c CH |
51 | static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
52 | struct list_head *list); | |
f6c80cff KB |
53 | static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx, |
54 | struct io_comp_batch *iob, unsigned int flags); | |
3e08773c | 55 | |
320ae51f | 56 | /* |
85fae294 YY |
57 | * Check if any of the ctx, dispatch list or elevator |
58 | * have pending work in this hardware queue. | |
320ae51f | 59 | */ |
79f720a7 | 60 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 61 | { |
79f720a7 JA |
62 | return !list_empty_careful(&hctx->dispatch) || |
63 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 64 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
65 | } |
66 | ||
320ae51f JA |
67 | /* |
68 | * Mark this ctx as having pending work in this hardware queue | |
69 | */ | |
70 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
71 | struct blk_mq_ctx *ctx) | |
72 | { | |
f31967f0 JA |
73 | const int bit = ctx->index_hw[hctx->type]; |
74 | ||
75 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
76 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
77 | } |
78 | ||
79 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
80 | struct blk_mq_ctx *ctx) | |
81 | { | |
f31967f0 JA |
82 | const int bit = ctx->index_hw[hctx->type]; |
83 | ||
84 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
85 | } |
86 | ||
f299b7c7 | 87 | struct mq_inflight { |
8446fe92 | 88 | struct block_device *part; |
a2e80f6f | 89 | unsigned int inflight[2]; |
f299b7c7 JA |
90 | }; |
91 | ||
2dd6532e | 92 | static bool blk_mq_check_inflight(struct request *rq, void *priv) |
f299b7c7 JA |
93 | { |
94 | struct mq_inflight *mi = priv; | |
95 | ||
b81c14ca HW |
96 | if (rq->part && blk_do_io_stat(rq) && |
97 | (!mi->part->bd_partno || rq->part == mi->part) && | |
b0d97557 | 98 | blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) |
bb4e6b14 | 99 | mi->inflight[rq_data_dir(rq)]++; |
7baa8572 JA |
100 | |
101 | return true; | |
f299b7c7 JA |
102 | } |
103 | ||
8446fe92 CH |
104 | unsigned int blk_mq_in_flight(struct request_queue *q, |
105 | struct block_device *part) | |
f299b7c7 | 106 | { |
a2e80f6f | 107 | struct mq_inflight mi = { .part = part }; |
f299b7c7 | 108 | |
f299b7c7 | 109 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 | 110 | |
a2e80f6f | 111 | return mi.inflight[0] + mi.inflight[1]; |
bf0ddaba OS |
112 | } |
113 | ||
8446fe92 CH |
114 | void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, |
115 | unsigned int inflight[2]) | |
bf0ddaba | 116 | { |
a2e80f6f | 117 | struct mq_inflight mi = { .part = part }; |
bf0ddaba | 118 | |
bb4e6b14 | 119 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
a2e80f6f PB |
120 | inflight[0] = mi.inflight[0]; |
121 | inflight[1] = mi.inflight[1]; | |
bf0ddaba OS |
122 | } |
123 | ||
1671d522 | 124 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 125 | { |
7996a8b5 BL |
126 | mutex_lock(&q->mq_freeze_lock); |
127 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 128 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 129 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 130 | if (queue_is_mq(q)) |
055f6e18 | 131 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
132 | } else { |
133 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 134 | } |
f3af020b | 135 | } |
1671d522 | 136 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 137 | |
6bae363e | 138 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 139 | { |
3ef28e83 | 140 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 141 | } |
6bae363e | 142 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 143 | |
f91328c4 KB |
144 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
145 | unsigned long timeout) | |
146 | { | |
147 | return wait_event_timeout(q->mq_freeze_wq, | |
148 | percpu_ref_is_zero(&q->q_usage_counter), | |
149 | timeout); | |
150 | } | |
151 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 152 | |
f3af020b TH |
153 | /* |
154 | * Guarantee no request is in use, so we can change any data structure of | |
155 | * the queue afterward. | |
156 | */ | |
3ef28e83 | 157 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 158 | { |
3ef28e83 DW |
159 | /* |
160 | * In the !blk_mq case we are only calling this to kill the | |
161 | * q_usage_counter, otherwise this increases the freeze depth | |
162 | * and waits for it to return to zero. For this reason there is | |
163 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
164 | * exported to drivers as the only user for unfreeze is blk_mq. | |
165 | */ | |
1671d522 | 166 | blk_freeze_queue_start(q); |
f3af020b TH |
167 | blk_mq_freeze_queue_wait(q); |
168 | } | |
3ef28e83 DW |
169 | |
170 | void blk_mq_freeze_queue(struct request_queue *q) | |
171 | { | |
172 | /* | |
173 | * ...just an alias to keep freeze and unfreeze actions balanced | |
174 | * in the blk_mq_* namespace | |
175 | */ | |
176 | blk_freeze_queue(q); | |
177 | } | |
c761d96b | 178 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 179 | |
aec89dc5 | 180 | void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic) |
320ae51f | 181 | { |
7996a8b5 | 182 | mutex_lock(&q->mq_freeze_lock); |
aec89dc5 CH |
183 | if (force_atomic) |
184 | q->q_usage_counter.data->force_atomic = true; | |
7996a8b5 BL |
185 | q->mq_freeze_depth--; |
186 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
187 | if (!q->mq_freeze_depth) { | |
bdd63160 | 188 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 189 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 190 | } |
7996a8b5 | 191 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 192 | } |
aec89dc5 CH |
193 | |
194 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
195 | { | |
196 | __blk_mq_unfreeze_queue(q, false); | |
197 | } | |
b4c6a028 | 198 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 199 | |
852ec809 BVA |
200 | /* |
201 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
202 | * mpt3sas driver such that this function can be removed. | |
203 | */ | |
204 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
205 | { | |
e70feb8b ML |
206 | unsigned long flags; |
207 | ||
208 | spin_lock_irqsave(&q->queue_lock, flags); | |
209 | if (!q->quiesce_depth++) | |
210 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
211 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
852ec809 BVA |
212 | } |
213 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
214 | ||
6a83e74d | 215 | /** |
9ef4d020 | 216 | * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done |
483239c7 | 217 | * @set: tag_set to wait on |
6a83e74d | 218 | * |
9ef4d020 | 219 | * Note: it is driver's responsibility for making sure that quiesce has |
483239c7 CH |
220 | * been started on or more of the request_queues of the tag_set. This |
221 | * function only waits for the quiesce on those request_queues that had | |
222 | * the quiesce flag set using blk_mq_quiesce_queue_nowait. | |
6a83e74d | 223 | */ |
483239c7 | 224 | void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set) |
6a83e74d | 225 | { |
483239c7 CH |
226 | if (set->flags & BLK_MQ_F_BLOCKING) |
227 | synchronize_srcu(set->srcu); | |
704b914f | 228 | else |
6a83e74d BVA |
229 | synchronize_rcu(); |
230 | } | |
9ef4d020 ML |
231 | EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); |
232 | ||
233 | /** | |
234 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished | |
235 | * @q: request queue. | |
236 | * | |
237 | * Note: this function does not prevent that the struct request end_io() | |
238 | * callback function is invoked. Once this function is returned, we make | |
239 | * sure no dispatch can happen until the queue is unquiesced via | |
240 | * blk_mq_unquiesce_queue(). | |
241 | */ | |
242 | void blk_mq_quiesce_queue(struct request_queue *q) | |
243 | { | |
244 | blk_mq_quiesce_queue_nowait(q); | |
8537380b CH |
245 | /* nothing to wait for non-mq queues */ |
246 | if (queue_is_mq(q)) | |
483239c7 | 247 | blk_mq_wait_quiesce_done(q->tag_set); |
9ef4d020 | 248 | } |
6a83e74d BVA |
249 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); |
250 | ||
e4e73913 ML |
251 | /* |
252 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
253 | * @q: request queue. | |
254 | * | |
255 | * This function recovers queue into the state before quiescing | |
256 | * which is done by blk_mq_quiesce_queue. | |
257 | */ | |
258 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
259 | { | |
e70feb8b ML |
260 | unsigned long flags; |
261 | bool run_queue = false; | |
262 | ||
263 | spin_lock_irqsave(&q->queue_lock, flags); | |
264 | if (WARN_ON_ONCE(q->quiesce_depth <= 0)) { | |
265 | ; | |
266 | } else if (!--q->quiesce_depth) { | |
267 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); | |
268 | run_queue = true; | |
269 | } | |
270 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
f4560ffe | 271 | |
1d9e9bc6 | 272 | /* dispatch requests which are inserted during quiescing */ |
e70feb8b ML |
273 | if (run_queue) |
274 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
275 | } |
276 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
277 | ||
414dd48e CL |
278 | void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set) |
279 | { | |
280 | struct request_queue *q; | |
281 | ||
282 | mutex_lock(&set->tag_list_lock); | |
283 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
284 | if (!blk_queue_skip_tagset_quiesce(q)) | |
285 | blk_mq_quiesce_queue_nowait(q); | |
286 | } | |
287 | blk_mq_wait_quiesce_done(set); | |
288 | mutex_unlock(&set->tag_list_lock); | |
289 | } | |
290 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset); | |
291 | ||
292 | void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set) | |
293 | { | |
294 | struct request_queue *q; | |
295 | ||
296 | mutex_lock(&set->tag_list_lock); | |
297 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
298 | if (!blk_queue_skip_tagset_quiesce(q)) | |
299 | blk_mq_unquiesce_queue(q); | |
300 | } | |
301 | mutex_unlock(&set->tag_list_lock); | |
302 | } | |
303 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset); | |
304 | ||
aed3ea94 JA |
305 | void blk_mq_wake_waiters(struct request_queue *q) |
306 | { | |
307 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 308 | unsigned long i; |
aed3ea94 JA |
309 | |
310 | queue_for_each_hw_ctx(q, hctx, i) | |
311 | if (blk_mq_hw_queue_mapped(hctx)) | |
312 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
313 | } | |
314 | ||
52fdbbcc CH |
315 | void blk_rq_init(struct request_queue *q, struct request *rq) |
316 | { | |
317 | memset(rq, 0, sizeof(*rq)); | |
318 | ||
319 | INIT_LIST_HEAD(&rq->queuelist); | |
320 | rq->q = q; | |
321 | rq->__sector = (sector_t) -1; | |
322 | INIT_HLIST_NODE(&rq->hash); | |
323 | RB_CLEAR_NODE(&rq->rb_node); | |
324 | rq->tag = BLK_MQ_NO_TAG; | |
325 | rq->internal_tag = BLK_MQ_NO_TAG; | |
326 | rq->start_time_ns = ktime_get_ns(); | |
327 | rq->part = NULL; | |
328 | blk_crypto_rq_set_defaults(rq); | |
329 | } | |
330 | EXPORT_SYMBOL(blk_rq_init); | |
331 | ||
5c17f45e CZ |
332 | /* Set start and alloc time when the allocated request is actually used */ |
333 | static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns) | |
334 | { | |
335 | if (blk_mq_need_time_stamp(rq)) | |
336 | rq->start_time_ns = ktime_get_ns(); | |
337 | else | |
338 | rq->start_time_ns = 0; | |
339 | ||
340 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME | |
341 | if (blk_queue_rq_alloc_time(rq->q)) | |
342 | rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns; | |
343 | else | |
344 | rq->alloc_time_ns = 0; | |
345 | #endif | |
346 | } | |
347 | ||
e4cdf1a1 | 348 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
5c17f45e | 349 | struct blk_mq_tags *tags, unsigned int tag) |
320ae51f | 350 | { |
605f784e PB |
351 | struct blk_mq_ctx *ctx = data->ctx; |
352 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
353 | struct request_queue *q = data->q; | |
e4cdf1a1 | 354 | struct request *rq = tags->static_rqs[tag]; |
c3a148d2 | 355 | |
c7b84d42 JA |
356 | rq->q = q; |
357 | rq->mq_ctx = ctx; | |
358 | rq->mq_hctx = hctx; | |
359 | rq->cmd_flags = data->cmd_flags; | |
360 | ||
361 | if (data->flags & BLK_MQ_REQ_PM) | |
362 | data->rq_flags |= RQF_PM; | |
363 | if (blk_queue_io_stat(q)) | |
364 | data->rq_flags |= RQF_IO_STAT; | |
365 | rq->rq_flags = data->rq_flags; | |
366 | ||
dd6216bb | 367 | if (data->rq_flags & RQF_SCHED_TAGS) { |
56f8da64 JA |
368 | rq->tag = BLK_MQ_NO_TAG; |
369 | rq->internal_tag = tag; | |
dd6216bb CH |
370 | } else { |
371 | rq->tag = tag; | |
372 | rq->internal_tag = BLK_MQ_NO_TAG; | |
e4cdf1a1 | 373 | } |
c7b84d42 | 374 | rq->timeout = 0; |
e4cdf1a1 | 375 | |
af76e555 | 376 | rq->part = NULL; |
544ccc8d | 377 | rq->io_start_time_ns = 0; |
3d244306 | 378 | rq->stats_sectors = 0; |
af76e555 CH |
379 | rq->nr_phys_segments = 0; |
380 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
381 | rq->nr_integrity_segments = 0; | |
382 | #endif | |
af76e555 CH |
383 | rq->end_io = NULL; |
384 | rq->end_io_data = NULL; | |
af76e555 | 385 | |
4f266f2b PB |
386 | blk_crypto_rq_set_defaults(rq); |
387 | INIT_LIST_HEAD(&rq->queuelist); | |
388 | /* tag was already set */ | |
389 | WRITE_ONCE(rq->deadline, 0); | |
0a467d0f | 390 | req_ref_set(rq, 1); |
7ea4d8a4 | 391 | |
dd6216bb | 392 | if (rq->rq_flags & RQF_USE_SCHED) { |
7ea4d8a4 CH |
393 | struct elevator_queue *e = data->q->elevator; |
394 | ||
4f266f2b PB |
395 | INIT_HLIST_NODE(&rq->hash); |
396 | RB_CLEAR_NODE(&rq->rb_node); | |
397 | ||
dd6216bb | 398 | if (e->type->ops.prepare_request) |
7ea4d8a4 | 399 | e->type->ops.prepare_request(rq); |
7ea4d8a4 CH |
400 | } |
401 | ||
e4cdf1a1 | 402 | return rq; |
5dee8577 CH |
403 | } |
404 | ||
349302da | 405 | static inline struct request * |
5c17f45e | 406 | __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data) |
349302da JA |
407 | { |
408 | unsigned int tag, tag_offset; | |
fe6134f6 | 409 | struct blk_mq_tags *tags; |
349302da | 410 | struct request *rq; |
fe6134f6 | 411 | unsigned long tag_mask; |
349302da JA |
412 | int i, nr = 0; |
413 | ||
fe6134f6 JA |
414 | tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset); |
415 | if (unlikely(!tag_mask)) | |
349302da JA |
416 | return NULL; |
417 | ||
fe6134f6 JA |
418 | tags = blk_mq_tags_from_data(data); |
419 | for (i = 0; tag_mask; i++) { | |
420 | if (!(tag_mask & (1UL << i))) | |
349302da JA |
421 | continue; |
422 | tag = tag_offset + i; | |
a22c00be | 423 | prefetch(tags->static_rqs[tag]); |
fe6134f6 | 424 | tag_mask &= ~(1UL << i); |
5c17f45e | 425 | rq = blk_mq_rq_ctx_init(data, tags, tag); |
013a7f95 | 426 | rq_list_add(data->cached_rq, rq); |
c5fc7b93 | 427 | nr++; |
349302da | 428 | } |
b8643d68 CZ |
429 | if (!(data->rq_flags & RQF_SCHED_TAGS)) |
430 | blk_mq_add_active_requests(data->hctx, nr); | |
c5fc7b93 JA |
431 | /* caller already holds a reference, add for remainder */ |
432 | percpu_ref_get_many(&data->q->q_usage_counter, nr - 1); | |
349302da JA |
433 | data->nr_tags -= nr; |
434 | ||
013a7f95 | 435 | return rq_list_pop(data->cached_rq); |
349302da JA |
436 | } |
437 | ||
b90cfaed | 438 | static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) |
d2c0d383 | 439 | { |
e6e7abff | 440 | struct request_queue *q = data->q; |
6f816b4b | 441 | u64 alloc_time_ns = 0; |
47c122e3 | 442 | struct request *rq; |
600c3b0c | 443 | unsigned int tag; |
d2c0d383 | 444 | |
6f816b4b TH |
445 | /* alloc_time includes depth and tag waits */ |
446 | if (blk_queue_rq_alloc_time(q)) | |
447 | alloc_time_ns = ktime_get_ns(); | |
448 | ||
f9afca4d | 449 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 450 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 | 451 | |
781dd830 | 452 | if (q->elevator) { |
dd6216bb CH |
453 | /* |
454 | * All requests use scheduler tags when an I/O scheduler is | |
455 | * enabled for the queue. | |
456 | */ | |
457 | data->rq_flags |= RQF_SCHED_TAGS; | |
781dd830 | 458 | |
d2c0d383 | 459 | /* |
8d663f34 | 460 | * Flush/passthrough requests are special and go directly to the |
dd6216bb | 461 | * dispatch list. |
d2c0d383 | 462 | */ |
be4c4278 | 463 | if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH && |
dd6216bb CH |
464 | !blk_op_is_passthrough(data->cmd_flags)) { |
465 | struct elevator_mq_ops *ops = &q->elevator->type->ops; | |
466 | ||
467 | WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED); | |
468 | ||
469 | data->rq_flags |= RQF_USE_SCHED; | |
470 | if (ops->limit_depth) | |
471 | ops->limit_depth(data->cmd_flags, data); | |
472 | } | |
d2c0d383 CH |
473 | } |
474 | ||
bf0beec0 | 475 | retry: |
600c3b0c CH |
476 | data->ctx = blk_mq_get_ctx(q); |
477 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
dd6216bb | 478 | if (!(data->rq_flags & RQF_SCHED_TAGS)) |
600c3b0c CH |
479 | blk_mq_tag_busy(data->hctx); |
480 | ||
99e48cd6 JG |
481 | if (data->flags & BLK_MQ_REQ_RESERVED) |
482 | data->rq_flags |= RQF_RESV; | |
483 | ||
349302da JA |
484 | /* |
485 | * Try batched alloc if we want more than 1 tag. | |
486 | */ | |
487 | if (data->nr_tags > 1) { | |
5c17f45e CZ |
488 | rq = __blk_mq_alloc_requests_batch(data); |
489 | if (rq) { | |
490 | blk_mq_rq_time_init(rq, alloc_time_ns); | |
349302da | 491 | return rq; |
5c17f45e | 492 | } |
349302da JA |
493 | data->nr_tags = 1; |
494 | } | |
495 | ||
bf0beec0 ML |
496 | /* |
497 | * Waiting allocations only fail because of an inactive hctx. In that | |
498 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
499 | * should have migrated us to an online CPU by now. | |
500 | */ | |
e4cdf1a1 | 501 | tag = blk_mq_get_tag(data); |
bf0beec0 ML |
502 | if (tag == BLK_MQ_NO_TAG) { |
503 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
504 | return NULL; | |
bf0beec0 | 505 | /* |
349302da JA |
506 | * Give up the CPU and sleep for a random short time to |
507 | * ensure that thread using a realtime scheduling class | |
508 | * are migrated off the CPU, and thus off the hctx that | |
509 | * is going away. | |
bf0beec0 ML |
510 | */ |
511 | msleep(3); | |
512 | goto retry; | |
513 | } | |
47c122e3 | 514 | |
b8643d68 CZ |
515 | if (!(data->rq_flags & RQF_SCHED_TAGS)) |
516 | blk_mq_inc_active_requests(data->hctx); | |
5c17f45e CZ |
517 | rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag); |
518 | blk_mq_rq_time_init(rq, alloc_time_ns); | |
519 | return rq; | |
d2c0d383 CH |
520 | } |
521 | ||
4b6a5d9c JA |
522 | static struct request *blk_mq_rq_cache_fill(struct request_queue *q, |
523 | struct blk_plug *plug, | |
524 | blk_opf_t opf, | |
525 | blk_mq_req_flags_t flags) | |
320ae51f | 526 | { |
e6e7abff CH |
527 | struct blk_mq_alloc_data data = { |
528 | .q = q, | |
529 | .flags = flags, | |
16458cf3 | 530 | .cmd_flags = opf, |
4b6a5d9c JA |
531 | .nr_tags = plug->nr_ios, |
532 | .cached_rq = &plug->cached_rq, | |
e6e7abff | 533 | }; |
bd166ef1 | 534 | struct request *rq; |
320ae51f | 535 | |
4b6a5d9c JA |
536 | if (blk_queue_enter(q, flags)) |
537 | return NULL; | |
538 | ||
539 | plug->nr_ios = 1; | |
320ae51f | 540 | |
b90cfaed | 541 | rq = __blk_mq_alloc_requests(&data); |
4b6a5d9c JA |
542 | if (unlikely(!rq)) |
543 | blk_queue_exit(q); | |
544 | return rq; | |
545 | } | |
546 | ||
547 | static struct request *blk_mq_alloc_cached_request(struct request_queue *q, | |
548 | blk_opf_t opf, | |
549 | blk_mq_req_flags_t flags) | |
550 | { | |
551 | struct blk_plug *plug = current->plug; | |
552 | struct request *rq; | |
553 | ||
554 | if (!plug) | |
555 | return NULL; | |
40467282 | 556 | |
4b6a5d9c JA |
557 | if (rq_list_empty(plug->cached_rq)) { |
558 | if (plug->nr_ios == 1) | |
559 | return NULL; | |
560 | rq = blk_mq_rq_cache_fill(q, plug, opf, flags); | |
40467282 JC |
561 | if (!rq) |
562 | return NULL; | |
563 | } else { | |
564 | rq = rq_list_peek(&plug->cached_rq); | |
565 | if (!rq || rq->q != q) | |
566 | return NULL; | |
4b6a5d9c | 567 | |
40467282 JC |
568 | if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type) |
569 | return NULL; | |
570 | if (op_is_flush(rq->cmd_flags) != op_is_flush(opf)) | |
571 | return NULL; | |
572 | ||
573 | plug->cached_rq = rq_list_next(rq); | |
5c17f45e | 574 | blk_mq_rq_time_init(rq, 0); |
40467282 | 575 | } |
4b6a5d9c | 576 | |
4b6a5d9c JA |
577 | rq->cmd_flags = opf; |
578 | INIT_LIST_HEAD(&rq->queuelist); | |
579 | return rq; | |
580 | } | |
581 | ||
582 | struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf, | |
583 | blk_mq_req_flags_t flags) | |
584 | { | |
585 | struct request *rq; | |
586 | ||
587 | rq = blk_mq_alloc_cached_request(q, opf, flags); | |
588 | if (!rq) { | |
589 | struct blk_mq_alloc_data data = { | |
590 | .q = q, | |
591 | .flags = flags, | |
592 | .cmd_flags = opf, | |
593 | .nr_tags = 1, | |
594 | }; | |
595 | int ret; | |
596 | ||
597 | ret = blk_queue_enter(q, flags); | |
598 | if (ret) | |
599 | return ERR_PTR(ret); | |
600 | ||
601 | rq = __blk_mq_alloc_requests(&data); | |
602 | if (!rq) | |
603 | goto out_queue_exit; | |
604 | } | |
0c4de0f3 CH |
605 | rq->__data_len = 0; |
606 | rq->__sector = (sector_t) -1; | |
607 | rq->bio = rq->biotail = NULL; | |
320ae51f | 608 | return rq; |
a5ea5811 CH |
609 | out_queue_exit: |
610 | blk_queue_exit(q); | |
611 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 612 | } |
4bb659b1 | 613 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 614 | |
cd6ce148 | 615 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
16458cf3 | 616 | blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 617 | { |
e6e7abff CH |
618 | struct blk_mq_alloc_data data = { |
619 | .q = q, | |
620 | .flags = flags, | |
16458cf3 | 621 | .cmd_flags = opf, |
47c122e3 | 622 | .nr_tags = 1, |
e6e7abff | 623 | }; |
600c3b0c | 624 | u64 alloc_time_ns = 0; |
e3c5a78c | 625 | struct request *rq; |
6d2809d5 | 626 | unsigned int cpu; |
600c3b0c | 627 | unsigned int tag; |
1f5bd336 ML |
628 | int ret; |
629 | ||
600c3b0c CH |
630 | /* alloc_time includes depth and tag waits */ |
631 | if (blk_queue_rq_alloc_time(q)) | |
632 | alloc_time_ns = ktime_get_ns(); | |
633 | ||
1f5bd336 ML |
634 | /* |
635 | * If the tag allocator sleeps we could get an allocation for a | |
636 | * different hardware context. No need to complicate the low level | |
637 | * allocator for this for the rare use case of a command tied to | |
638 | * a specific queue. | |
639 | */ | |
6ee858a3 KS |
640 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) || |
641 | WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED))) | |
1f5bd336 ML |
642 | return ERR_PTR(-EINVAL); |
643 | ||
644 | if (hctx_idx >= q->nr_hw_queues) | |
645 | return ERR_PTR(-EIO); | |
646 | ||
3a0a5299 | 647 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
648 | if (ret) |
649 | return ERR_PTR(ret); | |
650 | ||
c8712c6a CH |
651 | /* |
652 | * Check if the hardware context is actually mapped to anything. | |
653 | * If not tell the caller that it should skip this queue. | |
654 | */ | |
a5ea5811 | 655 | ret = -EXDEV; |
4e5cc99e | 656 | data.hctx = xa_load(&q->hctx_table, hctx_idx); |
e6e7abff | 657 | if (!blk_mq_hw_queue_mapped(data.hctx)) |
a5ea5811 | 658 | goto out_queue_exit; |
e6e7abff | 659 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
14dc7a18 BVA |
660 | if (cpu >= nr_cpu_ids) |
661 | goto out_queue_exit; | |
e6e7abff | 662 | data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 663 | |
dd6216bb CH |
664 | if (q->elevator) |
