]>
Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
75bb4625 JA |
2 | /* |
3 | * Block multiqueue core code | |
4 | * | |
5 | * Copyright (C) 2013-2014 Jens Axboe | |
6 | * Copyright (C) 2013-2014 Christoph Hellwig | |
7 | */ | |
320ae51f JA |
8 | #include <linux/kernel.h> |
9 | #include <linux/module.h> | |
10 | #include <linux/backing-dev.h> | |
11 | #include <linux/bio.h> | |
12 | #include <linux/blkdev.h> | |
fe45e630 | 13 | #include <linux/blk-integrity.h> |
f75782e4 | 14 | #include <linux/kmemleak.h> |
320ae51f JA |
15 | #include <linux/mm.h> |
16 | #include <linux/init.h> | |
17 | #include <linux/slab.h> | |
18 | #include <linux/workqueue.h> | |
19 | #include <linux/smp.h> | |
e41d12f5 | 20 | #include <linux/interrupt.h> |
320ae51f | 21 | #include <linux/llist.h> |
320ae51f JA |
22 | #include <linux/cpu.h> |
23 | #include <linux/cache.h> | |
105ab3d8 | 24 | #include <linux/sched/topology.h> |
174cd4b1 | 25 | #include <linux/sched/signal.h> |
320ae51f | 26 | #include <linux/delay.h> |
aedcd72f | 27 | #include <linux/crash_dump.h> |
88c7b2b7 | 28 | #include <linux/prefetch.h> |
a892c8d5 | 29 | #include <linux/blk-crypto.h> |
82d981d4 | 30 | #include <linux/part_stat.h> |
a46c2702 | 31 | #include <linux/sched/isolation.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" |
320ae51f | 43 | |
f9ab4918 | 44 | static DEFINE_PER_CPU(struct llist_head, blk_cpu_done); |
660e802c | 45 | static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd); |
c3077b5d | 46 | |
710fa378 | 47 | static void blk_mq_insert_request(struct request *rq, blk_insert_t flags); |
360f2648 CH |
48 | static void blk_mq_request_bypass_insert(struct request *rq, |
49 | blk_insert_t flags); | |
94aa228c CH |
50 | static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
51 | struct list_head *list); | |
f6c80cff KB |
52 | static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx, |
53 | struct io_comp_batch *iob, unsigned int flags); | |
3e08773c | 54 | |
320ae51f | 55 | /* |
85fae294 YY |
56 | * Check if any of the ctx, dispatch list or elevator |
57 | * have pending work in this hardware queue. | |
320ae51f | 58 | */ |
79f720a7 | 59 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 60 | { |
79f720a7 JA |
61 | return !list_empty_careful(&hctx->dispatch) || |
62 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 63 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
64 | } |
65 | ||
320ae51f JA |
66 | /* |
67 | * Mark this ctx as having pending work in this hardware queue | |
68 | */ | |
69 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
70 | struct blk_mq_ctx *ctx) | |
71 | { | |
f31967f0 JA |
72 | const int bit = ctx->index_hw[hctx->type]; |
73 | ||
74 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
75 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
76 | } |
77 | ||
78 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
79 | struct blk_mq_ctx *ctx) | |
80 | { | |
f31967f0 JA |
81 | const int bit = ctx->index_hw[hctx->type]; |
82 | ||
83 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
84 | } |
85 | ||
f299b7c7 | 86 | struct mq_inflight { |
8446fe92 | 87 | struct block_device *part; |
a2e80f6f | 88 | unsigned int inflight[2]; |
f299b7c7 JA |
89 | }; |
90 | ||
2dd6532e | 91 | static bool blk_mq_check_inflight(struct request *rq, void *priv) |
f299b7c7 JA |
92 | { |
93 | struct mq_inflight *mi = priv; | |
94 | ||
b81c14ca HW |
95 | if (rq->part && blk_do_io_stat(rq) && |
96 | (!mi->part->bd_partno || rq->part == mi->part) && | |
b0d97557 | 97 | blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) |
bb4e6b14 | 98 | mi->inflight[rq_data_dir(rq)]++; |
7baa8572 JA |
99 | |
100 | return true; | |
f299b7c7 JA |
101 | } |
102 | ||
8446fe92 CH |
103 | unsigned int blk_mq_in_flight(struct request_queue *q, |
104 | struct block_device *part) | |
f299b7c7 | 105 | { |
a2e80f6f | 106 | struct mq_inflight mi = { .part = part }; |
f299b7c7 | 107 | |
f299b7c7 | 108 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 | 109 | |
a2e80f6f | 110 | return mi.inflight[0] + mi.inflight[1]; |
bf0ddaba OS |
111 | } |
112 | ||
8446fe92 CH |
113 | void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, |
114 | unsigned int inflight[2]) | |
bf0ddaba | 115 | { |
a2e80f6f | 116 | struct mq_inflight mi = { .part = part }; |
bf0ddaba | 117 | |
bb4e6b14 | 118 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
a2e80f6f PB |
119 | inflight[0] = mi.inflight[0]; |
120 | inflight[1] = mi.inflight[1]; | |
bf0ddaba OS |
121 | } |
122 | ||
1671d522 | 123 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 124 | { |
7996a8b5 BL |
125 | mutex_lock(&q->mq_freeze_lock); |
126 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 127 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 128 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 129 | if (queue_is_mq(q)) |
055f6e18 | 130 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
131 | } else { |
132 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 133 | } |
f3af020b | 134 | } |
1671d522 | 135 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 136 | |
6bae363e | 137 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 138 | { |
3ef28e83 | 139 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 140 | } |
6bae363e | 141 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 142 | |
f91328c4 KB |
143 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
144 | unsigned long timeout) | |
145 | { | |
146 | return wait_event_timeout(q->mq_freeze_wq, | |
147 | percpu_ref_is_zero(&q->q_usage_counter), | |
148 | timeout); | |
149 | } | |
150 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 151 | |
f3af020b TH |
152 | /* |
153 | * Guarantee no request is in use, so we can change any data structure of | |
154 | * the queue afterward. | |
155 | */ | |
3ef28e83 | 156 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 157 | { |
3ef28e83 DW |
158 | /* |
159 | * In the !blk_mq case we are only calling this to kill the | |
160 | * q_usage_counter, otherwise this increases the freeze depth | |
161 | * and waits for it to return to zero. For this reason there is | |
162 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
163 | * exported to drivers as the only user for unfreeze is blk_mq. | |
164 | */ | |
1671d522 | 165 | blk_freeze_queue_start(q); |
f3af020b TH |
166 | blk_mq_freeze_queue_wait(q); |
167 | } | |
3ef28e83 DW |
168 | |
169 | void blk_mq_freeze_queue(struct request_queue *q) | |
170 | { | |
171 | /* | |
172 | * ...just an alias to keep freeze and unfreeze actions balanced | |
173 | * in the blk_mq_* namespace | |
174 | */ | |
175 | blk_freeze_queue(q); | |
176 | } | |
c761d96b | 177 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 178 | |
aec89dc5 | 179 | void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic) |
320ae51f | 180 | { |
7996a8b5 | 181 | mutex_lock(&q->mq_freeze_lock); |
aec89dc5 CH |
182 | if (force_atomic) |
183 | q->q_usage_counter.data->force_atomic = true; | |
7996a8b5 BL |
184 | q->mq_freeze_depth--; |
185 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
186 | if (!q->mq_freeze_depth) { | |
bdd63160 | 187 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 188 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 189 | } |
7996a8b5 | 190 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 191 | } |
aec89dc5 CH |
192 | |
193 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
194 | { | |
195 | __blk_mq_unfreeze_queue(q, false); | |
196 | } | |
b4c6a028 | 197 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 198 | |
852ec809 BVA |
199 | /* |
200 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
201 | * mpt3sas driver such that this function can be removed. | |
202 | */ | |
203 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
204 | { | |
e70feb8b ML |
205 | unsigned long flags; |
206 | ||
207 | spin_lock_irqsave(&q->queue_lock, flags); | |
208 | if (!q->quiesce_depth++) | |
209 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
210 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
852ec809 BVA |
211 | } |
212 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
213 | ||
6a83e74d | 214 | /** |
9ef4d020 | 215 | * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done |
483239c7 | 216 | * @set: tag_set to wait on |
6a83e74d | 217 | * |
9ef4d020 | 218 | * Note: it is driver's responsibility for making sure that quiesce has |
483239c7 CH |
219 | * been started on or more of the request_queues of the tag_set. This |
220 | * function only waits for the quiesce on those request_queues that had | |
221 | * the quiesce flag set using blk_mq_quiesce_queue_nowait. | |
6a83e74d | 222 | */ |
483239c7 | 223 | void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set) |
6a83e74d | 224 | { |
483239c7 CH |
225 | if (set->flags & BLK_MQ_F_BLOCKING) |
226 | synchronize_srcu(set->srcu); | |
704b914f | 227 | else |
6a83e74d BVA |
228 | synchronize_rcu(); |
229 | } | |
9ef4d020 ML |
230 | EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); |
231 | ||
232 | /** | |
233 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished | |
234 | * @q: request queue. | |
235 | * | |
236 | * Note: this function does not prevent that the struct request end_io() | |
237 | * callback function is invoked. Once this function is returned, we make | |
238 | * sure no dispatch can happen until the queue is unquiesced via | |
239 | * blk_mq_unquiesce_queue(). | |
240 | */ | |
241 | void blk_mq_quiesce_queue(struct request_queue *q) | |
242 | { | |
243 | blk_mq_quiesce_queue_nowait(q); | |
8537380b CH |
244 | /* nothing to wait for non-mq queues */ |
245 | if (queue_is_mq(q)) | |
483239c7 | 246 | blk_mq_wait_quiesce_done(q->tag_set); |
9ef4d020 | 247 | } |
6a83e74d BVA |
248 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); |
249 | ||
e4e73913 ML |
250 | /* |
251 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
252 | * @q: request queue. | |
253 | * | |
254 | * This function recovers queue into the state before quiescing | |
255 | * which is done by blk_mq_quiesce_queue. | |
256 | */ | |
257 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
258 | { | |
e70feb8b ML |
259 | unsigned long flags; |
260 | bool run_queue = false; | |
261 | ||
262 | spin_lock_irqsave(&q->queue_lock, flags); | |
263 | if (WARN_ON_ONCE(q->quiesce_depth <= 0)) { | |
264 | ; | |
265 | } else if (!--q->quiesce_depth) { | |
266 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); | |
267 | run_queue = true; | |
268 | } | |
269 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
f4560ffe | 270 | |
1d9e9bc6 | 271 | /* dispatch requests which are inserted during quiescing */ |
e70feb8b ML |
272 | if (run_queue) |
273 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
274 | } |
275 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
276 | ||
414dd48e CL |
277 | void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set) |
278 | { | |
279 | struct request_queue *q; | |
280 | ||
281 | mutex_lock(&set->tag_list_lock); | |
282 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
283 | if (!blk_queue_skip_tagset_quiesce(q)) | |
284 | blk_mq_quiesce_queue_nowait(q); | |
285 | } | |
286 | blk_mq_wait_quiesce_done(set); | |
287 | mutex_unlock(&set->tag_list_lock); | |
288 | } | |
289 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset); | |
290 | ||
291 | void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set) | |
292 | { | |
293 | struct request_queue *q; | |
294 | ||
295 | mutex_lock(&set->tag_list_lock); | |
296 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
297 | if (!blk_queue_skip_tagset_quiesce(q)) | |
298 | blk_mq_unquiesce_queue(q); | |
299 | } | |
300 | mutex_unlock(&set->tag_list_lock); | |
301 | } | |
302 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset); | |
303 | ||
aed3ea94 JA |
304 | void blk_mq_wake_waiters(struct request_queue *q) |
305 | { | |
306 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 307 | unsigned long i; |
aed3ea94 JA |
308 | |
309 | queue_for_each_hw_ctx(q, hctx, i) | |
310 | if (blk_mq_hw_queue_mapped(hctx)) | |
311 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
312 | } | |
313 | ||
52fdbbcc CH |
314 | void blk_rq_init(struct request_queue *q, struct request *rq) |
315 | { | |
316 | memset(rq, 0, sizeof(*rq)); | |
317 | ||
318 | INIT_LIST_HEAD(&rq->queuelist); | |
319 | rq->q = q; | |
320 | rq->__sector = (sector_t) -1; | |
321 | INIT_HLIST_NODE(&rq->hash); | |
322 | RB_CLEAR_NODE(&rq->rb_node); | |
323 | rq->tag = BLK_MQ_NO_TAG; | |
324 | rq->internal_tag = BLK_MQ_NO_TAG; | |
08420cf7 | 325 | rq->start_time_ns = blk_time_get_ns(); |
52fdbbcc CH |
326 | rq->part = NULL; |
327 | blk_crypto_rq_set_defaults(rq); | |
328 | } | |
329 | EXPORT_SYMBOL(blk_rq_init); | |
330 | ||
5c17f45e CZ |
331 | /* Set start and alloc time when the allocated request is actually used */ |
332 | static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns) | |
333 | { | |
334 | if (blk_mq_need_time_stamp(rq)) | |
08420cf7 | 335 | rq->start_time_ns = blk_time_get_ns(); |
5c17f45e CZ |
336 | else |
337 | rq->start_time_ns = 0; | |
338 | ||
339 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME | |
340 | if (blk_queue_rq_alloc_time(rq->q)) | |
341 | rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns; | |
342 | else | |
343 | rq->alloc_time_ns = 0; | |
344 | #endif | |
345 | } | |
346 | ||
e4cdf1a1 | 347 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
5c17f45e | 348 | struct blk_mq_tags *tags, unsigned int tag) |
320ae51f | 349 | { |
605f784e PB |
350 | struct blk_mq_ctx *ctx = data->ctx; |
351 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
352 | struct request_queue *q = data->q; | |
e4cdf1a1 | 353 | struct request *rq = tags->static_rqs[tag]; |
c3a148d2 | 354 | |
c7b84d42 JA |
355 | rq->q = q; |
356 | rq->mq_ctx = ctx; | |
357 | rq->mq_hctx = hctx; | |
358 | rq->cmd_flags = data->cmd_flags; | |
359 | ||
360 | if (data->flags & BLK_MQ_REQ_PM) | |
361 | data->rq_flags |= RQF_PM; | |
362 | if (blk_queue_io_stat(q)) | |
363 | data->rq_flags |= RQF_IO_STAT; | |
364 | rq->rq_flags = data->rq_flags; | |
365 | ||
dd6216bb | 366 | if (data->rq_flags & RQF_SCHED_TAGS) { |
56f8da64 JA |
367 | rq->tag = BLK_MQ_NO_TAG; |
368 | rq->internal_tag = tag; | |
dd6216bb CH |
369 | } else { |
370 | rq->tag = tag; | |
371 | rq->internal_tag = BLK_MQ_NO_TAG; | |
e4cdf1a1 | 372 | } |
c7b84d42 | 373 | rq->timeout = 0; |
e4cdf1a1 | 374 | |
af76e555 | 375 | rq->part = NULL; |
544ccc8d | 376 | rq->io_start_time_ns = 0; |
3d244306 | 377 | rq->stats_sectors = 0; |
af76e555 CH |
378 | rq->nr_phys_segments = 0; |
379 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
380 | rq->nr_integrity_segments = 0; | |
381 | #endif | |
af76e555 CH |
382 | rq->end_io = NULL; |
383 | rq->end_io_data = NULL; | |
af76e555 | 384 | |
4f266f2b PB |
385 | blk_crypto_rq_set_defaults(rq); |
386 | INIT_LIST_HEAD(&rq->queuelist); | |
387 | /* tag was already set */ | |
388 | WRITE_ONCE(rq->deadline, 0); | |
0a467d0f | 389 | req_ref_set(rq, 1); |
7ea4d8a4 | 390 | |
dd6216bb | 391 | if (rq->rq_flags & RQF_USE_SCHED) { |
7ea4d8a4 CH |
392 | struct elevator_queue *e = data->q->elevator; |
393 | ||
4f266f2b PB |
394 | INIT_HLIST_NODE(&rq->hash); |
395 | RB_CLEAR_NODE(&rq->rb_node); | |
396 | ||
dd6216bb | 397 | if (e->type->ops.prepare_request) |
7ea4d8a4 | 398 | e->type->ops.prepare_request(rq); |
7ea4d8a4 CH |
399 | } |
400 | ||
e4cdf1a1 | 401 | return rq; |
5dee8577 CH |
402 | } |
403 | ||
349302da | 404 | static inline struct request * |
5c17f45e | 405 | __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data) |
349302da JA |
406 | { |
407 | unsigned int tag, tag_offset; | |
fe6134f6 | 408 | struct blk_mq_tags *tags; |
349302da | 409 | struct request *rq; |
fe6134f6 | 410 | unsigned long tag_mask; |
349302da JA |
411 | int i, nr = 0; |
412 | ||
fe6134f6 JA |
413 | tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset); |
414 | if (unlikely(!tag_mask)) | |
349302da JA |
415 | return NULL; |
416 | ||
fe6134f6 JA |
417 | tags = blk_mq_tags_from_data(data); |
418 | for (i = 0; tag_mask; i++) { | |
419 | if (!(tag_mask & (1UL << i))) | |
349302da JA |
420 | continue; |
421 | tag = tag_offset + i; | |
a22c00be | 422 | prefetch(tags->static_rqs[tag]); |
fe6134f6 | 423 | tag_mask &= ~(1UL << i); |
5c17f45e | 424 | rq = blk_mq_rq_ctx_init(data, tags, tag); |
013a7f95 | 425 | rq_list_add(data->cached_rq, rq); |
c5fc7b93 | 426 | nr++; |
349302da | 427 | } |
b8643d68 CZ |
428 | if (!(data->rq_flags & RQF_SCHED_TAGS)) |
429 | blk_mq_add_active_requests(data->hctx, nr); | |
c5fc7b93 JA |
430 | /* caller already holds a reference, add for remainder */ |
431 | percpu_ref_get_many(&data->q->q_usage_counter, nr - 1); | |
349302da JA |
432 | data->nr_tags -= nr; |
433 | ||
013a7f95 | 434 | return rq_list_pop(data->cached_rq); |
349302da JA |
435 | } |
436 | ||
b90cfaed | 437 | static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) |
d2c0d383 | 438 | { |
e6e7abff | 439 | struct request_queue *q = data->q; |
6f816b4b | 440 | u64 alloc_time_ns = 0; |
47c122e3 | 441 | struct request *rq; |
600c3b0c | 442 | unsigned int tag; |
d2c0d383 | 443 | |
6f816b4b TH |
444 | /* alloc_time includes depth and tag waits */ |
445 | if (blk_queue_rq_alloc_time(q)) | |
08420cf7 | 446 | alloc_time_ns = blk_time_get_ns(); |
6f816b4b | 447 | |
f9afca4d | 448 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 449 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 | 450 | |
781dd830 | 451 | if (q->elevator) { |
dd6216bb CH |
452 | /* |
453 | * All requests use scheduler tags when an I/O scheduler is | |
454 | * enabled for the queue. | |
455 | */ | |
456 | data->rq_flags |= RQF_SCHED_TAGS; | |
781dd830 | 457 | |
d2c0d383 | 458 | /* |
8d663f34 | 459 | * Flush/passthrough requests are special and go directly to the |
dd6216bb | 460 | * dispatch list. |
d2c0d383 | 461 | */ |
be4c4278 | 462 | if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH && |
dd6216bb CH |
463 | !blk_op_is_passthrough(data->cmd_flags)) { |
464 | struct elevator_mq_ops *ops = &q->elevator->type->ops; | |
465 | ||
466 | WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED); | |
467 | ||
468 | data->rq_flags |= RQF_USE_SCHED; | |
469 | if (ops->limit_depth) | |
470 | ops->limit_depth(data->cmd_flags, data); | |
471 | } | |
d2c0d383 CH |
472 | } |
473 | ||
bf0beec0 | 474 | retry: |
600c3b0c CH |
475 | data->ctx = blk_mq_get_ctx(q); |
476 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
dd6216bb | 477 | if (!(data->rq_flags & RQF_SCHED_TAGS)) |
600c3b0c CH |
478 | blk_mq_tag_busy(data->hctx); |
479 | ||
99e48cd6 JG |
480 | if (data->flags & BLK_MQ_REQ_RESERVED) |
481 | data->rq_flags |= RQF_RESV; | |
482 | ||
349302da JA |
483 | /* |
484 | * Try batched alloc if we want more than 1 tag. | |
485 | */ | |
486 | if (data->nr_tags > 1) { | |
5c17f45e CZ |
487 | rq = __blk_mq_alloc_requests_batch(data); |
488 | if (rq) { | |
489 | blk_mq_rq_time_init(rq, alloc_time_ns); | |
349302da | 490 | return rq; |
5c17f45e | 491 | } |
349302da JA |
492 | data->nr_tags = 1; |
493 | } | |
494 | ||
bf0beec0 ML |
495 | /* |
496 | * Waiting allocations only fail because of an inactive hctx. In that | |
497 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
498 | * should have migrated us to an online CPU by now. | |
499 | */ | |
e4cdf1a1 | 500 | tag = blk_mq_get_tag(data); |
bf0beec0 ML |
501 | if (tag == BLK_MQ_NO_TAG) { |
502 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
503 | return NULL; | |
bf0beec0 | 504 | /* |
349302da JA |
505 | * Give up the CPU and sleep for a random short time to |
506 | * ensure that thread using a realtime scheduling class | |
507 | * are migrated off the CPU, and thus off the hctx that | |
508 | * is going away. | |
bf0beec0 ML |
509 | */ |
510 | msleep(3); | |
511 | goto retry; | |
512 | } | |
47c122e3 | 513 | |
b8643d68 CZ |
514 | if (!(data->rq_flags & RQF_SCHED_TAGS)) |
515 | blk_mq_inc_active_requests(data->hctx); | |
5c17f45e CZ |
516 | rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag); |
517 | blk_mq_rq_time_init(rq, alloc_time_ns); | |
518 | return rq; | |
d2c0d383 CH |
519 | } |
520 | ||
4b6a5d9c JA |
521 | static struct request *blk_mq_rq_cache_fill(struct request_queue *q, |
522 | struct blk_plug *plug, | |
523 | blk_opf_t opf, | |
524 | blk_mq_req_flags_t flags) | |
320ae51f | 525 | { |
e6e7abff CH |
526 | struct blk_mq_alloc_data data = { |
527 | .q = q, | |
528 | .flags = flags, | |
16458cf3 | 529 | .cmd_flags = opf, |
4b6a5d9c JA |
530 | .nr_tags = plug->nr_ios, |
531 | .cached_rq = &plug->cached_rq, | |
e6e7abff | 532 | }; |
bd166ef1 | 533 | struct request *rq; |
320ae51f | 534 | |
4b6a5d9c JA |
535 | if (blk_queue_enter(q, flags)) |
536 | return NULL; | |
537 | ||
538 | plug->nr_ios = 1; | |
320ae51f | 539 | |
b90cfaed | 540 | rq = __blk_mq_alloc_requests(&data); |
4b6a5d9c JA |
541 | if (unlikely(!rq)) |
542 | blk_queue_exit(q); | |
543 | return rq; | |
544 | } | |
545 | ||
546 | static struct request *blk_mq_alloc_cached_request(struct request_queue *q, | |
547 | blk_opf_t opf, | |
548 | blk_mq_req_flags_t flags) | |
549 | { | |
550 | struct blk_plug *plug = current->plug; | |
551 | struct request *rq; | |
552 | ||
553 | if (!plug) | |
554 | return NULL; | |
40467282 | 555 | |
4b6a5d9c JA |
556 | if (rq_list_empty(plug->cached_rq)) { |
557 | if (plug->nr_ios == 1) | |
558 | return NULL; | |
559 | rq = blk_mq_rq_cache_fill(q, plug, opf, flags); | |
40467282 JC |
560 | if (!rq) |
561 | return NULL; | |
562 | } else { | |
563 | rq = rq_list_peek(&plug->cached_rq); | |
564 | if (!rq || rq->q != q) | |
565 | return NULL; | |
4b6a5d9c | 566 | |
40467282 JC |
567 | if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type) |
568 | return NULL; | |
569 | if (op_is_flush(rq->cmd_flags) != op_is_flush(opf)) | |
570 | return NULL; | |
571 | ||
572 | plug->cached_rq = rq_list_next(rq); | |
5c17f45e | 573 | blk_mq_rq_time_init(rq, 0); |
40467282 | 574 | } |
4b6a5d9c | 575 | |
4b6a5d9c JA |
576 | rq->cmd_flags = opf; |
577 | INIT_LIST_HEAD(&rq->queuelist); | |
578 | return rq; | |
579 | } | |
580 | ||
581 | struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf, | |
582 | blk_mq_req_flags_t flags) | |
583 | { | |
584 | struct request *rq; | |
585 | ||
586 | rq = blk_mq_alloc_cached_request(q, opf, flags); | |
587 | if (!rq) { | |
588 | struct blk_mq_alloc_data data = { | |
589 | .q = q, | |
590 | .flags = flags, | |
591 | .cmd_flags = opf, | |
592 | .nr_tags = 1, | |
593 | }; | |
594 | int ret; | |
595 | ||
596 | ret = blk_queue_enter(q, flags); | |
597 | if (ret) | |
598 | return ERR_PTR(ret); | |
599 | ||
600 | rq = __blk_mq_alloc_requests(&data); | |
601 | if (!rq) | |
602 | goto out_queue_exit; | |
603 | } | |
0c4de0f3 CH |
604 | rq->__data_len = 0; |
605 | rq->__sector = (sector_t) -1; | |
606 | rq->bio = rq->biotail = NULL; | |
320ae51f | 607 | return rq; |
a5ea5811 CH |
608 | out_queue_exit: |
609 | blk_queue_exit(q); | |
610 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 611 | } |
4bb659b1 | 612 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 613 | |
cd6ce148 | 614 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
16458cf3 | 615 | blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 616 | { |
e6e7abff CH |
617 | struct blk_mq_alloc_data data = { |
618 | .q = q, | |
619 | .flags = flags, | |
16458cf3 | 620 | .cmd_flags = opf, |
47c122e3 | 621 | .nr_tags = 1, |
e6e7abff | 622 | }; |
600c3b0c | 623 | u64 alloc_time_ns = 0; |
e3c5a78c | 624 | struct request *rq; |
6d2809d5 | 625 | unsigned int cpu; |
600c3b0c | 626 | unsigned int tag; |
1f5bd336 ML |