665 | data.rq_flags |= RQF_SCHED_TAGS; | |
781dd830 | 666 | else |
dd6216bb | 667 | blk_mq_tag_busy(data.hctx); |
600c3b0c | 668 | |
99e48cd6 JG |
669 | if (flags & BLK_MQ_REQ_RESERVED) |
670 | data.rq_flags |= RQF_RESV; | |
671 | ||
a5ea5811 | 672 | ret = -EWOULDBLOCK; |
600c3b0c CH |
673 | tag = blk_mq_get_tag(&data); |
674 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 675 | goto out_queue_exit; |
b8643d68 CZ |
676 | if (!(data.rq_flags & RQF_SCHED_TAGS)) |
677 | blk_mq_inc_active_requests(data.hctx); | |
5c17f45e CZ |
678 | rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag); |
679 | blk_mq_rq_time_init(rq, alloc_time_ns); | |
e3c5a78c JG |
680 | rq->__data_len = 0; |
681 | rq->__sector = (sector_t) -1; | |
682 | rq->bio = rq->biotail = NULL; | |
683 | return rq; | |
600c3b0c | 684 | |
a5ea5811 CH |
685 | out_queue_exit: |
686 | blk_queue_exit(q); | |
687 | return ERR_PTR(ret); | |
1f5bd336 ML |
688 | } |
689 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
690 | ||
e5c0ca13 CZ |
691 | static void blk_mq_finish_request(struct request *rq) |
692 | { | |
693 | struct request_queue *q = rq->q; | |
694 | ||
695 | if (rq->rq_flags & RQF_USE_SCHED) { | |
696 | q->elevator->type->ops.finish_request(rq); | |
697 | /* | |
698 | * For postflush request that may need to be | |
699 | * completed twice, we should clear this flag | |
700 | * to avoid double finish_request() on the rq. | |
701 | */ | |
702 | rq->rq_flags &= ~RQF_USE_SCHED; | |
703 | } | |
704 | } | |
705 | ||
12f5b931 KB |
706 | static void __blk_mq_free_request(struct request *rq) |
707 | { | |
708 | struct request_queue *q = rq->q; | |
709 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 710 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
711 | const int sched_tag = rq->internal_tag; |
712 | ||
a892c8d5 | 713 | blk_crypto_free_request(rq); |
986d413b | 714 | blk_pm_mark_last_busy(rq); |
ea4f995e | 715 | rq->mq_hctx = NULL; |
ddad5933 | 716 | |
b8643d68 CZ |
717 | if (rq->tag != BLK_MQ_NO_TAG) { |
718 | blk_mq_dec_active_requests(hctx); | |
cae740a0 | 719 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
b8643d68 | 720 | } |
76647368 | 721 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 722 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
723 | blk_mq_sched_restart(hctx); |
724 | blk_queue_exit(q); | |
725 | } | |
726 | ||
6af54051 | 727 | void blk_mq_free_request(struct request *rq) |
320ae51f | 728 | { |
320ae51f | 729 | struct request_queue *q = rq->q; |
6af54051 | 730 | |
e5c0ca13 | 731 | blk_mq_finish_request(rq); |
320ae51f | 732 | |
7beb2f84 | 733 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 734 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 735 | |
a7905043 | 736 | rq_qos_done(q, rq); |
0d2602ca | 737 | |
12f5b931 | 738 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 739 | if (req_ref_put_and_test(rq)) |
12f5b931 | 740 | __blk_mq_free_request(rq); |
320ae51f | 741 | } |
1a3b595a | 742 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 743 | |
47c122e3 | 744 | void blk_mq_free_plug_rqs(struct blk_plug *plug) |
320ae51f | 745 | { |
013a7f95 | 746 | struct request *rq; |
fe1f4526 | 747 | |
c5fc7b93 | 748 | while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) |
47c122e3 | 749 | blk_mq_free_request(rq); |
47c122e3 | 750 | } |
522a7775 | 751 | |
22350ad7 CH |
752 | void blk_dump_rq_flags(struct request *rq, char *msg) |
753 | { | |
754 | printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, | |
f3fa33ac | 755 | rq->q->disk ? rq->q->disk->disk_name : "?", |
16458cf3 | 756 | (__force unsigned long long) rq->cmd_flags); |
22350ad7 CH |
757 | |
758 | printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", | |
759 | (unsigned long long)blk_rq_pos(rq), | |
760 | blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); | |
761 | printk(KERN_INFO " bio %p, biotail %p, len %u\n", | |
762 | rq->bio, rq->biotail, blk_rq_bytes(rq)); | |
763 | } | |
764 | EXPORT_SYMBOL(blk_dump_rq_flags); | |
765 | ||
9be3e06f JA |
766 | static void req_bio_endio(struct request *rq, struct bio *bio, |
767 | unsigned int nbytes, blk_status_t error) | |
768 | { | |
478eb72b | 769 | if (unlikely(error)) { |
9be3e06f | 770 | bio->bi_status = error; |
478eb72b | 771 | } else if (req_op(rq) == REQ_OP_ZONE_APPEND) { |
9be3e06f JA |
772 | /* |
773 | * Partial zone append completions cannot be supported as the | |
774 | * BIO fragments may end up not being written sequentially. | |
748dc0b6 DLM |
775 | * For such case, force the completed nbytes to be equal to |
776 | * the BIO size so that bio_advance() sets the BIO remaining | |
777 | * size to 0 and we end up calling bio_endio() before returning. | |
9be3e06f | 778 | */ |
748dc0b6 | 779 | if (bio->bi_iter.bi_size != nbytes) { |
9be3e06f | 780 | bio->bi_status = BLK_STS_IOERR; |
748dc0b6 DLM |
781 | nbytes = bio->bi_iter.bi_size; |
782 | } else { | |
9be3e06f | 783 | bio->bi_iter.bi_sector = rq->__sector; |
748dc0b6 | 784 | } |
9be3e06f JA |
785 | } |
786 | ||
478eb72b PB |
787 | bio_advance(bio, nbytes); |
788 | ||
789 | if (unlikely(rq->rq_flags & RQF_QUIET)) | |
790 | bio_set_flag(bio, BIO_QUIET); | |
9be3e06f JA |
791 | /* don't actually finish bio if it's part of flush sequence */ |
792 | if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ)) | |
793 | bio_endio(bio); | |
794 | } | |
795 | ||
796 | static void blk_account_io_completion(struct request *req, unsigned int bytes) | |
797 | { | |
798 | if (req->part && blk_do_io_stat(req)) { | |
799 | const int sgrp = op_stat_group(req_op(req)); | |
800 | ||
801 | part_stat_lock(); | |
802 | part_stat_add(req->part, sectors[sgrp], bytes >> 9); | |
803 | part_stat_unlock(); | |
804 | } | |
805 | } | |
806 | ||
0d7a29a2 CH |
807 | static void blk_print_req_error(struct request *req, blk_status_t status) |
808 | { | |
809 | printk_ratelimited(KERN_ERR | |
810 | "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x " | |
811 | "phys_seg %u prio class %u\n", | |
812 | blk_status_to_str(status), | |
f3fa33ac | 813 | req->q->disk ? req->q->disk->disk_name : "?", |
16458cf3 BVA |
814 | blk_rq_pos(req), (__force u32)req_op(req), |
815 | blk_op_str(req_op(req)), | |
816 | (__force u32)(req->cmd_flags & ~REQ_OP_MASK), | |
0d7a29a2 CH |
817 | req->nr_phys_segments, |
818 | IOPRIO_PRIO_CLASS(req->ioprio)); | |
819 | } | |
820 | ||
5581a5dd JA |
821 | /* |
822 | * Fully end IO on a request. Does not support partial completions, or | |
823 | * errors. | |
824 | */ | |
825 | static void blk_complete_request(struct request *req) | |
826 | { | |
827 | const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0; | |
828 | int total_bytes = blk_rq_bytes(req); | |
829 | struct bio *bio = req->bio; | |
830 | ||
831 | trace_block_rq_complete(req, BLK_STS_OK, total_bytes); | |
832 | ||
833 | if (!bio) | |
834 | return; | |
835 | ||
836 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
837 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ) | |
838 | req->q->integrity.profile->complete_fn(req, total_bytes); | |
839 | #endif | |
840 | ||
9cd1e566 EB |
841 | /* |
842 | * Upper layers may call blk_crypto_evict_key() anytime after the last | |
843 | * bio_endio(). Therefore, the keyslot must be released before that. | |
844 | */ | |
845 | blk_crypto_rq_put_keyslot(req); | |
846 | ||
5581a5dd JA |
847 | blk_account_io_completion(req, total_bytes); |
848 | ||
849 | do { | |
850 | struct bio *next = bio->bi_next; | |
851 | ||
852 | /* Completion has already been traced */ | |
853 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
a12821d5 PR |
854 | |
855 | if (req_op(req) == REQ_OP_ZONE_APPEND) | |
856 | bio->bi_iter.bi_sector = req->__sector; | |
857 | ||
5581a5dd JA |
858 | if (!is_flush) |
859 | bio_endio(bio); | |
860 | bio = next; | |
861 | } while (bio); | |
862 | ||
863 | /* | |
864 | * Reset counters so that the request stacking driver | |
865 | * can find how many bytes remain in the request | |
866 | * later. | |
867 | */ | |
ab3e1d3b JA |
868 | if (!req->end_io) { |
869 | req->bio = NULL; | |
870 | req->__data_len = 0; | |
871 | } | |
5581a5dd JA |
872 | } |
873 | ||
9be3e06f JA |
874 | /** |
875 | * blk_update_request - Complete multiple bytes without completing the request | |
876 | * @req: the request being processed | |
877 | * @error: block status code | |
878 | * @nr_bytes: number of bytes to complete for @req | |
879 | * | |
880 | * Description: | |
881 | * Ends I/O on a number of bytes attached to @req, but doesn't complete | |
882 | * the request structure even if @req doesn't have leftover. | |
883 | * If @req has leftover, sets it up for the next range of segments. | |
884 | * | |
885 | * Passing the result of blk_rq_bytes() as @nr_bytes guarantees | |
886 | * %false return from this function. | |
887 | * | |
888 | * Note: | |
889 | * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function | |
890 | * except in the consistency check at the end of this function. | |
891 | * | |
892 | * Return: | |
893 | * %false - this request doesn't have any more data | |
894 | * %true - this request has more data | |
895 | **/ | |
896 | bool blk_update_request(struct request *req, blk_status_t error, | |
897 | unsigned int nr_bytes) | |
898 | { | |
899 | int total_bytes; | |
900 | ||
8a7d267b | 901 | trace_block_rq_complete(req, error, nr_bytes); |
9be3e06f JA |
902 | |
903 | if (!req->bio) | |
904 | return false; | |
905 | ||
906 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
907 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && | |
908 | error == BLK_STS_OK) | |
909 | req->q->integrity.profile->complete_fn(req, nr_bytes); | |
910 | #endif | |
911 | ||
9cd1e566 EB |
912 | /* |
913 | * Upper layers may call blk_crypto_evict_key() anytime after the last | |
914 | * bio_endio(). Therefore, the keyslot must be released before that. | |
915 | */ | |
916 | if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req)) | |
917 | __blk_crypto_rq_put_keyslot(req); | |
918 | ||
9be3e06f | 919 | if (unlikely(error && !blk_rq_is_passthrough(req) && |
3d973a76 CH |
920 | !(req->rq_flags & RQF_QUIET)) && |
921 | !test_bit(GD_DEAD, &req->q->disk->state)) { | |
9be3e06f | 922 | blk_print_req_error(req, error); |
d5869fdc YS |
923 | trace_block_rq_error(req, error, nr_bytes); |
924 | } | |
9be3e06f JA |
925 | |
926 | blk_account_io_completion(req, nr_bytes); | |
927 | ||
928 | total_bytes = 0; | |
929 | while (req->bio) { | |
930 | struct bio *bio = req->bio; | |
931 | unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); | |
932 | ||
933 | if (bio_bytes == bio->bi_iter.bi_size) | |
934 | req->bio = bio->bi_next; | |
935 | ||
936 | /* Completion has already been traced */ | |
937 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
938 | req_bio_endio(req, bio, bio_bytes, error); | |
939 | ||
940 | total_bytes += bio_bytes; | |
941 | nr_bytes -= bio_bytes; | |
942 | ||
943 | if (!nr_bytes) | |
944 | break; | |
945 | } | |
946 | ||
947 | /* | |
948 | * completely done | |
949 | */ | |
950 | if (!req->bio) { | |
951 | /* | |
952 | * Reset counters so that the request stacking driver | |
953 | * can find how many bytes remain in the request | |
954 | * later. | |
955 | */ | |
956 | req->__data_len = 0; | |
957 | return false; | |
958 | } | |
959 | ||
960 | req->__data_len -= total_bytes; | |
961 | ||
962 | /* update sector only for requests with clear definition of sector */ | |
963 | if (!blk_rq_is_passthrough(req)) | |
964 | req->__sector += total_bytes >> 9; | |
965 | ||
966 | /* mixed attributes always follow the first bio */ | |
967 | if (req->rq_flags & RQF_MIXED_MERGE) { | |
968 | req->cmd_flags &= ~REQ_FAILFAST_MASK; | |
969 | req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; | |
970 | } | |
971 | ||
972 | if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { | |
973 | /* | |
974 | * If total number of sectors is less than the first segment | |
975 | * size, something has gone terribly wrong. | |
976 | */ | |
977 | if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { | |
978 | blk_dump_rq_flags(req, "request botched"); | |
979 | req->__data_len = blk_rq_cur_bytes(req); | |
980 | } | |
981 | ||
982 | /* recalculate the number of segments */ | |
983 | req->nr_phys_segments = blk_recalc_rq_segments(req); | |
984 | } | |
985 | ||
986 | return true; | |
987 | } | |
988 | EXPORT_SYMBOL_GPL(blk_update_request); | |
989 | ||
450b7879 CH |
990 | static inline void blk_account_io_done(struct request *req, u64 now) |
991 | { | |
5a80bd07 HC |
992 | trace_block_io_done(req); |
993 | ||
450b7879 CH |
994 | /* |
995 | * Account IO completion. flush_rq isn't accounted as a | |
996 | * normal IO on queueing nor completion. Accounting the | |
997 | * containing request is enough. | |
998 | */ | |
999 | if (blk_do_io_stat(req) && req->part && | |
06965037 CK |
1000 | !(req->rq_flags & RQF_FLUSH_SEQ)) { |
1001 | const int sgrp = op_stat_group(req_op(req)); | |
450b7879 | 1002 | |
06965037 CK |
1003 | part_stat_lock(); |
1004 | update_io_ticks(req->part, jiffies, true); | |
1005 | part_stat_inc(req->part, ios[sgrp]); | |
1006 | part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns); | |
1007 | part_stat_unlock(); | |
1008 | } | |
450b7879 CH |
1009 | } |
1010 | ||
1011 | static inline void blk_account_io_start(struct request *req) | |
1012 | { | |
5a80bd07 HC |
1013 | trace_block_io_start(req); |
1014 | ||
e165fb4d CK |
1015 | if (blk_do_io_stat(req)) { |
1016 | /* | |
1017 | * All non-passthrough requests are created from a bio with one | |
1018 | * exception: when a flush command that is part of a flush sequence | |
1019 | * generated by the state machine in blk-flush.c is cloned onto the | |
1020 | * lower device by dm-multipath we can get here without a bio. | |
1021 | */ | |
1022 | if (req->bio) | |
1023 | req->part = req->bio->bi_bdev; | |
1024 | else | |
1025 | req->part = req->q->disk->part0; | |
1026 | ||
1027 | part_stat_lock(); | |
1028 | update_io_ticks(req->part, jiffies, false); | |
1029 | part_stat_unlock(); | |
1030 | } | |
450b7879 CH |
1031 | } |
1032 | ||
f794f335 | 1033 | static inline void __blk_mq_end_request_acct(struct request *rq, u64 now) |
320ae51f | 1034 | { |
54bdd67d | 1035 | if (rq->rq_flags & RQF_STATS) |
522a7775 | 1036 | blk_stat_add(rq, now); |
4bc6339a | 1037 | |
87890092 | 1038 | blk_mq_sched_completed_request(rq, now); |
522a7775 | 1039 | blk_account_io_done(rq, now); |
f794f335 | 1040 | } |
522a7775 | 1041 | |
f794f335 JA |
1042 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
1043 | { | |
1044 | if (blk_mq_need_time_stamp(rq)) | |
1045 | __blk_mq_end_request_acct(rq, ktime_get_ns()); | |
0d11e6ac | 1046 | |
e5c0ca13 CZ |
1047 | blk_mq_finish_request(rq); |
1048 | ||
91b63639 | 1049 | if (rq->end_io) { |
a7905043 | 1050 | rq_qos_done(rq->q, rq); |
de671d61 JA |
1051 | if (rq->end_io(rq, error) == RQ_END_IO_FREE) |
1052 | blk_mq_free_request(rq); | |
91b63639 | 1053 | } else { |
320ae51f | 1054 | blk_mq_free_request(rq); |
91b63639 | 1055 | } |
320ae51f | 1056 | } |
c8a446ad | 1057 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 1058 | |
2a842aca | 1059 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
1060 | { |
1061 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
1062 | BUG(); | |
c8a446ad | 1063 | __blk_mq_end_request(rq, error); |
63151a44 | 1064 | } |
c8a446ad | 1065 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 1066 | |
f794f335 JA |
1067 | #define TAG_COMP_BATCH 32 |
1068 | ||
1069 | static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx, | |
1070 | int *tag_array, int nr_tags) | |
1071 | { | |
1072 | struct request_queue *q = hctx->queue; | |
1073 | ||
b8643d68 | 1074 | blk_mq_sub_active_requests(hctx, nr_tags); |
3b87c6ea | 1075 | |
f794f335 JA |
1076 | blk_mq_put_tags(hctx->tags, tag_array, nr_tags); |
1077 | percpu_ref_put_many(&q->q_usage_counter, nr_tags); | |
1078 | } | |
1079 | ||
1080 | void blk_mq_end_request_batch(struct io_comp_batch *iob) | |
1081 | { | |
1082 | int tags[TAG_COMP_BATCH], nr_tags = 0; | |
02f7eab0 | 1083 | struct blk_mq_hw_ctx *cur_hctx = NULL; |
f794f335 JA |
1084 | struct request *rq; |
1085 | u64 now = 0; | |
1086 | ||
1087 | if (iob->need_ts) | |
1088 | now = ktime_get_ns(); | |
1089 | ||
1090 | while ((rq = rq_list_pop(&iob->req_list)) != NULL) { | |
1091 | prefetch(rq->bio); | |
1092 | prefetch(rq->rq_next); | |
1093 | ||
5581a5dd | 1094 | blk_complete_request(rq); |
f794f335 JA |
1095 | if (iob->need_ts) |
1096 | __blk_mq_end_request_acct(rq, now); | |
1097 | ||
e5c0ca13 CZ |
1098 | blk_mq_finish_request(rq); |
1099 | ||
98b26a0e JA |
1100 | rq_qos_done(rq->q, rq); |
1101 | ||
ab3e1d3b JA |
1102 | /* |
1103 | * If end_io handler returns NONE, then it still has | |
1104 | * ownership of the request. | |
1105 | */ | |
1106 | if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE) | |
1107 | continue; | |
1108 | ||
f794f335 | 1109 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 1110 | if (!req_ref_put_and_test(rq)) |
f794f335 JA |
1111 | continue; |
1112 | ||
1113 | blk_crypto_free_request(rq); | |
1114 | blk_pm_mark_last_busy(rq); | |
f794f335 | 1115 | |
02f7eab0 JA |
1116 | if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) { |
1117 | if (cur_hctx) | |
1118 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); | |
f794f335 | 1119 | nr_tags = 0; |
02f7eab0 | 1120 | cur_hctx = rq->mq_hctx; |
f794f335 JA |
1121 | } |
1122 | tags[nr_tags++] = rq->tag; | |
f794f335 JA |
1123 | } |
1124 | ||
1125 | if (nr_tags) | |
02f7eab0 | 1126 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); |
f794f335 JA |
1127 | } |
1128 | EXPORT_SYMBOL_GPL(blk_mq_end_request_batch); | |
1129 | ||
f9ab4918 | 1130 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 1131 | { |
f9ab4918 SAS |
1132 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
1133 | struct request *rq, *next; | |
c3077b5d | 1134 | |
f9ab4918 | 1135 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 1136 | rq->q->mq_ops->complete(rq); |
320ae51f | 1137 | } |
320ae51f | 1138 | |
f9ab4918 | 1139 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 1140 | { |
f9ab4918 | 1141 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
1142 | } |
1143 | ||
c3077b5d CH |
1144 | static int blk_softirq_cpu_dead(unsigned int cpu) |
1145 | { | |
f9ab4918 | 1146 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
1147 | return 0; |
1148 | } | |
1149 | ||
40d09b53 | 1150 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 1151 | { |
f9ab4918 | 1152 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
1153 | } |
1154 | ||
96339526 CH |
1155 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
1156 | { | |
1157 | int cpu = raw_smp_processor_id(); | |
1158 | ||
1159 | if (!IS_ENABLED(CONFIG_SMP) || | |
1160 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
1161 | return false; | |
71425189 SAS |
1162 | /* |
1163 | * With force threaded interrupts enabled, raising softirq from an SMP | |
1164 | * function call will always result in waking the ksoftirqd thread. | |
1165 | * This is probably worse than completing the request on a different | |
1166 | * cache domain. | |
1167 | */ | |
91cc470e | 1168 | if (force_irqthreads()) |
71425189 | 1169 | return false; |
96339526 CH |
1170 | |
1171 | /* same CPU or cache domain? Complete locally */ | |
1172 | if (cpu == rq->mq_ctx->cpu || | |
1173 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
1174 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
1175 | return false; | |
1176 | ||
1177 | /* don't try to IPI to an offline CPU */ | |
1178 | return cpu_online(rq->mq_ctx->cpu); | |
1179 | } | |
1180 | ||
f9ab4918 SAS |
1181 | static void blk_mq_complete_send_ipi(struct request *rq) |
1182 | { | |
f9ab4918 SAS |
1183 | unsigned int cpu; |
1184 | ||
1185 | cpu = rq->mq_ctx->cpu; | |
660e802c CZ |
1186 | if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu))) |
1187 | smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu)); | |
f9ab4918 SAS |
1188 | } |
1189 | ||
1190 | static void blk_mq_raise_softirq(struct request *rq) | |
1191 | { | |
1192 | struct llist_head *list; | |
1193 | ||
1194 | preempt_disable(); | |
1195 | list = this_cpu_ptr(&blk_cpu_done); | |
1196 | if (llist_add(&rq->ipi_list, list)) | |
1197 | raise_softirq(BLOCK_SOFTIRQ); | |
1198 | preempt_enable(); | |
1199 | } | |
1200 | ||
40d09b53 | 1201 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 1202 | { |
af78ff7c | 1203 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 1204 | |
4ab32bf3 | 1205 | /* |
f168420c LS |
1206 | * For request which hctx has only one ctx mapping, |
1207 | * or a polled request, always complete locally, | |
1208 | * it's pointless to redirect the completion. | |
4ab32bf3 | 1209 | */ |
30654614 ET |
1210 | if ((rq->mq_hctx->nr_ctx == 1 && |
1211 | rq->mq_ctx->cpu == raw_smp_processor_id()) || | |
1212 | rq->cmd_flags & REQ_POLLED) | |
40d09b53 | 1213 | return false; |
38535201 | 1214 | |
96339526 | 1215 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
1216 | blk_mq_complete_send_ipi(rq); |
1217 | return true; | |
3d6efbf6 | 1218 | } |
40d09b53 | 1219 | |
f9ab4918 SAS |
1220 | if (rq->q->nr_hw_queues == 1) { |
1221 | blk_mq_raise_softirq(rq); | |
1222 | return true; | |
1223 | } | |
1224 | return false; | |
40d09b53 CH |
1225 | } |
1226 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
1227 | ||
1228 | /** | |
1229 | * blk_mq_complete_request - end I/O on a request | |
1230 | * @rq: the request being processed | |
1231 | * | |
1232 | * Description: | |
1233 | * Complete a request by scheduling the ->complete_rq operation. | |
1234 | **/ | |
1235 | void blk_mq_complete_request(struct request *rq) | |
1236 | { | |
1237 | if (!blk_mq_complete_request_remote(rq)) | |
1238 | rq->q->mq_ops->complete(rq); | |
320ae51f | 1239 | } |
15f73f5b | 1240 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 1241 | |
105663f7 AA |
1242 | /** |
1243 | * blk_mq_start_request - Start processing a request | |
1244 | * @rq: Pointer to request to be started | |
1245 | * | |
1246 | * Function used by device drivers to notify the block layer that a request | |
1247 | * is going to be processed now, so blk layer can do proper initializations | |
1248 | * such as starting the timeout timer. | |
1249 | */ | |
e2490073 | 1250 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
1251 | { |
1252 | struct request_queue *q = rq->q; | |
1253 | ||
a54895fa | 1254 | trace_block_rq_issue(rq); |
320ae51f | 1255 | |
847c5bcd KK |
1256 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) && |
1257 | !blk_rq_is_passthrough(rq)) { | |
4cddeaca | 1258 | rq->io_start_time_ns = ktime_get_ns(); |
3d244306 | 1259 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 1260 | rq->rq_flags |= RQF_STATS; |
a7905043 | 1261 | rq_qos_issue(q, rq); |
cf43e6be JA |
1262 | } |
1263 | ||
1d9bd516 | 1264 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 1265 | |
1d9bd516 | 1266 | blk_add_timer(rq); |
12f5b931 | 1267 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
217b613a | 1268 | rq->mq_hctx->tags->rqs[rq->tag] = rq; |
49f5baa5 | 1269 | |
54d4e6ab MG |
1270 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
1271 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
1272 | q->integrity.profile->prepare_fn(rq); | |
1273 | #endif | |
3e08773c | 1274 | if (rq->bio && rq->bio->bi_opf & REQ_POLLED) |
f6c80cff | 1275 | WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num); |