627 | int ret; |
628 | ||
600c3b0c CH |
629 | /* alloc_time includes depth and tag waits */ |
630 | if (blk_queue_rq_alloc_time(q)) | |
08420cf7 | 631 | alloc_time_ns = blk_time_get_ns(); |
600c3b0c | 632 | |
1f5bd336 ML |
633 | /* |
634 | * If the tag allocator sleeps we could get an allocation for a | |
635 | * different hardware context. No need to complicate the low level | |
636 | * allocator for this for the rare use case of a command tied to | |
637 | * a specific queue. | |
638 | */ | |
6ee858a3 KS |
639 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) || |
640 | WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED))) | |
1f5bd336 ML |
641 | return ERR_PTR(-EINVAL); |
642 | ||
643 | if (hctx_idx >= q->nr_hw_queues) | |
644 | return ERR_PTR(-EIO); | |
645 | ||
3a0a5299 | 646 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
647 | if (ret) |
648 | return ERR_PTR(ret); | |
649 | ||
c8712c6a CH |
650 | /* |
651 | * Check if the hardware context is actually mapped to anything. | |
652 | * If not tell the caller that it should skip this queue. | |
653 | */ | |
a5ea5811 | 654 | ret = -EXDEV; |
4e5cc99e | 655 | data.hctx = xa_load(&q->hctx_table, hctx_idx); |
e6e7abff | 656 | if (!blk_mq_hw_queue_mapped(data.hctx)) |
a5ea5811 | 657 | goto out_queue_exit; |
e6e7abff | 658 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
14dc7a18 BVA |
659 | if (cpu >= nr_cpu_ids) |
660 | goto out_queue_exit; | |
e6e7abff | 661 | data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 662 | |
dd6216bb CH |
663 | if (q->elevator) |
664 | data.rq_flags |= RQF_SCHED_TAGS; | |
781dd830 | 665 | else |
dd6216bb | 666 | blk_mq_tag_busy(data.hctx); |
600c3b0c | 667 | |
99e48cd6 JG |
668 | if (flags & BLK_MQ_REQ_RESERVED) |
669 | data.rq_flags |= RQF_RESV; | |
670 | ||
a5ea5811 | 671 | ret = -EWOULDBLOCK; |
600c3b0c CH |
672 | tag = blk_mq_get_tag(&data); |
673 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 674 | goto out_queue_exit; |
b8643d68 CZ |
675 | if (!(data.rq_flags & RQF_SCHED_TAGS)) |
676 | blk_mq_inc_active_requests(data.hctx); | |
5c17f45e CZ |
677 | rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag); |
678 | blk_mq_rq_time_init(rq, alloc_time_ns); | |
e3c5a78c JG |
679 | rq->__data_len = 0; |
680 | rq->__sector = (sector_t) -1; | |
681 | rq->bio = rq->biotail = NULL; | |
682 | return rq; | |
600c3b0c | 683 | |
a5ea5811 CH |
684 | out_queue_exit: |
685 | blk_queue_exit(q); | |
686 | return ERR_PTR(ret); | |
1f5bd336 ML |
687 | } |
688 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
689 | ||
e5c0ca13 CZ |
690 | static void blk_mq_finish_request(struct request *rq) |
691 | { | |
692 | struct request_queue *q = rq->q; | |
693 | ||
694 | if (rq->rq_flags & RQF_USE_SCHED) { | |
695 | q->elevator->type->ops.finish_request(rq); | |
696 | /* | |
697 | * For postflush request that may need to be | |
698 | * completed twice, we should clear this flag | |
699 | * to avoid double finish_request() on the rq. | |
700 | */ | |
701 | rq->rq_flags &= ~RQF_USE_SCHED; | |
702 | } | |
703 | } | |
704 | ||
12f5b931 KB |
705 | static void __blk_mq_free_request(struct request *rq) |
706 | { | |
707 | struct request_queue *q = rq->q; | |
708 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 709 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
710 | const int sched_tag = rq->internal_tag; |
711 | ||
a892c8d5 | 712 | blk_crypto_free_request(rq); |
986d413b | 713 | blk_pm_mark_last_busy(rq); |
ea4f995e | 714 | rq->mq_hctx = NULL; |
ddad5933 | 715 | |
b8643d68 CZ |
716 | if (rq->tag != BLK_MQ_NO_TAG) { |
717 | blk_mq_dec_active_requests(hctx); | |
cae740a0 | 718 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
b8643d68 | 719 | } |
76647368 | 720 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 721 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
722 | blk_mq_sched_restart(hctx); |
723 | blk_queue_exit(q); | |
724 | } | |
725 | ||
6af54051 | 726 | void blk_mq_free_request(struct request *rq) |
320ae51f | 727 | { |
320ae51f | 728 | struct request_queue *q = rq->q; |
6af54051 | 729 | |
e5c0ca13 | 730 | blk_mq_finish_request(rq); |
320ae51f | 731 | |
7beb2f84 | 732 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 733 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 734 | |
a7905043 | 735 | rq_qos_done(q, rq); |
0d2602ca | 736 | |
12f5b931 | 737 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 738 | if (req_ref_put_and_test(rq)) |
12f5b931 | 739 | __blk_mq_free_request(rq); |
320ae51f | 740 | } |
1a3b595a | 741 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 742 | |
47c122e3 | 743 | void blk_mq_free_plug_rqs(struct blk_plug *plug) |
320ae51f | 744 | { |
013a7f95 | 745 | struct request *rq; |
fe1f4526 | 746 | |
c5fc7b93 | 747 | while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) |
47c122e3 | 748 | blk_mq_free_request(rq); |
47c122e3 | 749 | } |
522a7775 | 750 | |
22350ad7 CH |
751 | void blk_dump_rq_flags(struct request *rq, char *msg) |
752 | { | |
753 | printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, | |
f3fa33ac | 754 | rq->q->disk ? rq->q->disk->disk_name : "?", |
16458cf3 | 755 | (__force unsigned long long) rq->cmd_flags); |
22350ad7 CH |
756 | |
757 | printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", | |
758 | (unsigned long long)blk_rq_pos(rq), | |
759 | blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); | |
760 | printk(KERN_INFO " bio %p, biotail %p, len %u\n", | |
761 | rq->bio, rq->biotail, blk_rq_bytes(rq)); | |
762 | } | |
763 | EXPORT_SYMBOL(blk_dump_rq_flags); | |
764 | ||
9be3e06f JA |
765 | static void blk_account_io_completion(struct request *req, unsigned int bytes) |
766 | { | |
767 | if (req->part && blk_do_io_stat(req)) { | |
768 | const int sgrp = op_stat_group(req_op(req)); | |
769 | ||
770 | part_stat_lock(); | |
771 | part_stat_add(req->part, sectors[sgrp], bytes >> 9); | |
772 | part_stat_unlock(); | |
773 | } | |
774 | } | |
775 | ||
0d7a29a2 CH |
776 | static void blk_print_req_error(struct request *req, blk_status_t status) |
777 | { | |
778 | printk_ratelimited(KERN_ERR | |
779 | "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x " | |
780 | "phys_seg %u prio class %u\n", | |
781 | blk_status_to_str(status), | |
f3fa33ac | 782 | req->q->disk ? req->q->disk->disk_name : "?", |
16458cf3 BVA |
783 | blk_rq_pos(req), (__force u32)req_op(req), |
784 | blk_op_str(req_op(req)), | |
785 | (__force u32)(req->cmd_flags & ~REQ_OP_MASK), | |
0d7a29a2 CH |
786 | req->nr_phys_segments, |
787 | IOPRIO_PRIO_CLASS(req->ioprio)); | |
788 | } | |
789 | ||
5581a5dd JA |
790 | /* |
791 | * Fully end IO on a request. Does not support partial completions, or | |
792 | * errors. | |
793 | */ | |
794 | static void blk_complete_request(struct request *req) | |
795 | { | |
796 | const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0; | |
797 | int total_bytes = blk_rq_bytes(req); | |
798 | struct bio *bio = req->bio; | |
799 | ||
800 | trace_block_rq_complete(req, BLK_STS_OK, total_bytes); | |
801 | ||
802 | if (!bio) | |
803 | return; | |
804 | ||
805 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
806 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ) | |
807 | req->q->integrity.profile->complete_fn(req, total_bytes); | |
808 | #endif | |
809 | ||
9cd1e566 EB |
810 | /* |
811 | * Upper layers may call blk_crypto_evict_key() anytime after the last | |
812 | * bio_endio(). Therefore, the keyslot must be released before that. | |
813 | */ | |
814 | blk_crypto_rq_put_keyslot(req); | |
815 | ||
5581a5dd JA |
816 | blk_account_io_completion(req, total_bytes); |
817 | ||
818 | do { | |
819 | struct bio *next = bio->bi_next; | |
820 | ||
821 | /* Completion has already been traced */ | |
822 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
a12821d5 | 823 | |
a0508c36 | 824 | blk_zone_update_request_bio(req, bio); |
a12821d5 | 825 | |
5581a5dd JA |
826 | if (!is_flush) |
827 | bio_endio(bio); | |
828 | bio = next; | |
829 | } while (bio); | |
830 | ||
dd291d77 DLM |
831 | blk_zone_complete_request(req); |
832 | ||
5581a5dd JA |
833 | /* |
834 | * Reset counters so that the request stacking driver | |
835 | * can find how many bytes remain in the request | |
836 | * later. | |
837 | */ | |
ab3e1d3b JA |
838 | if (!req->end_io) { |
839 | req->bio = NULL; | |
840 | req->__data_len = 0; | |
841 | } | |
5581a5dd JA |
842 | } |
843 | ||
9be3e06f JA |
844 | /** |
845 | * blk_update_request - Complete multiple bytes without completing the request | |
846 | * @req: the request being processed | |
847 | * @error: block status code | |
848 | * @nr_bytes: number of bytes to complete for @req | |
849 | * | |
850 | * Description: | |
851 | * Ends I/O on a number of bytes attached to @req, but doesn't complete | |
852 | * the request structure even if @req doesn't have leftover. | |
853 | * If @req has leftover, sets it up for the next range of segments. | |
854 | * | |
855 | * Passing the result of blk_rq_bytes() as @nr_bytes guarantees | |
856 | * %false return from this function. | |
857 | * | |
858 | * Note: | |
859 | * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function | |
860 | * except in the consistency check at the end of this function. | |
861 | * | |
862 | * Return: | |
863 | * %false - this request doesn't have any more data | |
864 | * %true - this request has more data | |
865 | **/ | |
866 | bool blk_update_request(struct request *req, blk_status_t error, | |
867 | unsigned int nr_bytes) | |
868 | { | |
c0da26f9 DLM |
869 | bool is_flush = req->rq_flags & RQF_FLUSH_SEQ; |
870 | bool quiet = req->rq_flags & RQF_QUIET; | |
9be3e06f JA |
871 | int total_bytes; |
872 | ||
8a7d267b | 873 | trace_block_rq_complete(req, error, nr_bytes); |
9be3e06f JA |
874 | |
875 | if (!req->bio) | |
876 | return false; | |
877 | ||
878 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
879 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && | |
880 | error == BLK_STS_OK) | |
881 | req->q->integrity.profile->complete_fn(req, nr_bytes); | |
882 | #endif | |
883 | ||
9cd1e566 EB |
884 | /* |
885 | * Upper layers may call blk_crypto_evict_key() anytime after the last | |
886 | * bio_endio(). Therefore, the keyslot must be released before that. | |
887 | */ | |
888 | if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req)) | |
889 | __blk_crypto_rq_put_keyslot(req); | |
890 | ||
c0da26f9 DLM |
891 | if (unlikely(error && !blk_rq_is_passthrough(req) && !quiet) && |
892 | !test_bit(GD_DEAD, &req->q->disk->state)) { | |
9be3e06f | 893 | blk_print_req_error(req, error); |
d5869fdc YS |
894 | trace_block_rq_error(req, error, nr_bytes); |
895 | } | |
9be3e06f JA |
896 | |
897 | blk_account_io_completion(req, nr_bytes); | |
898 | ||
899 | total_bytes = 0; | |
900 | while (req->bio) { | |
901 | struct bio *bio = req->bio; | |
902 | unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); | |
903 | ||
c0da26f9 DLM |
904 | if (unlikely(error)) |
905 | bio->bi_status = error; | |
906 | ||
907 | if (bio_bytes == bio->bi_iter.bi_size) { | |
9be3e06f | 908 | req->bio = bio->bi_next; |
9b1ce7f0 | 909 | } else if (bio_is_zone_append(bio) && error == BLK_STS_OK) { |
c0da26f9 DLM |
910 | /* |
911 | * Partial zone append completions cannot be supported | |
912 | * as the BIO fragments may end up not being written | |
913 | * sequentially. | |
914 | */ | |
915 | bio->bi_status = BLK_STS_IOERR; | |
916 | } | |
9be3e06f JA |
917 | |
918 | /* Completion has already been traced */ | |
919 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
c0da26f9 DLM |
920 | if (unlikely(quiet)) |
921 | bio_set_flag(bio, BIO_QUIET); | |
922 | ||
923 | bio_advance(bio, bio_bytes); | |
924 | ||
925 | /* Don't actually finish bio if it's part of flush sequence */ | |
a0508c36 DLM |
926 | if (!bio->bi_iter.bi_size) { |
927 | blk_zone_update_request_bio(req, bio); | |
928 | if (!is_flush) | |
929 | bio_endio(bio); | |
c0da26f9 | 930 | } |
9be3e06f JA |
931 | |
932 | total_bytes += bio_bytes; | |
933 | nr_bytes -= bio_bytes; | |
934 | ||
935 | if (!nr_bytes) | |
936 | break; | |
937 | } | |
938 | ||
939 | /* | |
940 | * completely done | |
941 | */ | |
942 | if (!req->bio) { | |
dd291d77 | 943 | blk_zone_complete_request(req); |
9be3e06f JA |
944 | /* |
945 | * Reset counters so that the request stacking driver | |
946 | * can find how many bytes remain in the request | |
947 | * later. | |
948 | */ | |
949 | req->__data_len = 0; | |
950 | return false; | |
951 | } | |
952 | ||
953 | req->__data_len -= total_bytes; | |
954 | ||
955 | /* update sector only for requests with clear definition of sector */ | |
956 | if (!blk_rq_is_passthrough(req)) | |
957 | req->__sector += total_bytes >> 9; | |
958 | ||
959 | /* mixed attributes always follow the first bio */ | |
960 | if (req->rq_flags & RQF_MIXED_MERGE) { | |
961 | req->cmd_flags &= ~REQ_FAILFAST_MASK; | |
962 | req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; | |
963 | } | |
964 | ||
965 | if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { | |
966 | /* | |
967 | * If total number of sectors is less than the first segment | |
968 | * size, something has gone terribly wrong. | |
969 | */ | |
970 | if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { | |
971 | blk_dump_rq_flags(req, "request botched"); | |
972 | req->__data_len = blk_rq_cur_bytes(req); | |
973 | } | |
974 | ||
975 | /* recalculate the number of segments */ | |
976 | req->nr_phys_segments = blk_recalc_rq_segments(req); | |
977 | } | |
978 | ||
979 | return true; | |
980 | } | |
981 | EXPORT_SYMBOL_GPL(blk_update_request); | |
982 | ||
450b7879 CH |
983 | static inline void blk_account_io_done(struct request *req, u64 now) |
984 | { | |
5a80bd07 HC |
985 | trace_block_io_done(req); |
986 | ||
450b7879 CH |
987 | /* |
988 | * Account IO completion. flush_rq isn't accounted as a | |
989 | * normal IO on queueing nor completion. Accounting the | |
990 | * containing request is enough. | |
991 | */ | |
992 | if (blk_do_io_stat(req) && req->part && | |
06965037 CK |
993 | !(req->rq_flags & RQF_FLUSH_SEQ)) { |
994 | const int sgrp = op_stat_group(req_op(req)); | |
450b7879 | 995 | |
06965037 CK |
996 | part_stat_lock(); |
997 | update_io_ticks(req->part, jiffies, true); | |
998 | part_stat_inc(req->part, ios[sgrp]); | |
999 | part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns); | |
1000 | part_stat_unlock(); | |
1001 | } | |
450b7879 CH |
1002 | } |
1003 | ||
1004 | static inline void blk_account_io_start(struct request *req) | |
1005 | { | |
5a80bd07 HC |
1006 | trace_block_io_start(req); |
1007 | ||
e165fb4d CK |
1008 | if (blk_do_io_stat(req)) { |
1009 | /* | |
1010 | * All non-passthrough requests are created from a bio with one | |
1011 | * exception: when a flush command that is part of a flush sequence | |
1012 | * generated by the state machine in blk-flush.c is cloned onto the | |
1013 | * lower device by dm-multipath we can get here without a bio. | |
1014 | */ | |
1015 | if (req->bio) | |
1016 | req->part = req->bio->bi_bdev; | |
1017 | else | |
1018 | req->part = req->q->disk->part0; | |
1019 | ||
1020 | part_stat_lock(); | |
1021 | update_io_ticks(req->part, jiffies, false); | |
1022 | part_stat_unlock(); | |
1023 | } | |
450b7879 CH |
1024 | } |
1025 | ||
f794f335 | 1026 | static inline void __blk_mq_end_request_acct(struct request *rq, u64 now) |
320ae51f | 1027 | { |
54bdd67d | 1028 | if (rq->rq_flags & RQF_STATS) |
522a7775 | 1029 | blk_stat_add(rq, now); |
4bc6339a | 1030 | |
87890092 | 1031 | blk_mq_sched_completed_request(rq, now); |
522a7775 | 1032 | blk_account_io_done(rq, now); |
f794f335 | 1033 | } |
522a7775 | 1034 | |
f794f335 JA |
1035 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
1036 | { | |
1037 | if (blk_mq_need_time_stamp(rq)) | |
08420cf7 | 1038 | __blk_mq_end_request_acct(rq, blk_time_get_ns()); |
0d11e6ac | 1039 | |
e5c0ca13 CZ |
1040 | blk_mq_finish_request(rq); |
1041 | ||
91b63639 | 1042 | if (rq->end_io) { |
a7905043 | 1043 | rq_qos_done(rq->q, rq); |
de671d61 JA |
1044 | if (rq->end_io(rq, error) == RQ_END_IO_FREE) |
1045 | blk_mq_free_request(rq); | |
91b63639 | 1046 | } else { |
320ae51f | 1047 | blk_mq_free_request(rq); |
91b63639 | 1048 | } |
320ae51f | 1049 | } |
c8a446ad | 1050 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 1051 | |
2a842aca | 1052 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
1053 | { |
1054 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
1055 | BUG(); | |
c8a446ad | 1056 | __blk_mq_end_request(rq, error); |
63151a44 | 1057 | } |
c8a446ad | 1058 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 1059 | |
f794f335 JA |
1060 | #define TAG_COMP_BATCH 32 |
1061 | ||
1062 | static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx, | |
1063 | int *tag_array, int nr_tags) | |
1064 | { | |
1065 | struct request_queue *q = hctx->queue; | |
1066 | ||
b8643d68 | 1067 | blk_mq_sub_active_requests(hctx, nr_tags); |
3b87c6ea | 1068 | |
f794f335 JA |
1069 | blk_mq_put_tags(hctx->tags, tag_array, nr_tags); |
1070 | percpu_ref_put_many(&q->q_usage_counter, nr_tags); | |
1071 | } | |
1072 | ||
1073 | void blk_mq_end_request_batch(struct io_comp_batch *iob) | |
1074 | { | |
1075 | int tags[TAG_COMP_BATCH], nr_tags = 0; | |
02f7eab0 | 1076 | struct blk_mq_hw_ctx *cur_hctx = NULL; |
f794f335 JA |
1077 | struct request *rq; |
1078 | u64 now = 0; | |
1079 | ||
1080 | if (iob->need_ts) | |
08420cf7 | 1081 | now = blk_time_get_ns(); |
f794f335 JA |
1082 | |
1083 | while ((rq = rq_list_pop(&iob->req_list)) != NULL) { | |
1084 | prefetch(rq->bio); | |
1085 | prefetch(rq->rq_next); | |
1086 | ||
5581a5dd | 1087 | blk_complete_request(rq); |
f794f335 JA |
1088 | if (iob->need_ts) |
1089 | __blk_mq_end_request_acct(rq, now); | |
1090 | ||
e5c0ca13 CZ |
1091 | blk_mq_finish_request(rq); |
1092 | ||
98b26a0e JA |
1093 | rq_qos_done(rq->q, rq); |
1094 | ||
ab3e1d3b JA |
1095 | /* |
1096 | * If end_io handler returns NONE, then it still has | |
1097 | * ownership of the request. | |
1098 | */ | |
1099 | if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE) | |
1100 | continue; | |
1101 | ||
f794f335 | 1102 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 1103 | if (!req_ref_put_and_test(rq)) |
f794f335 JA |
1104 | continue; |
1105 | ||
1106 | blk_crypto_free_request(rq); | |
1107 | blk_pm_mark_last_busy(rq); | |
f794f335 | 1108 | |
02f7eab0 JA |
1109 | if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) { |
1110 | if (cur_hctx) | |
1111 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); | |
f794f335 | 1112 | nr_tags = 0; |
02f7eab0 | 1113 | cur_hctx = rq->mq_hctx; |
f794f335 JA |
1114 | } |
1115 | tags[nr_tags++] = rq->tag; | |
f794f335 JA |
1116 | } |
1117 | ||
1118 | if (nr_tags) | |
02f7eab0 | 1119 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); |
f794f335 JA |
1120 | } |
1121 | EXPORT_SYMBOL_GPL(blk_mq_end_request_batch); | |
1122 | ||
f9ab4918 | 1123 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 1124 | { |
f9ab4918 SAS |
1125 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
1126 | struct request *rq, *next; | |
c3077b5d | 1127 | |
f9ab4918 | 1128 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 1129 | rq->q->mq_ops->complete(rq); |
320ae51f | 1130 | } |
320ae51f | 1131 | |
f9ab4918 | 1132 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 1133 | { |
f9ab4918 | 1134 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
1135 | } |
1136 | ||
c3077b5d CH |
1137 | static int blk_softirq_cpu_dead(unsigned int cpu) |
1138 | { | |
f9ab4918 | 1139 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
1140 | return 0; |
1141 | } | |
1142 | ||
40d09b53 | 1143 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 1144 | { |
f9ab4918 | 1145 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
1146 | } |
1147 | ||
96339526 CH |
1148 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
1149 | { | |
1150 | int cpu = raw_smp_processor_id(); | |
1151 | ||
1152 | if (!IS_ENABLED(CONFIG_SMP) || | |
1153 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
1154 | return false; | |
71425189 SAS |
1155 | /* |
1156 | * With force threaded interrupts enabled, raising softirq from an SMP | |
1157 | * function call will always result in waking the ksoftirqd thread. | |
1158 | * This is probably worse than completing the request on a different | |
1159 | * cache domain. | |
1160 | */ | |
91cc470e | 1161 | if (force_irqthreads()) |
71425189 | 1162 | return false; |
96339526 | 1163 | |
af550e4c | 1164 | /* same CPU or cache domain and capacity? Complete locally */ |
96339526 CH |
1165 | if (cpu == rq->mq_ctx->cpu || |
1166 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
af550e4c QY |
1167 | cpus_share_cache(cpu, rq->mq_ctx->cpu) && |
1168 | cpus_equal_capacity(cpu, rq->mq_ctx->cpu))) | |
96339526 CH |
1169 | return false; |
1170 | ||
1171 | /* don't try to IPI to an offline CPU */ | |
1172 | return cpu_online(rq->mq_ctx->cpu); | |
1173 | } | |
1174 | ||
f9ab4918 SAS |
1175 | static void blk_mq_complete_send_ipi(struct request *rq) |
1176 | { | |
f9ab4918 SAS |
1177 | unsigned int cpu; |
1178 | ||
1179 | cpu = rq->mq_ctx->cpu; | |
660e802c CZ |
1180 | if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu))) |
1181 | smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu)); | |
f9ab4918 SAS |
1182 | } |
1183 | ||
1184 | static void blk_mq_raise_softirq(struct request *rq) | |
1185 | { | |
1186 | struct llist_head *list; | |
1187 | ||
1188 | preempt_disable(); | |
1189 | list = this_cpu_ptr(&blk_cpu_done); | |
1190 | if (llist_add(&rq->ipi_list, list)) | |
1191 | raise_softirq(BLOCK_SOFTIRQ); | |
1192 | preempt_enable(); | |
1193 | } | |
1194 | ||
40d09b53 | 1195 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 1196 | { |
af78ff7c | 1197 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 1198 | |
4ab32bf3 | 1199 | /* |
f168420c LS |
1200 | * For request which hctx has only one ctx mapping, |
1201 | * or a polled request, always complete locally, | |
1202 | * it's pointless to redirect the completion. | |
4ab32bf3 | 1203 | */ |
30654614 ET |
1204 | if ((rq->mq_hctx->nr_ctx == 1 && |
1205 | rq->mq_ctx->cpu == raw_smp_processor_id()) || | |
1206 | rq->cmd_flags & REQ_POLLED) | |
40d09b53 | 1207 | return false; |
38535201 | 1208 | |
96339526 | 1209 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
1210 | blk_mq_complete_send_ipi(rq); |
1211 | return true; | |
3d6efbf6 | 1212 | } |
40d09b53 | 1213 | |
f9ab4918 SAS |
1214 | if (rq->q->nr_hw_queues == 1) { |
1215 | blk_mq_raise_softirq(rq); | |
1216 | return true; | |
1217 | } | |
1218 | return false; | |
40d09b53 CH |
1219 | } |
1220 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
1221 | ||
1222 | /** | |
1223 | * blk_mq_complete_request - end I/O on a request | |
1224 | * @rq: the request being processed | |
1225 | * | |
1226 | * Description: | |
1227 | * Complete a request by scheduling the ->complete_rq operation. | |
1228 | **/ | |
1229 | void blk_mq_complete_request(struct request *rq) | |
1230 | { | |
1231 | if (!blk_mq_complete_request_remote(rq)) | |
1232 | rq->q->mq_ops->complete(rq); | |
320ae51f | 1233 | } |
15f73f5b | 1234 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 1235 | |
105663f7 AA |
1236 | /** |
1237 | * blk_mq_start_request - Start processing a request | |
1238 | * @rq: Pointer to request to be started | |
1239 | * | |
1240 | * Function used by device drivers to notify the block layer that a request | |
1241 | * is going to be processed now, so blk layer can do proper initializations | |
1242 | * such as starting the timeout timer. | |
1243 | */ | |
e2490073 | 1244 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
1245 | { |
1246 | struct request_queue *q = rq->q; | |
1247 | ||
a54895fa | 1248 | trace_block_rq_issue(rq); |
320ae51f | 1249 | |
847c5bcd KK |
1250 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) && |
1251 | !blk_rq_is_passthrough(rq)) { | |
08420cf7 | 1252 | rq->io_start_time_ns = blk_time_get_ns(); |
3d244306 | 1253 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 1254 | rq->rq_flags |= RQF_STATS; |
a7905043 | 1255 | rq_qos_issue(q, rq); |
cf43e6be JA |
1256 | } |
1257 | ||
1d9bd516 | 1258 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 1259 | |
1d9bd516 | 1260 | blk_add_timer(rq); |
12f5b931 | 1261 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
217b613a | 1262 | rq->mq_hctx->tags->rqs[rq->tag] = rq; |
49f5baa5 | 1263 | |
54d4e6ab MG |
1264 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
1265 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
1266 | q->integrity.profile->prepare_fn(rq); | |
1267 | #endif | |
3e08773c | 1268 | if (rq->bio && rq->bio->bi_opf & REQ_POLLED) |
f6c80cff | 1269 | WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num); |
320ae51f | 1270 | } |
e2490073 | 1271 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 1272 | |
a327c341 ML |
1273 | /* |
1274 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple | |
1275 | * queues. This is important for md arrays to benefit from merging | |
1276 | * requests. | |
1277 | */ | |
1278 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
1279 | { | |
1280 | if (plug->multiple_queues) | |
1281 | return BLK_MAX_REQUEST_COUNT * 2; | |
1282 | return BLK_MAX_REQUEST_COUNT; | |
1283 | } | |
1284 | ||
1285 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) | |
1286 | { | |
1287 | struct request *last = rq_list_peek(&plug->mq_list); | |
1288 | ||
1289 | if (!plug->rq_count) { | |
1290 | trace_block_plug(rq->q); | |
1291 | } else if (plug->rq_count >= blk_plug_max_rq_count(plug) || | |
1292 | (!blk_queue_nomerges(rq->q) && | |
1293 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1294 | blk_mq_flush_plug_list(plug, false); | |
878eb6e4 | 1295 | last = NULL; |
a327c341 ML |
1296 | trace_block_plug(rq->q); |
1297 | } | |
1298 | ||
1299 | if (!plug->multiple_queues && last && last->q != rq->q) | |
1300 | plug->multiple_queues = true; | |
c6b7a3a2 ML |
1301 | /* |
1302 | * Any request allocated from sched tags can't be issued to | |
1303 | * ->queue_rqs() directly | |
1304 | */ | |
1305 | if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS)) | |
a327c341 ML |
1306 | plug->has_elevator = true; |
1307 | rq->rq_next = NULL; | |
1308 | rq_list_add(&plug->mq_list, rq); | |
1309 | plug->rq_count++; | |
1310 | } | |
1311 | ||
4054cff9 CH |
1312 | /** |
1313 | * blk_execute_rq_nowait - insert a request to I/O scheduler for execution | |
4054cff9 CH |
1314 | * @rq: request to insert |
1315 | * @at_head: insert request at head or tail of queue | |
4054cff9 CH |
1316 | * |
1317 | * Description: | |
1318 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1319 | * for execution. Don't wait for completion. | |
1320 | * | |
1321 | * Note: | |
1322 | * This function will invoke @done directly if the queue is dead. | |
1323 | */ | |
e2e53086 | 1324 | void blk_execute_rq_nowait(struct request *rq, bool at_head) |
4054cff9 | 1325 | { |
f0dbe6e8 CH |
1326 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
1327 | ||
ae948fd6 CH |
1328 | WARN_ON(irqs_disabled()); |
1329 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 | 1330 | |
ae948fd6 | 1331 | blk_account_io_start(rq); |
110fdb44 | 1332 | |
f0dbe6e8 | 1333 | if (current->plug && !at_head) { |
ae948fd6 | 1334 | blk_add_rq_to_plug(current->plug, rq); |
f0dbe6e8 CH |
1335 | return; |
1336 | } | |
1337 | ||
710fa378 | 1338 | blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); |
65a558f6 | 1339 | blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING); |
4054cff9 CH |
1340 | } |
1341 | EXPORT_SYMBOL_GPL(blk_execute_rq_nowait); | |
1342 | ||
32ac5a9b CH |
1343 | struct blk_rq_wait { |
1344 | struct completion done; | |
1345 | blk_status_t ret; | |
1346 | }; | |
1347 | ||
de671d61 | 1348 | static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret) |
32ac5a9b CH |
1349 | { |
1350 | struct blk_rq_wait *wait = rq->end_io_data; | |
1351 | ||
1352 | wait->ret = ret; | |
1353 | complete(&wait->done); | |
de671d61 | 1354 | return RQ_END_IO_NONE; |
32ac5a9b CH |
1355 | } |
1356 | ||
c6e99ea4 | 1357 | bool blk_rq_is_poll(struct request *rq) |
4054cff9 CH |
1358 | { |
1359 | if (!rq->mq_hctx) | |
1360 | return false; | |
1361 | if (rq->mq_hctx->type != HCTX_TYPE_POLL) | |
1362 | return false; | |
4054cff9 CH |
1363 | return true; |
1364 | } | |
c6e99ea4 | 1365 | EXPORT_SYMBOL_GPL(blk_rq_is_poll); |
4054cff9 CH |
1366 | |
1367 | static void blk_rq_poll_completion(struct request *rq, struct completion *wait) | |
1368 | { | |
1369 | do { | |
f6c80cff | 1370 | blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0); |
4054cff9 CH |
1371 | cond_resched(); |
1372 | } while (!completion_done(wait)); | |
1373 | } | |
1374 | ||
1375 | /** | |
1376 | * blk_execute_rq - insert a request into queue for execution | |
4054cff9 CH |
1377 | * @rq: request to insert |
1378 | * @at_head: insert request at head or tail of queue | |
1379 | * | |
1380 | * Description: | |
1381 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1382 | * for execution and wait for completion. | |
1383 | * Return: The blk_status_t result provided to blk_mq_end_request(). | |
1384 | */ | |
b84ba30b | 1385 | blk_status_t blk_execute_rq(struct request *rq, bool at_head) |
4054cff9 | 1386 | { |
f0dbe6e8 | 1387 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
32ac5a9b CH |
1388 | struct blk_rq_wait wait = { |
1389 | .done = COMPLETION_INITIALIZER_ONSTACK(wait.done), | |
1390 | }; | |
4054cff9 | 1391 | |
ae948fd6 CH |
1392 | WARN_ON(irqs_disabled()); |
1393 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 CH |
1394 | |
1395 | rq->end_io_data = &wait; | |
ae948fd6 | 1396 | rq->end_io = blk_end_sync_rq; |
4054cff9 | 1397 | |
ae948fd6 | 1398 | blk_account_io_start(rq); |
710fa378 | 1399 | blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); |
f0dbe6e8 | 1400 | blk_mq_run_hw_queue(hctx, false); |
4054cff9 | 1401 | |
0eb4db47 | 1402 | if (blk_rq_is_poll(rq)) |
32ac5a9b | 1403 | blk_rq_poll_completion(rq, &wait.done); |
0eb4db47 KB |
1404 | else |
1405 | blk_wait_io(&wait.done); | |
4054cff9 | 1406 | |
32ac5a9b | 1407 | return wait.ret; |
4054cff9 CH |
1408 | } |
1409 | EXPORT_SYMBOL(blk_execute_rq); | |
1410 | ||
ed0791b2 | 1411 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
1412 | { |
1413 | struct request_queue *q = rq->q; | |
1414 | ||
923218f6 ML |
1415 | blk_mq_put_driver_tag(rq); |
1416 | ||
a54895fa | 1417 | trace_block_rq_requeue(rq); |
a7905043 | 1418 | rq_qos_requeue(q, rq); |
49f5baa5 | 1419 | |
12f5b931 KB |
1420 | if (blk_mq_request_started(rq)) { |
1421 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 1422 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 1423 | } |
320ae51f JA |
1424 | } |
1425 | ||
2b053aca | 1426 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 1427 | { |
214a4418 | 1428 | struct request_queue *q = rq->q; |
9a67aa52 | 1429 | unsigned long flags; |
214a4418 | 1430 | |
ed0791b2 | 1431 | __blk_mq_requeue_request(rq); |
ed0791b2 | 1432 | |
105976f5 ML |
1433 | /* this request will be re-inserted to io scheduler queue */ |
1434 | blk_mq_sched_requeue_request(rq); | |
1435 | ||
9a67aa52 CH |
1436 | spin_lock_irqsave(&q->requeue_lock, flags); |
1437 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
1438 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
214a4418 CH |
1439 | |
1440 | if (kick_requeue_list) | |
1441 | blk_mq_kick_requeue_list(q); | |
ed0791b2 CH |
1442 | } |
1443 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
1444 | ||
6fca6a61 CH |
1445 | static void blk_mq_requeue_work(struct work_struct *work) |
1446 | { | |
1447 | struct request_queue *q = | |
2849450a | 1448 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 | 1449 | LIST_HEAD(rq_list); |
9a67aa52 CH |
1450 | LIST_HEAD(flush_list); |
1451 | struct request *rq; | |
6fca6a61 | 1452 | |
18e9781d | 1453 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 1454 | list_splice_init(&q->requeue_list, &rq_list); |
9a67aa52 | 1455 | list_splice_init(&q->flush_list, &flush_list); |
18e9781d | 1456 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 | 1457 | |
9a67aa52 CH |
1458 | while (!list_empty(&rq_list)) { |
1459 | rq = list_entry(rq_list.next, struct request, queuelist); | |
aef1897c | 1460 | /* |
a1e948b8 CH |
1461 | * If RQF_DONTPREP ist set, the request has been started by the |
1462 | * driver already and might have driver-specific data allocated | |
1463 | * already. Insert it into the hctx dispatch list to avoid | |
1464 | * block layer merges for the request. | |
aef1897c | 1465 | */ |
a1e948b8 | 1466 | if (rq->rq_flags & RQF_DONTPREP) { |
a1e948b8 | 1467 | list_del_init(&rq->queuelist); |
2b597613 | 1468 | blk_mq_request_bypass_insert(rq, 0); |
9a67aa52 | 1469 | } else { |
a1e948b8 | 1470 | list_del_init(&rq->queuelist); |
710fa378 | 1471 | blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD); |
a1e948b8 | 1472 | } |
6fca6a61 CH |
1473 | } |
1474 | ||
9a67aa52 CH |
1475 | while (!list_empty(&flush_list)) { |
1476 | rq = list_entry(flush_list.next, struct request, queuelist); | |
6fca6a61 | 1477 | list_del_init(&rq->queuelist); |
710fa378 | 1478 | blk_mq_insert_request(rq, 0); |
6fca6a61 CH |
1479 | } |
1480 | ||
52d7f1b5 | 1481 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
1482 | } |
1483 | ||
6fca6a61 CH |
1484 | void blk_mq_kick_requeue_list(struct request_queue *q) |
1485 | { | |
ae943d20 | 1486 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
1487 | } |
1488 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
1489 | ||
2849450a MS |
1490 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
1491 | unsigned long msecs) | |
1492 | { | |
d4acf365 BVA |
1493 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
1494 | msecs_to_jiffies(msecs)); | |
2849450a MS |
1495 | } |
1496 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
1497 | ||
0e4237ae ML |
1498 | static bool blk_is_flush_data_rq(struct request *rq) |
1499 | { | |
1500 | return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq); | |
1501 | } | |
1502 | ||
2dd6532e | 1503 | static bool blk_mq_rq_inflight(struct request *rq, void *priv) |
ae879912 JA |
1504 | { |
1505 | /* | |
8ab30a33 JG |
1506 | * If we find a request that isn't idle we know the queue is busy |
1507 | * as it's checked in the iter. | |
1508 | * Return false to stop the iteration. | |
0e4237ae ML |
1509 | * |
1510 | * In case of queue quiesce, if one flush data request is completed, | |
1511 | * don't count it as inflight given the flush sequence is suspended, | |
1512 | * and the original flush data request is invisible to driver, just | |
1513 | * like other pending requests because of quiesce | |
ae879912 | 1514 | */ |
0e4237ae ML |
1515 | if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) && |
1516 | blk_is_flush_data_rq(rq) && | |
1517 | blk_mq_request_completed(rq))) { | |
ae879912 JA |
1518 | bool *busy = priv; |
1519 | ||
1520 | *busy = true; | |
1521 | return false; | |
1522 | } | |
1523 | ||
1524 | return true; | |
1525 | } | |
1526 | ||
3c94d83c | 1527 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
1528 | { |
1529 | bool busy = false; | |
1530 | ||
3c94d83c | 1531 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
1532 | return busy; |
1533 | } | |
3c94d83c | 1534 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 1535 | |
9bdb4833 | 1536 | static void blk_mq_rq_timed_out(struct request *req) |
320ae51f | 1537 | { |
da661267 | 1538 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
1539 | if (req->q->mq_ops->timeout) { |
1540 | enum blk_eh_timer_return ret; | |
1541 | ||
9bdb4833 | 1542 | ret = req->q->mq_ops->timeout(req); |
d1210d5a CH |
1543 | if (ret == BLK_EH_DONE) |
1544 | return; | |
1545 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 1546 | } |
d1210d5a CH |
1547 | |
1548 | blk_add_timer(req); | |
87ee7b11 | 1549 | } |
5b3f25fc | 1550 | |
82c22947 DJ |
1551 | struct blk_expired_data { |
1552 | bool has_timedout_rq; | |
1553 | unsigned long next; | |
1554 | unsigned long timeout_start; | |
1555 | }; | |
1556 | ||
1557 | static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired) | |
81481eb4 | 1558 | { |
12f5b931 | 1559 | unsigned long deadline; |
87ee7b11 | 1560 | |
12f5b931 KB |
1561 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
1562 | return false; | |
da661267 CH |
1563 | if (rq->rq_flags & RQF_TIMED_OUT) |
1564 | return false; | |
a7af0af3 | 1565 | |
079076b3 | 1566 | deadline = READ_ONCE(rq->deadline); |
82c22947 | 1567 | if (time_after_eq(expired->timeout_start, deadline)) |
12f5b931 | 1568 | return true; |
a7af0af3 | 1569 | |
82c22947 DJ |
1570 | if (expired->next == 0) |
1571 | expired->next = deadline; | |
1572 | else if (time_after(expired->next, deadline)) | |
1573 | expired->next = deadline; | |
12f5b931 | 1574 | return false; |
87ee7b11 JA |
1575 | } |
1576 | ||
2e315dc0 ML |
1577 | void blk_mq_put_rq_ref(struct request *rq) |
1578 | { | |
de671d61 JA |
1579 | if (is_flush_rq(rq)) { |
1580 | if (rq->end_io(rq, 0) == RQ_END_IO_FREE) | |
1581 | blk_mq_free_request(rq); | |
1582 | } else if (req_ref_put_and_test(rq)) { | |
2e315dc0 | 1583 | __blk_mq_free_request(rq); |
de671d61 | 1584 | } |
2e315dc0 ML |
1585 | } |
1586 | ||
2dd6532e | 1587 | static bool blk_mq_check_expired(struct request *rq, void *priv) |
1d9bd516 | 1588 | { |
82c22947 | 1589 | struct blk_expired_data *expired = priv; |
12f5b931 KB |
1590 | |
1591 | /* | |
c797b40c ML |
1592 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
1593 | * be reallocated underneath the timeout handler's processing, then | |
1594 | * the expire check is reliable. If the request is not expired, then | |
1595 | * it was completed and reallocated as a new request after returning | |
1596 | * from blk_mq_check_expired(). | |
1d9bd516 | 1597 | */ |
82c22947 DJ |
1598 | if (blk_mq_req_expired(rq, expired)) { |
1599 | expired->has_timedout_rq = true; | |
1600 | return false; | |
1601 | } | |
1602 | return true; | |
1603 | } | |
1604 | ||
1605 | static bool blk_mq_handle_expired(struct request *rq, void *priv) | |
1606 | { | |
1607 | struct blk_expired_data *expired = priv; | |
1608 | ||
1609 | if (blk_mq_req_expired(rq, expired)) | |
9bdb4833 | 1610 | blk_mq_rq_timed_out(rq); |
7baa8572 | 1611 | return true; |
1d9bd516 TH |
1612 | } |
1613 | ||
287922eb | 1614 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 1615 | { |
287922eb CH |
1616 | struct request_queue *q = |
1617 | container_of(work, struct request_queue, timeout_work); | |
82c22947 DJ |
1618 | struct blk_expired_data expired = { |
1619 | .timeout_start = jiffies, | |
1620 | }; | |
1d9bd516 | 1621 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 1622 | unsigned long i; |
320ae51f | 1623 | |
71f79fb3 GKB |
1624 | /* A deadlock might occur if a request is stuck requiring a |
1625 | * timeout at the same time a queue freeze is waiting | |
1626 | * completion, since the timeout code would not be able to | |
1627 | * acquire the queue reference here. | |
1628 | * | |
1629 | * That's why we don't use blk_queue_enter here; instead, we use | |
1630 | * percpu_ref_tryget directly, because we need to be able to | |
1631 | * obtain a reference even in the short window between the queue | |
1632 | * starting to freeze, by dropping the first reference in | |
1671d522 | 1633 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
1634 | * consumed, marked by the instant q_usage_counter reaches |
1635 | * zero. | |
1636 | */ | |
1637 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
1638 | return; |
1639 | ||
82c22947 DJ |
1640 | /* check if there is any timed-out request */ |
1641 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired); | |
1642 | if (expired.has_timedout_rq) { | |
1643 | /* | |
1644 | * Before walking tags, we must ensure any submit started | |
1645 | * before the current time has finished. Since the submit | |
1646 | * uses srcu or rcu, wait for a synchronization point to | |
1647 | * ensure all running submits have finished | |
1648 | */ | |
483239c7 | 1649 | blk_mq_wait_quiesce_done(q->tag_set); |
82c22947 DJ |
1650 | |
1651 | expired.next = 0; | |
1652 | blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired); | |
1653 | } | |
320ae51f | 1654 | |
82c22947 DJ |
1655 | if (expired.next != 0) { |
1656 | mod_timer(&q->timeout, expired.next); | |
0d2602ca | 1657 | } else { |
fcd36c36 BVA |
1658 | /* |
1659 | * Request timeouts are handled as a forward rolling timer. If | |
1660 | * we end up here it means that no requests are pending and | |
1661 | * also that no request has been pending for a while. Mark | |
1662 | * each hctx as idle. | |
1663 | */ | |
f054b56c ML |
1664 | queue_for_each_hw_ctx(q, hctx, i) { |
1665 | /* the hctx may be unmapped, so check it here */ | |
1666 | if (blk_mq_hw_queue_mapped(hctx)) | |
1667 | blk_mq_tag_idle(hctx); | |
1668 | } | |
0d2602ca | 1669 | } |
287922eb | 1670 | blk_queue_exit(q); |
320ae51f JA |
1671 | } |
1672 | ||
88459642 OS |
1673 | struct flush_busy_ctx_data { |
1674 | struct blk_mq_hw_ctx *hctx; | |
1675 | struct list_head *list; | |
1676 | }; | |
1677 | ||
1678 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1679 | { | |
1680 | struct flush_busy_ctx_data *flush_data = data; | |
1681 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1682 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1683 | enum hctx_type type = hctx->type; |
88459642 | 1684 | |
88459642 | 1685 | spin_lock(&ctx->lock); |
c16d6b5a | 1686 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1687 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1688 | spin_unlock(&ctx->lock); |
1689 | return true; | |
1690 | } | |
1691 | ||
1429d7c9 JA |
1692 | /* |
1693 | * Process software queues that have been marked busy, splicing them | |
1694 | * to the for-dispatch | |
1695 | */ | |
2c3ad667 | 1696 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1697 | { |
88459642 OS |
1698 | struct flush_busy_ctx_data data = { |
1699 | .hctx = hctx, | |
1700 | .list = list, | |
1701 | }; | |
1429d7c9 | 1702 | |
88459642 | 1703 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1704 | } |
2c3ad667 | 1705 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1706 | |
b347689f ML |
1707 | struct dispatch_rq_data { |
1708 | struct blk_mq_hw_ctx *hctx; | |
1709 | struct request *rq; | |
1710 | }; | |
1711 | ||
1712 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1713 | void *data) | |
1714 | { | |
1715 | struct dispatch_rq_data *dispatch_data = data; | |
1716 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1717 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1718 | enum hctx_type type = hctx->type; |
b347689f ML |
1719 | |
1720 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1721 | if (!list_empty(&ctx->rq_lists[type])) { |
1722 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1723 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1724 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1725 | sbitmap_clear_bit(sb, bitnr); |
1726 | } | |
1727 | spin_unlock(&ctx->lock); | |
1728 | ||
1729 | return !dispatch_data->rq; | |
1730 | } | |
1731 | ||
1732 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1733 | struct blk_mq_ctx *start) | |
1734 | { | |
f31967f0 | 1735 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1736 | struct dispatch_rq_data data = { |
1737 | .hctx = hctx, | |
1738 | .rq = NULL, | |
1739 | }; | |
1740 | ||
1741 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1742 | dispatch_rq_from_ctx, &data); | |
1743 | ||
1744 | return data.rq; | |
1745 | } | |
1746 | ||
b8643d68 | 1747 | bool __blk_mq_alloc_driver_tag(struct request *rq) |
570e9b73 | 1748 | { |
ae0f1a73 | 1749 | struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1750 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1751 | int tag; |
1752 | ||
568f2700 ML |
1753 | blk_mq_tag_busy(rq->mq_hctx); |
1754 | ||
570e9b73 | 1755 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
ae0f1a73 | 1756 | bt = &rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1757 | tag_offset = 0; |
28500850 ML |
1758 | } else { |
1759 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1760 | return false; | |
570e9b73 ML |
1761 | } |
1762 | ||
570e9b73 ML |
1763 | tag = __sbitmap_queue_get(bt); |
1764 | if (tag == BLK_MQ_NO_TAG) | |
1765 | return false; | |
1766 | ||
1767 | rq->tag = tag + tag_offset; | |
b8643d68 | 1768 | blk_mq_inc_active_requests(rq->mq_hctx); |
568f2700 | 1769 | return true; |
570e9b73 ML |
1770 | } |
1771 | ||
eb619fdb JA |
1772 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1773 | int flags, void *key) | |
da55f2cc OS |
1774 | { |
1775 | struct blk_mq_hw_ctx *hctx; | |
1776 | ||
1777 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1778 | ||
5815839b | 1779 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1780 | if (!list_empty(&wait->entry)) { |
1781 | struct sbitmap_queue *sbq; | |
1782 | ||
1783 | list_del_init(&wait->entry); | |
ae0f1a73 | 1784 | sbq = &hctx->tags->bitmap_tags; |
e8618575 JA |
1785 | atomic_dec(&sbq->ws_active); |
1786 | } | |
5815839b ML |
1787 | spin_unlock(&hctx->dispatch_wait_lock); |
1788 | ||
da55f2cc OS |
1789 | blk_mq_run_hw_queue(hctx, true); |
1790 | return 1; | |
1791 | } | |
1792 | ||
f906a6a0 JA |
1793 | /* |
1794 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1795 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1796 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1797 | * marking us as waiting. |
1798 | */ | |
2278d69f | 1799 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1800 | struct request *rq) |
da55f2cc | 1801 | { |
98b99e94 | 1802 | struct sbitmap_queue *sbq; |
5815839b | 1803 | struct wait_queue_head *wq; |
f906a6a0 JA |
1804 | wait_queue_entry_t *wait; |
1805 | bool ret; | |
da55f2cc | 1806 | |
47df9ce9 KS |
1807 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
1808 | !(blk_mq_is_shared_tags(hctx->flags))) { | |
684b7324 | 1809 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1810 | |
c27d53fb BVA |
1811 | /* |
1812 | * It's possible that a tag was freed in the window between the | |
1813 | * allocation failure and adding the hardware queue to the wait | |
1814 | * queue. | |
1815 | * | |
1816 | * Don't clear RESTART here, someone else could have set it. | |
1817 | * At most this will cost an extra queue run. | |
1818 | */ | |
8ab6bb9e | 1819 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1820 | } |
eb619fdb | 1821 | |
2278d69f | 1822 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1823 | if (!list_empty_careful(&wait->entry)) |
1824 | return false; | |
1825 | ||
98b99e94 KS |
1826 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) |
1827 | sbq = &hctx->tags->breserved_tags; | |
1828 | else | |
1829 | sbq = &hctx->tags->bitmap_tags; | |
e8618575 | 1830 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1831 | |
1832 | spin_lock_irq(&wq->lock); | |
1833 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1834 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1835 | spin_unlock(&hctx->dispatch_wait_lock); |
1836 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1837 | return false; |
eb619fdb JA |
1838 | } |
1839 | ||