320ae51f | 1276 | } |
e2490073 | 1277 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 1278 | |
a327c341 ML |
1279 | /* |
1280 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple | |
1281 | * queues. This is important for md arrays to benefit from merging | |
1282 | * requests. | |
1283 | */ | |
1284 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
1285 | { | |
1286 | if (plug->multiple_queues) | |
1287 | return BLK_MAX_REQUEST_COUNT * 2; | |
1288 | return BLK_MAX_REQUEST_COUNT; | |
1289 | } | |
1290 | ||
1291 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) | |
1292 | { | |
1293 | struct request *last = rq_list_peek(&plug->mq_list); | |
1294 | ||
1295 | if (!plug->rq_count) { | |
1296 | trace_block_plug(rq->q); | |
1297 | } else if (plug->rq_count >= blk_plug_max_rq_count(plug) || | |
1298 | (!blk_queue_nomerges(rq->q) && | |
1299 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1300 | blk_mq_flush_plug_list(plug, false); | |
878eb6e4 | 1301 | last = NULL; |
a327c341 ML |
1302 | trace_block_plug(rq->q); |
1303 | } | |
1304 | ||
1305 | if (!plug->multiple_queues && last && last->q != rq->q) | |
1306 | plug->multiple_queues = true; | |
c6b7a3a2 ML |
1307 | /* |
1308 | * Any request allocated from sched tags can't be issued to | |
1309 | * ->queue_rqs() directly | |
1310 | */ | |
1311 | if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS)) | |
a327c341 ML |
1312 | plug->has_elevator = true; |
1313 | rq->rq_next = NULL; | |
1314 | rq_list_add(&plug->mq_list, rq); | |
1315 | plug->rq_count++; | |
1316 | } | |
1317 | ||
4054cff9 CH |
1318 | /** |
1319 | * blk_execute_rq_nowait - insert a request to I/O scheduler for execution | |
4054cff9 CH |
1320 | * @rq: request to insert |
1321 | * @at_head: insert request at head or tail of queue | |
4054cff9 CH |
1322 | * |
1323 | * Description: | |
1324 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1325 | * for execution. Don't wait for completion. | |
1326 | * | |
1327 | * Note: | |
1328 | * This function will invoke @done directly if the queue is dead. | |
1329 | */ | |
e2e53086 | 1330 | void blk_execute_rq_nowait(struct request *rq, bool at_head) |
4054cff9 | 1331 | { |
f0dbe6e8 CH |
1332 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
1333 | ||
ae948fd6 CH |
1334 | WARN_ON(irqs_disabled()); |
1335 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 | 1336 | |
ae948fd6 | 1337 | blk_account_io_start(rq); |
110fdb44 PR |
1338 | |
1339 | /* | |
1340 | * As plugging can be enabled for passthrough requests on a zoned | |
1341 | * device, directly accessing the plug instead of using blk_mq_plug() | |
1342 | * should not have any consequences. | |
1343 | */ | |
f0dbe6e8 | 1344 | if (current->plug && !at_head) { |
ae948fd6 | 1345 | blk_add_rq_to_plug(current->plug, rq); |
f0dbe6e8 CH |
1346 | return; |
1347 | } | |
1348 | ||
710fa378 | 1349 | blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); |
65a558f6 | 1350 | blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING); |
4054cff9 CH |
1351 | } |
1352 | EXPORT_SYMBOL_GPL(blk_execute_rq_nowait); | |
1353 | ||
32ac5a9b CH |
1354 | struct blk_rq_wait { |
1355 | struct completion done; | |
1356 | blk_status_t ret; | |
1357 | }; | |
1358 | ||
de671d61 | 1359 | static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret) |
32ac5a9b CH |
1360 | { |
1361 | struct blk_rq_wait *wait = rq->end_io_data; | |
1362 | ||
1363 | wait->ret = ret; | |
1364 | complete(&wait->done); | |
de671d61 | 1365 | return RQ_END_IO_NONE; |
32ac5a9b CH |
1366 | } |
1367 | ||
c6e99ea4 | 1368 | bool blk_rq_is_poll(struct request *rq) |
4054cff9 CH |
1369 | { |
1370 | if (!rq->mq_hctx) | |
1371 | return false; | |
1372 | if (rq->mq_hctx->type != HCTX_TYPE_POLL) | |
1373 | return false; | |
4054cff9 CH |
1374 | return true; |
1375 | } | |
c6e99ea4 | 1376 | EXPORT_SYMBOL_GPL(blk_rq_is_poll); |
4054cff9 CH |
1377 | |
1378 | static void blk_rq_poll_completion(struct request *rq, struct completion *wait) | |
1379 | { | |
1380 | do { | |
f6c80cff | 1381 | blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0); |
4054cff9 CH |
1382 | cond_resched(); |
1383 | } while (!completion_done(wait)); | |
1384 | } | |
1385 | ||
1386 | /** | |
1387 | * blk_execute_rq - insert a request into queue for execution | |
4054cff9 CH |
1388 | * @rq: request to insert |
1389 | * @at_head: insert request at head or tail of queue | |
1390 | * | |
1391 | * Description: | |
1392 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1393 | * for execution and wait for completion. | |
1394 | * Return: The blk_status_t result provided to blk_mq_end_request(). | |
1395 | */ | |
b84ba30b | 1396 | blk_status_t blk_execute_rq(struct request *rq, bool at_head) |
4054cff9 | 1397 | { |
f0dbe6e8 | 1398 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
32ac5a9b CH |
1399 | struct blk_rq_wait wait = { |
1400 | .done = COMPLETION_INITIALIZER_ONSTACK(wait.done), | |
1401 | }; | |
4054cff9 | 1402 | |
ae948fd6 CH |
1403 | WARN_ON(irqs_disabled()); |
1404 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 CH |
1405 | |
1406 | rq->end_io_data = &wait; | |
ae948fd6 | 1407 | rq->end_io = blk_end_sync_rq; |
4054cff9 | 1408 | |
ae948fd6 | 1409 | blk_account_io_start(rq); |
710fa378 | 1410 | blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); |
f0dbe6e8 | 1411 | blk_mq_run_hw_queue(hctx, false); |
4054cff9 | 1412 | |
ae948fd6 | 1413 | if (blk_rq_is_poll(rq)) { |
32ac5a9b | 1414 | blk_rq_poll_completion(rq, &wait.done); |
ae948fd6 CH |
1415 | } else { |
1416 | /* | |
1417 | * Prevent hang_check timer from firing at us during very long | |
1418 | * I/O | |
1419 | */ | |
1420 | unsigned long hang_check = sysctl_hung_task_timeout_secs; | |
1421 | ||
1422 | if (hang_check) | |
32ac5a9b | 1423 | while (!wait_for_completion_io_timeout(&wait.done, |
ae948fd6 CH |
1424 | hang_check * (HZ/2))) |
1425 | ; | |
1426 | else | |
32ac5a9b | 1427 | wait_for_completion_io(&wait.done); |
ae948fd6 | 1428 | } |
4054cff9 | 1429 | |
32ac5a9b | 1430 | return wait.ret; |
4054cff9 CH |
1431 | } |
1432 | EXPORT_SYMBOL(blk_execute_rq); | |
1433 | ||
ed0791b2 | 1434 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
1435 | { |
1436 | struct request_queue *q = rq->q; | |
1437 | ||
923218f6 ML |
1438 | blk_mq_put_driver_tag(rq); |
1439 | ||
a54895fa | 1440 | trace_block_rq_requeue(rq); |
a7905043 | 1441 | rq_qos_requeue(q, rq); |
49f5baa5 | 1442 | |
12f5b931 KB |
1443 | if (blk_mq_request_started(rq)) { |
1444 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 1445 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 1446 | } |
320ae51f JA |
1447 | } |
1448 | ||
2b053aca | 1449 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 1450 | { |
214a4418 | 1451 | struct request_queue *q = rq->q; |
9a67aa52 | 1452 | unsigned long flags; |
214a4418 | 1453 | |
ed0791b2 | 1454 | __blk_mq_requeue_request(rq); |
ed0791b2 | 1455 | |
105976f5 ML |
1456 | /* this request will be re-inserted to io scheduler queue */ |
1457 | blk_mq_sched_requeue_request(rq); | |
1458 | ||
9a67aa52 CH |
1459 | spin_lock_irqsave(&q->requeue_lock, flags); |
1460 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
1461 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
214a4418 CH |
1462 | |
1463 | if (kick_requeue_list) | |
1464 | blk_mq_kick_requeue_list(q); | |
ed0791b2 CH |
1465 | } |
1466 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
1467 | ||
6fca6a61 CH |
1468 | static void blk_mq_requeue_work(struct work_struct *work) |
1469 | { | |
1470 | struct request_queue *q = | |
2849450a | 1471 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 | 1472 | LIST_HEAD(rq_list); |
9a67aa52 CH |
1473 | LIST_HEAD(flush_list); |
1474 | struct request *rq; | |
6fca6a61 | 1475 | |
18e9781d | 1476 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 1477 | list_splice_init(&q->requeue_list, &rq_list); |
9a67aa52 | 1478 | list_splice_init(&q->flush_list, &flush_list); |
18e9781d | 1479 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 | 1480 | |
9a67aa52 CH |
1481 | while (!list_empty(&rq_list)) { |
1482 | rq = list_entry(rq_list.next, struct request, queuelist); | |
aef1897c | 1483 | /* |
a1e948b8 CH |
1484 | * If RQF_DONTPREP ist set, the request has been started by the |
1485 | * driver already and might have driver-specific data allocated | |
1486 | * already. Insert it into the hctx dispatch list to avoid | |
1487 | * block layer merges for the request. | |
aef1897c | 1488 | */ |
a1e948b8 | 1489 | if (rq->rq_flags & RQF_DONTPREP) { |
a1e948b8 | 1490 | list_del_init(&rq->queuelist); |
2b597613 | 1491 | blk_mq_request_bypass_insert(rq, 0); |
9a67aa52 | 1492 | } else { |
a1e948b8 | 1493 | list_del_init(&rq->queuelist); |
710fa378 | 1494 | blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD); |
a1e948b8 | 1495 | } |
6fca6a61 CH |
1496 | } |
1497 | ||
9a67aa52 CH |
1498 | while (!list_empty(&flush_list)) { |
1499 | rq = list_entry(flush_list.next, struct request, queuelist); | |
6fca6a61 | 1500 | list_del_init(&rq->queuelist); |
710fa378 | 1501 | blk_mq_insert_request(rq, 0); |
6fca6a61 CH |
1502 | } |
1503 | ||
52d7f1b5 | 1504 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
1505 | } |
1506 | ||
6fca6a61 CH |
1507 | void blk_mq_kick_requeue_list(struct request_queue *q) |
1508 | { | |
ae943d20 | 1509 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
1510 | } |
1511 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
1512 | ||
2849450a MS |
1513 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
1514 | unsigned long msecs) | |
1515 | { | |
d4acf365 BVA |
1516 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
1517 | msecs_to_jiffies(msecs)); | |
2849450a MS |
1518 | } |
1519 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
1520 | ||
0e4237ae ML |
1521 | static bool blk_is_flush_data_rq(struct request *rq) |
1522 | { | |
1523 | return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq); | |
1524 | } | |
1525 | ||
2dd6532e | 1526 | static bool blk_mq_rq_inflight(struct request *rq, void *priv) |
ae879912 JA |
1527 | { |
1528 | /* | |
8ab30a33 JG |
1529 | * If we find a request that isn't idle we know the queue is busy |
1530 | * as it's checked in the iter. | |
1531 | * Return false to stop the iteration. | |
0e4237ae ML |
1532 | * |
1533 | * In case of queue quiesce, if one flush data request is completed, | |
1534 | * don't count it as inflight given the flush sequence is suspended, | |
1535 | * and the original flush data request is invisible to driver, just | |
1536 | * like other pending requests because of quiesce | |
ae879912 | 1537 | */ |
0e4237ae ML |
1538 | if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) && |
1539 | blk_is_flush_data_rq(rq) && | |
1540 | blk_mq_request_completed(rq))) { | |
ae879912 JA |
1541 | bool *busy = priv; |
1542 | ||
1543 | *busy = true; | |
1544 | return false; | |
1545 | } | |
1546 | ||
1547 | return true; | |
1548 | } | |
1549 | ||
3c94d83c | 1550 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
1551 | { |
1552 | bool busy = false; | |
1553 | ||
3c94d83c | 1554 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
1555 | return busy; |
1556 | } | |
3c94d83c | 1557 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 1558 | |
9bdb4833 | 1559 | static void blk_mq_rq_timed_out(struct request *req) |
320ae51f | 1560 | { |
da661267 | 1561 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
1562 | if (req->q->mq_ops->timeout) { |
1563 | enum blk_eh_timer_return ret; | |
1564 | ||
9bdb4833 | 1565 | ret = req->q->mq_ops->timeout(req); |
d1210d5a CH |
1566 | if (ret == BLK_EH_DONE) |
1567 | return; | |
1568 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 1569 | } |
d1210d5a CH |
1570 | |
1571 | blk_add_timer(req); | |
87ee7b11 | 1572 | } |
5b3f25fc | 1573 | |
82c22947 DJ |
1574 | struct blk_expired_data { |
1575 | bool has_timedout_rq; | |
1576 | unsigned long next; | |
1577 | unsigned long timeout_start; | |
1578 | }; | |
1579 | ||
1580 | static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired) | |
81481eb4 | 1581 | { |
12f5b931 | 1582 | unsigned long deadline; |
87ee7b11 | 1583 | |
12f5b931 KB |
1584 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
1585 | return false; | |
da661267 CH |
1586 | if (rq->rq_flags & RQF_TIMED_OUT) |
1587 | return false; | |
a7af0af3 | 1588 | |
079076b3 | 1589 | deadline = READ_ONCE(rq->deadline); |
82c22947 | 1590 | if (time_after_eq(expired->timeout_start, deadline)) |
12f5b931 | 1591 | return true; |
a7af0af3 | 1592 | |
82c22947 DJ |
1593 | if (expired->next == 0) |
1594 | expired->next = deadline; | |
1595 | else if (time_after(expired->next, deadline)) | |
1596 | expired->next = deadline; | |
12f5b931 | 1597 | return false; |
87ee7b11 JA |
1598 | } |
1599 | ||
2e315dc0 ML |
1600 | void blk_mq_put_rq_ref(struct request *rq) |
1601 | { | |
de671d61 JA |
1602 | if (is_flush_rq(rq)) { |
1603 | if (rq->end_io(rq, 0) == RQ_END_IO_FREE) | |
1604 | blk_mq_free_request(rq); | |
1605 | } else if (req_ref_put_and_test(rq)) { | |
2e315dc0 | 1606 | __blk_mq_free_request(rq); |
de671d61 | 1607 | } |
2e315dc0 ML |
1608 | } |
1609 | ||
2dd6532e | 1610 | static bool blk_mq_check_expired(struct request *rq, void *priv) |
1d9bd516 | 1611 | { |
82c22947 | 1612 | struct blk_expired_data *expired = priv; |
12f5b931 KB |
1613 | |
1614 | /* | |
c797b40c ML |
1615 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
1616 | * be reallocated underneath the timeout handler's processing, then | |
1617 | * the expire check is reliable. If the request is not expired, then | |
1618 | * it was completed and reallocated as a new request after returning | |
1619 | * from blk_mq_check_expired(). | |
1d9bd516 | 1620 | */ |
82c22947 DJ |
1621 | if (blk_mq_req_expired(rq, expired)) { |
1622 | expired->has_timedout_rq = true; | |
1623 | return false; | |
1624 | } | |
1625 | return true; | |
1626 | } | |
1627 | ||
1628 | static bool blk_mq_handle_expired(struct request *rq, void *priv) | |
1629 | { | |
1630 | struct blk_expired_data *expired = priv; | |
1631 | ||
1632 | if (blk_mq_req_expired(rq, expired)) | |
9bdb4833 | 1633 | blk_mq_rq_timed_out(rq); |
7baa8572 | 1634 | return true; |
1d9bd516 TH |
1635 | } |
1636 | ||
287922eb | 1637 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 1638 | { |
287922eb CH |
1639 | struct request_queue *q = |
1640 | container_of(work, struct request_queue, timeout_work); | |
82c22947 DJ |
1641 | struct blk_expired_data expired = { |
1642 | .timeout_start = jiffies, | |
1643 | }; | |
1d9bd516 | 1644 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 1645 | unsigned long i; |
320ae51f | 1646 | |
71f79fb3 GKB |
1647 | /* A deadlock might occur if a request is stuck requiring a |
1648 | * timeout at the same time a queue freeze is waiting | |
1649 | * completion, since the timeout code would not be able to | |
1650 | * acquire the queue reference here. | |
1651 | * | |
1652 | * That's why we don't use blk_queue_enter here; instead, we use | |
1653 | * percpu_ref_tryget directly, because we need to be able to | |
1654 | * obtain a reference even in the short window between the queue | |
1655 | * starting to freeze, by dropping the first reference in | |
1671d522 | 1656 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
1657 | * consumed, marked by the instant q_usage_counter reaches |
1658 | * zero. | |
1659 | */ | |
1660 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
1661 | return; |
1662 | ||
82c22947 DJ |
1663 | /* check if there is any timed-out request */ |
1664 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired); | |
1665 | if (expired.has_timedout_rq) { | |
1666 | /* | |
1667 | * Before walking tags, we must ensure any submit started | |
1668 | * before the current time has finished. Since the submit | |
1669 | * uses srcu or rcu, wait for a synchronization point to | |
1670 | * ensure all running submits have finished | |
1671 | */ | |
483239c7 | 1672 | blk_mq_wait_quiesce_done(q->tag_set); |
82c22947 DJ |
1673 | |
1674 | expired.next = 0; | |
1675 | blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired); | |
1676 | } | |
320ae51f | 1677 | |
82c22947 DJ |
1678 | if (expired.next != 0) { |
1679 | mod_timer(&q->timeout, expired.next); | |
0d2602ca | 1680 | } else { |
fcd36c36 BVA |
1681 | /* |
1682 | * Request timeouts are handled as a forward rolling timer. If | |
1683 | * we end up here it means that no requests are pending and | |
1684 | * also that no request has been pending for a while. Mark | |
1685 | * each hctx as idle. | |
1686 | */ | |
f054b56c ML |
1687 | queue_for_each_hw_ctx(q, hctx, i) { |
1688 | /* the hctx may be unmapped, so check it here */ | |
1689 | if (blk_mq_hw_queue_mapped(hctx)) | |
1690 | blk_mq_tag_idle(hctx); | |
1691 | } | |
0d2602ca | 1692 | } |
287922eb | 1693 | blk_queue_exit(q); |
320ae51f JA |
1694 | } |
1695 | ||
88459642 OS |
1696 | struct flush_busy_ctx_data { |
1697 | struct blk_mq_hw_ctx *hctx; | |
1698 | struct list_head *list; | |
1699 | }; | |
1700 | ||
1701 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1702 | { | |
1703 | struct flush_busy_ctx_data *flush_data = data; | |
1704 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1705 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1706 | enum hctx_type type = hctx->type; |
88459642 | 1707 | |
88459642 | 1708 | spin_lock(&ctx->lock); |
c16d6b5a | 1709 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1710 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1711 | spin_unlock(&ctx->lock); |
1712 | return true; | |
1713 | } | |
1714 | ||
1429d7c9 JA |
1715 | /* |
1716 | * Process software queues that have been marked busy, splicing them | |
1717 | * to the for-dispatch | |
1718 | */ | |
2c3ad667 | 1719 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1720 | { |
88459642 OS |
1721 | struct flush_busy_ctx_data data = { |
1722 | .hctx = hctx, | |
1723 | .list = list, | |
1724 | }; | |
1429d7c9 | 1725 | |
88459642 | 1726 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1727 | } |
2c3ad667 | 1728 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1729 | |
b347689f ML |
1730 | struct dispatch_rq_data { |
1731 | struct blk_mq_hw_ctx *hctx; | |
1732 | struct request *rq; | |
1733 | }; | |
1734 | ||
1735 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1736 | void *data) | |
1737 | { | |
1738 | struct dispatch_rq_data *dispatch_data = data; | |
1739 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1740 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1741 | enum hctx_type type = hctx->type; |
b347689f ML |
1742 | |
1743 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1744 | if (!list_empty(&ctx->rq_lists[type])) { |
1745 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1746 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1747 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1748 | sbitmap_clear_bit(sb, bitnr); |
1749 | } | |
1750 | spin_unlock(&ctx->lock); | |
1751 | ||
1752 | return !dispatch_data->rq; | |
1753 | } | |
1754 | ||
1755 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1756 | struct blk_mq_ctx *start) | |
1757 | { | |
f31967f0 | 1758 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1759 | struct dispatch_rq_data data = { |
1760 | .hctx = hctx, | |
1761 | .rq = NULL, | |
1762 | }; | |
1763 | ||
1764 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1765 | dispatch_rq_from_ctx, &data); | |
1766 | ||
1767 | return data.rq; | |
1768 | } | |
1769 | ||
b8643d68 | 1770 | bool __blk_mq_alloc_driver_tag(struct request *rq) |
570e9b73 | 1771 | { |
ae0f1a73 | 1772 | struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1773 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1774 | int tag; |
1775 | ||
568f2700 ML |
1776 | blk_mq_tag_busy(rq->mq_hctx); |
1777 | ||
570e9b73 | 1778 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
ae0f1a73 | 1779 | bt = &rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1780 | tag_offset = 0; |
28500850 ML |
1781 | } else { |
1782 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1783 | return false; | |
570e9b73 ML |
1784 | } |
1785 | ||
570e9b73 ML |
1786 | tag = __sbitmap_queue_get(bt); |
1787 | if (tag == BLK_MQ_NO_TAG) | |
1788 | return false; | |
1789 | ||
1790 | rq->tag = tag + tag_offset; | |
b8643d68 | 1791 | blk_mq_inc_active_requests(rq->mq_hctx); |
568f2700 | 1792 | return true; |
570e9b73 ML |
1793 | } |
1794 | ||
eb619fdb JA |
1795 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1796 | int flags, void *key) | |
da55f2cc OS |
1797 | { |
1798 | struct blk_mq_hw_ctx *hctx; | |
1799 | ||
1800 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1801 | ||
5815839b | 1802 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1803 | if (!list_empty(&wait->entry)) { |
1804 | struct sbitmap_queue *sbq; | |
1805 | ||
1806 | list_del_init(&wait->entry); | |
ae0f1a73 | 1807 | sbq = &hctx->tags->bitmap_tags; |
e8618575 JA |
1808 | atomic_dec(&sbq->ws_active); |
1809 | } | |
5815839b ML |
1810 | spin_unlock(&hctx->dispatch_wait_lock); |
1811 | ||
da55f2cc OS |
1812 | blk_mq_run_hw_queue(hctx, true); |
1813 | return 1; | |
1814 | } | |
1815 | ||
f906a6a0 JA |
1816 | /* |
1817 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1818 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1819 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1820 | * marking us as waiting. |
1821 | */ | |
2278d69f | 1822 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1823 | struct request *rq) |
da55f2cc | 1824 | { |
98b99e94 | 1825 | struct sbitmap_queue *sbq; |
5815839b | 1826 | struct wait_queue_head *wq; |
f906a6a0 JA |
1827 | wait_queue_entry_t *wait; |
1828 | bool ret; | |
da55f2cc | 1829 | |
47df9ce9 KS |
1830 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
1831 | !(blk_mq_is_shared_tags(hctx->flags))) { | |
684b7324 | 1832 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1833 | |
c27d53fb BVA |
1834 | /* |
1835 | * It's possible that a tag was freed in the window between the | |
1836 | * allocation failure and adding the hardware queue to the wait | |
1837 | * queue. | |
1838 | * | |
1839 | * Don't clear RESTART here, someone else could have set it. | |
1840 | * At most this will cost an extra queue run. | |
1841 | */ | |
8ab6bb9e | 1842 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1843 | } |
eb619fdb | 1844 | |
2278d69f | 1845 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1846 | if (!list_empty_careful(&wait->entry)) |
1847 | return false; | |
1848 | ||
98b99e94 KS |
1849 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) |
1850 | sbq = &hctx->tags->breserved_tags; | |
1851 | else | |
1852 | sbq = &hctx->tags->bitmap_tags; | |
e8618575 | 1853 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1854 | |
1855 | spin_lock_irq(&wq->lock); | |
1856 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1857 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1858 | spin_unlock(&hctx->dispatch_wait_lock); |
1859 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1860 | return false; |
eb619fdb JA |
1861 | } |
1862 | ||
e8618575 | 1863 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1864 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1865 | __add_wait_queue(wq, wait); | |
c27d53fb | 1866 | |
5266caaf ML |
1867 | /* |
1868 | * Add one explicit barrier since blk_mq_get_driver_tag() may | |
1869 | * not imply barrier in case of failure. | |
1870 | * | |
1871 | * Order adding us to wait queue and allocating driver tag. | |
1872 | * | |
1873 | * The pair is the one implied in sbitmap_queue_wake_up() which | |
1874 | * orders clearing sbitmap tag bits and waitqueue_active() in | |
1875 | * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless | |
1876 | * | |
1877 | * Otherwise, re-order of adding wait queue and getting driver tag | |
1878 | * may cause __sbitmap_queue_wake_up() to wake up nothing because | |
1879 | * the waitqueue_active() may not observe us in wait queue. | |
1880 | */ | |
1881 | smp_mb(); | |
1882 | ||
da55f2cc | 1883 | /* |
eb619fdb JA |
1884 | * It's possible that a tag was freed in the window between the |
1885 | * allocation failure and adding the hardware queue to the wait | |
1886 | * queue. | |
da55f2cc | 1887 | */ |
8ab6bb9e | 1888 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1889 | if (!ret) { |
5815839b ML |
1890 | spin_unlock(&hctx->dispatch_wait_lock); |
1891 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1892 | return false; |
eb619fdb | 1893 | } |
c27d53fb BVA |
1894 | |
1895 | /* | |
1896 | * We got a tag, remove ourselves from the wait queue to ensure | |
1897 | * someone else gets the wakeup. | |
1898 | */ | |
c27d53fb | 1899 | list_del_init(&wait->entry); |
e8618575 | 1900 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1901 | spin_unlock(&hctx->dispatch_wait_lock); |
1902 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1903 | |
1904 | return true; | |
da55f2cc OS |
1905 | } |
1906 | ||
6e768717 ML |
1907 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1908 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1909 | /* | |
1910 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1911 | * - EWMA is one simple way to compute running average value | |
1912 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1913 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1914 | * factor doesn't matter because EWMA decreases exponentially | |
1915 | */ | |
1916 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1917 | { | |
1918 | unsigned int ewma; | |
1919 | ||
6e768717 ML |
1920 | ewma = hctx->dispatch_busy; |
1921 | ||
1922 | if (!ewma && !busy) | |
1923 | return; | |
1924 | ||
1925 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1926 | if (busy) | |
1927 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1928 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1929 | ||
1930 | hctx->dispatch_busy = ewma; | |
1931 | } | |
1932 | ||
86ff7c2a ML |
1933 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1934 | ||
c92a4103 JT |
1935 | static void blk_mq_handle_dev_resource(struct request *rq, |
1936 | struct list_head *list) | |
1937 | { | |
c92a4103 JT |
1938 | list_add(&rq->queuelist, list); |
1939 | __blk_mq_requeue_request(rq); | |
1940 | } | |
1941 | ||
0512a75b KB |
1942 | static void blk_mq_handle_zone_resource(struct request *rq, |
1943 | struct list_head *zone_list) | |
1944 | { | |
1945 | /* | |
1946 | * If we end up here it is because we cannot dispatch a request to a | |
1947 | * specific zone due to LLD level zone-write locking or other zone | |
1948 | * related resource not being available. In this case, set the request | |
1949 | * aside in zone_list for retrying it later. | |
1950 | */ | |
1951 | list_add(&rq->queuelist, zone_list); | |
1952 | __blk_mq_requeue_request(rq); | |
1953 | } | |
1954 | ||
75383524 ML |
1955 | enum prep_dispatch { |
1956 | PREP_DISPATCH_OK, | |
1957 | PREP_DISPATCH_NO_TAG, | |
1958 | PREP_DISPATCH_NO_BUDGET, | |
1959 | }; | |
1960 | ||
1961 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1962 | bool need_budget) | |
1963 | { | |
1964 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1965 | int budget_token = -1; |
75383524 | 1966 | |
2a5a24aa ML |
1967 | if (need_budget) { |
1968 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1969 | if (budget_token < 0) { | |
1970 | blk_mq_put_driver_tag(rq); | |
1971 | return PREP_DISPATCH_NO_BUDGET; | |
1972 | } | |
1973 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1974 | } |
1975 | ||
1976 | if (!blk_mq_get_driver_tag(rq)) { | |
1977 | /* | |
1978 | * The initial allocation attempt failed, so we need to | |
1979 | * rerun the hardware queue when a tag is freed. The | |
1980 | * waitqueue takes care of that. If the queue is run | |
1981 | * before we add this entry back on the dispatch list, | |
1982 | * we'll re-run it below. | |
1983 | */ | |
1984 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1985 | /* |
1986 | * All budgets not got from this function will be put | |
1987 | * together during handling partial dispatch | |
1988 | */ | |
1989 | if (need_budget) | |
2a5a24aa | 1990 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1991 | return PREP_DISPATCH_NO_TAG; |
1992 | } | |
1993 | } | |
1994 | ||
1995 | return PREP_DISPATCH_OK; | |
1996 | } | |
1997 | ||
1fd40b5e ML |
1998 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1999 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 2000 | struct list_head *list) |
1fd40b5e | 2001 | { |
2a5a24aa | 2002 | struct request *rq; |
1fd40b5e | 2003 | |
2a5a24aa ML |
2004 | list_for_each_entry(rq, list, queuelist) { |
2005 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 2006 | |
2a5a24aa ML |
2007 | if (budget_token >= 0) |
2008 | blk_mq_put_dispatch_budget(q, budget_token); | |
2009 | } | |
1fd40b5e ML |
2010 | } |
2011 | ||
34c9f547 KS |
2012 | /* |
2013 | * blk_mq_commit_rqs will notify driver using bd->last that there is no | |
2014 | * more requests. (See comment in struct blk_mq_ops for commit_rqs for | |
2015 | * details) | |
2016 | * Attention, we should explicitly call this in unusual cases: | |
2017 | * 1) did not queue everything initially scheduled to queue | |
2018 | * 2) the last attempt to queue a request failed | |
2019 | */ | |
2020 | static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued, | |
2021 | bool from_schedule) | |
2022 | { | |
2023 | if (hctx->queue->mq_ops->commit_rqs && queued) { | |
2024 | trace_block_unplug(hctx->queue, queued, !from_schedule); | |
2025 | hctx->queue->mq_ops->commit_rqs(hctx); | |
2026 | } | |
2027 | } | |
2028 | ||
1f57f8d4 JA |
2029 | /* |
2030 | * Returns true if we did some work AND can potentially do more. | |
2031 | */ | |
445874e8 | 2032 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 2033 | unsigned int nr_budgets) |
320ae51f | 2034 | { |
75383524 | 2035 | enum prep_dispatch prep; |
445874e8 | 2036 | struct request_queue *q = hctx->queue; |
f1ce99f7 | 2037 | struct request *rq; |
4ea58fe4 | 2038 | int queued; |
86ff7c2a | 2039 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 2040 | LIST_HEAD(zone_list); |
9586e67b | 2041 | bool needs_resource = false; |
320ae51f | 2042 | |
81380ca1 OS |
2043 | if (list_empty(list)) |
2044 | return false; | |
2045 | ||
320ae51f JA |
2046 | /* |
2047 | * Now process all the entries, sending them to the driver. | |
2048 | */ | |
4ea58fe4 | 2049 | queued = 0; |
81380ca1 | 2050 | do { |
74c45052 | 2051 | struct blk_mq_queue_data bd; |
320ae51f | 2052 | |
f04c3df3 | 2053 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 2054 | |
445874e8 | 2055 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 2056 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 2057 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 2058 | break; |
de148297 | 2059 | |
320ae51f | 2060 | list_del_init(&rq->queuelist); |
320ae51f | 2061 | |
74c45052 | 2062 | bd.rq = rq; |
f1ce99f7 | 2063 | bd.last = list_empty(list); |
74c45052 | 2064 | |
1fd40b5e ML |
2065 | /* |
2066 | * once the request is queued to lld, no need to cover the | |
2067 | * budget any more | |
2068 | */ | |
2069 | if (nr_budgets) | |
2070 | nr_budgets--; | |
74c45052 | 2071 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
2072 | switch (ret) { |
2073 | case BLK_STS_OK: | |
2074 | queued++; | |
320ae51f | 2075 | break; |
7bf13729 | 2076 | case BLK_STS_RESOURCE: |
9586e67b NA |
2077 | needs_resource = true; |
2078 | fallthrough; | |
7bf13729 ML |
2079 | case BLK_STS_DEV_RESOURCE: |
2080 | blk_mq_handle_dev_resource(rq, list); | |
2081 | goto out; | |
2082 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
2083 | /* |
2084 | * Move the request to zone_list and keep going through | |
2085 | * the dispatch list to find more requests the drive can | |
2086 | * accept. | |
2087 | */ | |
2088 | blk_mq_handle_zone_resource(rq, &zone_list); | |
9586e67b | 2089 | needs_resource = true; |
7bf13729 ML |
2090 | break; |
2091 | default: | |
e21ee5a6 | 2092 | blk_mq_end_request(rq, ret); |
320ae51f | 2093 | } |
81380ca1 | 2094 | } while (!list_empty(list)); |
7bf13729 | 2095 | out: |
0512a75b KB |
2096 | if (!list_empty(&zone_list)) |
2097 | list_splice_tail_init(&zone_list, list); | |
2098 | ||
632bfb63 | 2099 | /* If we didn't flush the entire list, we could have told the driver |
2100 | * there was more coming, but that turned out to be a lie. | |
2101 | */ | |
e4ef2e05 KS |
2102 | if (!list_empty(list) || ret != BLK_STS_OK) |
2103 | blk_mq_commit_rqs(hctx, queued, false); | |
2104 | ||
320ae51f JA |
2105 | /* |
2106 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
2107 | * that is where we will continue on next queue run. | |
2108 | */ | |
f04c3df3 | 2109 | if (!list_empty(list)) { |
86ff7c2a | 2110 | bool needs_restart; |
75383524 ML |
2111 | /* For non-shared tags, the RESTART check will suffice */ |
2112 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
47df9ce9 KS |
2113 | ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) || |
2114 | blk_mq_is_shared_tags(hctx->flags)); | |
86ff7c2a | 2115 | |
2a5a24aa ML |
2116 | if (nr_budgets) |
2117 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 2118 | |
320ae51f | 2119 | spin_lock(&hctx->lock); |
01e99aec | 2120 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 2121 | spin_unlock(&hctx->lock); |
f04c3df3 | 2122 | |
d7d8535f ML |
2123 | /* |
2124 | * Order adding requests to hctx->dispatch and checking | |
2125 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
2126 | * in blk_mq_sched_restart(). Avoid restart code path to | |
2127 | * miss the new added requests to hctx->dispatch, meantime | |
2128 | * SCHED_RESTART is observed here. | |
2129 | */ | |
2130 | smp_mb(); | |
2131 | ||
9ba52e58 | 2132 | /* |
710c785f BVA |
2133 | * If SCHED_RESTART was set by the caller of this function and |
2134 | * it is no longer set that means that it was cleared by another | |
2135 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 2136 | * |
eb619fdb JA |
2137 | * If 'no_tag' is set, that means that we failed getting |
2138 | * a driver tag with an I/O scheduler attached. If our dispatch | |
2139 | * waitqueue is no longer active, ensure that we run the queue | |
2140 | * AFTER adding our entries back to the list. | |
bd166ef1 | 2141 | * |
710c785f BVA |
2142 | * If no I/O scheduler has been configured it is possible that |
2143 | * the hardware queue got stopped and restarted before requests | |
2144 | * were pushed back onto the dispatch list. Rerun the queue to | |
2145 | * avoid starvation. Notes: | |
2146 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
2147 | * been stopped before rerunning a queue. | |
2148 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 2149 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 2150 | * and dm-rq. |
86ff7c2a ML |
2151 | * |
2152 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
2153 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 | 2154 | * that could otherwise occur if the queue is idle. We'll do |
9586e67b NA |
2155 | * similar if we couldn't get budget or couldn't lock a zone |
2156 | * and SCHED_RESTART is set. | |
bd166ef1 | 2157 | */ |
86ff7c2a | 2158 | needs_restart = blk_mq_sched_needs_restart(hctx); |
9586e67b NA |
2159 | if (prep == PREP_DISPATCH_NO_BUDGET) |
2160 | needs_resource = true; | |
86ff7c2a | 2161 | if (!needs_restart || |
eb619fdb | 2162 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 2163 | blk_mq_run_hw_queue(hctx, true); |
6d5e8d21 | 2164 | else if (needs_resource) |
86ff7c2a | 2165 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 2166 | |
6e768717 | 2167 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 2168 | return false; |
4ea58fe4 | 2169 | } |
f04c3df3 | 2170 | |
4ea58fe4 KS |
2171 | blk_mq_update_dispatch_busy(hctx, false); |
2172 | return true; | |
f04c3df3 JA |
2173 | } |
2174 | ||
f82ddf19 ML |
2175 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
2176 | { | |
2177 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
2178 | ||
2179 | if (cpu >= nr_cpu_ids) | |
2180 | cpu = cpumask_first(hctx->cpumask); | |
2181 | return cpu; | |
2182 | } | |
2183 | ||
506e931f JA |
2184 | /* |
2185 | * It'd be great if the workqueue API had a way to pass | |
2186 | * in a mask and had some smarts for more clever placement. | |
2187 | * For now we just round-robin here, switching for every | |
2188 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
2189 | */ | |
2190 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
2191 | { | |
7bed4595 | 2192 | bool tried = false; |
476f8c98 | 2193 | int next_cpu = hctx->next_cpu; |
7bed4595 | 2194 | |
b657d7e6 CH |
2195 | if (hctx->queue->nr_hw_queues == 1) |
2196 | return WORK_CPU_UNBOUND; | |
506e931f JA |
2197 | |
2198 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 2199 | select_cpu: |
476f8c98 | 2200 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 2201 | cpu_online_mask); |
506e931f | 2202 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 2203 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2204 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2205 | } | |
2206 | ||
7bed4595 ML |
2207 | /* |
2208 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
2209 | * and it should only happen in the path of handling CPU DEAD. | |
2210 | */ | |
476f8c98 | 2211 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
2212 | if (!tried) { |
2213 | tried = true; | |
2214 | goto select_cpu; | |
2215 | } | |
2216 | ||
2217 | /* | |
2218 | * Make sure to re-select CPU next time once after CPUs | |
2219 | * in hctx->cpumask become online again. | |
2220 | */ | |
476f8c98 | 2221 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
2222 | hctx->next_cpu_batch = 1; |
2223 | return WORK_CPU_UNBOUND; | |
2224 | } | |
476f8c98 ML |
2225 | |
2226 | hctx->next_cpu = next_cpu; | |
2227 | return next_cpu; | |
506e931f JA |
2228 | } |
2229 | ||
105663f7 | 2230 | /** |
1aa8d875 | 2231 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. |
105663f7 | 2232 | * @hctx: Pointer to the hardware queue to run. |
fa94ba8a | 2233 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 | 2234 | * |
1aa8d875 | 2235 | * Run a hardware queue asynchronously with a delay of @msecs. |
105663f7 | 2236 | */ |
1aa8d875 | 2237 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
320ae51f | 2238 | { |
5435c023 | 2239 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f | 2240 | return; |
ae943d20 BVA |
2241 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
2242 | msecs_to_jiffies(msecs)); | |
7587a5ae | 2243 | } |
7587a5ae BVA |
2244 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); |
2245 | ||
105663f7 AA |
2246 | /** |
2247 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
2248 | * @hctx: Pointer to the hardware queue to run. | |
2249 | * @async: If we want to run the queue asynchronously. | |
2250 | * | |
2251 | * Check if the request queue is not in a quiesced state and if there are | |
2252 | * pending requests to be sent. If this is true, run the queue to send requests | |
2253 | * to hardware. | |
2254 | */ | |
626fb735 | 2255 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 2256 | { |
24f5a90f ML |
2257 | bool need_run; |
2258 | ||
4d5bba5b CH |
2259 | /* |
2260 | * We can't run the queue inline with interrupts disabled. | |
2261 | */ | |
2262 | WARN_ON_ONCE(!async && in_interrupt()); | |
2263 | ||
65a558f6 BVA |
2264 | might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING); |
2265 | ||
24f5a90f ML |
2266 | /* |
2267 | * When queue is quiesced, we may be switching io scheduler, or | |
2268 | * updating nr_hw_queues, or other things, and we can't run queue | |
2269 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
2270 | * | |
2271 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
2272 | * quiesced. | |
2273 | */ | |
41adf531 | 2274 | __blk_mq_run_dispatch_ops(hctx->queue, false, |
2a904d00 ML |
2275 | need_run = !blk_queue_quiesced(hctx->queue) && |
2276 | blk_mq_hctx_has_pending(hctx)); | |
24f5a90f | 2277 | |
1aa8d875 CH |
2278 | if (!need_run) |
2279 | return; | |
2280 | ||
65a558f6 | 2281 | if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) { |
1aa8d875 CH |
2282 | blk_mq_delay_run_hw_queue(hctx, 0); |
2283 | return; | |
2284 | } | |
2285 | ||
4d5bba5b CH |
2286 | blk_mq_run_dispatch_ops(hctx->queue, |
2287 | blk_mq_sched_dispatch_requests(hctx)); | |
320ae51f | 2288 | } |
5b727272 | 2289 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 2290 | |
b6e68ee8 JK |
2291 | /* |
2292 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
2293 | * scheduler. | |
2294 | */ | |
2295 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
2296 | { | |
5d05426e | 2297 | struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); |
b6e68ee8 JK |
2298 | /* |
2299 | * If the IO scheduler does not respect hardware queues when | |
2300 | * dispatching, we just don't bother with multiple HW queues and | |
2301 | * dispatch from hctx for the current CPU since running multiple queues | |
2302 | * just causes lock contention inside the scheduler and pointless cache | |
2303 | * bouncing. | |
2304 | */ | |
51ab80f0 | 2305 | struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT]; |
5d05426e | 2306 | |
b6e68ee8 JK |
2307 | if (!blk_mq_hctx_stopped(hctx)) |
2308 | return hctx; | |
2309 | return NULL; | |
2310 | } | |
2311 | ||
105663f7 | 2312 | /** |
24f7bb88 | 2313 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
2314 | * @q: Pointer to the request queue to run. |
2315 | * @async: If we want to run the queue asynchronously. | |
2316 | */ | |
b94ec296 | 2317 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 2318 | { |
b6e68ee8 | 2319 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2320 | unsigned long i; |
320ae51f | 2321 | |
b6e68ee8 | 2322 | sq_hctx = NULL; |
4d337ceb | 2323 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2324 | sq_hctx = blk_mq_get_sq_hctx(q); |
320ae51f | 2325 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 2326 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 2327 | continue; |
b6e68ee8 JK |
2328 | /* |
2329 | * Dispatch from this hctx either if there's no hctx preferred | |
2330 | * by IO scheduler or if it has requests that bypass the | |
2331 | * scheduler. | |
2332 | */ | |
2333 | if (!sq_hctx || sq_hctx == hctx || | |
2334 | !list_empty_careful(&hctx->dispatch)) | |
2335 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
2336 | } |
2337 | } | |
b94ec296 | 2338 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 2339 | |
b9151e7b DA |
2340 | /** |
2341 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
2342 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 2343 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
2344 | */ |
2345 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
2346 | { | |
b6e68ee8 | 2347 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2348 | unsigned long i; |
b9151e7b | 2349 | |
b6e68ee8 | 2350 | sq_hctx = NULL; |
4d337ceb | 2351 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2352 | sq_hctx = blk_mq_get_sq_hctx(q); |
b9151e7b DA |
2353 | queue_for_each_hw_ctx(q, hctx, i) { |
2354 | if (blk_mq_hctx_stopped(hctx)) | |
2355 | continue; | |
8f5fea65 DJ |
2356 | /* |
2357 | * If there is already a run_work pending, leave the | |
2358 | * pending delay untouched. Otherwise, a hctx can stall | |
2359 | * if another hctx is re-delaying the other's work | |
2360 | * before the work executes. | |
2361 | */ | |
2362 | if (delayed_work_pending(&hctx->run_work)) | |
2363 | continue; | |
b6e68ee8 JK |
2364 | /* |
2365 | * Dispatch from this hctx either if there's no hctx preferred | |
2366 | * by IO scheduler or if it has requests that bypass the | |
2367 | * scheduler. | |
2368 | */ | |
2369 | if (!sq_hctx || sq_hctx == hctx || | |
2370 | !list_empty_careful(&hctx->dispatch)) | |
2371 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
2372 | } |
2373 | } | |
2374 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
2375 | ||
39a70c76 ML |
2376 | /* |
2377 | * This function is often used for pausing .queue_rq() by driver when | |
2378 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2379 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2380 | * |
2381 | * We do not guarantee that dispatch can be drained or blocked | |
2382 | * after blk_mq_stop_hw_queue() returns. Please use | |
2383 | * blk_mq_quiesce_queue() for that requirement. | |
2384 | */ | |
2719aa21 JA |
2385 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
2386 | { | |
641a9ed6 | 2387 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 2388 | |
641a9ed6 | 2389 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 2390 | } |
641a9ed6 | 2391 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 2392 | |
39a70c76 ML |
2393 | /* |
2394 | * This function is often used for pausing .queue_rq() by driver when | |
2395 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2396 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2397 | * |
2398 | * We do not guarantee that dispatch can be drained or blocked | |
2399 | * after blk_mq_stop_hw_queues() returns. Please use | |
2400 | * blk_mq_quiesce_queue() for that requirement. | |
2401 | */ | |
2719aa21 JA |
2402 | void blk_mq_stop_hw_queues(struct request_queue *q) |
2403 | { | |
641a9ed6 | 2404 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 2405 | unsigned long i; |
641a9ed6 ML |
2406 | |
2407 | queue_for_each_hw_ctx(q, hctx, i) | |
2408 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
2409 | } |
2410 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
2411 | ||
320ae51f JA |
2412 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
2413 | { | |
2414 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 2415 | |
65a558f6 | 2416 | blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING); |
320ae51f JA |
2417 | } |
2418 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
2419 | ||
2f268556 CH |
2420 | void blk_mq_start_hw_queues(struct request_queue *q) |
2421 | { | |
2422 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2423 | unsigned long i; |
2f268556 CH |
2424 | |
2425 | queue_for_each_hw_ctx(q, hctx, i) | |
2426 | blk_mq_start_hw_queue(hctx); | |
2427 | } | |
2428 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
2429 | ||
ae911c5e JA |
2430 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
2431 | { | |
2432 | if (!blk_mq_hctx_stopped(hctx)) | |
2433 | return; | |
2434 | ||
2435 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
2436 | blk_mq_run_hw_queue(hctx, async); | |
2437 | } | |
2438 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
2439 | ||
1b4a3258 | 2440 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
2441 | { |
2442 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2443 | unsigned long i; |