e8618575 | 1840 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1841 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1842 | __add_wait_queue(wq, wait); | |
c27d53fb | 1843 | |
5266caaf ML |
1844 | /* |
1845 | * Add one explicit barrier since blk_mq_get_driver_tag() may | |
1846 | * not imply barrier in case of failure. | |
1847 | * | |
1848 | * Order adding us to wait queue and allocating driver tag. | |
1849 | * | |
1850 | * The pair is the one implied in sbitmap_queue_wake_up() which | |
1851 | * orders clearing sbitmap tag bits and waitqueue_active() in | |
1852 | * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless | |
1853 | * | |
1854 | * Otherwise, re-order of adding wait queue and getting driver tag | |
1855 | * may cause __sbitmap_queue_wake_up() to wake up nothing because | |
1856 | * the waitqueue_active() may not observe us in wait queue. | |
1857 | */ | |
1858 | smp_mb(); | |
1859 | ||
da55f2cc | 1860 | /* |
eb619fdb JA |
1861 | * It's possible that a tag was freed in the window between the |
1862 | * allocation failure and adding the hardware queue to the wait | |
1863 | * queue. | |
da55f2cc | 1864 | */ |
8ab6bb9e | 1865 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1866 | if (!ret) { |
5815839b ML |
1867 | spin_unlock(&hctx->dispatch_wait_lock); |
1868 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1869 | return false; |
eb619fdb | 1870 | } |
c27d53fb BVA |
1871 | |
1872 | /* | |
1873 | * We got a tag, remove ourselves from the wait queue to ensure | |
1874 | * someone else gets the wakeup. | |
1875 | */ | |
c27d53fb | 1876 | list_del_init(&wait->entry); |
e8618575 | 1877 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1878 | spin_unlock(&hctx->dispatch_wait_lock); |
1879 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1880 | |
1881 | return true; | |
da55f2cc OS |
1882 | } |
1883 | ||
6e768717 ML |
1884 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1885 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1886 | /* | |
1887 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1888 | * - EWMA is one simple way to compute running average value | |
1889 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1890 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1891 | * factor doesn't matter because EWMA decreases exponentially | |
1892 | */ | |
1893 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1894 | { | |
1895 | unsigned int ewma; | |
1896 | ||
6e768717 ML |
1897 | ewma = hctx->dispatch_busy; |
1898 | ||
1899 | if (!ewma && !busy) | |
1900 | return; | |
1901 | ||
1902 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1903 | if (busy) | |
1904 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1905 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1906 | ||
1907 | hctx->dispatch_busy = ewma; | |
1908 | } | |
1909 | ||
86ff7c2a ML |
1910 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1911 | ||
c92a4103 JT |
1912 | static void blk_mq_handle_dev_resource(struct request *rq, |
1913 | struct list_head *list) | |
1914 | { | |
c92a4103 JT |
1915 | list_add(&rq->queuelist, list); |
1916 | __blk_mq_requeue_request(rq); | |
1917 | } | |
1918 | ||
75383524 ML |
1919 | enum prep_dispatch { |
1920 | PREP_DISPATCH_OK, | |
1921 | PREP_DISPATCH_NO_TAG, | |
1922 | PREP_DISPATCH_NO_BUDGET, | |
1923 | }; | |
1924 | ||
1925 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1926 | bool need_budget) | |
1927 | { | |
1928 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1929 | int budget_token = -1; |
75383524 | 1930 | |
2a5a24aa ML |
1931 | if (need_budget) { |
1932 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1933 | if (budget_token < 0) { | |
1934 | blk_mq_put_driver_tag(rq); | |
1935 | return PREP_DISPATCH_NO_BUDGET; | |
1936 | } | |
1937 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1938 | } |
1939 | ||
1940 | if (!blk_mq_get_driver_tag(rq)) { | |
1941 | /* | |
1942 | * The initial allocation attempt failed, so we need to | |
1943 | * rerun the hardware queue when a tag is freed. The | |
1944 | * waitqueue takes care of that. If the queue is run | |
1945 | * before we add this entry back on the dispatch list, | |
1946 | * we'll re-run it below. | |
1947 | */ | |
1948 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1949 | /* |
1950 | * All budgets not got from this function will be put | |
1951 | * together during handling partial dispatch | |
1952 | */ | |
1953 | if (need_budget) | |
2a5a24aa | 1954 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1955 | return PREP_DISPATCH_NO_TAG; |
1956 | } | |
1957 | } | |
1958 | ||
1959 | return PREP_DISPATCH_OK; | |
1960 | } | |
1961 | ||
1fd40b5e ML |
1962 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1963 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1964 | struct list_head *list) |
1fd40b5e | 1965 | { |
2a5a24aa | 1966 | struct request *rq; |
1fd40b5e | 1967 | |
2a5a24aa ML |
1968 | list_for_each_entry(rq, list, queuelist) { |
1969 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1970 | |
2a5a24aa ML |
1971 | if (budget_token >= 0) |
1972 | blk_mq_put_dispatch_budget(q, budget_token); | |
1973 | } | |
1fd40b5e ML |
1974 | } |
1975 | ||
34c9f547 KS |
1976 | /* |
1977 | * blk_mq_commit_rqs will notify driver using bd->last that there is no | |
1978 | * more requests. (See comment in struct blk_mq_ops for commit_rqs for | |
1979 | * details) | |
1980 | * Attention, we should explicitly call this in unusual cases: | |
1981 | * 1) did not queue everything initially scheduled to queue | |
1982 | * 2) the last attempt to queue a request failed | |
1983 | */ | |
1984 | static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued, | |
1985 | bool from_schedule) | |
1986 | { | |
1987 | if (hctx->queue->mq_ops->commit_rqs && queued) { | |
1988 | trace_block_unplug(hctx->queue, queued, !from_schedule); | |
1989 | hctx->queue->mq_ops->commit_rqs(hctx); | |
1990 | } | |
1991 | } | |
1992 | ||
1f57f8d4 JA |
1993 | /* |
1994 | * Returns true if we did some work AND can potentially do more. | |
1995 | */ | |
445874e8 | 1996 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 1997 | unsigned int nr_budgets) |
320ae51f | 1998 | { |
75383524 | 1999 | enum prep_dispatch prep; |
445874e8 | 2000 | struct request_queue *q = hctx->queue; |
f1ce99f7 | 2001 | struct request *rq; |
4ea58fe4 | 2002 | int queued; |
86ff7c2a | 2003 | blk_status_t ret = BLK_STS_OK; |
9586e67b | 2004 | bool needs_resource = false; |
320ae51f | 2005 | |
81380ca1 OS |
2006 | if (list_empty(list)) |
2007 | return false; | |
2008 | ||
320ae51f JA |
2009 | /* |
2010 | * Now process all the entries, sending them to the driver. | |
2011 | */ | |
4ea58fe4 | 2012 | queued = 0; |
81380ca1 | 2013 | do { |
74c45052 | 2014 | struct blk_mq_queue_data bd; |
320ae51f | 2015 | |
f04c3df3 | 2016 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 2017 | |
445874e8 | 2018 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 2019 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 2020 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 2021 | break; |
de148297 | 2022 | |
320ae51f | 2023 | list_del_init(&rq->queuelist); |
320ae51f | 2024 | |
74c45052 | 2025 | bd.rq = rq; |
f1ce99f7 | 2026 | bd.last = list_empty(list); |
74c45052 | 2027 | |
1fd40b5e ML |
2028 | /* |
2029 | * once the request is queued to lld, no need to cover the | |
2030 | * budget any more | |
2031 | */ | |
2032 | if (nr_budgets) | |
2033 | nr_budgets--; | |
74c45052 | 2034 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
2035 | switch (ret) { |
2036 | case BLK_STS_OK: | |
2037 | queued++; | |
320ae51f | 2038 | break; |
7bf13729 | 2039 | case BLK_STS_RESOURCE: |
9586e67b NA |
2040 | needs_resource = true; |
2041 | fallthrough; | |
7bf13729 ML |
2042 | case BLK_STS_DEV_RESOURCE: |
2043 | blk_mq_handle_dev_resource(rq, list); | |
2044 | goto out; | |
7bf13729 | 2045 | default: |
e21ee5a6 | 2046 | blk_mq_end_request(rq, ret); |
320ae51f | 2047 | } |
81380ca1 | 2048 | } while (!list_empty(list)); |
7bf13729 | 2049 | out: |
632bfb63 | 2050 | /* If we didn't flush the entire list, we could have told the driver |
2051 | * there was more coming, but that turned out to be a lie. | |
2052 | */ | |
e4ef2e05 KS |
2053 | if (!list_empty(list) || ret != BLK_STS_OK) |
2054 | blk_mq_commit_rqs(hctx, queued, false); | |
2055 | ||
320ae51f JA |
2056 | /* |
2057 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
2058 | * that is where we will continue on next queue run. | |
2059 | */ | |
f04c3df3 | 2060 | if (!list_empty(list)) { |
86ff7c2a | 2061 | bool needs_restart; |
75383524 ML |
2062 | /* For non-shared tags, the RESTART check will suffice */ |
2063 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
47df9ce9 KS |
2064 | ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) || |
2065 | blk_mq_is_shared_tags(hctx->flags)); | |
86ff7c2a | 2066 | |
2a5a24aa ML |
2067 | if (nr_budgets) |
2068 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 2069 | |
320ae51f | 2070 | spin_lock(&hctx->lock); |
01e99aec | 2071 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 2072 | spin_unlock(&hctx->lock); |
f04c3df3 | 2073 | |
d7d8535f ML |
2074 | /* |
2075 | * Order adding requests to hctx->dispatch and checking | |
2076 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
2077 | * in blk_mq_sched_restart(). Avoid restart code path to | |
2078 | * miss the new added requests to hctx->dispatch, meantime | |
2079 | * SCHED_RESTART is observed here. | |
2080 | */ | |
2081 | smp_mb(); | |
2082 | ||
9ba52e58 | 2083 | /* |
710c785f BVA |
2084 | * If SCHED_RESTART was set by the caller of this function and |
2085 | * it is no longer set that means that it was cleared by another | |
2086 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 2087 | * |
eb619fdb JA |
2088 | * If 'no_tag' is set, that means that we failed getting |
2089 | * a driver tag with an I/O scheduler attached. If our dispatch | |
2090 | * waitqueue is no longer active, ensure that we run the queue | |
2091 | * AFTER adding our entries back to the list. | |
bd166ef1 | 2092 | * |
710c785f BVA |
2093 | * If no I/O scheduler has been configured it is possible that |
2094 | * the hardware queue got stopped and restarted before requests | |
2095 | * were pushed back onto the dispatch list. Rerun the queue to | |
2096 | * avoid starvation. Notes: | |
2097 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
2098 | * been stopped before rerunning a queue. | |
2099 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 2100 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 2101 | * and dm-rq. |
86ff7c2a ML |
2102 | * |
2103 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
2104 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 | 2105 | * that could otherwise occur if the queue is idle. We'll do |
9586e67b NA |
2106 | * similar if we couldn't get budget or couldn't lock a zone |
2107 | * and SCHED_RESTART is set. | |
bd166ef1 | 2108 | */ |
86ff7c2a | 2109 | needs_restart = blk_mq_sched_needs_restart(hctx); |
9586e67b NA |
2110 | if (prep == PREP_DISPATCH_NO_BUDGET) |
2111 | needs_resource = true; | |
86ff7c2a | 2112 | if (!needs_restart || |
eb619fdb | 2113 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 2114 | blk_mq_run_hw_queue(hctx, true); |
6d5e8d21 | 2115 | else if (needs_resource) |
86ff7c2a | 2116 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 2117 | |
6e768717 | 2118 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 2119 | return false; |
4ea58fe4 | 2120 | } |
f04c3df3 | 2121 | |
4ea58fe4 KS |
2122 | blk_mq_update_dispatch_busy(hctx, false); |
2123 | return true; | |
f04c3df3 JA |
2124 | } |
2125 | ||
f82ddf19 ML |
2126 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
2127 | { | |
2128 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
2129 | ||
2130 | if (cpu >= nr_cpu_ids) | |
2131 | cpu = cpumask_first(hctx->cpumask); | |
2132 | return cpu; | |
2133 | } | |
2134 | ||
a46c2702 ML |
2135 | /* |
2136 | * ->next_cpu is always calculated from hctx->cpumask, so simply use | |
2137 | * it for speeding up the check | |
2138 | */ | |
2139 | static bool blk_mq_hctx_empty_cpumask(struct blk_mq_hw_ctx *hctx) | |
2140 | { | |
2141 | return hctx->next_cpu >= nr_cpu_ids; | |
2142 | } | |
2143 | ||
506e931f JA |
2144 | /* |
2145 | * It'd be great if the workqueue API had a way to pass | |
2146 | * in a mask and had some smarts for more clever placement. | |
2147 | * For now we just round-robin here, switching for every | |
2148 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
2149 | */ | |
2150 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
2151 | { | |
7bed4595 | 2152 | bool tried = false; |
476f8c98 | 2153 | int next_cpu = hctx->next_cpu; |
7bed4595 | 2154 | |
a46c2702 ML |
2155 | /* Switch to unbound if no allowable CPUs in this hctx */ |
2156 | if (hctx->queue->nr_hw_queues == 1 || blk_mq_hctx_empty_cpumask(hctx)) | |
b657d7e6 | 2157 | return WORK_CPU_UNBOUND; |
506e931f JA |
2158 | |
2159 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 2160 | select_cpu: |
476f8c98 | 2161 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 2162 | cpu_online_mask); |
506e931f | 2163 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 2164 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2165 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2166 | } | |
2167 | ||
7bed4595 ML |
2168 | /* |
2169 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
2170 | * and it should only happen in the path of handling CPU DEAD. | |
2171 | */ | |
476f8c98 | 2172 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
2173 | if (!tried) { |
2174 | tried = true; | |
2175 | goto select_cpu; | |
2176 | } | |
2177 | ||
2178 | /* | |
2179 | * Make sure to re-select CPU next time once after CPUs | |
2180 | * in hctx->cpumask become online again. | |
2181 | */ | |
476f8c98 | 2182 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
2183 | hctx->next_cpu_batch = 1; |
2184 | return WORK_CPU_UNBOUND; | |
2185 | } | |
476f8c98 ML |
2186 | |
2187 | hctx->next_cpu = next_cpu; | |
2188 | return next_cpu; | |
506e931f JA |
2189 | } |
2190 | ||
105663f7 | 2191 | /** |
1aa8d875 | 2192 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. |
105663f7 | 2193 | * @hctx: Pointer to the hardware queue to run. |
fa94ba8a | 2194 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 | 2195 | * |
1aa8d875 | 2196 | * Run a hardware queue asynchronously with a delay of @msecs. |
105663f7 | 2197 | */ |
1aa8d875 | 2198 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
320ae51f | 2199 | { |
5435c023 | 2200 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f | 2201 | return; |
ae943d20 BVA |
2202 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
2203 | msecs_to_jiffies(msecs)); | |
7587a5ae | 2204 | } |
7587a5ae BVA |
2205 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); |
2206 | ||
105663f7 AA |
2207 | /** |
2208 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
2209 | * @hctx: Pointer to the hardware queue to run. | |
2210 | * @async: If we want to run the queue asynchronously. | |
2211 | * | |
2212 | * Check if the request queue is not in a quiesced state and if there are | |
2213 | * pending requests to be sent. If this is true, run the queue to send requests | |
2214 | * to hardware. | |
2215 | */ | |
626fb735 | 2216 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 2217 | { |
24f5a90f ML |
2218 | bool need_run; |
2219 | ||
4d5bba5b CH |
2220 | /* |
2221 | * We can't run the queue inline with interrupts disabled. | |
2222 | */ | |
2223 | WARN_ON_ONCE(!async && in_interrupt()); | |
2224 | ||
65a558f6 BVA |
2225 | might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING); |
2226 | ||
24f5a90f ML |
2227 | /* |
2228 | * When queue is quiesced, we may be switching io scheduler, or | |
2229 | * updating nr_hw_queues, or other things, and we can't run queue | |
2230 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
2231 | * | |
2232 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
2233 | * quiesced. | |
2234 | */ | |
41adf531 | 2235 | __blk_mq_run_dispatch_ops(hctx->queue, false, |
2a904d00 ML |
2236 | need_run = !blk_queue_quiesced(hctx->queue) && |
2237 | blk_mq_hctx_has_pending(hctx)); | |
24f5a90f | 2238 | |
1aa8d875 CH |
2239 | if (!need_run) |
2240 | return; | |
2241 | ||
65a558f6 | 2242 | if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) { |
1aa8d875 CH |
2243 | blk_mq_delay_run_hw_queue(hctx, 0); |
2244 | return; | |
2245 | } | |
2246 | ||
4d5bba5b CH |
2247 | blk_mq_run_dispatch_ops(hctx->queue, |
2248 | blk_mq_sched_dispatch_requests(hctx)); | |
320ae51f | 2249 | } |
5b727272 | 2250 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 2251 | |
b6e68ee8 JK |
2252 | /* |
2253 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
2254 | * scheduler. | |
2255 | */ | |
2256 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
2257 | { | |
5d05426e | 2258 | struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); |
b6e68ee8 JK |
2259 | /* |
2260 | * If the IO scheduler does not respect hardware queues when | |
2261 | * dispatching, we just don't bother with multiple HW queues and | |
2262 | * dispatch from hctx for the current CPU since running multiple queues | |
2263 | * just causes lock contention inside the scheduler and pointless cache | |
2264 | * bouncing. | |
2265 | */ | |
51ab80f0 | 2266 | struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT]; |
5d05426e | 2267 | |
b6e68ee8 JK |
2268 | if (!blk_mq_hctx_stopped(hctx)) |
2269 | return hctx; | |
2270 | return NULL; | |
2271 | } | |
2272 | ||
105663f7 | 2273 | /** |
24f7bb88 | 2274 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
2275 | * @q: Pointer to the request queue to run. |
2276 | * @async: If we want to run the queue asynchronously. | |
2277 | */ | |
b94ec296 | 2278 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 2279 | { |
b6e68ee8 | 2280 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2281 | unsigned long i; |
320ae51f | 2282 | |
b6e68ee8 | 2283 | sq_hctx = NULL; |
4d337ceb | 2284 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2285 | sq_hctx = blk_mq_get_sq_hctx(q); |
320ae51f | 2286 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 2287 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 2288 | continue; |
b6e68ee8 JK |
2289 | /* |
2290 | * Dispatch from this hctx either if there's no hctx preferred | |
2291 | * by IO scheduler or if it has requests that bypass the | |
2292 | * scheduler. | |
2293 | */ | |
2294 | if (!sq_hctx || sq_hctx == hctx || | |
2295 | !list_empty_careful(&hctx->dispatch)) | |
2296 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
2297 | } |
2298 | } | |
b94ec296 | 2299 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 2300 | |
b9151e7b DA |
2301 | /** |
2302 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
2303 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 2304 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
2305 | */ |
2306 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
2307 | { | |
b6e68ee8 | 2308 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2309 | unsigned long i; |
b9151e7b | 2310 | |
b6e68ee8 | 2311 | sq_hctx = NULL; |
4d337ceb | 2312 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2313 | sq_hctx = blk_mq_get_sq_hctx(q); |
b9151e7b DA |
2314 | queue_for_each_hw_ctx(q, hctx, i) { |
2315 | if (blk_mq_hctx_stopped(hctx)) | |
2316 | continue; | |
8f5fea65 DJ |
2317 | /* |
2318 | * If there is already a run_work pending, leave the | |
2319 | * pending delay untouched. Otherwise, a hctx can stall | |
2320 | * if another hctx is re-delaying the other's work | |
2321 | * before the work executes. | |
2322 | */ | |
2323 | if (delayed_work_pending(&hctx->run_work)) | |
2324 | continue; | |
b6e68ee8 JK |
2325 | /* |
2326 | * Dispatch from this hctx either if there's no hctx preferred | |
2327 | * by IO scheduler or if it has requests that bypass the | |
2328 | * scheduler. | |
2329 | */ | |
2330 | if (!sq_hctx || sq_hctx == hctx || | |
2331 | !list_empty_careful(&hctx->dispatch)) | |
2332 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
2333 | } |
2334 | } | |
2335 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
2336 | ||
39a70c76 ML |
2337 | /* |
2338 | * This function is often used for pausing .queue_rq() by driver when | |
2339 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2340 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2341 | * |
2342 | * We do not guarantee that dispatch can be drained or blocked | |
2343 | * after blk_mq_stop_hw_queue() returns. Please use | |
2344 | * blk_mq_quiesce_queue() for that requirement. | |
2345 | */ | |
2719aa21 JA |
2346 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
2347 | { | |
641a9ed6 | 2348 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 2349 | |
641a9ed6 | 2350 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 2351 | } |
641a9ed6 | 2352 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 2353 | |
39a70c76 ML |
2354 | /* |
2355 | * This function is often used for pausing .queue_rq() by driver when | |
2356 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2357 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2358 | * |
2359 | * We do not guarantee that dispatch can be drained or blocked | |
2360 | * after blk_mq_stop_hw_queues() returns. Please use | |
2361 | * blk_mq_quiesce_queue() for that requirement. | |
2362 | */ | |
2719aa21 JA |
2363 | void blk_mq_stop_hw_queues(struct request_queue *q) |
2364 | { | |
641a9ed6 | 2365 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 2366 | unsigned long i; |
641a9ed6 ML |
2367 | |
2368 | queue_for_each_hw_ctx(q, hctx, i) | |
2369 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
2370 | } |
2371 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
2372 | ||
320ae51f JA |
2373 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
2374 | { | |
2375 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 2376 | |
65a558f6 | 2377 | blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING); |
320ae51f JA |
2378 | } |
2379 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
2380 | ||
2f268556 CH |
2381 | void blk_mq_start_hw_queues(struct request_queue *q) |
2382 | { | |
2383 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2384 | unsigned long i; |
2f268556 CH |
2385 | |
2386 | queue_for_each_hw_ctx(q, hctx, i) | |
2387 | blk_mq_start_hw_queue(hctx); | |
2388 | } | |
2389 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
2390 | ||
ae911c5e JA |
2391 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
2392 | { | |
2393 | if (!blk_mq_hctx_stopped(hctx)) | |
2394 | return; | |
2395 | ||
2396 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
2397 | blk_mq_run_hw_queue(hctx, async); | |
2398 | } | |
2399 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
2400 | ||
1b4a3258 | 2401 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
2402 | { |
2403 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2404 | unsigned long i; |
320ae51f | 2405 | |
ae911c5e | 2406 | queue_for_each_hw_ctx(q, hctx, i) |
65a558f6 BVA |
2407 | blk_mq_start_stopped_hw_queue(hctx, async || |
2408 | (hctx->flags & BLK_MQ_F_BLOCKING)); | |
320ae51f JA |
2409 | } |
2410 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
2411 | ||
70f4db63 | 2412 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f | 2413 | { |
c20a1a2c CH |
2414 | struct blk_mq_hw_ctx *hctx = |
2415 | container_of(work, struct blk_mq_hw_ctx, run_work.work); | |
7b607814 | 2416 | |
4d5bba5b CH |
2417 | blk_mq_run_dispatch_ops(hctx->queue, |
2418 | blk_mq_sched_dispatch_requests(hctx)); | |