320ae51f | 2444 | |
ae911c5e | 2445 | queue_for_each_hw_ctx(q, hctx, i) |
65a558f6 BVA |
2446 | blk_mq_start_stopped_hw_queue(hctx, async || |
2447 | (hctx->flags & BLK_MQ_F_BLOCKING)); | |
320ae51f JA |
2448 | } |
2449 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
2450 | ||
70f4db63 | 2451 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f | 2452 | { |
c20a1a2c CH |
2453 | struct blk_mq_hw_ctx *hctx = |
2454 | container_of(work, struct blk_mq_hw_ctx, run_work.work); | |
7b607814 | 2455 | |
4d5bba5b CH |
2456 | blk_mq_run_dispatch_ops(hctx->queue, |
2457 | blk_mq_sched_dispatch_requests(hctx)); | |
320ae51f JA |
2458 | } |
2459 | ||
105663f7 AA |
2460 | /** |
2461 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
2462 | * @rq: Pointer to request to be inserted. | |
2b597613 | 2463 | * @flags: BLK_MQ_INSERT_* |
105663f7 | 2464 | * |
157f377b JA |
2465 | * Should only be used carefully, when the caller knows we want to |
2466 | * bypass a potential IO scheduler on the target device. | |
2467 | */ | |
360f2648 | 2468 | static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags) |
157f377b | 2469 | { |
ea4f995e | 2470 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
2471 | |
2472 | spin_lock(&hctx->lock); | |
2b597613 | 2473 | if (flags & BLK_MQ_INSERT_AT_HEAD) |
01e99aec ML |
2474 | list_add(&rq->queuelist, &hctx->dispatch); |
2475 | else | |
2476 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b | 2477 | spin_unlock(&hctx->lock); |
157f377b JA |
2478 | } |
2479 | ||
05a93117 CH |
2480 | static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, |
2481 | struct blk_mq_ctx *ctx, struct list_head *list, | |
2482 | bool run_queue_async) | |
320ae51f | 2483 | { |
3f0cedc7 | 2484 | struct request *rq; |
c16d6b5a | 2485 | enum hctx_type type = hctx->type; |
3f0cedc7 | 2486 | |
94aa228c CH |
2487 | /* |
2488 | * Try to issue requests directly if the hw queue isn't busy to save an | |
2489 | * extra enqueue & dequeue to the sw queue. | |
2490 | */ | |
2491 | if (!hctx->dispatch_busy && !run_queue_async) { | |
2492 | blk_mq_run_dispatch_ops(hctx->queue, | |
2493 | blk_mq_try_issue_list_directly(hctx, list)); | |
2494 | if (list_empty(list)) | |
2495 | goto out; | |
2496 | } | |
2497 | ||
320ae51f JA |
2498 | /* |
2499 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
2500 | * offline now | |
2501 | */ | |
3f0cedc7 | 2502 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 2503 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 2504 | trace_block_rq_insert(rq); |
65a558f6 BVA |
2505 | if (rq->cmd_flags & REQ_NOWAIT) |
2506 | run_queue_async = true; | |
320ae51f | 2507 | } |
3f0cedc7 ML |
2508 | |
2509 | spin_lock(&ctx->lock); | |
c16d6b5a | 2510 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 2511 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 2512 | spin_unlock(&ctx->lock); |
94aa228c CH |
2513 | out: |
2514 | blk_mq_run_hw_queue(hctx, run_queue_async); | |
320ae51f JA |
2515 | } |
2516 | ||
710fa378 | 2517 | static void blk_mq_insert_request(struct request *rq, blk_insert_t flags) |
2bd215df CH |
2518 | { |
2519 | struct request_queue *q = rq->q; | |
2bd215df CH |
2520 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
2521 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2522 | ||
53548d2a CH |
2523 | if (blk_rq_is_passthrough(rq)) { |
2524 | /* | |
2525 | * Passthrough request have to be added to hctx->dispatch | |
2526 | * directly. The device may be in a situation where it can't | |
2527 | * handle FS request, and always returns BLK_STS_RESOURCE for | |
2528 | * them, which gets them added to hctx->dispatch. | |
2529 | * | |
2530 | * If a passthrough request is required to unblock the queues, | |
2531 | * and it is added to the scheduler queue, there is no chance to | |
2532 | * dispatch it given we prioritize requests in hctx->dispatch. | |
2533 | */ | |
2b597613 | 2534 | blk_mq_request_bypass_insert(rq, flags); |
be4c4278 | 2535 | } else if (req_op(rq) == REQ_OP_FLUSH) { |
2bd215df CH |
2536 | /* |
2537 | * Firstly normal IO request is inserted to scheduler queue or | |
2538 | * sw queue, meantime we add flush request to dispatch queue( | |
2539 | * hctx->dispatch) directly and there is at most one in-flight | |
2540 | * flush request for each hw queue, so it doesn't matter to add | |
2541 | * flush request to tail or front of the dispatch queue. | |
2542 | * | |
2543 | * Secondly in case of NCQ, flush request belongs to non-NCQ | |
2544 | * command, and queueing it will fail when there is any | |
2545 | * in-flight normal IO request(NCQ command). When adding flush | |
2546 | * rq to the front of hctx->dispatch, it is easier to introduce | |
2547 | * extra time to flush rq's latency because of S_SCHED_RESTART | |
2548 | * compared with adding to the tail of dispatch queue, then | |
2549 | * chance of flush merge is increased, and less flush requests | |
2550 | * will be issued to controller. It is observed that ~10% time | |
2551 | * is saved in blktests block/004 on disk attached to AHCI/NCQ | |
2552 | * drive when adding flush rq to the front of hctx->dispatch. | |
2553 | * | |
2554 | * Simply queue flush rq to the front of hctx->dispatch so that | |
2555 | * intensive flush workloads can benefit in case of NCQ HW. | |
2556 | */ | |
2b597613 | 2557 | blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD); |
53548d2a | 2558 | } else if (q->elevator) { |
2bd215df CH |
2559 | LIST_HEAD(list); |
2560 | ||
53548d2a CH |
2561 | WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG); |
2562 | ||
2bd215df | 2563 | list_add(&rq->queuelist, &list); |
93fffe16 | 2564 | q->elevator->type->ops.insert_requests(hctx, &list, flags); |
2bd215df | 2565 | } else { |
4ec5c055 CH |
2566 | trace_block_rq_insert(rq); |
2567 | ||
2bd215df | 2568 | spin_lock(&ctx->lock); |
710fa378 | 2569 | if (flags & BLK_MQ_INSERT_AT_HEAD) |
4ec5c055 CH |
2570 | list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]); |
2571 | else | |
2572 | list_add_tail(&rq->queuelist, | |
2573 | &ctx->rq_lists[hctx->type]); | |
a88db1e0 | 2574 | blk_mq_hctx_mark_pending(hctx, ctx); |
2bd215df CH |
2575 | spin_unlock(&ctx->lock); |
2576 | } | |
320ae51f JA |
2577 | } |
2578 | ||
14ccb66b CH |
2579 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
2580 | unsigned int nr_segs) | |
320ae51f | 2581 | { |
93f221ae EB |
2582 | int err; |
2583 | ||
f924cdde CH |
2584 | if (bio->bi_opf & REQ_RAHEAD) |
2585 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
2586 | ||
2587 | rq->__sector = bio->bi_iter.bi_sector; | |
14ccb66b | 2588 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
2589 | |
2590 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
2591 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
2592 | WARN_ON_ONCE(err); | |
4b570521 | 2593 | |
b5af37ab | 2594 | blk_account_io_start(rq); |
320ae51f JA |
2595 | } |
2596 | ||
0f95549c | 2597 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2598 | struct request *rq, bool last) |
f984df1f | 2599 | { |
f984df1f | 2600 | struct request_queue *q = rq->q; |
f984df1f SL |
2601 | struct blk_mq_queue_data bd = { |
2602 | .rq = rq, | |
be94f058 | 2603 | .last = last, |
f984df1f | 2604 | }; |
f06345ad | 2605 | blk_status_t ret; |
0f95549c | 2606 | |
0f95549c MS |
2607 | /* |
2608 | * For OK queue, we are done. For error, caller may kill it. | |
2609 | * Any other error (busy), just add it to our list as we | |
2610 | * previously would have done. | |
2611 | */ | |
2612 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
2613 | switch (ret) { | |
2614 | case BLK_STS_OK: | |
6ce3dd6e | 2615 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2616 | break; |
2617 | case BLK_STS_RESOURCE: | |
86ff7c2a | 2618 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 2619 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
2620 | __blk_mq_requeue_request(rq); |
2621 | break; | |
2622 | default: | |
6ce3dd6e | 2623 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2624 | break; |
2625 | } | |
2626 | ||
2627 | return ret; | |
2628 | } | |
2629 | ||
2b71b877 | 2630 | static bool blk_mq_get_budget_and_tag(struct request *rq) |
0f95549c | 2631 | { |
2a5a24aa | 2632 | int budget_token; |
d964f04a | 2633 | |
2b71b877 | 2634 | budget_token = blk_mq_get_dispatch_budget(rq->q); |
2a5a24aa | 2635 | if (budget_token < 0) |
2b71b877 | 2636 | return false; |
2a5a24aa | 2637 | blk_mq_set_rq_budget_token(rq, budget_token); |
8ab6bb9e | 2638 | if (!blk_mq_get_driver_tag(rq)) { |
2b71b877 CH |
2639 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
2640 | return false; | |
88022d72 | 2641 | } |
2b71b877 | 2642 | return true; |
fd9c40f6 BVA |
2643 | } |
2644 | ||
105663f7 AA |
2645 | /** |
2646 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2647 | * @hctx: Pointer of the associated hardware queue. | |
2648 | * @rq: Pointer to request to be sent. | |
105663f7 AA |
2649 | * |
2650 | * If the device has enough resources to accept a new request now, send the | |
2651 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2652 | * we can try send it another time in the future. Requests inserted at this | |
2653 | * queue have higher priority. | |
2654 | */ | |
fd9c40f6 | 2655 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2656 | struct request *rq) |
fd9c40f6 | 2657 | { |
e1f44ac0 CH |
2658 | blk_status_t ret; |
2659 | ||
2660 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { | |
710fa378 | 2661 | blk_mq_insert_request(rq, 0); |
e1f44ac0 CH |
2662 | return; |
2663 | } | |
2664 | ||
dd6216bb | 2665 | if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) { |
710fa378 | 2666 | blk_mq_insert_request(rq, 0); |
65a558f6 | 2667 | blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT); |
e1f44ac0 CH |
2668 | return; |
2669 | } | |
fd9c40f6 | 2670 | |
e1f44ac0 CH |
2671 | ret = __blk_mq_issue_directly(hctx, rq, true); |
2672 | switch (ret) { | |
2673 | case BLK_STS_OK: | |
2674 | break; | |
2675 | case BLK_STS_RESOURCE: | |
2676 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2677 | blk_mq_request_bypass_insert(rq, 0); |
2394395c | 2678 | blk_mq_run_hw_queue(hctx, false); |
e1f44ac0 CH |
2679 | break; |
2680 | default: | |
fd9c40f6 | 2681 | blk_mq_end_request(rq, ret); |
e1f44ac0 CH |
2682 | break; |
2683 | } | |
fd9c40f6 BVA |
2684 | } |
2685 | ||
06c8c691 | 2686 | static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) |
fd9c40f6 | 2687 | { |
e1f44ac0 CH |
2688 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
2689 | ||
2690 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { | |
710fa378 | 2691 | blk_mq_insert_request(rq, 0); |
e1f44ac0 CH |
2692 | return BLK_STS_OK; |
2693 | } | |
2694 | ||
2695 | if (!blk_mq_get_budget_and_tag(rq)) | |
2696 | return BLK_STS_RESOURCE; | |
2697 | return __blk_mq_issue_directly(hctx, rq, last); | |
5eb6126e CH |
2698 | } |
2699 | ||
3e368fb0 | 2700 | static void blk_mq_plug_issue_direct(struct blk_plug *plug) |
b84c5b50 CH |
2701 | { |
2702 | struct blk_mq_hw_ctx *hctx = NULL; | |
2703 | struct request *rq; | |
2704 | int queued = 0; | |
0d617a83 | 2705 | blk_status_t ret = BLK_STS_OK; |
b84c5b50 CH |
2706 | |
2707 | while ((rq = rq_list_pop(&plug->mq_list))) { | |
2708 | bool last = rq_list_empty(plug->mq_list); | |
b84c5b50 CH |
2709 | |
2710 | if (hctx != rq->mq_hctx) { | |
34c9f547 KS |
2711 | if (hctx) { |
2712 | blk_mq_commit_rqs(hctx, queued, false); | |
2713 | queued = 0; | |
2714 | } | |
b84c5b50 CH |
2715 | hctx = rq->mq_hctx; |
2716 | } | |
2717 | ||
2718 | ret = blk_mq_request_issue_directly(rq, last); | |
2719 | switch (ret) { | |
2720 | case BLK_STS_OK: | |
2721 | queued++; | |
2722 | break; | |
2723 | case BLK_STS_RESOURCE: | |
2724 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2725 | blk_mq_request_bypass_insert(rq, 0); |
2394395c | 2726 | blk_mq_run_hw_queue(hctx, false); |
0d617a83 | 2727 | goto out; |
b84c5b50 CH |
2728 | default: |
2729 | blk_mq_end_request(rq, ret); | |
b84c5b50 CH |
2730 | break; |
2731 | } | |
2732 | } | |
2733 | ||
0d617a83 KS |
2734 | out: |
2735 | if (ret != BLK_STS_OK) | |
34c9f547 | 2736 | blk_mq_commit_rqs(hctx, queued, false); |
b84c5b50 CH |
2737 | } |
2738 | ||
518579a9 KB |
2739 | static void __blk_mq_flush_plug_list(struct request_queue *q, |
2740 | struct blk_plug *plug) | |
2741 | { | |
2742 | if (blk_queue_quiesced(q)) | |
2743 | return; | |
2744 | q->mq_ops->queue_rqs(&plug->mq_list); | |
2745 | } | |
2746 | ||
26fed4ac JA |
2747 | static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched) |
2748 | { | |
2749 | struct blk_mq_hw_ctx *this_hctx = NULL; | |
2750 | struct blk_mq_ctx *this_ctx = NULL; | |
2751 | struct request *requeue_list = NULL; | |
34e0a279 | 2752 | struct request **requeue_lastp = &requeue_list; |
26fed4ac | 2753 | unsigned int depth = 0; |
d97217e7 | 2754 | bool is_passthrough = false; |
26fed4ac JA |
2755 | LIST_HEAD(list); |
2756 | ||
2757 | do { | |
2758 | struct request *rq = rq_list_pop(&plug->mq_list); | |
2759 | ||
2760 | if (!this_hctx) { | |
2761 | this_hctx = rq->mq_hctx; | |
2762 | this_ctx = rq->mq_ctx; | |
d97217e7 ML |
2763 | is_passthrough = blk_rq_is_passthrough(rq); |
2764 | } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx || | |
2765 | is_passthrough != blk_rq_is_passthrough(rq)) { | |
34e0a279 | 2766 | rq_list_add_tail(&requeue_lastp, rq); |
26fed4ac JA |
2767 | continue; |
2768 | } | |
34e0a279 | 2769 | list_add(&rq->queuelist, &list); |
26fed4ac JA |
2770 | depth++; |
2771 | } while (!rq_list_empty(plug->mq_list)); | |
2772 | ||
2773 | plug->mq_list = requeue_list; | |
2774 | trace_block_unplug(this_hctx->queue, depth, !from_sched); | |
05a93117 CH |
2775 | |
2776 | percpu_ref_get(&this_hctx->queue->q_usage_counter); | |
d97217e7 | 2777 | /* passthrough requests should never be issued to the I/O scheduler */ |
2293cae7 ML |
2778 | if (is_passthrough) { |
2779 | spin_lock(&this_hctx->lock); | |
2780 | list_splice_tail_init(&list, &this_hctx->dispatch); | |
2781 | spin_unlock(&this_hctx->lock); | |
2782 | blk_mq_run_hw_queue(this_hctx, from_sched); | |
2783 | } else if (this_hctx->queue->elevator) { | |
05a93117 | 2784 | this_hctx->queue->elevator->type->ops.insert_requests(this_hctx, |
93fffe16 | 2785 | &list, 0); |
05a93117 CH |
2786 | blk_mq_run_hw_queue(this_hctx, from_sched); |
2787 | } else { | |
2788 | blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched); | |
2789 | } | |
2790 | percpu_ref_put(&this_hctx->queue->q_usage_counter); | |
26fed4ac JA |
2791 | } |
2792 | ||
b84c5b50 CH |
2793 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
2794 | { | |
3c67d44d | 2795 | struct request *rq; |
b84c5b50 | 2796 | |
70904263 RL |
2797 | /* |
2798 | * We may have been called recursively midway through handling | |
2799 | * plug->mq_list via a schedule() in the driver's queue_rq() callback. | |
2800 | * To avoid mq_list changing under our feet, clear rq_count early and | |
2801 | * bail out specifically if rq_count is 0 rather than checking | |
2802 | * whether the mq_list is empty. | |
2803 | */ | |
2804 | if (plug->rq_count == 0) | |
b84c5b50 CH |
2805 | return; |
2806 | plug->rq_count = 0; | |
2807 | ||
2808 | if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) { | |
3c67d44d JA |
2809 | struct request_queue *q; |
2810 | ||
2811 | rq = rq_list_peek(&plug->mq_list); | |
2812 | q = rq->q; | |
2813 | ||
2814 | /* | |
2815 | * Peek first request and see if we have a ->queue_rqs() hook. | |
2816 | * If we do, we can dispatch the whole plug list in one go. We | |
2817 | * already know at this point that all requests belong to the | |
2818 | * same queue, caller must ensure that's the case. | |
3c67d44d | 2819 | */ |
434097ee | 2820 | if (q->mq_ops->queue_rqs) { |
3c67d44d | 2821 | blk_mq_run_dispatch_ops(q, |
518579a9 | 2822 | __blk_mq_flush_plug_list(q, plug)); |
3c67d44d JA |
2823 | if (rq_list_empty(plug->mq_list)) |
2824 | return; | |
2825 | } | |
73f3760e ML |
2826 | |
2827 | blk_mq_run_dispatch_ops(q, | |
3e368fb0 | 2828 | blk_mq_plug_issue_direct(plug)); |
b84c5b50 CH |
2829 | if (rq_list_empty(plug->mq_list)) |
2830 | return; | |
2831 | } | |
2832 | ||
b84c5b50 | 2833 | do { |
26fed4ac | 2834 | blk_mq_dispatch_plug_list(plug, from_schedule); |
b84c5b50 | 2835 | } while (!rq_list_empty(plug->mq_list)); |
b84c5b50 CH |
2836 | } |
2837 | ||
94aa228c | 2838 | static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
6ce3dd6e ML |
2839 | struct list_head *list) |
2840 | { | |
536167d4 | 2841 | int queued = 0; |
984ce0a7 | 2842 | blk_status_t ret = BLK_STS_OK; |
536167d4 | 2843 | |
6ce3dd6e | 2844 | while (!list_empty(list)) { |
6ce3dd6e ML |
2845 | struct request *rq = list_first_entry(list, struct request, |
2846 | queuelist); | |
2847 | ||
2848 | list_del_init(&rq->queuelist); | |
fd9c40f6 | 2849 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
27e8b2bb KS |
2850 | switch (ret) { |
2851 | case BLK_STS_OK: | |
536167d4 | 2852 | queued++; |
27e8b2bb KS |
2853 | break; |
2854 | case BLK_STS_RESOURCE: | |
2855 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2856 | blk_mq_request_bypass_insert(rq, 0); |
2394395c CH |
2857 | if (list_empty(list)) |
2858 | blk_mq_run_hw_queue(hctx, false); | |
27e8b2bb KS |
2859 | goto out; |
2860 | default: | |
2861 | blk_mq_end_request(rq, ret); | |
2862 | break; | |
2863 | } | |
6ce3dd6e | 2864 | } |
d666ba98 | 2865 | |
27e8b2bb | 2866 | out: |
984ce0a7 KS |
2867 | if (ret != BLK_STS_OK) |
2868 | blk_mq_commit_rqs(hctx, queued, false); | |
6ce3dd6e ML |
2869 | } |
2870 | ||
b131f201 | 2871 | static bool blk_mq_attempt_bio_merge(struct request_queue *q, |
0c5bcc92 | 2872 | struct bio *bio, unsigned int nr_segs) |
900e0807 JA |
2873 | { |
2874 | if (!blk_queue_nomerges(q) && bio_mergeable(bio)) { | |
0c5bcc92 | 2875 | if (blk_attempt_plug_merge(q, bio, nr_segs)) |
900e0807 JA |
2876 | return true; |
2877 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) | |
2878 | return true; | |
2879 | } | |
2880 | return false; | |
2881 | } | |
2882 | ||
71539717 JA |
2883 | static struct request *blk_mq_get_new_requests(struct request_queue *q, |
2884 | struct blk_plug *plug, | |
0a5aa8d1 SK |
2885 | struct bio *bio, |
2886 | unsigned int nsegs) | |
71539717 JA |
2887 | { |
2888 | struct blk_mq_alloc_data data = { | |
2889 | .q = q, | |
2890 | .nr_tags = 1, | |
9d497e29 | 2891 | .cmd_flags = bio->bi_opf, |
71539717 JA |
2892 | }; |
2893 | struct request *rq; | |
2894 | ||
0a5aa8d1 | 2895 | if (blk_mq_attempt_bio_merge(q, bio, nsegs)) |
b0077e26 | 2896 | return NULL; |
0a5aa8d1 SK |
2897 | |
2898 | rq_qos_throttle(q, bio); | |
2899 | ||
71539717 JA |
2900 | if (plug) { |
2901 | data.nr_tags = plug->nr_ios; | |
2902 | plug->nr_ios = 1; | |
2903 | data.cached_rq = &plug->cached_rq; | |
2904 | } | |
2905 | ||
2906 | rq = __blk_mq_alloc_requests(&data); | |
373b5416 JA |
2907 | if (rq) |
2908 | return rq; | |
71539717 JA |
2909 | rq_qos_cleanup(q, bio); |
2910 | if (bio->bi_opf & REQ_NOWAIT) | |
2911 | bio_wouldblock_error(bio); | |
2912 | return NULL; | |
2913 | } | |
2914 | ||
309ce674 CH |
2915 | /* |
2916 | * Check if we can use the passed on request for submitting the passed in bio, | |
2917 | * and remove it from the request list if it can be used. | |
2918 | */ | |
2919 | static bool blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug, | |
b0077e26 | 2920 | struct bio *bio) |
71539717 | 2921 | { |
b0077e26 CH |
2922 | enum hctx_type type = blk_mq_get_hctx_type(bio->bi_opf); |
2923 | enum hctx_type hctx_type = rq->mq_hctx->type; | |
71539717 | 2924 | |
b0077e26 | 2925 | WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq); |
0a5aa8d1 | 2926 | |
77465647 PB |
2927 | if (type != hctx_type && |
2928 | !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT)) | |
b0077e26 CH |
2929 | return false; |
2930 | if (op_is_flush(rq->cmd_flags) != op_is_flush(bio->bi_opf)) | |
2931 | return false; | |
5b13bc8a | 2932 | |
2645672f JA |
2933 | /* |
2934 | * If any qos ->throttle() end up blocking, we will have flushed the | |
2935 | * plug and hence killed the cached_rq list as well. Pop this entry | |
2936 | * before we throttle. | |
2937 | */ | |
5b13bc8a | 2938 | plug->cached_rq = rq_list_next(rq); |
b0077e26 | 2939 | rq_qos_throttle(rq->q, bio); |
2645672f | 2940 | |
5c17f45e | 2941 | blk_mq_rq_time_init(rq, 0); |
b0077e26 | 2942 | rq->cmd_flags = bio->bi_opf; |
5b13bc8a | 2943 | INIT_LIST_HEAD(&rq->queuelist); |
b0077e26 | 2944 | return true; |
71539717 JA |
2945 | } |
2946 | ||
82b74cac JK |
2947 | static void bio_set_ioprio(struct bio *bio) |
2948 | { | |
a78418e6 JK |
2949 | /* Nobody set ioprio so far? Initialize it based on task's nice value */ |
2950 | if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE) | |
2951 | bio->bi_ioprio = get_current_ioprio(); | |
82b74cac JK |
2952 | blkcg_set_ioprio(bio); |
2953 | } | |
2954 | ||
105663f7 | 2955 | /** |
c62b37d9 | 2956 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2957 | * @bio: Bio pointer. |
2958 | * | |
2959 | * Builds up a request structure from @q and @bio and send to the device. The | |
2960 | * request may not be queued directly to hardware if: | |
2961 | * * This request can be merged with another one | |
2962 | * * We want to place request at plug queue for possible future merging | |
2963 | * * There is an IO scheduler active at this queue | |
2964 | * | |
2965 | * It will not queue the request if there is an error with the bio, or at the | |
2966 | * request creation. | |
105663f7 | 2967 | */ |
3e08773c | 2968 | void blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2969 | { |
ed6cddef | 2970 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
6deacb3b | 2971 | struct blk_plug *plug = blk_mq_plug(bio); |
ef295ecf | 2972 | const int is_sync = op_is_sync(bio->bi_opf); |
f0dbe6e8 | 2973 | struct blk_mq_hw_ctx *hctx; |
b0077e26 | 2974 | struct request *rq = NULL; |
abd45c15 | 2975 | unsigned int nr_segs = 1; |
a892c8d5 | 2976 | blk_status_t ret; |
07068d5b | 2977 | |
51d798cd | 2978 | bio = blk_queue_bounce(bio, q); |
9c6227e0 JK |
2979 | bio_set_ioprio(bio); |
2980 | ||
b0077e26 CH |
2981 | if (plug) { |
2982 | rq = rq_list_peek(&plug->cached_rq); | |
2983 | if (rq && rq->q != q) | |
2984 | rq = NULL; | |
2985 | } | |
2986 | if (rq) { | |
7b4f36cd JA |
2987 | if (unlikely(bio_may_exceed_limits(bio, &q->limits))) { |
2988 | bio = __bio_split_to_limits(bio, &q->limits, &nr_segs); | |
2989 | if (!bio) | |
2990 | return; | |
2991 | } | |
b0077e26 CH |
2992 | if (!bio_integrity_prep(bio)) |
2993 | return; | |
2994 | if (blk_mq_attempt_bio_merge(q, bio, nr_segs)) | |
0a5aa8d1 | 2995 | return; |
309ce674 | 2996 | if (blk_mq_use_cached_rq(rq, plug, bio)) |
b0077e26 CH |
2997 | goto done; |
2998 | percpu_ref_get(&q->q_usage_counter); | |
2999 | } else { | |
3000 | if (unlikely(bio_queue_enter(bio))) | |
5b13bc8a | 3001 | return; |
7b4f36cd JA |
3002 | if (unlikely(bio_may_exceed_limits(bio, &q->limits))) { |
3003 | bio = __bio_split_to_limits(bio, &q->limits, &nr_segs); | |
3004 | if (!bio) | |
3005 | goto fail; | |
3006 | } | |
b0077e26 CH |
3007 | if (!bio_integrity_prep(bio)) |
3008 | goto fail; | |
3009 | } | |
3010 | ||
3011 | rq = blk_mq_get_new_requests(q, plug, bio, nr_segs); | |
3012 | if (unlikely(!rq)) { | |
3013 | fail: | |
3014 | blk_queue_exit(q); | |
3015 | return; | |
5b13bc8a | 3016 | } |
87760e5e | 3017 | |
b0077e26 | 3018 | done: |
e8a676d6 | 3019 | trace_block_getrq(bio); |
d6f1dda2 | 3020 | |
c1c80384 | 3021 | rq_qos_track(q, rq, bio); |
07068d5b | 3022 | |
970d168d BVA |
3023 | blk_mq_bio_to_request(rq, bio, nr_segs); |
3024 | ||
9cd1e566 | 3025 | ret = blk_crypto_rq_get_keyslot(rq); |
a892c8d5 ST |
3026 | if (ret != BLK_STS_OK) { |
3027 | bio->bi_status = ret; | |
3028 | bio_endio(bio); | |
3029 | blk_mq_free_request(rq); | |
3e08773c | 3030 | return; |
a892c8d5 ST |
3031 | } |
3032 | ||
360f2648 | 3033 | if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq)) |
d92ca9d8 CH |
3034 | return; |
3035 | ||
f0dbe6e8 | 3036 | if (plug) { |
ce5b009c | 3037 | blk_add_rq_to_plug(plug, rq); |
f0dbe6e8 CH |
3038 | return; |
3039 | } | |
3040 | ||
3041 | hctx = rq->mq_hctx; | |
dd6216bb | 3042 | if ((rq->rq_flags & RQF_USE_SCHED) || |
f0dbe6e8 | 3043 | (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) { |
710fa378 | 3044 | blk_mq_insert_request(rq, 0); |
f0dbe6e8 CH |
3045 | blk_mq_run_hw_queue(hctx, true); |
3046 | } else { | |
3047 | blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq)); | |
3048 | } | |
320ae51f JA |
3049 | } |
3050 | ||
248c7933 | 3051 | #ifdef CONFIG_BLK_MQ_STACKING |
06c8c691 | 3052 | /** |
a5efda3c | 3053 | * blk_insert_cloned_request - Helper for stacking drivers to submit a request |
a5efda3c | 3054 | * @rq: the request being queued |
06c8c691 | 3055 | */ |
28db4711 | 3056 | blk_status_t blk_insert_cloned_request(struct request *rq) |
06c8c691 | 3057 | { |
28db4711 | 3058 | struct request_queue *q = rq->q; |
06c8c691 | 3059 | unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); |
49d24398 | 3060 | unsigned int max_segments = blk_rq_get_max_segments(rq); |
a5efda3c | 3061 | blk_status_t ret; |
06c8c691 CH |
3062 | |
3063 | if (blk_rq_sectors(rq) > max_sectors) { | |
3064 | /* | |
3065 | * SCSI device does not have a good way to return if | |
3066 | * Write Same/Zero is actually supported. If a device rejects | |
3067 | * a non-read/write command (discard, write same,etc.) the | |
3068 | * low-level device driver will set the relevant queue limit to | |
3069 | * 0 to prevent blk-lib from issuing more of the offending | |
3070 | * operations. Commands queued prior to the queue limit being | |
3071 | * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O | |
3072 | * errors being propagated to upper layers. | |
3073 | */ | |
3074 | if (max_sectors == 0) | |
3075 | return BLK_STS_NOTSUPP; | |
3076 | ||
3077 | printk(KERN_ERR "%s: over max size limit. (%u > %u)\n", | |
3078 | __func__, blk_rq_sectors(rq), max_sectors); | |
3079 | return BLK_STS_IOERR; | |
3080 | } | |
3081 | ||
3082 | /* | |
3083 | * The queue settings related to segment counting may differ from the | |
3084 | * original queue. | |
3085 | */ | |
3086 | rq->nr_phys_segments = blk_recalc_rq_segments(rq); | |
49d24398 US |
3087 | if (rq->nr_phys_segments > max_segments) { |
3088 | printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n", | |
3089 | __func__, rq->nr_phys_segments, max_segments); | |
06c8c691 CH |
3090 | return BLK_STS_IOERR; |
3091 | } | |
3092 | ||
28db4711 | 3093 | if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq))) |
06c8c691 CH |
3094 | return BLK_STS_IOERR; |
3095 | ||
5b8562f0 EB |
3096 | ret = blk_crypto_rq_get_keyslot(rq); |
3097 | if (ret != BLK_STS_OK) | |
3098 | return ret; | |
06c8c691 CH |
3099 | |
3100 | blk_account_io_start(rq); | |
3101 | ||
3102 | /* | |
3103 | * Since we have a scheduler attached on the top device, | |
3104 | * bypass a potential scheduler on the bottom device for | |
3105 | * insert. | |
3106 | */ | |
28db4711 | 3107 | blk_mq_run_dispatch_ops(q, |
4cafe86c | 3108 | ret = blk_mq_request_issue_directly(rq, true)); |
592ee119 YK |
3109 | if (ret) |
3110 | blk_account_io_done(rq, ktime_get_ns()); | |
4cafe86c | 3111 | return ret; |
06c8c691 CH |
3112 | } |
3113 | EXPORT_SYMBOL_GPL(blk_insert_cloned_request); | |
3114 | ||
3115 | /** | |
3116 | * blk_rq_unprep_clone - Helper function to free all bios in a cloned request | |
3117 | * @rq: the clone request to be cleaned up | |
3118 | * | |
3119 | * Description: | |
3120 | * Free all bios in @rq for a cloned request. | |
3121 | */ | |
3122 | void blk_rq_unprep_clone(struct request *rq) | |
3123 | { | |
3124 | struct bio *bio; | |
3125 | ||
3126 | while ((bio = rq->bio) != NULL) { | |
3127 | rq->bio = bio->bi_next; | |
3128 | ||
3129 | bio_put(bio); | |
3130 | } | |
3131 | } | |
3132 | EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); | |
3133 | ||
3134 | /** | |
3135 | * blk_rq_prep_clone - Helper function to setup clone request | |
3136 | * @rq: the request to be setup | |
3137 | * @rq_src: original request to be cloned | |
3138 | * @bs: bio_set that bios for clone are allocated from | |
3139 | * @gfp_mask: memory allocation mask for bio | |
3140 | * @bio_ctr: setup function to be called for each clone bio. | |
3141 | * Returns %0 for success, non %0 for failure. | |
3142 | * @data: private data to be passed to @bio_ctr | |
3143 | * | |
3144 | * Description: | |
3145 | * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. | |
3146 | * Also, pages which the original bios are pointing to are not copied | |
3147 | * and the cloned bios just point same pages. | |
3148 | * So cloned bios must be completed before original bios, which means | |
3149 | * the caller must complete @rq before @rq_src. | |
3150 | */ | |
3151 | int blk_rq_prep_clone(struct request *rq, struct request *rq_src, | |
3152 | struct bio_set *bs, gfp_t gfp_mask, | |
3153 | int (*bio_ctr)(struct bio *, struct bio *, void *), | |
3154 | void *data) | |
3155 | { | |
3156 | struct bio *bio, *bio_src; | |
3157 | ||
3158 | if (!bs) | |
3159 | bs = &fs_bio_set; | |
3160 | ||
3161 | __rq_for_each_bio(bio_src, rq_src) { | |
abfc426d CH |
3162 | bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask, |
3163 | bs); | |
06c8c691 CH |
3164 | if (!bio) |
3165 | goto free_and_out; | |
3166 | ||
3167 | if (bio_ctr && bio_ctr(bio, bio_src, data)) | |
3168 | goto free_and_out; | |
3169 | ||
3170 | if (rq->bio) { | |
3171 | rq->biotail->bi_next = bio; | |
3172 | rq->biotail = bio; | |
3173 | } else { | |
3174 | rq->bio = rq->biotail = bio; | |
3175 | } | |
3176 | bio = NULL; | |
3177 | } | |
3178 | ||
3179 | /* Copy attributes of the original request to the clone request. */ | |
3180 | rq->__sector = blk_rq_pos(rq_src); | |
3181 | rq->__data_len = blk_rq_bytes(rq_src); | |
3182 | if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) { | |
3183 | rq->rq_flags |= RQF_SPECIAL_PAYLOAD; | |
3184 | rq->special_vec = rq_src->special_vec; | |
3185 | } | |
3186 | rq->nr_phys_segments = rq_src->nr_phys_segments; | |
3187 | rq->ioprio = rq_src->ioprio; | |
3188 | ||
3189 | if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0) | |
3190 | goto free_and_out; | |
3191 | ||
3192 | return 0; | |
3193 | ||
3194 | free_and_out: | |
3195 | if (bio) | |
3196 | bio_put(bio); | |
3197 | blk_rq_unprep_clone(rq); | |
3198 | ||
3199 | return -ENOMEM; | |
3200 | } | |
3201 | EXPORT_SYMBOL_GPL(blk_rq_prep_clone); | |
248c7933 | 3202 | #endif /* CONFIG_BLK_MQ_STACKING */ |
06c8c691 | 3203 | |
f2b8f3ce CH |
3204 | /* |
3205 | * Steal bios from a request and add them to a bio list. | |
3206 | * The request must not have been partially completed before. | |
3207 | */ | |
3208 | void blk_steal_bios(struct bio_list *list, struct request *rq) | |
3209 | { | |
3210 | if (rq->bio) { | |
3211 | if (list->tail) | |
3212 | list->tail->bi_next = rq->bio; | |
3213 | else | |
3214 | list->head = rq->bio; | |
3215 | list->tail = rq->biotail; | |
3216 | ||
3217 | rq->bio = NULL; | |
3218 | rq->biotail = NULL; | |
3219 | } | |
3220 | ||
3221 | rq->__data_len = 0; | |
3222 | } | |
3223 | EXPORT_SYMBOL_GPL(blk_steal_bios); | |
3224 | ||
bd63141d ML |
3225 | static size_t order_to_size(unsigned int order) |
3226 | { | |
3227 | return (size_t)PAGE_SIZE << order; | |
3228 | } | |
3229 | ||
3230 | /* called before freeing request pool in @tags */ | |
f32e4eaf JG |
3231 | static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, |
3232 | struct blk_mq_tags *tags) | |
bd63141d | 3233 | { |
bd63141d ML |
3234 | struct page *page; |
3235 | unsigned long flags; | |
3236 | ||
76dd2980 YK |
3237 | /* |
3238 | * There is no need to clear mapping if driver tags is not initialized | |
3239 | * or the mapping belongs to the driver tags. | |
3240 | */ | |
3241 | if (!drv_tags || drv_tags == tags) | |
4f245d5b JG |
3242 | return; |
3243 | ||
bd63141d ML |
3244 | list_for_each_entry(page, &tags->page_list, lru) { |
3245 | unsigned long start = (unsigned long)page_address(page); | |
3246 | unsigned long end = start + order_to_size(page->private); | |
3247 | int i; | |
3248 | ||
f32e4eaf | 3249 | for (i = 0; i < drv_tags->nr_tags; i++) { |
bd63141d ML |
3250 | struct request *rq = drv_tags->rqs[i]; |
3251 | unsigned long rq_addr = (unsigned long)rq; | |
3252 | ||
3253 | if (rq_addr >= start && rq_addr < end) { | |
0a467d0f | 3254 | WARN_ON_ONCE(req_ref_read(rq) != 0); |
bd63141d ML |
3255 | cmpxchg(&drv_tags->rqs[i], rq, NULL); |
3256 | } | |
3257 | } | |
3258 | } | |
3259 | ||
3260 | /* | |
3261 | * Wait until all pending iteration is done. | |
3262 | * | |
3263 | * Request reference is cleared and it is guaranteed to be observed | |
3264 | * after the ->lock is released. | |
3265 | */ | |
3266 | spin_lock_irqsave(&drv_tags->lock, flags); | |
3267 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
3268 | } | |
3269 | ||
cc71a6f4 JA |
3270 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
3271 | unsigned int hctx_idx) | |
95363efd | 3272 | { |
f32e4eaf | 3273 | struct blk_mq_tags *drv_tags; |
e9b267d9 | 3274 | struct page *page; |
320ae51f | 3275 | |
e02657ea ML |
3276 | if (list_empty(&tags->page_list)) |
3277 | return; | |
3278 | ||
079a2e3e JG |
3279 | if (blk_mq_is_shared_tags(set->flags)) |
3280 | drv_tags = set->shared_tags; | |
e155b0c2 JG |
3281 | else |
3282 | drv_tags = set->tags[hctx_idx]; | |
f32e4eaf | 3283 | |
65de57bb | 3284 | if (tags->static_rqs && set->ops->exit_request) { |
e9b267d9 | 3285 | int i; |
320ae51f | 3286 | |
24d2f903 | 3287 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
3288 | struct request *rq = tags->static_rqs[i]; |
3289 | ||
3290 | if (!rq) | |
e9b267d9 | 3291 | continue; |
d6296d39 | 3292 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 3293 | tags->static_rqs[i] = NULL; |
e9b267d9 | 3294 | } |
320ae51f | 3295 | } |
320ae51f | 3296 | |
f32e4eaf | 3297 | blk_mq_clear_rq_mapping(drv_tags, tags); |
bd63141d | 3298 | |
24d2f903 CH |
3299 | while (!list_empty(&tags->page_list)) { |
3300 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 3301 | list_del_init(&page->lru); |
f75782e4 CM |
3302 | /* |
3303 | * Remove kmemleak object previously allocated in | |
273938bf | 3304 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
3305 | */ |
3306 | kmemleak_free(page_address(page)); | |
320ae51f JA |
3307 | __free_pages(page, page->private); |
3308 | } | |
cc71a6f4 | 3309 | } |
320ae51f | 3310 | |
e155b0c2 | 3311 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
cc71a6f4 | 3312 | { |
24d2f903 | 3313 | kfree(tags->rqs); |
cc71a6f4 | 3314 | tags->rqs = NULL; |
2af8cbe3 JA |
3315 | kfree(tags->static_rqs); |
3316 | tags->static_rqs = NULL; | |
320ae51f | 3317 | |
e155b0c2 | 3318 | blk_mq_free_tags(tags); |
320ae51f JA |
3319 | } |
3320 | ||
4d805131 ML |
3321 | static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set, |
3322 | unsigned int hctx_idx) | |
3323 | { | |
3324 | int i; | |
3325 | ||
3326 | for (i = 0; i < set->nr_maps; i++) { | |
3327 | unsigned int start = set->map[i].queue_offset; | |
3328 | unsigned int end = start + set->map[i].nr_queues; | |
3329 | ||
3330 | if (hctx_idx >= start && hctx_idx < end) | |
3331 | break; | |
3332 | } | |
3333 | ||
3334 | if (i >= set->nr_maps) | |
3335 | i = HCTX_TYPE_DEFAULT; | |
3336 | ||
3337 | return i; | |
3338 | } | |
3339 | ||
3340 | static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set, | |
3341 | unsigned int hctx_idx) | |
3342 | { | |
3343 | enum hctx_type type = hctx_idx_to_type(set, hctx_idx); | |
3344 | ||
3345 | return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx); | |
3346 | } | |
3347 | ||
63064be1 JG |
3348 | static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
3349 | unsigned int hctx_idx, | |
3350 | unsigned int nr_tags, | |
e155b0c2 | 3351 | unsigned int reserved_tags) |
320ae51f | 3352 | { |
4d805131 | 3353 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
24d2f903 | 3354 | struct blk_mq_tags *tags; |
320ae51f | 3355 | |
59f082e4 SL |
3356 | if (node == NUMA_NO_NODE) |
3357 | node = set->numa_node; | |
3358 | ||
e155b0c2 JG |
3359 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
3360 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
3361 | if (!tags) |
3362 | return NULL; | |
320ae51f | 3363 | |
590b5b7d | 3364 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 3365 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 3366 | node); |
7edfd681 JC |
3367 | if (!tags->rqs) |
3368 | goto err_free_tags; | |
320ae51f | 3369 | |
590b5b7d KC |
3370 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
3371 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
3372 | node); | |
7edfd681 JC |
3373 | if (!tags->static_rqs) |
3374 | goto err_free_rqs; | |
2af8cbe3 | 3375 | |
cc71a6f4 | 3376 | return tags; |
7edfd681 JC |
3377 | |
3378 | err_free_rqs: | |
3379 | kfree(tags->rqs); | |
3380 | err_free_tags: | |
3381 | blk_mq_free_tags(tags); | |
3382 | return NULL; | |
cc71a6f4 JA |
3383 | } |
3384 | ||
1d9bd516 TH |
3385 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
3386 | unsigned int hctx_idx, int node) | |
3387 | { | |
3388 | int ret; | |
3389 | ||
3390 | if (set->ops->init_request) { | |
3391 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
3392 | if (ret) | |
3393 | return ret; | |
3394 | } | |
3395 | ||
12f5b931 | 3396 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
3397 | return 0; |
3398 | } | |
3399 | ||
63064be1 JG |
3400 | static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, |
3401 | struct blk_mq_tags *tags, | |
3402 | unsigned int hctx_idx, unsigned int depth) | |
cc71a6f4 JA |
3403 | { |
3404 | unsigned int i, j, entries_per_page, max_order = 4; | |
4d805131 | 3405 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
cc71a6f4 | 3406 | size_t rq_size, left; |
59f082e4 | 3407 | |
59f082e4 SL |
3408 | if (node == NUMA_NO_NODE) |
3409 | node = set->numa_node; | |
cc71a6f4 JA |
3410 | |
3411 | INIT_LIST_HEAD(&tags->page_list); | |
3412 | ||
320ae51f JA |
3413 | /* |
3414 | * rq_size is the size of the request plus driver payload, rounded | |
3415 | * to the cacheline size | |
3416 | */ | |
24d2f903 | 3417 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 3418 | cache_line_size()); |
cc71a6f4 | 3419 | left = rq_size * depth; |
320ae51f | 3420 | |
cc71a6f4 | 3421 | for (i = 0; i < depth; ) { |
320ae51f JA |
3422 | int this_order = max_order; |
3423 | struct page *page; | |
3424 | int to_do; | |
3425 | void *p; | |
3426 | ||
b3a834b1 | 3427 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
3428 | this_order--; |
3429 | ||
3430 | do { | |
59f082e4 | 3431 | page = alloc_pages_node(node, |
36e1f3d1 | 3432 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 3433 | this_order); |
320ae51f JA |
3434 | if (page) |
3435 | break; | |
3436 | if (!this_order--) | |
3437 | break; | |
3438 | if (order_to_size(this_order) < rq_size) | |
3439 | break; | |
3440 | } while (1); | |
3441 | ||
3442 | if (!page) | |
24d2f903 | 3443 | goto fail; |
320ae51f JA |
3444 | |
3445 | page->private = this_order; | |
24d2f903 | 3446 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
3447 | |
3448 | p = page_address(page); | |
f75782e4 CM |
3449 | /* |
3450 | * Allow kmemleak to scan these pages as they contain pointers | |
3451 | * to additional allocations like via ops->init_request(). | |
3452 | */ | |
36e1f3d1 | 3453 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 3454 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 3455 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
3456 | left -= to_do * rq_size; |
3457 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
3458 | struct request *rq = p; |
3459 | ||
3460 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
3461 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
3462 | tags->static_rqs[i] = NULL; | |
3463 | goto fail; | |
e9b267d9 CH |
3464 | } |
3465 | ||
320ae51f JA |
3466 | p += rq_size; |
3467 | i++; | |
3468 | } | |
3469 | } | |
cc71a6f4 | 3470 | return 0; |
320ae51f | 3471 | |
24d2f903 | 3472 | fail: |
cc71a6f4 JA |
3473 | blk_mq_free_rqs(set, tags, hctx_idx); |
3474 | return -ENOMEM; | |
320ae51f JA |
3475 | } |
3476 | ||
bf0beec0 ML |
3477 | struct rq_iter_data { |
3478 | struct blk_mq_hw_ctx *hctx; | |
3479 | bool has_rq; | |
3480 | }; | |
3481 | ||
2dd6532e | 3482 | static bool blk_mq_has_request(struct request *rq, void *data) |
bf0beec0 ML |
3483 | { |
3484 | struct rq_iter_data *iter_data = data; | |
3485 | ||
3486 | if (rq->mq_hctx != iter_data->hctx) | |
3487 | return true; | |
3488 | iter_data->has_rq = true; | |
3489 | return false; | |
3490 | } | |
3491 | ||
3492 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
3493 | { | |
3494 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
3495 | hctx->sched_tags : hctx->tags; | |
3496 | struct rq_iter_data data = { | |
3497 | .hctx = hctx, | |
3498 | }; | |
3499 | ||
3500 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
3501 | return data.has_rq; | |
3502 | } | |
3503 | ||
3504 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
3505 | struct blk_mq_hw_ctx *hctx) | |
3506 | { | |
9b51d9d8 | 3507 | if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu) |
bf0beec0 ML |
3508 | return false; |
3509 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
3510 | return false; | |
3511 | return true; | |
3512 | } | |
3513 | ||
3514 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
3515 | { | |
3516 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3517 | struct blk_mq_hw_ctx, cpuhp_online); | |
3518 | ||
3519 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
3520 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
3521 | return 0; | |
3522 | ||
3523 | /* | |
3524 | * Prevent new request from being allocated on the current hctx. | |
3525 | * | |
3526 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
3527 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
3528 | * seen once we return from the tag allocator. | |
3529 | */ | |
3530 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3531 | smp_mb__after_atomic(); | |
3532 | ||
3533 | /* | |
3534 | * Try to grab a reference to the queue and wait for any outstanding | |
3535 | * requests. If we could not grab a reference the queue has been | |
3536 | * frozen and there are no requests. | |
3537 | */ | |
3538 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
3539 | while (blk_mq_hctx_has_requests(hctx)) | |
3540 | msleep(5); | |
3541 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
3542 | } | |
3543 | ||
3544 | return 0; | |
3545 | } | |
3546 | ||
3547 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
3548 | { | |
3549 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3550 | struct blk_mq_hw_ctx, cpuhp_online); | |
3551 | ||
3552 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
3553 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3554 | return 0; | |
3555 | } | |
3556 | ||
e57690fe JA |
3557 | /* |
3558 | * 'cpu' is going away. splice any existing rq_list entries from this | |
3559 | * software queue to the hw queue dispatch list, and ensure that it | |
3560 | * gets run. | |
3561 | */ | |
9467f859 | 3562 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 3563 | { |
9467f859 | 3564 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
3565 | struct blk_mq_ctx *ctx; |
3566 | LIST_HEAD(tmp); | |
c16d6b5a | 3567 | enum hctx_type type; |
484b4061 | 3568 | |
9467f859 | 3569 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
3570 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
3571 | return 0; | |
3572 | ||
e57690fe | 3573 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 3574 | type = hctx->type; |
484b4061 JA |
3575 | |
3576 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
3577 | if (!list_empty(&ctx->rq_lists[type])) { |
3578 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
3579 | blk_mq_hctx_clear_pending(hctx, ctx); |
3580 | } | |
3581 | spin_unlock(&ctx->lock); | |
3582 | ||
3583 | if (list_empty(&tmp)) | |
9467f859 | 3584 | return 0; |
484b4061 | 3585 | |
e57690fe JA |
3586 | spin_lock(&hctx->lock); |
3587 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
3588 | spin_unlock(&hctx->lock); | |
484b4061 JA |
3589 | |
3590 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 3591 | return 0; |
484b4061 JA |
3592 | } |
3593 | ||
9467f859 | 3594 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 3595 | { |
bf0beec0 ML |
3596 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3597 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3598 | &hctx->cpuhp_online); | |
9467f859 TG |
3599 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
3600 | &hctx->cpuhp_dead); | |
484b4061 JA |
3601 | } |
3602 | ||
364b6181 ML |
3603 | /* |
3604 | * Before freeing hw queue, clearing the flush request reference in | |
3605 | * tags->rqs[] for avoiding potential UAF. | |
3606 | */ | |
3607 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
3608 | unsigned int queue_depth, struct request *flush_rq) | |
3609 | { | |
3610 | int i; | |
3611 | unsigned long flags; | |
3612 | ||
3613 | /* The hw queue may not be mapped yet */ | |
3614 | if (!tags) | |
3615 | return; | |
3616 | ||
0a467d0f | 3617 | WARN_ON_ONCE(req_ref_read(flush_rq) != 0); |
364b6181 ML |
3618 | |
3619 | for (i = 0; i < queue_depth; i++) | |
3620 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
3621 | ||
3622 | /* | |
3623 | * Wait until all pending iteration is done. | |
3624 | * | |
3625 | * Request reference is cleared and it is guaranteed to be observed | |
3626 | * after the ->lock is released. | |
3627 | */ | |
3628 | spin_lock_irqsave(&tags->lock, flags); | |
3629 | spin_unlock_irqrestore(&tags->lock, flags); | |
3630 | } | |
3631 | ||
c3b4afca | 3632 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
3633 | static void blk_mq_exit_hctx(struct request_queue *q, |
3634 | struct blk_mq_tag_set *set, | |
3635 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
3636 | { | |
364b6181 ML |
3637 | struct request *flush_rq = hctx->fq->flush_rq; |
3638 | ||
8ab0b7dc ML |
3639 | if (blk_mq_hw_queue_mapped(hctx)) |
3640 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 3641 | |
6cfeadbf ML |
3642 | if (blk_queue_init_done(q)) |
3643 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], | |
3644 | set->queue_depth, flush_rq); | |
f70ced09 | 3645 | if (set->ops->exit_request) |
364b6181 | 3646 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 3647 | |
08e98fc6 ML |
3648 | if (set->ops->exit_hctx) |
3649 | set->ops->exit_hctx(hctx, hctx_idx); | |
3650 | ||
9467f859 | 3651 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 | 3652 | |
4e5cc99e ML |
3653 | xa_erase(&q->hctx_table, hctx_idx); |
3654 | ||
2f8f1336 ML |
3655 | spin_lock(&q->unused_hctx_lock); |
3656 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