320ae51f JA |
2419 | } |
2420 | ||
105663f7 AA |
2421 | /** |
2422 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
2423 | * @rq: Pointer to request to be inserted. | |
2b597613 | 2424 | * @flags: BLK_MQ_INSERT_* |
105663f7 | 2425 | * |
157f377b JA |
2426 | * Should only be used carefully, when the caller knows we want to |
2427 | * bypass a potential IO scheduler on the target device. | |
2428 | */ | |
360f2648 | 2429 | static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags) |
157f377b | 2430 | { |
ea4f995e | 2431 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
2432 | |
2433 | spin_lock(&hctx->lock); | |
2b597613 | 2434 | if (flags & BLK_MQ_INSERT_AT_HEAD) |
01e99aec ML |
2435 | list_add(&rq->queuelist, &hctx->dispatch); |
2436 | else | |
2437 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b | 2438 | spin_unlock(&hctx->lock); |
157f377b JA |
2439 | } |
2440 | ||
05a93117 CH |
2441 | static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, |
2442 | struct blk_mq_ctx *ctx, struct list_head *list, | |
2443 | bool run_queue_async) | |
320ae51f | 2444 | { |
3f0cedc7 | 2445 | struct request *rq; |
c16d6b5a | 2446 | enum hctx_type type = hctx->type; |
3f0cedc7 | 2447 | |
94aa228c CH |
2448 | /* |
2449 | * Try to issue requests directly if the hw queue isn't busy to save an | |
2450 | * extra enqueue & dequeue to the sw queue. | |
2451 | */ | |
2452 | if (!hctx->dispatch_busy && !run_queue_async) { | |
2453 | blk_mq_run_dispatch_ops(hctx->queue, | |
2454 | blk_mq_try_issue_list_directly(hctx, list)); | |
2455 | if (list_empty(list)) | |
2456 | goto out; | |
2457 | } | |
2458 | ||
320ae51f JA |
2459 | /* |
2460 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
2461 | * offline now | |
2462 | */ | |
3f0cedc7 | 2463 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 2464 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 2465 | trace_block_rq_insert(rq); |
65a558f6 BVA |
2466 | if (rq->cmd_flags & REQ_NOWAIT) |
2467 | run_queue_async = true; | |
320ae51f | 2468 | } |
3f0cedc7 ML |
2469 | |
2470 | spin_lock(&ctx->lock); | |
c16d6b5a | 2471 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 2472 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 2473 | spin_unlock(&ctx->lock); |
94aa228c CH |
2474 | out: |
2475 | blk_mq_run_hw_queue(hctx, run_queue_async); | |
320ae51f JA |
2476 | } |
2477 | ||
710fa378 | 2478 | static void blk_mq_insert_request(struct request *rq, blk_insert_t flags) |
2bd215df CH |
2479 | { |
2480 | struct request_queue *q = rq->q; | |
2bd215df CH |
2481 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
2482 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2483 | ||
53548d2a CH |
2484 | if (blk_rq_is_passthrough(rq)) { |
2485 | /* | |
2486 | * Passthrough request have to be added to hctx->dispatch | |
2487 | * directly. The device may be in a situation where it can't | |
2488 | * handle FS request, and always returns BLK_STS_RESOURCE for | |
2489 | * them, which gets them added to hctx->dispatch. | |
2490 | * | |
2491 | * If a passthrough request is required to unblock the queues, | |
2492 | * and it is added to the scheduler queue, there is no chance to | |
2493 | * dispatch it given we prioritize requests in hctx->dispatch. | |
2494 | */ | |
2b597613 | 2495 | blk_mq_request_bypass_insert(rq, flags); |
be4c4278 | 2496 | } else if (req_op(rq) == REQ_OP_FLUSH) { |
2bd215df CH |
2497 | /* |
2498 | * Firstly normal IO request is inserted to scheduler queue or | |
2499 | * sw queue, meantime we add flush request to dispatch queue( | |
2500 | * hctx->dispatch) directly and there is at most one in-flight | |
2501 | * flush request for each hw queue, so it doesn't matter to add | |
2502 | * flush request to tail or front of the dispatch queue. | |
2503 | * | |
2504 | * Secondly in case of NCQ, flush request belongs to non-NCQ | |
2505 | * command, and queueing it will fail when there is any | |
2506 | * in-flight normal IO request(NCQ command). When adding flush | |
2507 | * rq to the front of hctx->dispatch, it is easier to introduce | |
2508 | * extra time to flush rq's latency because of S_SCHED_RESTART | |
2509 | * compared with adding to the tail of dispatch queue, then | |
2510 | * chance of flush merge is increased, and less flush requests | |
2511 | * will be issued to controller. It is observed that ~10% time | |
2512 | * is saved in blktests block/004 on disk attached to AHCI/NCQ | |
2513 | * drive when adding flush rq to the front of hctx->dispatch. | |
2514 | * | |
2515 | * Simply queue flush rq to the front of hctx->dispatch so that | |
2516 | * intensive flush workloads can benefit in case of NCQ HW. | |
2517 | */ | |
2b597613 | 2518 | blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD); |
53548d2a | 2519 | } else if (q->elevator) { |
2bd215df CH |
2520 | LIST_HEAD(list); |
2521 | ||
53548d2a CH |
2522 | WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG); |
2523 | ||
2bd215df | 2524 | list_add(&rq->queuelist, &list); |
93fffe16 | 2525 | q->elevator->type->ops.insert_requests(hctx, &list, flags); |
2bd215df | 2526 | } else { |
4ec5c055 CH |
2527 | trace_block_rq_insert(rq); |
2528 | ||
2bd215df | 2529 | spin_lock(&ctx->lock); |
710fa378 | 2530 | if (flags & BLK_MQ_INSERT_AT_HEAD) |
4ec5c055 CH |
2531 | list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]); |
2532 | else | |
2533 | list_add_tail(&rq->queuelist, | |
2534 | &ctx->rq_lists[hctx->type]); | |
a88db1e0 | 2535 | blk_mq_hctx_mark_pending(hctx, ctx); |
2bd215df CH |
2536 | spin_unlock(&ctx->lock); |
2537 | } | |
320ae51f JA |
2538 | } |
2539 | ||
14ccb66b CH |
2540 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
2541 | unsigned int nr_segs) | |
320ae51f | 2542 | { |
93f221ae EB |
2543 | int err; |
2544 | ||
f924cdde CH |
2545 | if (bio->bi_opf & REQ_RAHEAD) |
2546 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
2547 | ||
2548 | rq->__sector = bio->bi_iter.bi_sector; | |
44981351 | 2549 | rq->write_hint = bio->bi_write_hint; |
14ccb66b | 2550 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
2551 | |
2552 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
2553 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
2554 | WARN_ON_ONCE(err); | |
4b570521 | 2555 | |
b5af37ab | 2556 | blk_account_io_start(rq); |
320ae51f JA |
2557 | } |
2558 | ||
0f95549c | 2559 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2560 | struct request *rq, bool last) |
f984df1f | 2561 | { |
f984df1f | 2562 | struct request_queue *q = rq->q; |
f984df1f SL |
2563 | struct blk_mq_queue_data bd = { |
2564 | .rq = rq, | |
be94f058 | 2565 | .last = last, |
f984df1f | 2566 | }; |
f06345ad | 2567 | blk_status_t ret; |
0f95549c | 2568 | |
0f95549c MS |
2569 | /* |
2570 | * For OK queue, we are done. For error, caller may kill it. | |
2571 | * Any other error (busy), just add it to our list as we | |
2572 | * previously would have done. | |
2573 | */ | |
2574 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
2575 | switch (ret) { | |
2576 | case BLK_STS_OK: | |
6ce3dd6e | 2577 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2578 | break; |
2579 | case BLK_STS_RESOURCE: | |
86ff7c2a | 2580 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 2581 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
2582 | __blk_mq_requeue_request(rq); |
2583 | break; | |
2584 | default: | |
6ce3dd6e | 2585 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2586 | break; |
2587 | } | |
2588 | ||
2589 | return ret; | |
2590 | } | |
2591 | ||
2b71b877 | 2592 | static bool blk_mq_get_budget_and_tag(struct request *rq) |
0f95549c | 2593 | { |
2a5a24aa | 2594 | int budget_token; |
d964f04a | 2595 | |
2b71b877 | 2596 | budget_token = blk_mq_get_dispatch_budget(rq->q); |
2a5a24aa | 2597 | if (budget_token < 0) |
2b71b877 | 2598 | return false; |
2a5a24aa | 2599 | blk_mq_set_rq_budget_token(rq, budget_token); |
8ab6bb9e | 2600 | if (!blk_mq_get_driver_tag(rq)) { |
2b71b877 CH |
2601 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
2602 | return false; | |
88022d72 | 2603 | } |
2b71b877 | 2604 | return true; |
fd9c40f6 BVA |
2605 | } |
2606 | ||
105663f7 AA |
2607 | /** |
2608 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2609 | * @hctx: Pointer of the associated hardware queue. | |
2610 | * @rq: Pointer to request to be sent. | |
105663f7 AA |
2611 | * |
2612 | * If the device has enough resources to accept a new request now, send the | |
2613 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2614 | * we can try send it another time in the future. Requests inserted at this | |
2615 | * queue have higher priority. | |
2616 | */ | |
fd9c40f6 | 2617 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2618 | struct request *rq) |
fd9c40f6 | 2619 | { |
e1f44ac0 CH |
2620 | blk_status_t ret; |
2621 | ||
2622 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { | |
710fa378 | 2623 | blk_mq_insert_request(rq, 0); |
e1f44ac0 CH |
2624 | return; |
2625 | } | |
2626 | ||
dd6216bb | 2627 | if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) { |
710fa378 | 2628 | blk_mq_insert_request(rq, 0); |
65a558f6 | 2629 | blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT); |
e1f44ac0 CH |
2630 | return; |
2631 | } | |
fd9c40f6 | 2632 | |
e1f44ac0 CH |
2633 | ret = __blk_mq_issue_directly(hctx, rq, true); |
2634 | switch (ret) { | |
2635 | case BLK_STS_OK: | |
2636 | break; | |
2637 | case BLK_STS_RESOURCE: | |
2638 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2639 | blk_mq_request_bypass_insert(rq, 0); |
2394395c | 2640 | blk_mq_run_hw_queue(hctx, false); |
e1f44ac0 CH |
2641 | break; |
2642 | default: | |
fd9c40f6 | 2643 | blk_mq_end_request(rq, ret); |
e1f44ac0 CH |
2644 | break; |
2645 | } | |
fd9c40f6 BVA |
2646 | } |
2647 | ||
06c8c691 | 2648 | static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) |
fd9c40f6 | 2649 | { |
e1f44ac0 CH |
2650 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
2651 | ||
2652 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { | |
710fa378 | 2653 | blk_mq_insert_request(rq, 0); |
e1f44ac0 CH |
2654 | return BLK_STS_OK; |
2655 | } | |
2656 | ||
2657 | if (!blk_mq_get_budget_and_tag(rq)) | |
2658 | return BLK_STS_RESOURCE; | |
2659 | return __blk_mq_issue_directly(hctx, rq, last); | |
5eb6126e CH |
2660 | } |
2661 | ||
3e368fb0 | 2662 | static void blk_mq_plug_issue_direct(struct blk_plug *plug) |
b84c5b50 CH |
2663 | { |
2664 | struct blk_mq_hw_ctx *hctx = NULL; | |
2665 | struct request *rq; | |
2666 | int queued = 0; | |
0d617a83 | 2667 | blk_status_t ret = BLK_STS_OK; |
b84c5b50 CH |
2668 | |
2669 | while ((rq = rq_list_pop(&plug->mq_list))) { | |
2670 | bool last = rq_list_empty(plug->mq_list); | |
b84c5b50 CH |
2671 | |
2672 | if (hctx != rq->mq_hctx) { | |
34c9f547 KS |
2673 | if (hctx) { |
2674 | blk_mq_commit_rqs(hctx, queued, false); | |
2675 | queued = 0; | |
2676 | } | |
b84c5b50 CH |
2677 | hctx = rq->mq_hctx; |
2678 | } | |
2679 | ||
2680 | ret = blk_mq_request_issue_directly(rq, last); | |
2681 | switch (ret) { | |
2682 | case BLK_STS_OK: | |
2683 | queued++; | |
2684 | break; | |
2685 | case BLK_STS_RESOURCE: | |
2686 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2687 | blk_mq_request_bypass_insert(rq, 0); |
2394395c | 2688 | blk_mq_run_hw_queue(hctx, false); |
0d617a83 | 2689 | goto out; |
b84c5b50 CH |
2690 | default: |
2691 | blk_mq_end_request(rq, ret); | |
b84c5b50 CH |
2692 | break; |
2693 | } | |
2694 | } | |
2695 | ||
0d617a83 KS |
2696 | out: |
2697 | if (ret != BLK_STS_OK) | |
34c9f547 | 2698 | blk_mq_commit_rqs(hctx, queued, false); |
b84c5b50 CH |
2699 | } |
2700 | ||
518579a9 KB |
2701 | static void __blk_mq_flush_plug_list(struct request_queue *q, |
2702 | struct blk_plug *plug) | |
2703 | { | |
2704 | if (blk_queue_quiesced(q)) | |
2705 | return; | |
2706 | q->mq_ops->queue_rqs(&plug->mq_list); | |
2707 | } | |
2708 | ||
26fed4ac JA |
2709 | static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched) |
2710 | { | |
2711 | struct blk_mq_hw_ctx *this_hctx = NULL; | |
2712 | struct blk_mq_ctx *this_ctx = NULL; | |
2713 | struct request *requeue_list = NULL; | |
34e0a279 | 2714 | struct request **requeue_lastp = &requeue_list; |
26fed4ac | 2715 | unsigned int depth = 0; |
d97217e7 | 2716 | bool is_passthrough = false; |
26fed4ac JA |
2717 | LIST_HEAD(list); |
2718 | ||
2719 | do { | |
2720 | struct request *rq = rq_list_pop(&plug->mq_list); | |
2721 | ||
2722 | if (!this_hctx) { | |
2723 | this_hctx = rq->mq_hctx; | |
2724 | this_ctx = rq->mq_ctx; | |
d97217e7 ML |
2725 | is_passthrough = blk_rq_is_passthrough(rq); |
2726 | } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx || | |
2727 | is_passthrough != blk_rq_is_passthrough(rq)) { | |
34e0a279 | 2728 | rq_list_add_tail(&requeue_lastp, rq); |
26fed4ac JA |
2729 | continue; |
2730 | } | |
34e0a279 | 2731 | list_add(&rq->queuelist, &list); |
26fed4ac JA |
2732 | depth++; |
2733 | } while (!rq_list_empty(plug->mq_list)); | |
2734 | ||
2735 | plug->mq_list = requeue_list; | |
2736 | trace_block_unplug(this_hctx->queue, depth, !from_sched); | |
05a93117 CH |
2737 | |
2738 | percpu_ref_get(&this_hctx->queue->q_usage_counter); | |
d97217e7 | 2739 | /* passthrough requests should never be issued to the I/O scheduler */ |
2293cae7 ML |
2740 | if (is_passthrough) { |
2741 | spin_lock(&this_hctx->lock); | |
2742 | list_splice_tail_init(&list, &this_hctx->dispatch); | |
2743 | spin_unlock(&this_hctx->lock); | |
2744 | blk_mq_run_hw_queue(this_hctx, from_sched); | |
2745 | } else if (this_hctx->queue->elevator) { | |
05a93117 | 2746 | this_hctx->queue->elevator->type->ops.insert_requests(this_hctx, |
93fffe16 | 2747 | &list, 0); |
05a93117 CH |
2748 | blk_mq_run_hw_queue(this_hctx, from_sched); |
2749 | } else { | |
2750 | blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched); | |
2751 | } | |
2752 | percpu_ref_put(&this_hctx->queue->q_usage_counter); | |
26fed4ac JA |
2753 | } |
2754 | ||
b84c5b50 CH |
2755 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
2756 | { | |
3c67d44d | 2757 | struct request *rq; |
b84c5b50 | 2758 | |
70904263 RL |
2759 | /* |
2760 | * We may have been called recursively midway through handling | |
2761 | * plug->mq_list via a schedule() in the driver's queue_rq() callback. | |
2762 | * To avoid mq_list changing under our feet, clear rq_count early and | |
2763 | * bail out specifically if rq_count is 0 rather than checking | |
2764 | * whether the mq_list is empty. | |
2765 | */ | |
2766 | if (plug->rq_count == 0) | |
b84c5b50 CH |
2767 | return; |
2768 | plug->rq_count = 0; | |
2769 | ||
2770 | if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) { | |
3c67d44d JA |
2771 | struct request_queue *q; |
2772 | ||
2773 | rq = rq_list_peek(&plug->mq_list); | |
2774 | q = rq->q; | |
2775 | ||
2776 | /* | |
2777 | * Peek first request and see if we have a ->queue_rqs() hook. | |
2778 | * If we do, we can dispatch the whole plug list in one go. We | |
2779 | * already know at this point that all requests belong to the | |
2780 | * same queue, caller must ensure that's the case. | |
3c67d44d | 2781 | */ |
434097ee | 2782 | if (q->mq_ops->queue_rqs) { |
3c67d44d | 2783 | blk_mq_run_dispatch_ops(q, |
518579a9 | 2784 | __blk_mq_flush_plug_list(q, plug)); |
3c67d44d JA |
2785 | if (rq_list_empty(plug->mq_list)) |
2786 | return; | |
2787 | } | |
73f3760e ML |
2788 | |
2789 | blk_mq_run_dispatch_ops(q, | |
3e368fb0 | 2790 | blk_mq_plug_issue_direct(plug)); |
b84c5b50 CH |
2791 | if (rq_list_empty(plug->mq_list)) |
2792 | return; | |
2793 | } | |
2794 | ||
b84c5b50 | 2795 | do { |
26fed4ac | 2796 | blk_mq_dispatch_plug_list(plug, from_schedule); |
b84c5b50 | 2797 | } while (!rq_list_empty(plug->mq_list)); |
b84c5b50 CH |
2798 | } |
2799 | ||
94aa228c | 2800 | static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
6ce3dd6e ML |
2801 | struct list_head *list) |
2802 | { | |
536167d4 | 2803 | int queued = 0; |
984ce0a7 | 2804 | blk_status_t ret = BLK_STS_OK; |
536167d4 | 2805 | |
6ce3dd6e | 2806 | while (!list_empty(list)) { |
6ce3dd6e ML |
2807 | struct request *rq = list_first_entry(list, struct request, |
2808 | queuelist); | |
2809 | ||
2810 | list_del_init(&rq->queuelist); | |
fd9c40f6 | 2811 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
27e8b2bb KS |
2812 | switch (ret) { |
2813 | case BLK_STS_OK: | |
536167d4 | 2814 | queued++; |
27e8b2bb KS |
2815 | break; |
2816 | case BLK_STS_RESOURCE: | |
2817 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2818 | blk_mq_request_bypass_insert(rq, 0); |
2394395c CH |
2819 | if (list_empty(list)) |
2820 | blk_mq_run_hw_queue(hctx, false); | |
27e8b2bb KS |
2821 | goto out; |
2822 | default: | |
2823 | blk_mq_end_request(rq, ret); | |
2824 | break; | |
2825 | } | |
6ce3dd6e | 2826 | } |
d666ba98 | 2827 | |
27e8b2bb | 2828 | out: |
984ce0a7 KS |
2829 | if (ret != BLK_STS_OK) |
2830 | blk_mq_commit_rqs(hctx, queued, false); | |
6ce3dd6e ML |
2831 | } |
2832 | ||
b131f201 | 2833 | static bool blk_mq_attempt_bio_merge(struct request_queue *q, |
0c5bcc92 | 2834 | struct bio *bio, unsigned int nr_segs) |
900e0807 JA |
2835 | { |
2836 | if (!blk_queue_nomerges(q) && bio_mergeable(bio)) { | |
0c5bcc92 | 2837 | if (blk_attempt_plug_merge(q, bio, nr_segs)) |
900e0807 JA |
2838 | return true; |
2839 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) | |
2840 | return true; | |
2841 | } | |
2842 | return false; | |
2843 | } | |
2844 | ||
71539717 JA |
2845 | static struct request *blk_mq_get_new_requests(struct request_queue *q, |
2846 | struct blk_plug *plug, | |
0a5aa8d1 SK |
2847 | struct bio *bio, |
2848 | unsigned int nsegs) | |
71539717 JA |
2849 | { |
2850 | struct blk_mq_alloc_data data = { | |
2851 | .q = q, | |
2852 | .nr_tags = 1, | |
9d497e29 | 2853 | .cmd_flags = bio->bi_opf, |
71539717 JA |
2854 | }; |
2855 | struct request *rq; | |
2856 | ||
0a5aa8d1 SK |
2857 | rq_qos_throttle(q, bio); |
2858 | ||
71539717 JA |
2859 | if (plug) { |
2860 | data.nr_tags = plug->nr_ios; | |
2861 | plug->nr_ios = 1; | |
2862 | data.cached_rq = &plug->cached_rq; | |
2863 | } | |
2864 | ||
2865 | rq = __blk_mq_alloc_requests(&data); | |
373b5416 JA |
2866 | if (rq) |
2867 | return rq; | |
71539717 JA |
2868 | rq_qos_cleanup(q, bio); |
2869 | if (bio->bi_opf & REQ_NOWAIT) | |
2870 | bio_wouldblock_error(bio); | |
2871 | return NULL; | |
2872 | } | |
2873 | ||
309ce674 | 2874 | /* |
337e89fe | 2875 | * Check if there is a suitable cached request and return it. |
309ce674 | 2876 | */ |
337e89fe CH |
2877 | static struct request *blk_mq_peek_cached_request(struct blk_plug *plug, |
2878 | struct request_queue *q, blk_opf_t opf) | |
71539717 | 2879 | { |
337e89fe CH |
2880 | enum hctx_type type = blk_mq_get_hctx_type(opf); |
2881 | struct request *rq; | |
71539717 | 2882 | |
337e89fe CH |
2883 | if (!plug) |
2884 | return NULL; | |
2885 | rq = rq_list_peek(&plug->cached_rq); | |
2886 | if (!rq || rq->q != q) | |
2887 | return NULL; | |
2888 | if (type != rq->mq_hctx->type && | |
2889 | (type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT)) | |
2890 | return NULL; | |
2891 | if (op_is_flush(rq->cmd_flags) != op_is_flush(opf)) | |
2892 | return NULL; | |
2893 | return rq; | |
2894 | } | |
0a5aa8d1 | 2895 | |
337e89fe CH |
2896 | static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug, |
2897 | struct bio *bio) | |
2898 | { | |
2899 | WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq); | |
5b13bc8a | 2900 | |
2645672f JA |
2901 | /* |
2902 | * If any qos ->throttle() end up blocking, we will have flushed the | |
2903 | * plug and hence killed the cached_rq list as well. Pop this entry | |
2904 | * before we throttle. | |
2905 | */ | |
5b13bc8a | 2906 | plug->cached_rq = rq_list_next(rq); |
b0077e26 | 2907 | rq_qos_throttle(rq->q, bio); |
2645672f | 2908 | |
5c17f45e | 2909 | blk_mq_rq_time_init(rq, 0); |
b0077e26 | 2910 | rq->cmd_flags = bio->bi_opf; |
5b13bc8a | 2911 | INIT_LIST_HEAD(&rq->queuelist); |
71539717 JA |
2912 | } |
2913 | ||
105663f7 | 2914 | /** |
c62b37d9 | 2915 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2916 | * @bio: Bio pointer. |
2917 | * | |
2918 | * Builds up a request structure from @q and @bio and send to the device. The | |
2919 | * request may not be queued directly to hardware if: | |
2920 | * * This request can be merged with another one | |
2921 | * * We want to place request at plug queue for possible future merging | |
2922 | * * There is an IO scheduler active at this queue | |
2923 | * | |
2924 | * It will not queue the request if there is an error with the bio, or at the | |
2925 | * request creation. | |
105663f7 | 2926 | */ |
3e08773c | 2927 | void blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2928 | { |
ed6cddef | 2929 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
99a9476b | 2930 | struct blk_plug *plug = current->plug; |
ef295ecf | 2931 | const int is_sync = op_is_sync(bio->bi_opf); |
f0dbe6e8 | 2932 | struct blk_mq_hw_ctx *hctx; |
abd45c15 | 2933 | unsigned int nr_segs = 1; |
72e84e90 | 2934 | struct request *rq; |
a892c8d5 | 2935 | blk_status_t ret; |
07068d5b | 2936 | |
dd291d77 DLM |
2937 | /* |
2938 | * If the plug has a cached request for this queue, try to use it. | |
2939 | */ | |
2940 | rq = blk_mq_peek_cached_request(plug, q, bio->bi_opf); | |
2941 | ||
2942 | /* | |
2943 | * A BIO that was released from a zone write plug has already been | |
2944 | * through the preparation in this function, already holds a reference | |
2945 | * on the queue usage counter, and is the only write BIO in-flight for | |
2946 | * the target zone. Go straight to preparing a request for it. | |
2947 | */ | |
2948 | if (bio_zone_write_plugging(bio)) { | |
2949 | nr_segs = bio->__bi_nr_segments; | |
2950 | if (rq) | |
2951 | blk_queue_exit(q); | |
2952 | goto new_request; | |
2953 | } | |
2954 | ||
51d798cd | 2955 | bio = blk_queue_bounce(bio, q); |
9c6227e0 | 2956 | |
72e84e90 | 2957 | /* |
72e84e90 CH |
2958 | * The cached request already holds a q_usage_counter reference and we |
2959 | * don't have to acquire a new one if we use it. | |
2960 | */ | |
72e84e90 | 2961 | if (!rq) { |
b0077e26 | 2962 | if (unlikely(bio_queue_enter(bio))) |
5b13bc8a | 2963 | return; |
b0077e26 CH |
2964 | } |
2965 | ||
337e89fe CH |
2966 | if (unlikely(bio_may_exceed_limits(bio, &q->limits))) { |
2967 | bio = __bio_split_to_limits(bio, &q->limits, &nr_segs); | |
2968 | if (!bio) | |
0f299da5 | 2969 | goto queue_exit; |
b0077e26 | 2970 | } |
337e89fe CH |
2971 | if (!bio_integrity_prep(bio)) |
2972 | goto queue_exit; | |
b0077e26 | 2973 | |
0f299da5 CH |
2974 | if (blk_mq_attempt_bio_merge(q, bio, nr_segs)) |
2975 | goto queue_exit; | |
2976 | ||
dd291d77 DLM |
2977 | if (blk_queue_is_zoned(q) && blk_zone_plug_bio(bio, nr_segs)) |
2978 | goto queue_exit; | |
2979 | ||
2980 | new_request: | |
72e84e90 CH |
2981 | if (!rq) { |
2982 | rq = blk_mq_get_new_requests(q, plug, bio, nr_segs); | |
2983 | if (unlikely(!rq)) | |
2984 | goto queue_exit; | |
2985 | } else { | |
2986 | blk_mq_use_cached_rq(rq, plug, bio); | |
5b13bc8a | 2987 | } |
87760e5e | 2988 | |
e8a676d6 | 2989 | trace_block_getrq(bio); |
d6f1dda2 | 2990 | |
c1c80384 | 2991 | rq_qos_track(q, rq, bio); |
07068d5b | 2992 | |
970d168d BVA |
2993 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2994 | ||
9cd1e566 | 2995 | ret = blk_crypto_rq_get_keyslot(rq); |
a892c8d5 ST |
2996 | if (ret != BLK_STS_OK) { |
2997 | bio->bi_status = ret; | |
2998 | bio_endio(bio); | |
dd291d77 | 2999 | blk_zone_complete_request(rq); |
a892c8d5 | 3000 | blk_mq_free_request(rq); |
3e08773c | 3001 | return; |
a892c8d5 ST |
3002 | } |
3003 | ||