3657 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
3658 | } |
3659 | ||
624dbe47 ML |
3660 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
3661 | struct blk_mq_tag_set *set, int nr_queue) | |
3662 | { | |
3663 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3664 | unsigned long i; |
624dbe47 ML |
3665 | |
3666 | queue_for_each_hw_ctx(q, hctx, i) { | |
3667 | if (i == nr_queue) | |
3668 | break; | |
08e98fc6 | 3669 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 3670 | } |
624dbe47 ML |
3671 | } |
3672 | ||
08e98fc6 ML |
3673 | static int blk_mq_init_hctx(struct request_queue *q, |
3674 | struct blk_mq_tag_set *set, | |
3675 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 3676 | { |
7c6c5b7c ML |
3677 | hctx->queue_num = hctx_idx; |
3678 | ||
bf0beec0 ML |
3679 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3680 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3681 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
3682 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
3683 | ||
3684 | hctx->tags = set->tags[hctx_idx]; | |
3685 | ||
3686 | if (set->ops->init_hctx && | |
3687 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
3688 | goto unregister_cpu_notifier; | |
08e98fc6 | 3689 | |
7c6c5b7c ML |
3690 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
3691 | hctx->numa_node)) | |
3692 | goto exit_hctx; | |
4e5cc99e ML |
3693 | |
3694 | if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL)) | |
3695 | goto exit_flush_rq; | |
3696 | ||
7c6c5b7c ML |
3697 | return 0; |
3698 | ||
4e5cc99e ML |
3699 | exit_flush_rq: |
3700 | if (set->ops->exit_request) | |
3701 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); | |
7c6c5b7c ML |
3702 | exit_hctx: |
3703 | if (set->ops->exit_hctx) | |
3704 | set->ops->exit_hctx(hctx, hctx_idx); | |
3705 | unregister_cpu_notifier: | |
3706 | blk_mq_remove_cpuhp(hctx); | |
3707 | return -1; | |
3708 | } | |
3709 | ||
3710 | static struct blk_mq_hw_ctx * | |
3711 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
3712 | int node) | |
3713 | { | |
3714 | struct blk_mq_hw_ctx *hctx; | |
3715 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
3716 | ||
704b914f | 3717 | hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node); |
7c6c5b7c ML |
3718 | if (!hctx) |
3719 | goto fail_alloc_hctx; | |
3720 | ||
3721 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
3722 | goto free_hctx; | |
3723 | ||
3724 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 3725 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
3726 | node = set->numa_node; |
3727 | hctx->numa_node = node; | |
08e98fc6 | 3728 | |
9f993737 | 3729 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
3730 | spin_lock_init(&hctx->lock); |
3731 | INIT_LIST_HEAD(&hctx->dispatch); | |
3732 | hctx->queue = q; | |
51db1c37 | 3733 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 3734 | |
2f8f1336 ML |
3735 | INIT_LIST_HEAD(&hctx->hctx_list); |
3736 | ||
320ae51f | 3737 | /* |
08e98fc6 ML |
3738 | * Allocate space for all possible cpus to avoid allocation at |
3739 | * runtime | |
320ae51f | 3740 | */ |
d904bfa7 | 3741 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 3742 | gfp, node); |
08e98fc6 | 3743 | if (!hctx->ctxs) |
7c6c5b7c | 3744 | goto free_cpumask; |
320ae51f | 3745 | |
5b202853 | 3746 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 3747 | gfp, node, false, false)) |
08e98fc6 | 3748 | goto free_ctxs; |
08e98fc6 | 3749 | hctx->nr_ctx = 0; |
320ae51f | 3750 | |
5815839b | 3751 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
3752 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
3753 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
3754 | ||
754a1572 | 3755 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 3756 | if (!hctx->fq) |
7c6c5b7c | 3757 | goto free_bitmap; |
320ae51f | 3758 | |
7c6c5b7c | 3759 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 3760 | |
7c6c5b7c | 3761 | return hctx; |
320ae51f | 3762 | |
08e98fc6 | 3763 | free_bitmap: |
88459642 | 3764 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
3765 | free_ctxs: |
3766 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
3767 | free_cpumask: |
3768 | free_cpumask_var(hctx->cpumask); | |
3769 | free_hctx: | |
3770 | kfree(hctx); | |
3771 | fail_alloc_hctx: | |
3772 | return NULL; | |
08e98fc6 | 3773 | } |
320ae51f | 3774 | |
320ae51f JA |
3775 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
3776 | unsigned int nr_hw_queues) | |
3777 | { | |
b3c661b1 JA |
3778 | struct blk_mq_tag_set *set = q->tag_set; |
3779 | unsigned int i, j; | |
320ae51f JA |
3780 | |
3781 | for_each_possible_cpu(i) { | |
3782 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
3783 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 3784 | int k; |
320ae51f | 3785 | |
320ae51f JA |
3786 | __ctx->cpu = i; |
3787 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
3788 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
3789 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
3790 | ||
320ae51f JA |
3791 | __ctx->queue = q; |
3792 | ||
320ae51f JA |
3793 | /* |
3794 | * Set local node, IFF we have more than one hw queue. If | |
3795 | * not, we remain on the home node of the device | |
3796 | */ | |
b3c661b1 JA |
3797 | for (j = 0; j < set->nr_maps; j++) { |
3798 | hctx = blk_mq_map_queue_type(q, j, i); | |
3799 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 3800 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 3801 | } |
320ae51f JA |
3802 | } |
3803 | } | |
3804 | ||
63064be1 JG |
3805 | struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3806 | unsigned int hctx_idx, | |
3807 | unsigned int depth) | |
cc71a6f4 | 3808 | { |
63064be1 JG |
3809 | struct blk_mq_tags *tags; |
3810 | int ret; | |
cc71a6f4 | 3811 | |
e155b0c2 | 3812 | tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); |
63064be1 JG |
3813 | if (!tags) |
3814 | return NULL; | |
cc71a6f4 | 3815 | |
63064be1 JG |
3816 | ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); |
3817 | if (ret) { | |
e155b0c2 | 3818 | blk_mq_free_rq_map(tags); |
63064be1 JG |
3819 | return NULL; |
3820 | } | |
cc71a6f4 | 3821 | |
63064be1 | 3822 | return tags; |
cc71a6f4 JA |
3823 | } |
3824 | ||
63064be1 JG |
3825 | static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3826 | int hctx_idx) | |
cc71a6f4 | 3827 | { |
079a2e3e JG |
3828 | if (blk_mq_is_shared_tags(set->flags)) { |
3829 | set->tags[hctx_idx] = set->shared_tags; | |
1c0706a7 | 3830 | |
e155b0c2 | 3831 | return true; |
bd166ef1 | 3832 | } |
e155b0c2 | 3833 | |
63064be1 JG |
3834 | set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, |
3835 | set->queue_depth); | |
3836 | ||
3837 | return set->tags[hctx_idx]; | |
cc71a6f4 JA |
3838 | } |
3839 | ||
645db34e JG |
3840 | void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3841 | struct blk_mq_tags *tags, | |
3842 | unsigned int hctx_idx) | |
cc71a6f4 | 3843 | { |
645db34e JG |
3844 | if (tags) { |
3845 | blk_mq_free_rqs(set, tags, hctx_idx); | |
e155b0c2 | 3846 | blk_mq_free_rq_map(tags); |
bd166ef1 | 3847 | } |
cc71a6f4 JA |
3848 | } |
3849 | ||
e155b0c2 JG |
3850 | static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3851 | unsigned int hctx_idx) | |
3852 | { | |
079a2e3e | 3853 | if (!blk_mq_is_shared_tags(set->flags)) |
e155b0c2 JG |
3854 | blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); |
3855 | ||
3856 | set->tags[hctx_idx] = NULL; | |
cc71a6f4 JA |
3857 | } |
3858 | ||
4b855ad3 | 3859 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 3860 | { |
4f481208 ML |
3861 | unsigned int j, hctx_idx; |
3862 | unsigned long i; | |
320ae51f JA |
3863 | struct blk_mq_hw_ctx *hctx; |
3864 | struct blk_mq_ctx *ctx; | |
2a34c087 | 3865 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
3866 | |
3867 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 3868 | cpumask_clear(hctx->cpumask); |
320ae51f | 3869 | hctx->nr_ctx = 0; |
d416c92c | 3870 | hctx->dispatch_from = NULL; |
320ae51f JA |
3871 | } |
3872 | ||
3873 | /* | |
4b855ad3 | 3874 | * Map software to hardware queues. |
4412efec ML |
3875 | * |
3876 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 3877 | */ |
20e4d813 | 3878 | for_each_possible_cpu(i) { |
4412efec | 3879 | |
897bb0c7 | 3880 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 3881 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
3882 | if (!set->map[j].nr_queues) { |
3883 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3884 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 3885 | continue; |
bb94aea1 | 3886 | } |
fd689871 ML |
3887 | hctx_idx = set->map[j].mq_map[i]; |
3888 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
3889 | if (!set->tags[hctx_idx] && | |
63064be1 | 3890 | !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { |
fd689871 ML |
3891 | /* |
3892 | * If tags initialization fail for some hctx, | |
3893 | * that hctx won't be brought online. In this | |
3894 | * case, remap the current ctx to hctx[0] which | |
3895 | * is guaranteed to always have tags allocated | |
3896 | */ | |
3897 | set->map[j].mq_map[i] = 0; | |
3898 | } | |
e5edd5f2 | 3899 | |
b3c661b1 | 3900 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 3901 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
3902 | /* |
3903 | * If the CPU is already set in the mask, then we've | |
3904 | * mapped this one already. This can happen if | |
3905 | * devices share queues across queue maps. | |
3906 | */ | |
3907 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
3908 | continue; | |
3909 | ||
3910 | cpumask_set_cpu(i, hctx->cpumask); | |
3911 | hctx->type = j; | |
3912 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
3913 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
3914 | ||
3915 | /* | |
3916 | * If the nr_ctx type overflows, we have exceeded the | |
3917 | * amount of sw queues we can support. | |
3918 | */ | |
3919 | BUG_ON(!hctx->nr_ctx); | |
3920 | } | |
bb94aea1 JW |
3921 | |
3922 | for (; j < HCTX_MAX_TYPES; j++) | |
3923 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3924 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 3925 | } |
506e931f JA |
3926 | |
3927 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
3928 | /* |
3929 | * If no software queues are mapped to this hardware queue, | |
3930 | * disable it and free the request entries. | |
3931 | */ | |
3932 | if (!hctx->nr_ctx) { | |
3933 | /* Never unmap queue 0. We need it as a | |
3934 | * fallback in case of a new remap fails | |
3935 | * allocation | |
3936 | */ | |
e155b0c2 JG |
3937 | if (i) |
3938 | __blk_mq_free_map_and_rqs(set, i); | |
4412efec ML |
3939 | |
3940 | hctx->tags = NULL; | |
3941 | continue; | |
3942 | } | |
484b4061 | 3943 | |
2a34c087 ML |
3944 | hctx->tags = set->tags[i]; |
3945 | WARN_ON(!hctx->tags); | |
3946 | ||
889fa31f CY |
3947 | /* |
3948 | * Set the map size to the number of mapped software queues. | |
3949 | * This is more accurate and more efficient than looping | |
3950 | * over all possibly mapped software queues. | |
3951 | */ | |
88459642 | 3952 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 3953 | |
484b4061 JA |
3954 | /* |
3955 | * Initialize batch roundrobin counts | |
3956 | */ | |
f82ddf19 | 3957 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
3958 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
3959 | } | |
320ae51f JA |
3960 | } |
3961 | ||
8e8320c9 JA |
3962 | /* |
3963 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
3964 | * the queue isn't live yet. | |
3965 | */ | |
2404e607 | 3966 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
3967 | { |
3968 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3969 | unsigned long i; |
0d2602ca | 3970 | |
2404e607 | 3971 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 3972 | if (shared) { |
51db1c37 | 3973 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
3974 | } else { |
3975 | blk_mq_tag_idle(hctx); | |
51db1c37 | 3976 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 3977 | } |
2404e607 JM |
3978 | } |
3979 | } | |
3980 | ||
655ac300 HR |
3981 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
3982 | bool shared) | |
2404e607 JM |
3983 | { |
3984 | struct request_queue *q; | |
0d2602ca | 3985 | |
705cda97 BVA |
3986 | lockdep_assert_held(&set->tag_list_lock); |
3987 | ||
0d2602ca JA |
3988 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3989 | blk_mq_freeze_queue(q); | |
2404e607 | 3990 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
3991 | blk_mq_unfreeze_queue(q); |
3992 | } | |
3993 | } | |
3994 | ||
3995 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
3996 | { | |
3997 | struct blk_mq_tag_set *set = q->tag_set; | |
3998 | ||
0d2602ca | 3999 | mutex_lock(&set->tag_list_lock); |
08c875cb | 4000 | list_del(&q->tag_set_list); |
2404e607 JM |
4001 | if (list_is_singular(&set->tag_list)) { |
4002 | /* just transitioned to unshared */ | |
51db1c37 | 4003 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 4004 | /* update existing queue */ |
655ac300 | 4005 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 4006 | } |
0d2602ca | 4007 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 4008 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
4009 | } |
4010 | ||
4011 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
4012 | struct request_queue *q) | |
4013 | { | |
0d2602ca | 4014 | mutex_lock(&set->tag_list_lock); |
2404e607 | 4015 | |
ff821d27 JA |
4016 | /* |
4017 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
4018 | */ | |
4019 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
4020 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
4021 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 4022 | /* update existing queue */ |
655ac300 | 4023 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 4024 | } |
51db1c37 | 4025 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 4026 | queue_set_hctx_shared(q, true); |
08c875cb | 4027 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 4028 | |
0d2602ca JA |
4029 | mutex_unlock(&set->tag_list_lock); |
4030 | } | |
4031 | ||
1db4909e ML |
4032 | /* All allocations will be freed in release handler of q->mq_kobj */ |
4033 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
4034 | { | |
4035 | struct blk_mq_ctxs *ctxs; | |
4036 | int cpu; | |
4037 | ||
4038 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
4039 | if (!ctxs) | |
4040 | return -ENOMEM; | |
4041 | ||
4042 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
4043 | if (!ctxs->queue_ctx) | |
4044 | goto fail; | |
4045 | ||
4046 | for_each_possible_cpu(cpu) { | |
4047 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
4048 | ctx->ctxs = ctxs; | |
4049 | } | |
4050 | ||
4051 | q->mq_kobj = &ctxs->kobj; | |
4052 | q->queue_ctx = ctxs->queue_ctx; | |
4053 | ||
4054 | return 0; | |
4055 | fail: | |
4056 | kfree(ctxs); | |
4057 | return -ENOMEM; | |
4058 | } | |
4059 | ||
e09aae7e ML |
4060 | /* |
4061 | * It is the actual release handler for mq, but we do it from | |
4062 | * request queue's release handler for avoiding use-after-free | |
4063 | * and headache because q->mq_kobj shouldn't have been introduced, | |
4064 | * but we can't group ctx/kctx kobj without it. | |
4065 | */ | |
4066 | void blk_mq_release(struct request_queue *q) | |
4067 | { | |
2f8f1336 | 4068 | struct blk_mq_hw_ctx *hctx, *next; |
4f481208 | 4069 | unsigned long i; |
e09aae7e | 4070 | |
2f8f1336 ML |
4071 | queue_for_each_hw_ctx(q, hctx, i) |
4072 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
4073 | ||
4074 | /* all hctx are in .unused_hctx_list now */ | |
4075 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
4076 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 4077 | kobject_put(&hctx->kobj); |
c3b4afca | 4078 | } |
e09aae7e | 4079 | |
4e5cc99e | 4080 | xa_destroy(&q->hctx_table); |
e09aae7e | 4081 | |
7ea5fe31 ML |
4082 | /* |
4083 | * release .mq_kobj and sw queue's kobject now because | |
4084 | * both share lifetime with request queue. | |
4085 | */ | |
4086 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
4087 | } |
4088 | ||
5ec780a6 | 4089 | static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
2f227bb9 | 4090 | void *queuedata) |
b62c21b7 | 4091 | { |
26a9750a CH |
4092 | struct request_queue *q; |
4093 | int ret; | |
b62c21b7 | 4094 | |
80bd4a7a | 4095 | q = blk_alloc_queue(set->numa_node); |
26a9750a | 4096 | if (!q) |
b62c21b7 | 4097 | return ERR_PTR(-ENOMEM); |
26a9750a CH |
4098 | q->queuedata = queuedata; |
4099 | ret = blk_mq_init_allocated_queue(set, q); | |
4100 | if (ret) { | |
6f8191fd | 4101 | blk_put_queue(q); |
26a9750a CH |
4102 | return ERR_PTR(ret); |
4103 | } | |
b62c21b7 MS |
4104 | return q; |
4105 | } | |
2f227bb9 CH |
4106 | |
4107 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
4108 | { | |
4109 | return blk_mq_init_queue_data(set, NULL); | |
4110 | } | |
b62c21b7 MS |
4111 | EXPORT_SYMBOL(blk_mq_init_queue); |
4112 | ||
6f8191fd CH |
4113 | /** |
4114 | * blk_mq_destroy_queue - shutdown a request queue | |
4115 | * @q: request queue to shutdown | |
4116 | * | |
81ea42b9 BVA |
4117 | * This shuts down a request queue allocated by blk_mq_init_queue(). All future |
4118 | * requests will be failed with -ENODEV. The caller is responsible for dropping | |
4119 | * the reference from blk_mq_init_queue() by calling blk_put_queue(). | |
6f8191fd CH |
4120 | * |
4121 | * Context: can sleep | |
4122 | */ | |
4123 | void blk_mq_destroy_queue(struct request_queue *q) | |
4124 | { | |
4125 | WARN_ON_ONCE(!queue_is_mq(q)); | |
4126 | WARN_ON_ONCE(blk_queue_registered(q)); | |
4127 | ||
4128 | might_sleep(); | |
4129 | ||
4130 | blk_queue_flag_set(QUEUE_FLAG_DYING, q); | |
4131 | blk_queue_start_drain(q); | |
56c1ee92 | 4132 | blk_mq_freeze_queue_wait(q); |
6f8191fd CH |
4133 | |
4134 | blk_sync_queue(q); | |
4135 | blk_mq_cancel_work_sync(q); | |
4136 | blk_mq_exit_queue(q); | |
6f8191fd CH |
4137 | } |
4138 | EXPORT_SYMBOL(blk_mq_destroy_queue); | |
4139 | ||
4dcc4874 CH |
4140 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, |
4141 | struct lock_class_key *lkclass) | |
9316a9ed JA |
4142 | { |
4143 | struct request_queue *q; | |
b461dfc4 | 4144 | struct gendisk *disk; |
9316a9ed | 4145 | |
b461dfc4 CH |
4146 | q = blk_mq_init_queue_data(set, queuedata); |
4147 | if (IS_ERR(q)) | |
4148 | return ERR_CAST(q); | |
9316a9ed | 4149 | |
4a1fa41d | 4150 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 | 4151 | if (!disk) { |
0a3e5cc7 | 4152 | blk_mq_destroy_queue(q); |
2b3f056f | 4153 | blk_put_queue(q); |
b461dfc4 | 4154 | return ERR_PTR(-ENOMEM); |
9316a9ed | 4155 | } |
6f8191fd | 4156 | set_bit(GD_OWNS_QUEUE, &disk->state); |
b461dfc4 | 4157 | return disk; |
9316a9ed | 4158 | } |
b461dfc4 | 4159 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 4160 | |
6f8191fd CH |
4161 | struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q, |
4162 | struct lock_class_key *lkclass) | |
4163 | { | |
22c17e27 CH |
4164 | struct gendisk *disk; |
4165 | ||
6f8191fd CH |
4166 | if (!blk_get_queue(q)) |
4167 | return NULL; | |
22c17e27 CH |
4168 | disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass); |
4169 | if (!disk) | |
4170 | blk_put_queue(q); | |
4171 | return disk; | |
6f8191fd CH |
4172 | } |
4173 | EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue); | |
4174 | ||
34d11ffa JW |
4175 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
4176 | struct blk_mq_tag_set *set, struct request_queue *q, | |
4177 | int hctx_idx, int node) | |
4178 | { | |
2f8f1336 | 4179 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 4180 | |
2f8f1336 ML |
4181 | /* reuse dead hctx first */ |
4182 | spin_lock(&q->unused_hctx_lock); | |
4183 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
4184 | if (tmp->numa_node == node) { | |
4185 | hctx = tmp; | |
4186 | break; | |
4187 | } | |
4188 | } | |
4189 | if (hctx) | |
4190 | list_del_init(&hctx->hctx_list); | |
4191 | spin_unlock(&q->unused_hctx_lock); | |
4192 | ||
4193 | if (!hctx) | |
4194 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 4195 | if (!hctx) |
7c6c5b7c | 4196 | goto fail; |
34d11ffa | 4197 | |
7c6c5b7c ML |
4198 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
4199 | goto free_hctx; | |
34d11ffa JW |
4200 | |
4201 | return hctx; | |
7c6c5b7c ML |
4202 | |
4203 | free_hctx: | |
4204 | kobject_put(&hctx->kobj); | |
4205 | fail: | |
4206 | return NULL; | |
34d11ffa JW |
4207 | } |
4208 | ||
868f2f0b KB |
4209 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
4210 | struct request_queue *q) | |
320ae51f | 4211 | { |
4e5cc99e ML |
4212 | struct blk_mq_hw_ctx *hctx; |
4213 | unsigned long i, j; | |
ac0d6b92 | 4214 | |
fb350e0a ML |
4215 | /* protect against switching io scheduler */ |
4216 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 4217 | for (i = 0; i < set->nr_hw_queues; i++) { |
306f13ee | 4218 | int old_node; |
4d805131 | 4219 | int node = blk_mq_get_hctx_node(set, i); |
4e5cc99e | 4220 | struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i); |
868f2f0b | 4221 | |
306f13ee ML |
4222 | if (old_hctx) { |
4223 | old_node = old_hctx->numa_node; | |
4224 | blk_mq_exit_hctx(q, set, old_hctx, i); | |
4225 | } | |
868f2f0b | 4226 | |
4e5cc99e | 4227 | if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) { |
306f13ee | 4228 | if (!old_hctx) |
34d11ffa | 4229 | break; |
306f13ee ML |
4230 | pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n", |
4231 | node, old_node); | |
4e5cc99e ML |
4232 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node); |
4233 | WARN_ON_ONCE(!hctx); | |
868f2f0b | 4234 | } |
320ae51f | 4235 | } |
e01ad46d JW |
4236 | /* |
4237 | * Increasing nr_hw_queues fails. Free the newly allocated | |
4238 | * hctxs and keep the previous q->nr_hw_queues. | |
4239 | */ | |
4240 | if (i != set->nr_hw_queues) { | |
4241 | j = q->nr_hw_queues; | |
e01ad46d JW |
4242 | } else { |
4243 | j = i; | |
e01ad46d JW |
4244 | q->nr_hw_queues = set->nr_hw_queues; |
4245 | } | |
34d11ffa | 4246 | |
4e5cc99e ML |
4247 | xa_for_each_start(&q->hctx_table, j, hctx, j) |
4248 | blk_mq_exit_hctx(q, set, hctx, j); | |
fb350e0a | 4249 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
4250 | } |
4251 | ||
42ee3061 ML |
4252 | static void blk_mq_update_poll_flag(struct request_queue *q) |
4253 | { | |
4254 | struct blk_mq_tag_set *set = q->tag_set; | |
4255 | ||
4256 | if (set->nr_maps > HCTX_TYPE_POLL && | |
4257 | set->map[HCTX_TYPE_POLL].nr_queues) | |
4258 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); | |
4259 | else | |
4260 | blk_queue_flag_clear(QUEUE_FLAG_POLL, q); | |
4261 | } | |
4262 | ||
26a9750a CH |
4263 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
4264 | struct request_queue *q) | |
868f2f0b | 4265 | { |
66841672 ML |
4266 | /* mark the queue as mq asap */ |
4267 | q->mq_ops = set->ops; | |
4268 | ||
1db4909e | 4269 | if (blk_mq_alloc_ctxs(q)) |
54bdd67d | 4270 | goto err_exit; |
868f2f0b | 4271 | |
737f98cf ML |
4272 | /* init q->mq_kobj and sw queues' kobjects */ |
4273 | blk_mq_sysfs_init(q); | |
4274 | ||
2f8f1336 ML |
4275 | INIT_LIST_HEAD(&q->unused_hctx_list); |
4276 | spin_lock_init(&q->unused_hctx_lock); | |
4277 | ||
4e5cc99e ML |
4278 | xa_init(&q->hctx_table); |
4279 | ||
868f2f0b KB |
4280 | blk_mq_realloc_hw_ctxs(set, q); |
4281 | if (!q->nr_hw_queues) | |
4282 | goto err_hctxs; | |