096bc7ea DLM |
3004 | if (bio_zone_write_plugging(bio)) |
3005 | blk_zone_write_plug_init_request(rq); | |
3006 | ||
360f2648 | 3007 | if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq)) |
d92ca9d8 CH |
3008 | return; |
3009 | ||
f0dbe6e8 | 3010 | if (plug) { |
ce5b009c | 3011 | blk_add_rq_to_plug(plug, rq); |
f0dbe6e8 CH |
3012 | return; |
3013 | } | |
3014 | ||
3015 | hctx = rq->mq_hctx; | |
dd6216bb | 3016 | if ((rq->rq_flags & RQF_USE_SCHED) || |
f0dbe6e8 | 3017 | (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) { |
710fa378 | 3018 | blk_mq_insert_request(rq, 0); |
f0dbe6e8 CH |
3019 | blk_mq_run_hw_queue(hctx, true); |
3020 | } else { | |
3021 | blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq)); | |
3022 | } | |
0f299da5 CH |
3023 | return; |
3024 | ||
3025 | queue_exit: | |
72e84e90 CH |
3026 | /* |
3027 | * Don't drop the queue reference if we were trying to use a cached | |
3028 | * request and thus didn't acquire one. | |
3029 | */ | |
3030 | if (!rq) | |
3031 | blk_queue_exit(q); | |
320ae51f JA |
3032 | } |
3033 | ||
248c7933 | 3034 | #ifdef CONFIG_BLK_MQ_STACKING |
06c8c691 | 3035 | /** |
a5efda3c | 3036 | * blk_insert_cloned_request - Helper for stacking drivers to submit a request |
a5efda3c | 3037 | * @rq: the request being queued |
06c8c691 | 3038 | */ |
28db4711 | 3039 | blk_status_t blk_insert_cloned_request(struct request *rq) |
06c8c691 | 3040 | { |
28db4711 | 3041 | struct request_queue *q = rq->q; |
06c8c691 | 3042 | unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); |
49d24398 | 3043 | unsigned int max_segments = blk_rq_get_max_segments(rq); |
a5efda3c | 3044 | blk_status_t ret; |
06c8c691 CH |
3045 | |
3046 | if (blk_rq_sectors(rq) > max_sectors) { | |
3047 | /* | |
3048 | * SCSI device does not have a good way to return if | |
3049 | * Write Same/Zero is actually supported. If a device rejects | |
3050 | * a non-read/write command (discard, write same,etc.) the | |
3051 | * low-level device driver will set the relevant queue limit to | |
3052 | * 0 to prevent blk-lib from issuing more of the offending | |
3053 | * operations. Commands queued prior to the queue limit being | |
3054 | * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O | |
3055 | * errors being propagated to upper layers. | |
3056 | */ | |
3057 | if (max_sectors == 0) | |
3058 | return BLK_STS_NOTSUPP; | |
3059 | ||
3060 | printk(KERN_ERR "%s: over max size limit. (%u > %u)\n", | |
3061 | __func__, blk_rq_sectors(rq), max_sectors); | |
3062 | return BLK_STS_IOERR; | |
3063 | } | |
3064 | ||
3065 | /* | |
3066 | * The queue settings related to segment counting may differ from the | |
3067 | * original queue. | |
3068 | */ | |
3069 | rq->nr_phys_segments = blk_recalc_rq_segments(rq); | |
49d24398 US |
3070 | if (rq->nr_phys_segments > max_segments) { |
3071 | printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n", | |
3072 | __func__, rq->nr_phys_segments, max_segments); | |
06c8c691 CH |
3073 | return BLK_STS_IOERR; |
3074 | } | |
3075 | ||
28db4711 | 3076 | if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq))) |
06c8c691 CH |
3077 | return BLK_STS_IOERR; |
3078 | ||
5b8562f0 EB |
3079 | ret = blk_crypto_rq_get_keyslot(rq); |
3080 | if (ret != BLK_STS_OK) | |
3081 | return ret; | |
06c8c691 CH |
3082 | |
3083 | blk_account_io_start(rq); | |
3084 | ||
3085 | /* | |
3086 | * Since we have a scheduler attached on the top device, | |
3087 | * bypass a potential scheduler on the bottom device for | |
3088 | * insert. | |
3089 | */ | |
28db4711 | 3090 | blk_mq_run_dispatch_ops(q, |
4cafe86c | 3091 | ret = blk_mq_request_issue_directly(rq, true)); |
592ee119 | 3092 | if (ret) |
08420cf7 | 3093 | blk_account_io_done(rq, blk_time_get_ns()); |
4cafe86c | 3094 | return ret; |
06c8c691 CH |
3095 | } |
3096 | EXPORT_SYMBOL_GPL(blk_insert_cloned_request); | |
3097 | ||
3098 | /** | |
3099 | * blk_rq_unprep_clone - Helper function to free all bios in a cloned request | |
3100 | * @rq: the clone request to be cleaned up | |
3101 | * | |
3102 | * Description: | |
3103 | * Free all bios in @rq for a cloned request. | |
3104 | */ | |
3105 | void blk_rq_unprep_clone(struct request *rq) | |
3106 | { | |
3107 | struct bio *bio; | |
3108 | ||
3109 | while ((bio = rq->bio) != NULL) { | |
3110 | rq->bio = bio->bi_next; | |
3111 | ||
3112 | bio_put(bio); | |
3113 | } | |
3114 | } | |
3115 | EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); | |
3116 | ||
3117 | /** | |
3118 | * blk_rq_prep_clone - Helper function to setup clone request | |
3119 | * @rq: the request to be setup | |
3120 | * @rq_src: original request to be cloned | |
3121 | * @bs: bio_set that bios for clone are allocated from | |
3122 | * @gfp_mask: memory allocation mask for bio | |
3123 | * @bio_ctr: setup function to be called for each clone bio. | |
3124 | * Returns %0 for success, non %0 for failure. | |
3125 | * @data: private data to be passed to @bio_ctr | |
3126 | * | |
3127 | * Description: | |
3128 | * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. | |
3129 | * Also, pages which the original bios are pointing to are not copied | |
3130 | * and the cloned bios just point same pages. | |
3131 | * So cloned bios must be completed before original bios, which means | |
3132 | * the caller must complete @rq before @rq_src. | |
3133 | */ | |
3134 | int blk_rq_prep_clone(struct request *rq, struct request *rq_src, | |
3135 | struct bio_set *bs, gfp_t gfp_mask, | |
3136 | int (*bio_ctr)(struct bio *, struct bio *, void *), | |
3137 | void *data) | |
3138 | { | |
3139 | struct bio *bio, *bio_src; | |
3140 | ||
3141 | if (!bs) | |
3142 | bs = &fs_bio_set; | |
3143 | ||
3144 | __rq_for_each_bio(bio_src, rq_src) { | |
abfc426d CH |
3145 | bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask, |
3146 | bs); | |
06c8c691 CH |
3147 | if (!bio) |
3148 | goto free_and_out; | |
3149 | ||
3150 | if (bio_ctr && bio_ctr(bio, bio_src, data)) | |
3151 | goto free_and_out; | |
3152 | ||
3153 | if (rq->bio) { | |
3154 | rq->biotail->bi_next = bio; | |
3155 | rq->biotail = bio; | |
3156 | } else { | |
3157 | rq->bio = rq->biotail = bio; | |
3158 | } | |
3159 | bio = NULL; | |
3160 | } | |
3161 | ||
3162 | /* Copy attributes of the original request to the clone request. */ | |
3163 | rq->__sector = blk_rq_pos(rq_src); | |
3164 | rq->__data_len = blk_rq_bytes(rq_src); | |
3165 | if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) { | |
3166 | rq->rq_flags |= RQF_SPECIAL_PAYLOAD; | |
3167 | rq->special_vec = rq_src->special_vec; | |
3168 | } | |
3169 | rq->nr_phys_segments = rq_src->nr_phys_segments; | |
3170 | rq->ioprio = rq_src->ioprio; | |
44981351 | 3171 | rq->write_hint = rq_src->write_hint; |
06c8c691 CH |
3172 | |
3173 | if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0) | |
3174 | goto free_and_out; | |
3175 | ||
3176 | return 0; | |
3177 | ||
3178 | free_and_out: | |
3179 | if (bio) | |
3180 | bio_put(bio); | |
3181 | blk_rq_unprep_clone(rq); | |
3182 | ||
3183 | return -ENOMEM; | |
3184 | } | |
3185 | EXPORT_SYMBOL_GPL(blk_rq_prep_clone); | |
248c7933 | 3186 | #endif /* CONFIG_BLK_MQ_STACKING */ |
06c8c691 | 3187 | |
f2b8f3ce CH |
3188 | /* |
3189 | * Steal bios from a request and add them to a bio list. | |
3190 | * The request must not have been partially completed before. | |
3191 | */ | |
3192 | void blk_steal_bios(struct bio_list *list, struct request *rq) | |
3193 | { | |
3194 | if (rq->bio) { | |
3195 | if (list->tail) | |
3196 | list->tail->bi_next = rq->bio; | |
3197 | else | |
3198 | list->head = rq->bio; | |
3199 | list->tail = rq->biotail; | |
3200 | ||
3201 | rq->bio = NULL; | |
3202 | rq->biotail = NULL; | |
3203 | } | |
3204 | ||
3205 | rq->__data_len = 0; | |
3206 | } | |
3207 | EXPORT_SYMBOL_GPL(blk_steal_bios); | |
3208 | ||
bd63141d ML |
3209 | static size_t order_to_size(unsigned int order) |
3210 | { | |
3211 | return (size_t)PAGE_SIZE << order; | |
3212 | } | |
3213 | ||
3214 | /* called before freeing request pool in @tags */ | |
f32e4eaf JG |
3215 | static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, |
3216 | struct blk_mq_tags *tags) | |
bd63141d | 3217 | { |
bd63141d ML |
3218 | struct page *page; |
3219 | unsigned long flags; | |
3220 | ||
76dd2980 YK |
3221 | /* |
3222 | * There is no need to clear mapping if driver tags is not initialized | |
3223 | * or the mapping belongs to the driver tags. | |
3224 | */ | |
3225 | if (!drv_tags || drv_tags == tags) | |
4f245d5b JG |
3226 | return; |
3227 | ||
bd63141d ML |
3228 | list_for_each_entry(page, &tags->page_list, lru) { |
3229 | unsigned long start = (unsigned long)page_address(page); | |
3230 | unsigned long end = start + order_to_size(page->private); | |
3231 | int i; | |
3232 | ||
f32e4eaf | 3233 | for (i = 0; i < drv_tags->nr_tags; i++) { |
bd63141d ML |
3234 | struct request *rq = drv_tags->rqs[i]; |
3235 | unsigned long rq_addr = (unsigned long)rq; | |
3236 | ||
3237 | if (rq_addr >= start && rq_addr < end) { | |
0a467d0f | 3238 | WARN_ON_ONCE(req_ref_read(rq) != 0); |
bd63141d ML |
3239 | cmpxchg(&drv_tags->rqs[i], rq, NULL); |
3240 | } | |
3241 | } | |
3242 | } | |
3243 | ||
3244 | /* | |
3245 | * Wait until all pending iteration is done. | |
3246 | * | |
3247 | * Request reference is cleared and it is guaranteed to be observed | |
3248 | * after the ->lock is released. | |
3249 | */ | |
3250 | spin_lock_irqsave(&drv_tags->lock, flags); | |
3251 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
3252 | } | |
3253 | ||
cc71a6f4 JA |
3254 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
3255 | unsigned int hctx_idx) | |
95363efd | 3256 | { |
f32e4eaf | 3257 | struct blk_mq_tags *drv_tags; |
e9b267d9 | 3258 | struct page *page; |
320ae51f | 3259 | |
e02657ea ML |
3260 | if (list_empty(&tags->page_list)) |
3261 | return; | |
3262 | ||
079a2e3e JG |
3263 | if (blk_mq_is_shared_tags(set->flags)) |
3264 | drv_tags = set->shared_tags; | |
e155b0c2 JG |
3265 | else |
3266 | drv_tags = set->tags[hctx_idx]; | |
f32e4eaf | 3267 | |
65de57bb | 3268 | if (tags->static_rqs && set->ops->exit_request) { |
e9b267d9 | 3269 | int i; |
320ae51f | 3270 | |
24d2f903 | 3271 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
3272 | struct request *rq = tags->static_rqs[i]; |
3273 | ||
3274 | if (!rq) | |
e9b267d9 | 3275 | continue; |
d6296d39 | 3276 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 3277 | tags->static_rqs[i] = NULL; |
e9b267d9 | 3278 | } |
320ae51f | 3279 | } |
320ae51f | 3280 | |
f32e4eaf | 3281 | blk_mq_clear_rq_mapping(drv_tags, tags); |
bd63141d | 3282 | |
24d2f903 CH |
3283 | while (!list_empty(&tags->page_list)) { |
3284 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 3285 | list_del_init(&page->lru); |
f75782e4 CM |
3286 | /* |
3287 | * Remove kmemleak object previously allocated in | |
273938bf | 3288 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
3289 | */ |
3290 | kmemleak_free(page_address(page)); | |
320ae51f JA |
3291 | __free_pages(page, page->private); |
3292 | } | |
cc71a6f4 | 3293 | } |
320ae51f | 3294 | |
e155b0c2 | 3295 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
cc71a6f4 | 3296 | { |
24d2f903 | 3297 | kfree(tags->rqs); |
cc71a6f4 | 3298 | tags->rqs = NULL; |
2af8cbe3 JA |
3299 | kfree(tags->static_rqs); |
3300 | tags->static_rqs = NULL; | |
320ae51f | 3301 | |
e155b0c2 | 3302 | blk_mq_free_tags(tags); |
320ae51f JA |
3303 | } |
3304 | ||
4d805131 ML |
3305 | static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set, |
3306 | unsigned int hctx_idx) | |
3307 | { | |
3308 | int i; | |
3309 | ||
3310 | for (i = 0; i < set->nr_maps; i++) { | |
3311 | unsigned int start = set->map[i].queue_offset; | |
3312 | unsigned int end = start + set->map[i].nr_queues; | |
3313 | ||
3314 | if (hctx_idx >= start && hctx_idx < end) | |
3315 | break; | |
3316 | } | |
3317 | ||
3318 | if (i >= set->nr_maps) | |
3319 | i = HCTX_TYPE_DEFAULT; | |
3320 | ||
3321 | return i; | |
3322 | } | |
3323 | ||
3324 | static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set, | |
3325 | unsigned int hctx_idx) | |
3326 | { | |
3327 | enum hctx_type type = hctx_idx_to_type(set, hctx_idx); | |
3328 | ||
3329 | return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx); | |
3330 | } | |
3331 | ||
63064be1 JG |
3332 | static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
3333 | unsigned int hctx_idx, | |
3334 | unsigned int nr_tags, | |
e155b0c2 | 3335 | unsigned int reserved_tags) |
320ae51f | 3336 | { |
4d805131 | 3337 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
24d2f903 | 3338 | struct blk_mq_tags *tags; |
320ae51f | 3339 | |
59f082e4 SL |
3340 | if (node == NUMA_NO_NODE) |
3341 | node = set->numa_node; | |
3342 | ||
e155b0c2 JG |
3343 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
3344 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
3345 | if (!tags) |
3346 | return NULL; | |
320ae51f | 3347 | |
590b5b7d | 3348 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 3349 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 3350 | node); |
7edfd681 JC |
3351 | if (!tags->rqs) |
3352 | goto err_free_tags; | |
320ae51f | 3353 | |
590b5b7d KC |
3354 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
3355 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
3356 | node); | |
7edfd681 JC |
3357 | if (!tags->static_rqs) |
3358 | goto err_free_rqs; | |
2af8cbe3 | 3359 | |
cc71a6f4 | 3360 | return tags; |
7edfd681 JC |
3361 | |
3362 | err_free_rqs: | |
3363 | kfree(tags->rqs); | |
3364 | err_free_tags: | |
3365 | blk_mq_free_tags(tags); | |
3366 | return NULL; | |
cc71a6f4 JA |
3367 | } |
3368 | ||
1d9bd516 TH |
3369 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
3370 | unsigned int hctx_idx, int node) | |
3371 | { | |
3372 | int ret; | |
3373 | ||
3374 | if (set->ops->init_request) { | |
3375 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
3376 | if (ret) | |
3377 | return ret; | |
3378 | } | |
3379 | ||
12f5b931 | 3380 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
3381 | return 0; |
3382 | } | |
3383 | ||
63064be1 JG |
3384 | static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, |
3385 | struct blk_mq_tags *tags, | |
3386 | unsigned int hctx_idx, unsigned int depth) | |
cc71a6f4 JA |
3387 | { |
3388 | unsigned int i, j, entries_per_page, max_order = 4; | |
4d805131 | 3389 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
cc71a6f4 | 3390 | size_t rq_size, left; |
59f082e4 | 3391 | |
59f082e4 SL |
3392 | if (node == NUMA_NO_NODE) |
3393 | node = set->numa_node; | |
cc71a6f4 JA |
3394 | |
3395 | INIT_LIST_HEAD(&tags->page_list); | |
3396 | ||
320ae51f JA |
3397 | /* |
3398 | * rq_size is the size of the request plus driver payload, rounded | |
3399 | * to the cacheline size | |
3400 | */ | |
24d2f903 | 3401 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 3402 | cache_line_size()); |
cc71a6f4 | 3403 | left = rq_size * depth; |
320ae51f | 3404 | |
cc71a6f4 | 3405 | for (i = 0; i < depth; ) { |
320ae51f JA |
3406 | int this_order = max_order; |
3407 | struct page *page; | |
3408 | int to_do; | |
3409 | void *p; | |
3410 | ||
b3a834b1 | 3411 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
3412 | this_order--; |
3413 | ||
3414 | do { | |
59f082e4 | 3415 | page = alloc_pages_node(node, |
36e1f3d1 | 3416 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 3417 | this_order); |
320ae51f JA |
3418 | if (page) |
3419 | break; | |
3420 | if (!this_order--) | |
3421 | break; | |
3422 | if (order_to_size(this_order) < rq_size) | |
3423 | break; | |
3424 | } while (1); | |
3425 | ||
3426 | if (!page) | |
24d2f903 | 3427 | goto fail; |
320ae51f JA |
3428 | |
3429 | page->private = this_order; | |
24d2f903 | 3430 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
3431 | |
3432 | p = page_address(page); | |
f75782e4 CM |
3433 | /* |
3434 | * Allow kmemleak to scan these pages as they contain pointers | |
3435 | * to additional allocations like via ops->init_request(). | |
3436 | */ | |
36e1f3d1 | 3437 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 3438 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 3439 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
3440 | left -= to_do * rq_size; |
3441 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
3442 | struct request *rq = p; |
3443 | ||
3444 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
3445 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
3446 | tags->static_rqs[i] = NULL; | |
3447 | goto fail; | |
e9b267d9 CH |
3448 | } |
3449 | ||
320ae51f JA |
3450 | p += rq_size; |
3451 | i++; | |
3452 | } | |
3453 | } | |
cc71a6f4 | 3454 | return 0; |
320ae51f | 3455 | |
24d2f903 | 3456 | fail: |
cc71a6f4 JA |
3457 | blk_mq_free_rqs(set, tags, hctx_idx); |
3458 | return -ENOMEM; | |
320ae51f JA |
3459 | } |
3460 | ||
bf0beec0 ML |
3461 | struct rq_iter_data { |
3462 | struct blk_mq_hw_ctx *hctx; | |
3463 | bool has_rq; | |
3464 | }; | |
3465 | ||
2dd6532e | 3466 | static bool blk_mq_has_request(struct request *rq, void *data) |
bf0beec0 ML |
3467 | { |
3468 | struct rq_iter_data *iter_data = data; | |
3469 | ||
3470 | if (rq->mq_hctx != iter_data->hctx) | |
3471 | return true; | |
3472 | iter_data->has_rq = true; | |
3473 | return false; | |
3474 | } | |
3475 | ||
3476 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
3477 | { | |
3478 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
3479 | hctx->sched_tags : hctx->tags; | |
3480 | struct rq_iter_data data = { | |
3481 | .hctx = hctx, | |
3482 | }; | |
3483 | ||
3484 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
3485 | return data.has_rq; | |
3486 | } | |
3487 | ||
a46c2702 ML |
3488 | static bool blk_mq_hctx_has_online_cpu(struct blk_mq_hw_ctx *hctx, |
3489 | unsigned int this_cpu) | |
bf0beec0 | 3490 | { |
a46c2702 ML |
3491 | enum hctx_type type = hctx->type; |
3492 | int cpu; | |
3493 | ||
3494 | /* | |
3495 | * hctx->cpumask has to rule out isolated CPUs, but userspace still | |
3496 | * might submit IOs on these isolated CPUs, so use the queue map to | |
3497 | * check if all CPUs mapped to this hctx are offline | |
3498 | */ | |
3499 | for_each_online_cpu(cpu) { | |
3500 | struct blk_mq_hw_ctx *h = blk_mq_map_queue_type(hctx->queue, | |
3501 | type, cpu); | |
3502 | ||
3503 | if (h != hctx) | |
3504 | continue; | |
3505 | ||
3506 | /* this hctx has at least one online CPU */ | |
3507 | if (this_cpu != cpu) | |
3508 | return true; | |
3509 | } | |
3510 | ||
3511 | return false; | |
bf0beec0 ML |
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 | ||
a46c2702 | 3519 | if (blk_mq_hctx_has_online_cpu(hctx, cpu)) |
bf0beec0 ML |
3520 | return 0; |
3521 | ||
3522 | /* | |
3523 | * Prevent new request from being allocated on the current hctx. | |
3524 | * | |
3525 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
3526 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
3527 | * seen once we return from the tag allocator. | |
3528 | */ | |
3529 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3530 | smp_mb__after_atomic(); | |
3531 | ||
3532 | /* | |
3533 | * Try to grab a reference to the queue and wait for any outstanding | |
3534 | * requests. If we could not grab a reference the queue has been | |
3535 | * frozen and there are no requests. | |
3536 | */ | |
3537 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
3538 | while (blk_mq_hctx_has_requests(hctx)) | |
3539 | msleep(5); | |
3540 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
3541 | } | |
3542 | ||
3543 | return 0; | |
3544 | } | |
3545 | ||
3546 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
3547 | { | |
3548 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3549 | struct blk_mq_hw_ctx, cpuhp_online); | |
3550 | ||
3551 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
3552 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3553 | return 0; | |
3554 | } | |
3555 | ||
e57690fe JA |
3556 | /* |
3557 | * 'cpu' is going away. splice any existing rq_list entries from this | |
3558 | * software queue to the hw queue dispatch list, and ensure that it | |
3559 | * gets run. | |
3560 | */ | |
9467f859 | 3561 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 3562 | { |
9467f859 | 3563 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
3564 | struct blk_mq_ctx *ctx; |
3565 | LIST_HEAD(tmp); | |
c16d6b5a | 3566 | enum hctx_type type; |
484b4061 | 3567 | |
9467f859 | 3568 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
3569 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
3570 | return 0; | |
3571 | ||
e57690fe | 3572 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 3573 | type = hctx->type; |
484b4061 JA |
3574 | |
3575 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
3576 | if (!list_empty(&ctx->rq_lists[type])) { |
3577 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
3578 | blk_mq_hctx_clear_pending(hctx, ctx); |
3579 | } | |
3580 | spin_unlock(&ctx->lock); | |
3581 | ||
3582 | if (list_empty(&tmp)) | |
9467f859 | 3583 | return 0; |
484b4061 | 3584 | |
e57690fe JA |
3585 | spin_lock(&hctx->lock); |
3586 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
3587 | spin_unlock(&hctx->lock); | |
484b4061 JA |
3588 | |
3589 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 3590 | return 0; |
484b4061 JA |
3591 | } |
3592 | ||
9467f859 | 3593 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 3594 | { |
bf0beec0 ML |
3595 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3596 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3597 | &hctx->cpuhp_online); | |
9467f859 TG |
3598 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
3599 | &hctx->cpuhp_dead); | |
484b4061 JA |
3600 | } |
3601 | ||
364b6181 ML |
3602 | /* |
3603 | * Before freeing hw queue, clearing the flush request reference in | |
3604 | * tags->rqs[] for avoiding potential UAF. | |
3605 | */ | |
3606 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
3607 | unsigned int queue_depth, struct request *flush_rq) | |
3608 | { | |
3609 | int i; | |
3610 | unsigned long flags; | |
3611 | ||
3612 | /* The hw queue may not be mapped yet */ | |
3613 | if (!tags) | |
3614 | return; | |
3615 | ||
0a467d0f | 3616 | WARN_ON_ONCE(req_ref_read(flush_rq) != 0); |
364b6181 ML |
3617 | |
3618 | for (i = 0; i < queue_depth; i++) | |
3619 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
3620 | ||
3621 | /* | |
3622 | * Wait until all pending iteration is done. | |
3623 | * | |
3624 | * Request reference is cleared and it is guaranteed to be observed | |
3625 | * after the ->lock is released. | |
3626 | */ | |
3627 | spin_lock_irqsave(&tags->lock, flags); | |
3628 | spin_unlock_irqrestore(&tags->lock, flags); | |
3629 | } | |
3630 | ||
c3b4afca | 3631 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
3632 | static void blk_mq_exit_hctx(struct request_queue *q, |
3633 | struct blk_mq_tag_set *set, | |
3634 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
3635 | { | |
364b6181 ML |
3636 | struct request *flush_rq = hctx->fq->flush_rq; |
3637 | ||
8ab0b7dc ML |
3638 | if (blk_mq_hw_queue_mapped(hctx)) |
3639 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 3640 | |
6cfeadbf ML |
3641 | if (blk_queue_init_done(q)) |
3642 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], | |
3643 | set->queue_depth, flush_rq); | |
f70ced09 | 3644 | if (set->ops->exit_request) |
364b6181 | 3645 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 3646 | |
08e98fc6 ML |
3647 | if (set->ops->exit_hctx) |
3648 | set->ops->exit_hctx(hctx, hctx_idx); | |
3649 | ||
9467f859 | 3650 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 | 3651 | |
4e5cc99e ML |
3652 | xa_erase(&q->hctx_table, hctx_idx); |
3653 | ||
2f8f1336 ML |
3654 | spin_lock(&q->unused_hctx_lock); |
3655 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
3656 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
3657 | } |
3658 | ||
624dbe47 ML |
3659 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
3660 | struct blk_mq_tag_set *set, int nr_queue) | |
3661 | { | |
3662 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3663 | unsigned long i; |
624dbe47 ML |
3664 | |
3665 | queue_for_each_hw_ctx(q, hctx, i) { | |
3666 | if (i == nr_queue) | |
3667 | break; | |
08e98fc6 | 3668 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 3669 | } |
624dbe47 ML |
3670 | } |
3671 | ||
08e98fc6 ML |
3672 | static int blk_mq_init_hctx(struct request_queue *q, |
3673 | struct blk_mq_tag_set *set, | |
3674 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 3675 | { |
7c6c5b7c ML |
3676 | hctx->queue_num = hctx_idx; |
3677 | ||
bf0beec0 ML |
3678 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3679 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3680 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
3681 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
3682 | ||
3683 | hctx->tags = set->tags[hctx_idx]; | |
3684 | ||
3685 | if (set->ops->init_hctx && | |
3686 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
3687 | goto unregister_cpu_notifier; | |
08e98fc6 | 3688 | |
7c6c5b7c ML |
3689 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
3690 | hctx->numa_node)) | |
3691 | goto exit_hctx; | |
4e5cc99e ML |
3692 | |
3693 | if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL)) | |
3694 | goto exit_flush_rq; | |
3695 | ||
7c6c5b7c ML |
3696 | return 0; |
3697 | ||
4e5cc99e ML |
3698 | exit_flush_rq: |
3699 | if (set->ops->exit_request) | |
3700 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); | |
7c6c5b7c ML |
3701 | exit_hctx: |
3702 | if (set->ops->exit_hctx) | |
3703 | set->ops->exit_hctx(hctx, hctx_idx); | |
3704 | unregister_cpu_notifier: | |
3705 | blk_mq_remove_cpuhp(hctx); | |
3706 | return -1; | |
3707 | } | |
3708 | ||
3709 | static struct blk_mq_hw_ctx * | |
3710 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
3711 | int node) | |
3712 | { | |
3713 | struct blk_mq_hw_ctx *hctx; | |
3714 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
3715 | ||
704b914f | 3716 | hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node); |
7c6c5b7c ML |
3717 | if (!hctx) |
3718 | goto fail_alloc_hctx; | |
3719 | ||
3720 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
3721 | goto free_hctx; | |
3722 | ||
3723 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 3724 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
3725 | node = set->numa_node; |
3726 | hctx->numa_node = node; | |
08e98fc6 | 3727 | |
9f993737 | 3728 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
3729 | spin_lock_init(&hctx->lock); |
3730 | INIT_LIST_HEAD(&hctx->dispatch); | |
3731 | hctx->queue = q; | |
51db1c37 | 3732 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 3733 | |
2f8f1336 ML |
3734 | INIT_LIST_HEAD(&hctx->hctx_list); |
3735 | ||
320ae51f | 3736 | /* |
08e98fc6 ML |
3737 | * Allocate space for all possible cpus to avoid allocation at |
3738 | * runtime | |
320ae51f | 3739 | */ |
d904bfa7 | 3740 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 3741 | gfp, node); |
08e98fc6 | 3742 | if (!hctx->ctxs) |
7c6c5b7c | 3743 | goto free_cpumask; |
320ae51f | 3744 | |
5b202853 | 3745 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 3746 | gfp, node, false, false)) |
08e98fc6 | 3747 | goto free_ctxs; |
08e98fc6 | 3748 | hctx->nr_ctx = 0; |
320ae51f | 3749 | |
5815839b | 3750 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
3751 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
3752 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
3753 | ||
754a1572 | 3754 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 3755 | if (!hctx->fq) |
7c6c5b7c | 3756 | goto free_bitmap; |
320ae51f | 3757 | |
7c6c5b7c | 3758 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 3759 | |
7c6c5b7c | 3760 | return hctx; |
320ae51f | 3761 | |
08e98fc6 | 3762 | free_bitmap: |
88459642 | 3763 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
3764 | free_ctxs: |
3765 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
3766 | free_cpumask: |
3767 | free_cpumask_var(hctx->cpumask); | |
3768 | free_hctx: | |
3769 | kfree(hctx); | |
3770 | fail_alloc_hctx: | |
3771 | return NULL; | |
08e98fc6 | 3772 | } |
320ae51f | 3773 | |
320ae51f JA |
3774 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
3775 | unsigned int nr_hw_queues) | |
3776 | { | |
b3c661b1 JA |
3777 | struct blk_mq_tag_set *set = q->tag_set; |
3778 | unsigned int i, j; | |
320ae51f JA |
3779 | |
3780 | for_each_possible_cpu(i) { | |
3781 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
3782 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 3783 | int k; |
320ae51f | 3784 | |
320ae51f JA |
3785 | __ctx->cpu = i; |
3786 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
3787 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
3788 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
3789 | ||
320ae51f JA |
3790 | __ctx->queue = q; |
3791 | ||
320ae51f JA |
3792 | /* |
3793 | * Set local node, IFF we have more than one hw queue. If | |
3794 | * not, we remain on the home node of the device | |
3795 | */ | |
b3c661b1 JA |
3796 | for (j = 0; j < set->nr_maps; j++) { |
3797 | hctx = blk_mq_map_queue_type(q, j, i); | |
3798 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 3799 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 3800 | } |
320ae51f JA |
3801 | } |
3802 | } | |
3803 | ||
63064be1 JG |
3804 | struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3805 | unsigned int hctx_idx, | |
3806 | unsigned int depth) | |
cc71a6f4 | 3807 | { |
63064be1 JG |
3808 | struct blk_mq_tags *tags; |
3809 | int ret; | |
cc71a6f4 | 3810 | |
e155b0c2 | 3811 | tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); |
63064be1 JG |
3812 | if (!tags) |
3813 | return NULL; | |
cc71a6f4 | 3814 | |
63064be1 JG |
3815 | ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); |
3816 | if (ret) { | |
e155b0c2 | 3817 | blk_mq_free_rq_map(tags); |
63064be1 JG |
3818 | return NULL; |
3819 | } | |
cc71a6f4 | 3820 | |
63064be1 | 3821 | return tags; |
cc71a6f4 JA |
3822 | } |
3823 | ||
63064be1 JG |
3824 | static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3825 | int hctx_idx) | |
cc71a6f4 | 3826 | { |
079a2e3e JG |
3827 | if (blk_mq_is_shared_tags(set->flags)) { |
3828 | set->tags[hctx_idx] = set->shared_tags; | |
1c0706a7 | 3829 | |
e155b0c2 | 3830 | return true; |
bd166ef1 | 3831 | } |
e155b0c2 | 3832 | |
63064be1 JG |
3833 | set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, |
3834 | set->queue_depth); | |
3835 | ||
3836 | return set->tags[hctx_idx]; | |
cc71a6f4 JA |
3837 | } |
3838 | ||
645db34e JG |
3839 | void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3840 | struct blk_mq_tags *tags, | |
3841 | unsigned int hctx_idx) | |
cc71a6f4 | 3842 | { |
645db34e JG |
3843 | if (tags) { |
3844 | blk_mq_free_rqs(set, tags, hctx_idx); | |
e155b0c2 | 3845 | blk_mq_free_rq_map(tags); |
bd166ef1 | 3846 | } |
cc71a6f4 JA |
3847 | } |
3848 | ||
e155b0c2 JG |
3849 | static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3850 | unsigned int hctx_idx) | |
3851 | { | |
079a2e3e | 3852 | if (!blk_mq_is_shared_tags(set->flags)) |
e155b0c2 JG |
3853 | blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); |
3854 | ||
3855 | set->tags[hctx_idx] = NULL; | |
cc71a6f4 JA |
3856 | } |
3857 | ||
4b855ad3 | 3858 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 3859 | { |
4f481208 ML |
3860 | unsigned int j, hctx_idx; |
3861 | unsigned long i; | |
320ae51f JA |
3862 | struct blk_mq_hw_ctx *hctx; |
3863 | struct blk_mq_ctx *ctx; | |
2a34c087 | 3864 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
3865 | |
3866 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 3867 | cpumask_clear(hctx->cpumask); |
320ae51f | 3868 | hctx->nr_ctx = 0; |
d416c92c | 3869 | hctx->dispatch_from = NULL; |
320ae51f JA |
3870 | } |
3871 | ||
3872 | /* | |
4b855ad3 | 3873 | * Map software to hardware queues. |
4412efec ML |
3874 | * |
3875 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 3876 | */ |
20e4d813 | 3877 | for_each_possible_cpu(i) { |
4412efec | 3878 | |
897bb0c7 | 3879 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 3880 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
3881 | if (!set->map[j].nr_queues) { |
3882 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3883 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 3884 | continue; |
bb94aea1 | 3885 | } |
fd689871 ML |
3886 | hctx_idx = set->map[j].mq_map[i]; |
3887 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
3888 | if (!set->tags[hctx_idx] && | |
63064be1 | 3889 | !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { |
fd689871 ML |
3890 | /* |
3891 | * If tags initialization fail for some hctx, | |
3892 | * that hctx won't be brought online. In this | |
3893 | * case, remap the current ctx to hctx[0] which | |
3894 | * is guaranteed to always have tags allocated | |
3895 | */ | |
3896 | set->map[j].mq_map[i] = 0; | |
3897 | } | |
e5edd5f2 | 3898 | |
b3c661b1 | 3899 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 3900 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
3901 | /* |
3902 | * If the CPU is already set in the mask, then we've | |
3903 | * mapped this one already. This can happen if | |
3904 | * devices share queues across queue maps. | |
3905 | */ | |
3906 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
3907 | continue; | |
3908 | ||
3909 | cpumask_set_cpu(i, hctx->cpumask); | |
3910 | hctx->type = j; | |
3911 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
3912 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
3913 | ||
3914 | /* | |
3915 | * If the nr_ctx type overflows, we have exceeded the | |
3916 | * amount of sw queues we can support. | |
3917 | */ | |
3918 | BUG_ON(!hctx->nr_ctx); | |
3919 | } | |
bb94aea1 JW |
3920 | |
3921 | for (; j < HCTX_MAX_TYPES; j++) | |
3922 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3923 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 3924 | } |
506e931f JA |
3925 | |
3926 | queue_for_each_hw_ctx(q, hctx, i) { | |
a46c2702 ML |
3927 | int cpu; |
3928 | ||
4412efec ML |
3929 | /* |
3930 | * If no software queues are mapped to this hardware queue, | |
3931 | * disable it and free the request entries. | |
3932 | */ | |
3933 | if (!hctx->nr_ctx) { | |
3934 | /* Never unmap queue 0. We need it as a | |
3935 | * fallback in case of a new remap fails | |
3936 | * allocation | |
3937 | */ | |
e155b0c2 JG |
3938 | if (i) |
3939 | __blk_mq_free_map_and_rqs(set, i); | |
4412efec ML |
3940 | |
3941 | hctx->tags = NULL; | |
3942 | continue; | |
3943 | } | |
484b4061 | 3944 | |
2a34c087 ML |
3945 | hctx->tags = set->tags[i]; |
3946 | WARN_ON(!hctx->tags); | |
3947 | ||
889fa31f CY |
3948 | /* |
3949 | * Set the map size to the number of mapped software queues. | |
3950 | * This is more accurate and more efficient than looping | |
3951 | * over all possibly mapped software queues. | |
3952 | */ | |
88459642 | 3953 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 3954 | |
a46c2702 ML |
3955 | /* |
3956 | * Rule out isolated CPUs from hctx->cpumask to avoid | |
3957 | * running block kworker on isolated CPUs | |
3958 | */ | |
3959 | for_each_cpu(cpu, hctx->cpumask) { | |
3960 | if (cpu_is_isolated(cpu)) | |
3961 | cpumask_clear_cpu(cpu, hctx->cpumask); | |
3962 | } | |
3963 | ||
484b4061 JA |
3964 | /* |
3965 | * Initialize batch roundrobin counts | |
3966 | */ | |
f82ddf19 | 3967 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
3968 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
3969 | } | |
320ae51f JA |
3970 | } |
3971 | ||
8e8320c9 JA |
3972 | /* |
3973 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
3974 | * the queue isn't live yet. | |
3975 | */ | |
2404e607 | 3976 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
3977 | { |
3978 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3979 | unsigned long i; |
0d2602ca | 3980 | |
2404e607 | 3981 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 3982 | if (shared) { |
51db1c37 | 3983 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
3984 | } else { |
3985 | blk_mq_tag_idle(hctx); | |
51db1c37 | 3986 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 3987 | } |
2404e607 JM |
3988 | } |
3989 | } | |
3990 | ||
655ac300 HR |
3991 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
3992 | bool shared) | |
2404e607 JM |
3993 | { |
3994 | struct request_queue *q; | |
0d2602ca | 3995 | |
705cda97 BVA |
3996 | lockdep_assert_held(&set->tag_list_lock); |
3997 | ||
0d2602ca JA |
3998 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3999 | blk_mq_freeze_queue(q); | |
2404e607 | 4000 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
4001 | blk_mq_unfreeze_queue(q); |
4002 | } | |
4003 | } | |
4004 | ||
4005 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
4006 | { | |
4007 | struct blk_mq_tag_set *set = q->tag_set; | |
4008 | ||
0d2602ca | 4009 | mutex_lock(&set->tag_list_lock); |
08c875cb | 4010 | list_del(&q->tag_set_list); |
2404e607 JM |
4011 | if (list_is_singular(&set->tag_list)) { |
4012 | /* just transitioned to unshared */ | |
51db1c37 | 4013 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 4014 | /* update existing queue */ |
655ac300 | 4015 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 4016 | } |
0d2602ca | 4017 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 4018 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
4019 | } |
4020 | ||
4021 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
4022 | struct request_queue *q) | |
4023 | { | |
0d2602ca | 4024 | mutex_lock(&set->tag_list_lock); |
2404e607 | 4025 | |
ff821d27 JA |
4026 | /* |
4027 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
4028 | */ | |
4029 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
4030 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
4031 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 4032 | /* update existing queue */ |
655ac300 | 4033 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 4034 | } |
51db1c37 | 4035 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 4036 | queue_set_hctx_shared(q, true); |
08c875cb | 4037 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 4038 | |
0d2602ca JA |
4039 | mutex_unlock(&set->tag_list_lock); |
4040 | } | |
4041 | ||
1db4909e ML |
4042 | /* All allocations will be freed in release handler of q->mq_kobj */ |
4043 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
4044 | { | |
4045 | struct blk_mq_ctxs *ctxs; | |
4046 | int cpu; | |
4047 | ||
4048 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
4049 | if (!ctxs) | |
4050 | return -ENOMEM; | |
4051 | ||
4052 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
4053 | if (!ctxs->queue_ctx) | |
4054 | goto fail; | |
4055 | ||
4056 | for_each_possible_cpu(cpu) { | |
4057 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
4058 | ctx->ctxs = ctxs; | |
4059 | } | |
4060 | ||
4061 | q->mq_kobj = &ctxs->kobj; | |
4062 | q->queue_ctx = ctxs->queue_ctx; | |
4063 | ||
4064 | return 0; | |
4065 | fail: | |
4066 | kfree(ctxs); | |
4067 | return -ENOMEM; | |
4068 | } | |
4069 | ||
e09aae7e ML |
4070 | /* |
4071 | * It is the actual release handler for mq, but we do it from | |
4072 | * request queue's release handler for avoiding use-after-free | |
4073 | * and headache because q->mq_kobj shouldn't have been introduced, | |
4074 | * but we can't group ctx/kctx kobj without it. | |
4075 | */ | |
4076 | void blk_mq_release(struct request_queue *q) | |
4077 | { | |
2f8f1336 | 4078 | struct blk_mq_hw_ctx *hctx, *next; |
4f481208 | 4079 | unsigned long i; |
e09aae7e | 4080 | |
2f8f1336 ML |
4081 | queue_for_each_hw_ctx(q, hctx, i) |
4082 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
4083 | ||
4084 | /* all hctx are in .unused_hctx_list now */ | |
4085 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
4086 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 4087 | kobject_put(&hctx->kobj); |
c3b4afca | 4088 | } |
e09aae7e | 4089 | |
4e5cc99e | 4090 | xa_destroy(&q->hctx_table); |
e09aae7e | 4091 | |
7ea5fe31 ML |
4092 | /* |
4093 | * release .mq_kobj and sw queue's kobject now because | |
4094 | * both share lifetime with request queue. | |
4095 | */ | |
4096 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
4097 | } |
4098 | ||
9ac4dd8c CH |
4099 | struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set, |
4100 | struct queue_limits *lim, void *queuedata) | |
b62c21b7 | 4101 | { |
9ac4dd8c | 4102 | struct queue_limits default_lim = { }; |
26a9750a CH |
4103 | struct request_queue *q; |
4104 | int ret; | |
b62c21b7 | 4105 | |
9ac4dd8c | 4106 | q = blk_alloc_queue(lim ? lim : &default_lim, set->numa_node); |
ad751ba1 CH |
4107 | if (IS_ERR(q)) |
4108 | return q; | |
26a9750a CH |
4109 | q->queuedata = queuedata; |
4110 | ret = blk_mq_init_allocated_queue(set, q); | |
4111 | if (ret) { | |
6f8191fd | 4112 | blk_put_queue(q); |
26a9750a CH |
4113 | return ERR_PTR(ret); |
4114 | } | |
b62c21b7 MS |
4115 | return q; |
4116 | } | |
9ac4dd8c | 4117 | EXPORT_SYMBOL(blk_mq_alloc_queue); |
b62c21b7 | 4118 | |
6f8191fd CH |
4119 | /** |
4120 | * blk_mq_destroy_queue - shutdown a request queue | |
4121 | * @q: request queue to shutdown | |
4122 | * | |
9ac4dd8c | 4123 | * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future |
81ea42b9 | 4124 | * requests will be failed with -ENODEV. The caller is responsible for dropping |
9ac4dd8c | 4125 | * the reference from blk_mq_alloc_queue() by calling blk_put_queue(). |
6f8191fd CH |
4126 | * |
4127 | * Context: can sleep | |
4128 | */ | |
4129 | void blk_mq_destroy_queue(struct request_queue *q) | |
4130 | { | |
4131 | WARN_ON_ONCE(!queue_is_mq(q)); | |
4132 | WARN_ON_ONCE(blk_queue_registered(q)); | |
4133 | ||
4134 | might_sleep(); | |
4135 | ||
4136 | blk_queue_flag_set(QUEUE_FLAG_DYING, q); | |
4137 | blk_queue_start_drain(q); | |
56c1ee92 | 4138 | blk_mq_freeze_queue_wait(q); |
6f8191fd CH |
4139 | |
4140 | blk_sync_queue(q); | |
4141 | blk_mq_cancel_work_sync(q); | |
4142 | blk_mq_exit_queue(q); | |
6f8191fd CH |
4143 | } |
4144 | EXPORT_SYMBOL(blk_mq_destroy_queue); | |
4145 | ||
27e32cd2 CH |
4146 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, |
4147 | struct queue_limits *lim, void *queuedata, | |
4dcc4874 | 4148 | struct lock_class_key *lkclass) |
9316a9ed JA |
4149 | { |
4150 | struct request_queue *q; | |
b461dfc4 | 4151 | struct gendisk *disk; |
9316a9ed | 4152 | |
27e32cd2 | 4153 | q = blk_mq_alloc_queue(set, lim, queuedata); |
b461dfc4 CH |
4154 | if (IS_ERR(q)) |
4155 | return ERR_CAST(q); | |
9316a9ed | 4156 | |
4a1fa41d | 4157 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 | 4158 | if (!disk) { |
0a3e5cc7 | 4159 | blk_mq_destroy_queue(q); |
2b3f056f | 4160 | blk_put_queue(q); |
b461dfc4 | 4161 | return ERR_PTR(-ENOMEM); |
9316a9ed | 4162 | } |
6f8191fd | 4163 | set_bit(GD_OWNS_QUEUE, &disk->state); |
b461dfc4 | 4164 | return disk; |
9316a9ed | 4165 | } |
b461dfc4 | 4166 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 4167 | |
6f8191fd CH |
4168 | struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q, |
4169 | struct lock_class_key *lkclass) | |
4170 | { | |
22c17e27 CH |
4171 | struct gendisk *disk; |
4172 | ||
6f8191fd CH |
4173 | if (!blk_get_queue(q)) |
4174 | return NULL; | |
22c17e27 CH |
4175 | disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass); |
4176 | if (!disk) | |
4177 | blk_put_queue(q); | |
4178 | return disk; | |
6f8191fd CH |
4179 | } |
4180 | EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue); | |
4181 | ||
34d11ffa JW |
4182 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
4183 | struct blk_mq_tag_set *set, struct request_queue *q, | |
4184 | int hctx_idx, int node) | |
4185 | { | |
2f8f1336 | 4186 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 4187 | |
2f8f1336 ML |
4188 | /* reuse dead hctx first */ |
4189 | spin_lock(&q->unused_hctx_lock); | |
4190 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
4191 | if (tmp->numa_node == node) { | |
4192 | hctx = tmp; | |
4193 | break; | |
4194 | } | |
4195 | } | |
4196 | if (hctx) | |
4197 | list_del_init(&hctx->hctx_list); | |
4198 | spin_unlock(&q->unused_hctx_lock); | |
4199 | ||
4200 | if (!hctx) | |
4201 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 4202 | if (!hctx) |
7c6c5b7c | 4203 | goto fail; |
34d11ffa | 4204 | |
7c6c5b7c ML |
4205 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
4206 | goto free_hctx; | |
34d11ffa JW |
4207 | |
4208 | return hctx; | |
7c6c5b7c ML |
4209 | |
4210 | free_hctx: | |
4211 | kobject_put(&hctx->kobj); | |
4212 | fail: | |
4213 | return NULL; | |
34d11ffa JW |
4214 | } |
4215 | ||
868f2f0b KB |
4216 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
4217 | struct request_queue *q) | |
320ae51f | 4218 | { |
4e5cc99e ML |
4219 | struct blk_mq_hw_ctx *hctx; |
4220 | unsigned long i, j; | |
ac0d6b92 | 4221 | |
fb350e0a ML |
4222 | /* protect against switching io scheduler */ |
4223 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 4224 | for (i = 0; i < set->nr_hw_queues; i++) { |
306f13ee | 4225 | int old_node; |
4d805131 | 4226 | int node = blk_mq_get_hctx_node(set, i); |
4e5cc99e | 4227 | struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i); |
868f2f0b | 4228 | |
306f13ee ML |
4229 | if (old_hctx) { |
4230 | old_node = old_hctx->numa_node; | |
4231 | blk_mq_exit_hctx(q, set, old_hctx, i); | |
4232 | } | |
868f2f0b | 4233 | |
4e5cc99e | 4234 | if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) { |
306f13ee | 4235 | if (!old_hctx) |
34d11ffa | 4236 | break; |
306f13ee ML |
4237 | pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n", |
4238 | node, old_node); | |
4e5cc99e ML |
4239 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node); |
4240 | WARN_ON_ONCE(!hctx); | |
868f2f0b | 4241 | } |
320ae51f | 4242 | } |
e01ad46d JW |
4243 | /* |
4244 | * Increasing nr_hw_queues fails. Free the newly allocated | |
4245 | * hctxs and keep the previous q->nr_hw_queues. | |
4246 | */ | |
4247 | if (i != set->nr_hw_queues) { | |
4248 | j = q->nr_hw_queues; | |
e01ad46d JW |
4249 | } else { |
4250 | j = i; | |
e01ad46d JW |
4251 | q->nr_hw_queues = set->nr_hw_queues; |
4252 | } | |
34d11ffa | 4253 | |
4e5cc99e ML |
4254 | xa_for_each_start(&q->hctx_table, j, hctx, j) |
4255 | blk_mq_exit_hctx(q, set, hctx, j); | |
fb350e0a | 4256 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
4257 | } |
4258 | ||
42ee3061 ML |
4259 | static void blk_mq_update_poll_flag(struct request_queue *q) |
4260 | { | |
4261 | struct blk_mq_tag_set *set = q->tag_set; | |
4262 | ||
4263 | if (set->nr_maps > HCTX_TYPE_POLL && | |
4264 | set->map[HCTX_TYPE_POLL].nr_queues) | |
4265 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); | |
4266 | else | |
4267 | blk_queue_flag_clear(QUEUE_FLAG_POLL, q); | |
4268 | } | |
4269 | ||
26a9750a CH |
4270 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
4271 | struct request_queue *q) | |
868f2f0b | 4272 | { |
66841672 ML |
4273 | /* mark the queue as mq asap */ |
4274 | q->mq_ops = set->ops; | |
4275 | ||
1db4909e | 4276 | if (blk_mq_alloc_ctxs(q)) |
54bdd67d | 4277 | goto err_exit; |
868f2f0b | 4278 | |
737f98cf ML |
4279 | /* init q->mq_kobj and sw queues' kobjects */ |
4280 | blk_mq_sysfs_init(q); | |
4281 | ||