320ae51f | 4283 | |
287922eb | 4284 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 4285 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 4286 | |
a8908939 | 4287 | q->tag_set = set; |
320ae51f | 4288 | |
94eddfbe | 4289 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
42ee3061 | 4290 | blk_mq_update_poll_flag(q); |
320ae51f | 4291 | |
2849450a | 4292 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
9a67aa52 | 4293 | INIT_LIST_HEAD(&q->flush_list); |
6fca6a61 CH |
4294 | INIT_LIST_HEAD(&q->requeue_list); |
4295 | spin_lock_init(&q->requeue_lock); | |
4296 | ||
eba71768 JA |
4297 | q->nr_requests = set->queue_depth; |
4298 | ||
24d2f903 | 4299 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 4300 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 4301 | blk_mq_map_swqueue(q); |
26a9750a | 4302 | return 0; |
18741986 | 4303 | |
320ae51f | 4304 | err_hctxs: |
943f45b9 | 4305 | blk_mq_release(q); |
c7de5726 ML |
4306 | err_exit: |
4307 | q->mq_ops = NULL; | |
26a9750a | 4308 | return -ENOMEM; |
320ae51f | 4309 | } |
b62c21b7 | 4310 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 4311 | |
c7e2d94b ML |
4312 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
4313 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 4314 | { |
630ef623 | 4315 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 4316 | |
630ef623 | 4317 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 4318 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
4319 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
4320 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 4321 | } |
320ae51f | 4322 | |
a5164405 JA |
4323 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
4324 | { | |
4325 | int i; | |
4326 | ||
079a2e3e JG |
4327 | if (blk_mq_is_shared_tags(set->flags)) { |
4328 | set->shared_tags = blk_mq_alloc_map_and_rqs(set, | |
e155b0c2 JG |
4329 | BLK_MQ_NO_HCTX_IDX, |
4330 | set->queue_depth); | |
079a2e3e | 4331 | if (!set->shared_tags) |
e155b0c2 JG |
4332 | return -ENOMEM; |
4333 | } | |
4334 | ||
8229cca8 | 4335 | for (i = 0; i < set->nr_hw_queues; i++) { |
63064be1 | 4336 | if (!__blk_mq_alloc_map_and_rqs(set, i)) |
a5164405 | 4337 | goto out_unwind; |
8229cca8 XT |
4338 | cond_resched(); |
4339 | } | |
a5164405 JA |
4340 | |
4341 | return 0; | |
4342 | ||
4343 | out_unwind: | |
4344 | while (--i >= 0) | |
e155b0c2 JG |
4345 | __blk_mq_free_map_and_rqs(set, i); |
4346 | ||
079a2e3e JG |
4347 | if (blk_mq_is_shared_tags(set->flags)) { |
4348 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 | 4349 | BLK_MQ_NO_HCTX_IDX); |
645db34e | 4350 | } |
a5164405 | 4351 | |
a5164405 JA |
4352 | return -ENOMEM; |
4353 | } | |
4354 | ||
4355 | /* | |
4356 | * Allocate the request maps associated with this tag_set. Note that this | |
4357 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
4358 | * will be updated to reflect the allocated depth. | |
4359 | */ | |
63064be1 | 4360 | static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) |
a5164405 JA |
4361 | { |
4362 | unsigned int depth; | |
4363 | int err; | |
4364 | ||
4365 | depth = set->queue_depth; | |
4366 | do { | |
4367 | err = __blk_mq_alloc_rq_maps(set); | |
4368 | if (!err) | |
4369 | break; | |
4370 | ||
4371 | set->queue_depth >>= 1; | |
4372 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
4373 | err = -ENOMEM; | |
4374 | break; | |
4375 | } | |
4376 | } while (set->queue_depth); | |
4377 | ||
4378 | if (!set->queue_depth || err) { | |
4379 | pr_err("blk-mq: failed to allocate request map\n"); | |
4380 | return -ENOMEM; | |
4381 | } | |
4382 | ||
4383 | if (depth != set->queue_depth) | |
4384 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
4385 | depth, set->queue_depth); | |
4386 | ||
4387 | return 0; | |
4388 | } | |
4389 | ||
a4e1d0b7 | 4390 | static void blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
ebe8bddb | 4391 | { |
6e66b493 BVA |
4392 | /* |
4393 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
4394 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
4395 | * number of hardware queues. | |
4396 | */ | |
4397 | if (set->nr_maps == 1) | |
4398 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
4399 | ||
59388702 | 4400 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
4401 | int i; |
4402 | ||
7d4901a9 ML |
4403 | /* |
4404 | * transport .map_queues is usually done in the following | |
4405 | * way: | |
4406 | * | |
4407 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
4408 | * mask = get_cpu_mask(queue) | |
4409 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 4410 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
4411 | * } |
4412 | * | |
4413 | * When we need to remap, the table has to be cleared for | |
4414 | * killing stale mapping since one CPU may not be mapped | |
4415 | * to any hw queue. | |
4416 | */ | |
b3c661b1 JA |
4417 | for (i = 0; i < set->nr_maps; i++) |
4418 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 4419 | |
a4e1d0b7 | 4420 | set->ops->map_queues(set); |
b3c661b1 JA |
4421 | } else { |
4422 | BUG_ON(set->nr_maps > 1); | |
a4e1d0b7 | 4423 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 4424 | } |
ebe8bddb OS |
4425 | } |
4426 | ||
f7e76dbc | 4427 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
ee9d5521 | 4428 | int new_nr_hw_queues) |
f7e76dbc BVA |
4429 | { |
4430 | struct blk_mq_tags **new_tags; | |
e1dd7bc9 | 4431 | int i; |
f7e76dbc | 4432 | |
6be6d112 | 4433 | if (set->nr_hw_queues >= new_nr_hw_queues) |
d4b2e0d4 | 4434 | goto done; |
f7e76dbc BVA |
4435 | |
4436 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
4437 | GFP_KERNEL, set->numa_node); | |
4438 | if (!new_tags) | |
4439 | return -ENOMEM; | |
4440 | ||
4441 | if (set->tags) | |
ee9d5521 | 4442 | memcpy(new_tags, set->tags, set->nr_hw_queues * |
f7e76dbc BVA |
4443 | sizeof(*set->tags)); |
4444 | kfree(set->tags); | |
4445 | set->tags = new_tags; | |
7222657e CZ |
4446 | |
4447 | for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) { | |
4448 | if (!__blk_mq_alloc_map_and_rqs(set, i)) { | |
4449 | while (--i >= set->nr_hw_queues) | |
4450 | __blk_mq_free_map_and_rqs(set, i); | |
4451 | return -ENOMEM; | |
4452 | } | |
4453 | cond_resched(); | |
4454 | } | |
4455 | ||
d4b2e0d4 | 4456 | done: |
f7e76dbc | 4457 | set->nr_hw_queues = new_nr_hw_queues; |
f7e76dbc BVA |
4458 | return 0; |
4459 | } | |
4460 | ||
a4391c64 JA |
4461 | /* |
4462 | * Alloc a tag set to be associated with one or more request queues. | |
4463 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 4464 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
4465 | * value will be stored in set->queue_depth. |
4466 | */ | |
24d2f903 CH |
4467 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
4468 | { | |
b3c661b1 | 4469 | int i, ret; |
da695ba2 | 4470 | |
205fb5f5 BVA |
4471 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
4472 | ||
24d2f903 CH |
4473 | if (!set->nr_hw_queues) |
4474 | return -EINVAL; | |
a4391c64 | 4475 | if (!set->queue_depth) |
24d2f903 CH |
4476 | return -EINVAL; |
4477 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
4478 | return -EINVAL; | |
4479 | ||
7d7e0f90 | 4480 | if (!set->ops->queue_rq) |
24d2f903 CH |
4481 | return -EINVAL; |
4482 | ||
de148297 ML |
4483 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
4484 | return -EINVAL; | |
4485 | ||
a4391c64 JA |
4486 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
4487 | pr_info("blk-mq: reduced tag depth to %u\n", | |
4488 | BLK_MQ_MAX_DEPTH); | |
4489 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
4490 | } | |
24d2f903 | 4491 | |
b3c661b1 JA |
4492 | if (!set->nr_maps) |
4493 | set->nr_maps = 1; | |
4494 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
4495 | return -EINVAL; | |
4496 | ||
6637fadf SL |
4497 | /* |
4498 | * If a crashdump is active, then we are potentially in a very | |
4499 | * memory constrained environment. Limit us to 1 queue and | |
4500 | * 64 tags to prevent using too much memory. | |
4501 | */ | |
4502 | if (is_kdump_kernel()) { | |
4503 | set->nr_hw_queues = 1; | |
59388702 | 4504 | set->nr_maps = 1; |
6637fadf SL |
4505 | set->queue_depth = min(64U, set->queue_depth); |
4506 | } | |
868f2f0b | 4507 | /* |
392546ae JA |
4508 | * There is no use for more h/w queues than cpus if we just have |
4509 | * a single map | |
868f2f0b | 4510 | */ |
392546ae | 4511 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4512 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 4513 | |
80bd4a7a CH |
4514 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4515 | set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL); | |
4516 | if (!set->srcu) | |
4517 | return -ENOMEM; | |
4518 | ret = init_srcu_struct(set->srcu); | |
4519 | if (ret) | |
4520 | goto out_free_srcu; | |
4521 | } | |
24d2f903 | 4522 | |
da695ba2 | 4523 | ret = -ENOMEM; |
5ee20298 CH |
4524 | set->tags = kcalloc_node(set->nr_hw_queues, |
4525 | sizeof(struct blk_mq_tags *), GFP_KERNEL, | |
4526 | set->numa_node); | |
4527 | if (!set->tags) | |
80bd4a7a | 4528 | goto out_cleanup_srcu; |
24d2f903 | 4529 | |
b3c661b1 JA |
4530 | for (i = 0; i < set->nr_maps; i++) { |
4531 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 4532 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
4533 | GFP_KERNEL, set->numa_node); |
4534 | if (!set->map[i].mq_map) | |
4535 | goto out_free_mq_map; | |
59388702 | 4536 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 4537 | } |
bdd17e75 | 4538 | |
a4e1d0b7 | 4539 | blk_mq_update_queue_map(set); |
da695ba2 | 4540 | |
63064be1 | 4541 | ret = blk_mq_alloc_set_map_and_rqs(set); |
da695ba2 | 4542 | if (ret) |
bdd17e75 | 4543 | goto out_free_mq_map; |
24d2f903 | 4544 | |
0d2602ca JA |
4545 | mutex_init(&set->tag_list_lock); |
4546 | INIT_LIST_HEAD(&set->tag_list); | |
4547 | ||
24d2f903 | 4548 | return 0; |
bdd17e75 CH |
4549 | |
4550 | out_free_mq_map: | |
b3c661b1 JA |
4551 | for (i = 0; i < set->nr_maps; i++) { |
4552 | kfree(set->map[i].mq_map); | |
4553 | set->map[i].mq_map = NULL; | |
4554 | } | |
5676e7b6 RE |
4555 | kfree(set->tags); |
4556 | set->tags = NULL; | |
80bd4a7a CH |
4557 | out_cleanup_srcu: |
4558 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4559 | cleanup_srcu_struct(set->srcu); | |
4560 | out_free_srcu: | |
4561 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4562 | kfree(set->srcu); | |
da695ba2 | 4563 | return ret; |
24d2f903 CH |
4564 | } |
4565 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
4566 | ||
cdb14e0f CH |
4567 | /* allocate and initialize a tagset for a simple single-queue device */ |
4568 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
4569 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
4570 | unsigned int set_flags) | |
4571 | { | |
4572 | memset(set, 0, sizeof(*set)); | |
4573 | set->ops = ops; | |
4574 | set->nr_hw_queues = 1; | |
4575 | set->nr_maps = 1; | |
4576 | set->queue_depth = queue_depth; | |
4577 | set->numa_node = NUMA_NO_NODE; | |
4578 | set->flags = set_flags; | |
4579 | return blk_mq_alloc_tag_set(set); | |
4580 | } | |
4581 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
4582 | ||
24d2f903 CH |
4583 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
4584 | { | |
b3c661b1 | 4585 | int i, j; |
24d2f903 | 4586 | |
f7e76dbc | 4587 | for (i = 0; i < set->nr_hw_queues; i++) |
e155b0c2 | 4588 | __blk_mq_free_map_and_rqs(set, i); |
484b4061 | 4589 | |
079a2e3e JG |
4590 | if (blk_mq_is_shared_tags(set->flags)) { |
4591 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 JG |
4592 | BLK_MQ_NO_HCTX_IDX); |
4593 | } | |
32bc15af | 4594 | |
b3c661b1 JA |
4595 | for (j = 0; j < set->nr_maps; j++) { |
4596 | kfree(set->map[j].mq_map); | |
4597 | set->map[j].mq_map = NULL; | |
4598 | } | |
bdd17e75 | 4599 | |
981bd189 | 4600 | kfree(set->tags); |
5676e7b6 | 4601 | set->tags = NULL; |
80bd4a7a CH |
4602 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4603 | cleanup_srcu_struct(set->srcu); | |
4604 | kfree(set->srcu); | |
4605 | } | |
24d2f903 CH |
4606 | } |
4607 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
4608 | ||
e3a2b3f9 JA |
4609 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
4610 | { | |
4611 | struct blk_mq_tag_set *set = q->tag_set; | |
4612 | struct blk_mq_hw_ctx *hctx; | |
4f481208 ML |
4613 | int ret; |
4614 | unsigned long i; | |
e3a2b3f9 | 4615 | |
bd166ef1 | 4616 | if (!set) |
e3a2b3f9 JA |
4617 | return -EINVAL; |
4618 | ||
e5fa8140 AZ |
4619 | if (q->nr_requests == nr) |
4620 | return 0; | |
4621 | ||
70f36b60 | 4622 | blk_mq_freeze_queue(q); |
24f5a90f | 4623 | blk_mq_quiesce_queue(q); |
70f36b60 | 4624 | |
e3a2b3f9 JA |
4625 | ret = 0; |
4626 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
4627 | if (!hctx->tags) |
4628 | continue; | |
bd166ef1 JA |
4629 | /* |
4630 | * If we're using an MQ scheduler, just update the scheduler | |
4631 | * queue depth. This is similar to what the old code would do. | |
4632 | */ | |
f6adcef5 | 4633 | if (hctx->sched_tags) { |
70f36b60 | 4634 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
f6adcef5 | 4635 | nr, true); |
f6adcef5 JG |
4636 | } else { |
4637 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
4638 | false); | |
70f36b60 | 4639 | } |
e3a2b3f9 JA |
4640 | if (ret) |
4641 | break; | |
77f1e0a5 JA |
4642 | if (q->elevator && q->elevator->type->ops.depth_updated) |
4643 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 4644 | } |
d97e594c | 4645 | if (!ret) { |
e3a2b3f9 | 4646 | q->nr_requests = nr; |
079a2e3e | 4647 | if (blk_mq_is_shared_tags(set->flags)) { |
8fa04464 | 4648 | if (q->elevator) |
079a2e3e | 4649 | blk_mq_tag_update_sched_shared_tags(q); |
8fa04464 | 4650 | else |
079a2e3e | 4651 | blk_mq_tag_resize_shared_tags(set, nr); |
8fa04464 | 4652 | } |
d97e594c | 4653 | } |
e3a2b3f9 | 4654 | |
24f5a90f | 4655 | blk_mq_unquiesce_queue(q); |
70f36b60 | 4656 | blk_mq_unfreeze_queue(q); |
70f36b60 | 4657 | |
e3a2b3f9 JA |
4658 | return ret; |
4659 | } | |
4660 | ||
d48ece20 JW |
4661 | /* |
4662 | * request_queue and elevator_type pair. | |
4663 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
4664 | * the elevator_type associated with a request_queue. | |
4665 | */ | |
4666 | struct blk_mq_qe_pair { | |
4667 | struct list_head node; | |
4668 | struct request_queue *q; | |
4669 | struct elevator_type *type; | |
4670 | }; | |
4671 | ||
4672 | /* | |
4673 | * Cache the elevator_type in qe pair list and switch the | |
4674 | * io scheduler to 'none' | |
4675 | */ | |
4676 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
4677 | struct request_queue *q) | |
4678 | { | |
4679 | struct blk_mq_qe_pair *qe; | |
4680 | ||
d48ece20 JW |
4681 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); |
4682 | if (!qe) | |
4683 | return false; | |
4684 | ||
5fd7a84a ML |
4685 | /* q->elevator needs protection from ->sysfs_lock */ |
4686 | mutex_lock(&q->sysfs_lock); | |
4687 | ||
24516565 ML |
4688 | /* the check has to be done with holding sysfs_lock */ |
4689 | if (!q->elevator) { | |
4690 | kfree(qe); | |
4691 | goto unlock; | |
4692 | } | |
4693 | ||
d48ece20 JW |
4694 | INIT_LIST_HEAD(&qe->node); |
4695 | qe->q = q; | |
4696 | qe->type = q->elevator->type; | |
dd6f7f17 CH |
4697 | /* keep a reference to the elevator module as we'll switch back */ |
4698 | __elevator_get(qe->type); | |
d48ece20 | 4699 | list_add(&qe->node, head); |
64b36075 | 4700 | elevator_disable(q); |
24516565 | 4701 | unlock: |
d48ece20 JW |
4702 | mutex_unlock(&q->sysfs_lock); |
4703 | ||
4704 | return true; | |
4705 | } | |
4706 | ||
4a3b666e JK |
4707 | static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head, |
4708 | struct request_queue *q) | |
d48ece20 JW |
4709 | { |
4710 | struct blk_mq_qe_pair *qe; | |
d48ece20 JW |
4711 | |
4712 | list_for_each_entry(qe, head, node) | |
4a3b666e JK |
4713 | if (qe->q == q) |
4714 | return qe; | |
d48ece20 | 4715 | |
4a3b666e JK |
4716 | return NULL; |
4717 | } | |
d48ece20 | 4718 | |
4a3b666e JK |
4719 | static void blk_mq_elv_switch_back(struct list_head *head, |
4720 | struct request_queue *q) | |
4721 | { | |
4722 | struct blk_mq_qe_pair *qe; | |
4723 | struct elevator_type *t; | |
4724 | ||
4725 | qe = blk_lookup_qe_pair(head, q); | |
4726 | if (!qe) | |
4727 | return; | |
4728 | t = qe->type; | |
d48ece20 JW |
4729 | list_del(&qe->node); |
4730 | kfree(qe); | |
4731 | ||
4732 | mutex_lock(&q->sysfs_lock); | |
8237c01f | 4733 | elevator_switch(q, t); |
8ed40ee3 JC |
4734 | /* drop the reference acquired in blk_mq_elv_switch_none */ |
4735 | elevator_put(t); | |
d48ece20 JW |
4736 | mutex_unlock(&q->sysfs_lock); |
4737 | } | |
4738 | ||
e4dc2b32 KB |
4739 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
4740 | int nr_hw_queues) | |
868f2f0b KB |
4741 | { |
4742 | struct request_queue *q; | |
d48ece20 | 4743 | LIST_HEAD(head); |
6be6d112 CZ |
4744 | int prev_nr_hw_queues = set->nr_hw_queues; |
4745 | int i; | |
868f2f0b | 4746 | |
705cda97 BVA |
4747 | lockdep_assert_held(&set->tag_list_lock); |
4748 | ||
392546ae | 4749 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4750 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
4751 | if (nr_hw_queues < 1) |
4752 | return; | |
4753 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
4754 | return; |
4755 | ||
4756 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4757 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
4758 | /* |
4759 | * Switch IO scheduler to 'none', cleaning up the data associated | |
4760 | * with the previous scheduler. We will switch back once we are done | |
4761 | * updating the new sw to hw queue mappings. | |
4762 | */ | |
4763 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4764 | if (!blk_mq_elv_switch_none(&head, q)) | |
4765 | goto switch_back; | |
868f2f0b | 4766 | |
477e19de JW |
4767 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4768 | blk_mq_debugfs_unregister_hctxs(q); | |
eaa870f9 | 4769 | blk_mq_sysfs_unregister_hctxs(q); |
477e19de JW |
4770 | } |
4771 | ||
ee9d5521 | 4772 | if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0) |
f7e76dbc BVA |
4773 | goto reregister; |
4774 | ||
e01ad46d | 4775 | fallback: |
aa880ad6 | 4776 | blk_mq_update_queue_map(set); |
868f2f0b KB |
4777 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4778 | blk_mq_realloc_hw_ctxs(set, q); | |
42ee3061 | 4779 | blk_mq_update_poll_flag(q); |
e01ad46d | 4780 | if (q->nr_hw_queues != set->nr_hw_queues) { |
a846a8e6 YB |
4781 | int i = prev_nr_hw_queues; |
4782 | ||
e01ad46d JW |
4783 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", |
4784 | nr_hw_queues, prev_nr_hw_queues); | |
a846a8e6 YB |
4785 | for (; i < set->nr_hw_queues; i++) |
4786 | __blk_mq_free_map_and_rqs(set, i); | |
4787 | ||
e01ad46d | 4788 | set->nr_hw_queues = prev_nr_hw_queues; |
e01ad46d JW |
4789 | goto fallback; |
4790 | } | |
477e19de JW |
4791 | blk_mq_map_swqueue(q); |
4792 | } | |
4793 | ||
f7e76dbc | 4794 | reregister: |
477e19de | 4795 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
eaa870f9 | 4796 | blk_mq_sysfs_register_hctxs(q); |
477e19de | 4797 | blk_mq_debugfs_register_hctxs(q); |
868f2f0b KB |
4798 | } |
4799 | ||
d48ece20 JW |
4800 | switch_back: |
4801 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4802 | blk_mq_elv_switch_back(&head, q); | |
4803 | ||
868f2f0b KB |
4804 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
4805 | blk_mq_unfreeze_queue(q); | |
6be6d112 CZ |
4806 | |
4807 | /* Free the excess tags when nr_hw_queues shrink. */ | |
4808 | for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++) | |
4809 | __blk_mq_free_map_and_rqs(set, i); | |
868f2f0b | 4810 | } |
e4dc2b32 KB |
4811 | |
4812 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
4813 | { | |
4814 | mutex_lock(&set->tag_list_lock); | |
4815 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
4816 | mutex_unlock(&set->tag_list_lock); | |
4817 | } | |
868f2f0b KB |
4818 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
4819 | ||
f6c80cff KB |
4820 | static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx, |
4821 | struct io_comp_batch *iob, unsigned int flags) | |
bbd7bb70 | 4822 | { |
c6699d6f CH |
4823 | long state = get_current_state(); |
4824 | int ret; | |
bbd7bb70 | 4825 | |
aa61bec3 | 4826 | do { |
5a72e899 | 4827 | ret = q->mq_ops->poll(hctx, iob); |
bbd7bb70 | 4828 | if (ret > 0) { |
849a3700 | 4829 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4830 | return ret; |
bbd7bb70 JA |
4831 | } |
4832 | ||
4833 | if (signal_pending_state(state, current)) | |
849a3700 | 4834 | __set_current_state(TASK_RUNNING); |
b03fbd4f | 4835 | if (task_is_running(current)) |
85f4d4b6 | 4836 | return 1; |
c6699d6f | 4837 | |
ef99b2d3 | 4838 | if (ret < 0 || (flags & BLK_POLL_ONESHOT)) |
bbd7bb70 JA |
4839 | break; |
4840 | cpu_relax(); | |
aa61bec3 | 4841 | } while (!need_resched()); |
bbd7bb70 | 4842 | |
67b4110f | 4843 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4844 | return 0; |
bbd7bb70 | 4845 | } |
1052b8ac | 4846 | |
f6c80cff KB |
4847 | int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, |
4848 | struct io_comp_batch *iob, unsigned int flags) | |
4849 | { | |
4850 | struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie); | |
4851 | ||
4852 | return blk_hctx_poll(q, hctx, iob, flags); | |
4853 | } | |
4854 | ||
4855 | int blk_rq_poll(struct request *rq, struct io_comp_batch *iob, | |
4856 | unsigned int poll_flags) | |
4857 | { | |
4858 | struct request_queue *q = rq->q; | |
4859 | int ret; | |
4860 | ||
4861 | if (!blk_rq_is_poll(rq)) | |
4862 | return 0; | |
4863 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
4864 | return 0; | |
4865 | ||
4866 | ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags); | |
4867 | blk_queue_exit(q); | |
4868 | ||
4869 | return ret; | |
4870 | } | |
4871 | EXPORT_SYMBOL_GPL(blk_rq_poll); | |
4872 | ||
9cf2bab6 JA |
4873 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4874 | { | |
4875 | return rq->mq_ctx->cpu; | |
4876 | } | |
4877 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4878 | ||
2a19b28f ML |
4879 | void blk_mq_cancel_work_sync(struct request_queue *q) |
4880 | { | |
219cf43c JC |
4881 | struct blk_mq_hw_ctx *hctx; |
4882 | unsigned long i; | |
2a19b28f | 4883 | |
219cf43c | 4884 | cancel_delayed_work_sync(&q->requeue_work); |
2a19b28f | 4885 | |
219cf43c JC |
4886 | queue_for_each_hw_ctx(q, hctx, i) |
4887 | cancel_delayed_work_sync(&hctx->run_work); | |
2a19b28f ML |
4888 | } |
4889 | ||
320ae51f JA |
4890 | static int __init blk_mq_init(void) |
4891 | { | |
c3077b5d CH |
4892 | int i; |
4893 | ||
4894 | for_each_possible_cpu(i) | |
f9ab4918 | 4895 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
660e802c CZ |
4896 | for_each_possible_cpu(i) |
4897 | INIT_CSD(&per_cpu(blk_cpu_csd, i), | |
4898 | __blk_mq_complete_request_remote, NULL); | |
c3077b5d CH |
4899 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4900 | ||
4901 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4902 | "block/softirq:dead", NULL, | |
4903 | blk_softirq_cpu_dead); | |
9467f859 TG |
4904 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4905 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4906 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4907 | blk_mq_hctx_notify_online, | |
4908 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4909 | return 0; |
4910 | } | |
4911 | subsys_initcall(blk_mq_init); |