2f8f1336 ML |
4282 | INIT_LIST_HEAD(&q->unused_hctx_list); |
4283 | spin_lock_init(&q->unused_hctx_lock); | |
4284 | ||
4e5cc99e ML |
4285 | xa_init(&q->hctx_table); |
4286 | ||
868f2f0b KB |
4287 | blk_mq_realloc_hw_ctxs(set, q); |
4288 | if (!q->nr_hw_queues) | |
4289 | goto err_hctxs; | |
320ae51f | 4290 | |
287922eb | 4291 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 4292 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 4293 | |
a8908939 | 4294 | q->tag_set = set; |
320ae51f | 4295 | |
94eddfbe | 4296 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
42ee3061 | 4297 | blk_mq_update_poll_flag(q); |
320ae51f | 4298 | |
2849450a | 4299 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
9a67aa52 | 4300 | INIT_LIST_HEAD(&q->flush_list); |
6fca6a61 CH |
4301 | INIT_LIST_HEAD(&q->requeue_list); |
4302 | spin_lock_init(&q->requeue_lock); | |
4303 | ||
eba71768 JA |
4304 | q->nr_requests = set->queue_depth; |
4305 | ||
24d2f903 | 4306 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 4307 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 4308 | blk_mq_map_swqueue(q); |
26a9750a | 4309 | return 0; |
18741986 | 4310 | |
320ae51f | 4311 | err_hctxs: |
943f45b9 | 4312 | blk_mq_release(q); |
c7de5726 ML |
4313 | err_exit: |
4314 | q->mq_ops = NULL; | |
26a9750a | 4315 | return -ENOMEM; |
320ae51f | 4316 | } |
b62c21b7 | 4317 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 4318 | |
c7e2d94b ML |
4319 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
4320 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 4321 | { |
630ef623 | 4322 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 4323 | |
630ef623 | 4324 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 4325 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
4326 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
4327 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 4328 | } |
320ae51f | 4329 | |
a5164405 JA |
4330 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
4331 | { | |
4332 | int i; | |
4333 | ||
079a2e3e JG |
4334 | if (blk_mq_is_shared_tags(set->flags)) { |
4335 | set->shared_tags = blk_mq_alloc_map_and_rqs(set, | |
e155b0c2 JG |
4336 | BLK_MQ_NO_HCTX_IDX, |
4337 | set->queue_depth); | |
079a2e3e | 4338 | if (!set->shared_tags) |
e155b0c2 JG |
4339 | return -ENOMEM; |
4340 | } | |
4341 | ||
8229cca8 | 4342 | for (i = 0; i < set->nr_hw_queues; i++) { |
63064be1 | 4343 | if (!__blk_mq_alloc_map_and_rqs(set, i)) |
a5164405 | 4344 | goto out_unwind; |
8229cca8 XT |
4345 | cond_resched(); |
4346 | } | |
a5164405 JA |
4347 | |
4348 | return 0; | |
4349 | ||
4350 | out_unwind: | |
4351 | while (--i >= 0) | |
e155b0c2 JG |
4352 | __blk_mq_free_map_and_rqs(set, i); |
4353 | ||
079a2e3e JG |
4354 | if (blk_mq_is_shared_tags(set->flags)) { |
4355 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 | 4356 | BLK_MQ_NO_HCTX_IDX); |
645db34e | 4357 | } |
a5164405 | 4358 | |
a5164405 JA |
4359 | return -ENOMEM; |
4360 | } | |
4361 | ||
4362 | /* | |
4363 | * Allocate the request maps associated with this tag_set. Note that this | |
4364 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
4365 | * will be updated to reflect the allocated depth. | |
4366 | */ | |
63064be1 | 4367 | static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) |
a5164405 JA |
4368 | { |
4369 | unsigned int depth; | |
4370 | int err; | |
4371 | ||
4372 | depth = set->queue_depth; | |
4373 | do { | |
4374 | err = __blk_mq_alloc_rq_maps(set); | |
4375 | if (!err) | |
4376 | break; | |
4377 | ||
4378 | set->queue_depth >>= 1; | |
4379 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
4380 | err = -ENOMEM; | |
4381 | break; | |
4382 | } | |
4383 | } while (set->queue_depth); | |
4384 | ||
4385 | if (!set->queue_depth || err) { | |
4386 | pr_err("blk-mq: failed to allocate request map\n"); | |
4387 | return -ENOMEM; | |
4388 | } | |
4389 | ||
4390 | if (depth != set->queue_depth) | |
4391 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
4392 | depth, set->queue_depth); | |
4393 | ||
4394 | return 0; | |
4395 | } | |
4396 | ||
a4e1d0b7 | 4397 | static void blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
ebe8bddb | 4398 | { |
6e66b493 BVA |
4399 | /* |
4400 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
4401 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
4402 | * number of hardware queues. | |
4403 | */ | |
4404 | if (set->nr_maps == 1) | |
4405 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
4406 | ||
ec30b461 | 4407 | if (set->ops->map_queues) { |
b3c661b1 JA |
4408 | int i; |
4409 | ||
7d4901a9 ML |
4410 | /* |
4411 | * transport .map_queues is usually done in the following | |
4412 | * way: | |
4413 | * | |
4414 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
4415 | * mask = get_cpu_mask(queue) | |
4416 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 4417 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
4418 | * } |
4419 | * | |
4420 | * When we need to remap, the table has to be cleared for | |
4421 | * killing stale mapping since one CPU may not be mapped | |
4422 | * to any hw queue. | |
4423 | */ | |
b3c661b1 JA |
4424 | for (i = 0; i < set->nr_maps; i++) |
4425 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 4426 | |
a4e1d0b7 | 4427 | set->ops->map_queues(set); |
b3c661b1 JA |
4428 | } else { |
4429 | BUG_ON(set->nr_maps > 1); | |
a4e1d0b7 | 4430 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 4431 | } |
ebe8bddb OS |
4432 | } |
4433 | ||
f7e76dbc | 4434 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
ee9d5521 | 4435 | int new_nr_hw_queues) |
f7e76dbc BVA |
4436 | { |
4437 | struct blk_mq_tags **new_tags; | |
e1dd7bc9 | 4438 | int i; |
f7e76dbc | 4439 | |
6be6d112 | 4440 | if (set->nr_hw_queues >= new_nr_hw_queues) |
d4b2e0d4 | 4441 | goto done; |
f7e76dbc BVA |
4442 | |
4443 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
4444 | GFP_KERNEL, set->numa_node); | |
4445 | if (!new_tags) | |
4446 | return -ENOMEM; | |
4447 | ||
4448 | if (set->tags) | |
ee9d5521 | 4449 | memcpy(new_tags, set->tags, set->nr_hw_queues * |
f7e76dbc BVA |
4450 | sizeof(*set->tags)); |
4451 | kfree(set->tags); | |
4452 | set->tags = new_tags; | |
7222657e CZ |
4453 | |
4454 | for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) { | |
4455 | if (!__blk_mq_alloc_map_and_rqs(set, i)) { | |
4456 | while (--i >= set->nr_hw_queues) | |
4457 | __blk_mq_free_map_and_rqs(set, i); | |
4458 | return -ENOMEM; | |
4459 | } | |
4460 | cond_resched(); | |
4461 | } | |
4462 | ||
d4b2e0d4 | 4463 | done: |
f7e76dbc | 4464 | set->nr_hw_queues = new_nr_hw_queues; |
f7e76dbc BVA |
4465 | return 0; |
4466 | } | |
4467 | ||
a4391c64 JA |
4468 | /* |
4469 | * Alloc a tag set to be associated with one or more request queues. | |
4470 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 4471 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
4472 | * value will be stored in set->queue_depth. |
4473 | */ | |
24d2f903 CH |
4474 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
4475 | { | |
b3c661b1 | 4476 | int i, ret; |
da695ba2 | 4477 | |
205fb5f5 BVA |
4478 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
4479 | ||
24d2f903 CH |
4480 | if (!set->nr_hw_queues) |
4481 | return -EINVAL; | |
a4391c64 | 4482 | if (!set->queue_depth) |
24d2f903 CH |
4483 | return -EINVAL; |
4484 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
4485 | return -EINVAL; | |
4486 | ||
7d7e0f90 | 4487 | if (!set->ops->queue_rq) |
24d2f903 CH |
4488 | return -EINVAL; |
4489 | ||
de148297 ML |
4490 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
4491 | return -EINVAL; | |
4492 | ||
a4391c64 JA |
4493 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
4494 | pr_info("blk-mq: reduced tag depth to %u\n", | |
4495 | BLK_MQ_MAX_DEPTH); | |
4496 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
4497 | } | |
24d2f903 | 4498 | |
b3c661b1 JA |
4499 | if (!set->nr_maps) |
4500 | set->nr_maps = 1; | |
4501 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
4502 | return -EINVAL; | |
4503 | ||
6637fadf SL |
4504 | /* |
4505 | * If a crashdump is active, then we are potentially in a very | |
ec30b461 ML |
4506 | * memory constrained environment. Limit us to 64 tags to prevent |
4507 | * using too much memory. | |
6637fadf | 4508 | */ |
ec30b461 | 4509 | if (is_kdump_kernel()) |
6637fadf | 4510 | set->queue_depth = min(64U, set->queue_depth); |
ec30b461 | 4511 | |
868f2f0b | 4512 | /* |
392546ae JA |
4513 | * There is no use for more h/w queues than cpus if we just have |
4514 | * a single map | |
868f2f0b | 4515 | */ |
392546ae | 4516 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4517 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 4518 | |
80bd4a7a CH |
4519 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4520 | set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL); | |
4521 | if (!set->srcu) | |
4522 | return -ENOMEM; | |
4523 | ret = init_srcu_struct(set->srcu); | |
4524 | if (ret) | |
4525 | goto out_free_srcu; | |
4526 | } | |
24d2f903 | 4527 | |
da695ba2 | 4528 | ret = -ENOMEM; |
5ee20298 CH |
4529 | set->tags = kcalloc_node(set->nr_hw_queues, |
4530 | sizeof(struct blk_mq_tags *), GFP_KERNEL, | |
4531 | set->numa_node); | |
4532 | if (!set->tags) | |
80bd4a7a | 4533 | goto out_cleanup_srcu; |
24d2f903 | 4534 | |
b3c661b1 JA |
4535 | for (i = 0; i < set->nr_maps; i++) { |
4536 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 4537 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
4538 | GFP_KERNEL, set->numa_node); |
4539 | if (!set->map[i].mq_map) | |
4540 | goto out_free_mq_map; | |
ec30b461 | 4541 | set->map[i].nr_queues = set->nr_hw_queues; |
b3c661b1 | 4542 | } |
bdd17e75 | 4543 | |
a4e1d0b7 | 4544 | blk_mq_update_queue_map(set); |
da695ba2 | 4545 | |
63064be1 | 4546 | ret = blk_mq_alloc_set_map_and_rqs(set); |
da695ba2 | 4547 | if (ret) |
bdd17e75 | 4548 | goto out_free_mq_map; |
24d2f903 | 4549 | |
0d2602ca JA |
4550 | mutex_init(&set->tag_list_lock); |
4551 | INIT_LIST_HEAD(&set->tag_list); | |
4552 | ||
24d2f903 | 4553 | return 0; |
bdd17e75 CH |
4554 | |
4555 | out_free_mq_map: | |
b3c661b1 JA |
4556 | for (i = 0; i < set->nr_maps; i++) { |
4557 | kfree(set->map[i].mq_map); | |
4558 | set->map[i].mq_map = NULL; | |
4559 | } | |
5676e7b6 RE |
4560 | kfree(set->tags); |
4561 | set->tags = NULL; | |
80bd4a7a CH |
4562 | out_cleanup_srcu: |
4563 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4564 | cleanup_srcu_struct(set->srcu); | |
4565 | out_free_srcu: | |
4566 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4567 | kfree(set->srcu); | |
da695ba2 | 4568 | return ret; |
24d2f903 CH |
4569 | } |
4570 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
4571 | ||
cdb14e0f CH |
4572 | /* allocate and initialize a tagset for a simple single-queue device */ |
4573 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
4574 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
4575 | unsigned int set_flags) | |
4576 | { | |
4577 | memset(set, 0, sizeof(*set)); | |
4578 | set->ops = ops; | |
4579 | set->nr_hw_queues = 1; | |
4580 | set->nr_maps = 1; | |
4581 | set->queue_depth = queue_depth; | |
4582 | set->numa_node = NUMA_NO_NODE; | |
4583 | set->flags = set_flags; | |
4584 | return blk_mq_alloc_tag_set(set); | |
4585 | } | |
4586 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
4587 | ||
24d2f903 CH |
4588 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
4589 | { | |
b3c661b1 | 4590 | int i, j; |
24d2f903 | 4591 | |
f7e76dbc | 4592 | for (i = 0; i < set->nr_hw_queues; i++) |
e155b0c2 | 4593 | __blk_mq_free_map_and_rqs(set, i); |
484b4061 | 4594 | |
079a2e3e JG |
4595 | if (blk_mq_is_shared_tags(set->flags)) { |
4596 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 JG |
4597 | BLK_MQ_NO_HCTX_IDX); |
4598 | } | |
32bc15af | 4599 | |
b3c661b1 JA |
4600 | for (j = 0; j < set->nr_maps; j++) { |
4601 | kfree(set->map[j].mq_map); | |
4602 | set->map[j].mq_map = NULL; | |
4603 | } | |
bdd17e75 | 4604 | |
981bd189 | 4605 | kfree(set->tags); |
5676e7b6 | 4606 | set->tags = NULL; |
80bd4a7a CH |
4607 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4608 | cleanup_srcu_struct(set->srcu); | |
4609 | kfree(set->srcu); | |
4610 | } | |
24d2f903 CH |
4611 | } |
4612 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
4613 | ||
e3a2b3f9 JA |
4614 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
4615 | { | |
4616 | struct blk_mq_tag_set *set = q->tag_set; | |
4617 | struct blk_mq_hw_ctx *hctx; | |
4f481208 ML |
4618 | int ret; |
4619 | unsigned long i; | |
e3a2b3f9 | 4620 | |
bd166ef1 | 4621 | if (!set) |
e3a2b3f9 JA |
4622 | return -EINVAL; |
4623 | ||
e5fa8140 AZ |
4624 | if (q->nr_requests == nr) |
4625 | return 0; | |
4626 | ||
70f36b60 | 4627 | blk_mq_freeze_queue(q); |
24f5a90f | 4628 | blk_mq_quiesce_queue(q); |
70f36b60 | 4629 | |
e3a2b3f9 JA |
4630 | ret = 0; |
4631 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
4632 | if (!hctx->tags) |
4633 | continue; | |
bd166ef1 JA |
4634 | /* |
4635 | * If we're using an MQ scheduler, just update the scheduler | |
4636 | * queue depth. This is similar to what the old code would do. | |
4637 | */ | |
f6adcef5 | 4638 | if (hctx->sched_tags) { |
70f36b60 | 4639 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
f6adcef5 | 4640 | nr, true); |
f6adcef5 JG |
4641 | } else { |
4642 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
4643 | false); | |
70f36b60 | 4644 | } |
e3a2b3f9 JA |
4645 | if (ret) |
4646 | break; | |
77f1e0a5 JA |
4647 | if (q->elevator && q->elevator->type->ops.depth_updated) |
4648 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 4649 | } |
d97e594c | 4650 | if (!ret) { |
e3a2b3f9 | 4651 | q->nr_requests = nr; |
079a2e3e | 4652 | if (blk_mq_is_shared_tags(set->flags)) { |
8fa04464 | 4653 | if (q->elevator) |
079a2e3e | 4654 | blk_mq_tag_update_sched_shared_tags(q); |
8fa04464 | 4655 | else |
079a2e3e | 4656 | blk_mq_tag_resize_shared_tags(set, nr); |
8fa04464 | 4657 | } |
d97e594c | 4658 | } |
e3a2b3f9 | 4659 | |
24f5a90f | 4660 | blk_mq_unquiesce_queue(q); |
70f36b60 | 4661 | blk_mq_unfreeze_queue(q); |
70f36b60 | 4662 | |
e3a2b3f9 JA |
4663 | return ret; |
4664 | } | |
4665 | ||
d48ece20 JW |
4666 | /* |
4667 | * request_queue and elevator_type pair. | |
4668 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
4669 | * the elevator_type associated with a request_queue. | |
4670 | */ | |
4671 | struct blk_mq_qe_pair { | |
4672 | struct list_head node; | |
4673 | struct request_queue *q; | |
4674 | struct elevator_type *type; | |
4675 | }; | |
4676 | ||
4677 | /* | |
4678 | * Cache the elevator_type in qe pair list and switch the | |
4679 | * io scheduler to 'none' | |
4680 | */ | |
4681 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
4682 | struct request_queue *q) | |
4683 | { | |
4684 | struct blk_mq_qe_pair *qe; | |
4685 | ||
d48ece20 JW |
4686 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); |
4687 | if (!qe) | |
4688 | return false; | |
4689 | ||
5fd7a84a ML |
4690 | /* q->elevator needs protection from ->sysfs_lock */ |
4691 | mutex_lock(&q->sysfs_lock); | |
4692 | ||
24516565 ML |
4693 | /* the check has to be done with holding sysfs_lock */ |
4694 | if (!q->elevator) { | |
4695 | kfree(qe); | |
4696 | goto unlock; | |
4697 | } | |
4698 | ||
d48ece20 JW |
4699 | INIT_LIST_HEAD(&qe->node); |
4700 | qe->q = q; | |
4701 | qe->type = q->elevator->type; | |
dd6f7f17 CH |
4702 | /* keep a reference to the elevator module as we'll switch back */ |
4703 | __elevator_get(qe->type); | |
d48ece20 | 4704 | list_add(&qe->node, head); |
64b36075 | 4705 | elevator_disable(q); |
24516565 | 4706 | unlock: |
d48ece20 JW |
4707 | mutex_unlock(&q->sysfs_lock); |
4708 | ||
4709 | return true; | |
4710 | } | |
4711 | ||
4a3b666e JK |
4712 | static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head, |
4713 | struct request_queue *q) | |
d48ece20 JW |
4714 | { |
4715 | struct blk_mq_qe_pair *qe; | |
d48ece20 JW |
4716 | |
4717 | list_for_each_entry(qe, head, node) | |
4a3b666e JK |
4718 | if (qe->q == q) |
4719 | return qe; | |
d48ece20 | 4720 | |
4a3b666e JK |
4721 | return NULL; |
4722 | } | |
d48ece20 | 4723 | |
4a3b666e JK |
4724 | static void blk_mq_elv_switch_back(struct list_head *head, |
4725 | struct request_queue *q) | |
4726 | { | |
4727 | struct blk_mq_qe_pair *qe; | |
4728 | struct elevator_type *t; | |
4729 | ||
4730 | qe = blk_lookup_qe_pair(head, q); | |
4731 | if (!qe) | |
4732 | return; | |
4733 | t = qe->type; | |
d48ece20 JW |
4734 | list_del(&qe->node); |
4735 | kfree(qe); | |
4736 | ||
4737 | mutex_lock(&q->sysfs_lock); | |
8237c01f | 4738 | elevator_switch(q, t); |
8ed40ee3 JC |
4739 | /* drop the reference acquired in blk_mq_elv_switch_none */ |
4740 | elevator_put(t); | |
d48ece20 JW |
4741 | mutex_unlock(&q->sysfs_lock); |
4742 | } | |
4743 | ||
e4dc2b32 KB |
4744 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
4745 | int nr_hw_queues) | |
868f2f0b KB |
4746 | { |
4747 | struct request_queue *q; | |
d48ece20 | 4748 | LIST_HEAD(head); |
6be6d112 CZ |
4749 | int prev_nr_hw_queues = set->nr_hw_queues; |
4750 | int i; | |
868f2f0b | 4751 | |
705cda97 BVA |
4752 | lockdep_assert_held(&set->tag_list_lock); |
4753 | ||
392546ae | 4754 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4755 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
4756 | if (nr_hw_queues < 1) |
4757 | return; | |
4758 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
4759 | return; |
4760 | ||
4761 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4762 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
4763 | /* |
4764 | * Switch IO scheduler to 'none', cleaning up the data associated | |
4765 | * with the previous scheduler. We will switch back once we are done | |
4766 | * updating the new sw to hw queue mappings. | |
4767 | */ | |
4768 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4769 | if (!blk_mq_elv_switch_none(&head, q)) | |
4770 | goto switch_back; | |
868f2f0b | 4771 | |
477e19de JW |
4772 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4773 | blk_mq_debugfs_unregister_hctxs(q); | |
eaa870f9 | 4774 | blk_mq_sysfs_unregister_hctxs(q); |
477e19de JW |
4775 | } |
4776 | ||
ee9d5521 | 4777 | if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0) |
f7e76dbc BVA |
4778 | goto reregister; |
4779 | ||
e01ad46d | 4780 | fallback: |
aa880ad6 | 4781 | blk_mq_update_queue_map(set); |
868f2f0b KB |
4782 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4783 | blk_mq_realloc_hw_ctxs(set, q); | |
42ee3061 | 4784 | blk_mq_update_poll_flag(q); |
e01ad46d | 4785 | if (q->nr_hw_queues != set->nr_hw_queues) { |
a846a8e6 YB |
4786 | int i = prev_nr_hw_queues; |
4787 | ||
e01ad46d JW |
4788 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", |
4789 | nr_hw_queues, prev_nr_hw_queues); | |
a846a8e6 YB |
4790 | for (; i < set->nr_hw_queues; i++) |
4791 | __blk_mq_free_map_and_rqs(set, i); | |
4792 | ||
e01ad46d | 4793 | set->nr_hw_queues = prev_nr_hw_queues; |
e01ad46d JW |
4794 | goto fallback; |
4795 | } | |
477e19de JW |
4796 | blk_mq_map_swqueue(q); |
4797 | } | |
4798 | ||
f7e76dbc | 4799 | reregister: |
477e19de | 4800 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
eaa870f9 | 4801 | blk_mq_sysfs_register_hctxs(q); |
477e19de | 4802 | blk_mq_debugfs_register_hctxs(q); |
868f2f0b KB |
4803 | } |
4804 | ||
d48ece20 JW |
4805 | switch_back: |
4806 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4807 | blk_mq_elv_switch_back(&head, q); | |
4808 | ||
868f2f0b KB |
4809 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
4810 | blk_mq_unfreeze_queue(q); | |
6be6d112 CZ |
4811 | |
4812 | /* Free the excess tags when nr_hw_queues shrink. */ | |
4813 | for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++) | |
4814 | __blk_mq_free_map_and_rqs(set, i); | |
868f2f0b | 4815 | } |
e4dc2b32 KB |
4816 | |
4817 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
4818 | { | |
4819 | mutex_lock(&set->tag_list_lock); | |
4820 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
4821 | mutex_unlock(&set->tag_list_lock); | |
4822 | } | |
868f2f0b KB |
4823 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
4824 | ||
f6c80cff KB |
4825 | static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx, |
4826 | struct io_comp_batch *iob, unsigned int flags) | |
bbd7bb70 | 4827 | { |
c6699d6f CH |
4828 | long state = get_current_state(); |
4829 | int ret; | |
bbd7bb70 | 4830 | |
aa61bec3 | 4831 | do { |
5a72e899 | 4832 | ret = q->mq_ops->poll(hctx, iob); |
bbd7bb70 | 4833 | if (ret > 0) { |
849a3700 | 4834 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4835 | return ret; |
bbd7bb70 JA |
4836 | } |
4837 | ||
4838 | if (signal_pending_state(state, current)) | |
849a3700 | 4839 | __set_current_state(TASK_RUNNING); |
b03fbd4f | 4840 | if (task_is_running(current)) |
85f4d4b6 | 4841 | return 1; |
c6699d6f | 4842 | |
ef99b2d3 | 4843 | if (ret < 0 || (flags & BLK_POLL_ONESHOT)) |
bbd7bb70 JA |
4844 | break; |
4845 | cpu_relax(); | |
aa61bec3 | 4846 | } while (!need_resched()); |
bbd7bb70 | 4847 | |
67b4110f | 4848 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4849 | return 0; |
bbd7bb70 | 4850 | } |
1052b8ac | 4851 | |
f6c80cff KB |
4852 | int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, |
4853 | struct io_comp_batch *iob, unsigned int flags) | |
4854 | { | |
4855 | struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie); | |
4856 | ||
4857 | return blk_hctx_poll(q, hctx, iob, flags); | |
4858 | } | |
4859 | ||
4860 | int blk_rq_poll(struct request *rq, struct io_comp_batch *iob, | |
4861 | unsigned int poll_flags) | |
4862 | { | |
4863 | struct request_queue *q = rq->q; | |
4864 | int ret; | |
4865 | ||
4866 | if (!blk_rq_is_poll(rq)) | |
4867 | return 0; | |
4868 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
4869 | return 0; | |
4870 | ||
4871 | ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags); | |
4872 | blk_queue_exit(q); | |
4873 | ||
4874 | return ret; | |
4875 | } | |
4876 | EXPORT_SYMBOL_GPL(blk_rq_poll); | |
4877 | ||
9cf2bab6 JA |
4878 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4879 | { | |
4880 | return rq->mq_ctx->cpu; | |
4881 | } | |
4882 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4883 | ||
2a19b28f ML |
4884 | void blk_mq_cancel_work_sync(struct request_queue *q) |
4885 | { | |
219cf43c JC |
4886 | struct blk_mq_hw_ctx *hctx; |
4887 | unsigned long i; | |
2a19b28f | 4888 | |
219cf43c | 4889 | cancel_delayed_work_sync(&q->requeue_work); |
2a19b28f | 4890 | |
219cf43c JC |
4891 | queue_for_each_hw_ctx(q, hctx, i) |
4892 | cancel_delayed_work_sync(&hctx->run_work); | |
2a19b28f ML |
4893 | } |
4894 | ||
320ae51f JA |
4895 | static int __init blk_mq_init(void) |
4896 | { | |
c3077b5d CH |
4897 | int i; |
4898 | ||
4899 | for_each_possible_cpu(i) | |
f9ab4918 | 4900 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
660e802c CZ |
4901 | for_each_possible_cpu(i) |
4902 | INIT_CSD(&per_cpu(blk_cpu_csd, i), | |
4903 | __blk_mq_complete_request_remote, NULL); | |
c3077b5d CH |
4904 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4905 | ||
4906 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4907 | "block/softirq:dead", NULL, | |
4908 | blk_softirq_cpu_dead); | |
9467f859 TG |
4909 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4910 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4911 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4912 | blk_mq_hctx_notify_online, | |
4913 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4914 | return 0; |
4915 | } | |
4916 | subsys_initcall(blk_mq_init); |