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 | 95 | if (rq->part && blk_do_io_stat(rq) && |
3f9b8fb4 | 96 | (!bdev_is_partition(mi->part) || 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 | |
6259151c BVA |
451 | retry: |
452 | data->ctx = blk_mq_get_ctx(q); | |
453 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
454 | ||
781dd830 | 455 | if (q->elevator) { |
dd6216bb CH |
456 | /* |
457 | * All requests use scheduler tags when an I/O scheduler is | |
458 | * enabled for the queue. | |
459 | */ | |
460 | data->rq_flags |= RQF_SCHED_TAGS; | |
781dd830 | 461 | |
d2c0d383 | 462 | /* |
8d663f34 | 463 | * Flush/passthrough requests are special and go directly to the |
dd6216bb | 464 | * dispatch list. |
d2c0d383 | 465 | */ |
be4c4278 | 466 | if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH && |
dd6216bb CH |
467 | !blk_op_is_passthrough(data->cmd_flags)) { |
468 | struct elevator_mq_ops *ops = &q->elevator->type->ops; | |
469 | ||
470 | WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED); | |
471 | ||
472 | data->rq_flags |= RQF_USE_SCHED; | |
473 | if (ops->limit_depth) | |
474 | ops->limit_depth(data->cmd_flags, data); | |
475 | } | |
6259151c | 476 | } else { |
600c3b0c | 477 | blk_mq_tag_busy(data->hctx); |
6259151c | 478 | } |
600c3b0c | 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 | ||
347bde9d DLM |
694 | blk_zone_finish_request(rq); |
695 | ||
e5c0ca13 CZ |
696 | if (rq->rq_flags & RQF_USE_SCHED) { |
697 | q->elevator->type->ops.finish_request(rq); | |
698 | /* | |
699 | * For postflush request that may need to be | |
700 | * completed twice, we should clear this flag | |
701 | * to avoid double finish_request() on the rq. | |
702 | */ | |
703 | rq->rq_flags &= ~RQF_USE_SCHED; | |
704 | } | |
705 | } | |
706 | ||
12f5b931 KB |
707 | static void __blk_mq_free_request(struct request *rq) |
708 | { | |
709 | struct request_queue *q = rq->q; | |
710 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 711 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
712 | const int sched_tag = rq->internal_tag; |
713 | ||
a892c8d5 | 714 | blk_crypto_free_request(rq); |
986d413b | 715 | blk_pm_mark_last_busy(rq); |
ea4f995e | 716 | rq->mq_hctx = NULL; |
ddad5933 | 717 | |
b8643d68 CZ |
718 | if (rq->tag != BLK_MQ_NO_TAG) { |
719 | blk_mq_dec_active_requests(hctx); | |
cae740a0 | 720 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
b8643d68 | 721 | } |
76647368 | 722 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 723 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
724 | blk_mq_sched_restart(hctx); |
725 | blk_queue_exit(q); | |
726 | } | |
727 | ||
6af54051 | 728 | void blk_mq_free_request(struct request *rq) |
320ae51f | 729 | { |
320ae51f | 730 | struct request_queue *q = rq->q; |
6af54051 | 731 | |
e5c0ca13 | 732 | blk_mq_finish_request(rq); |
320ae51f | 733 | |
7beb2f84 | 734 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 735 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 736 | |
a7905043 | 737 | rq_qos_done(q, rq); |
0d2602ca | 738 | |
12f5b931 | 739 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 740 | if (req_ref_put_and_test(rq)) |
12f5b931 | 741 | __blk_mq_free_request(rq); |
320ae51f | 742 | } |
1a3b595a | 743 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 744 | |
47c122e3 | 745 | void blk_mq_free_plug_rqs(struct blk_plug *plug) |
320ae51f | 746 | { |
013a7f95 | 747 | struct request *rq; |
fe1f4526 | 748 | |
c5fc7b93 | 749 | while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) |
47c122e3 | 750 | blk_mq_free_request(rq); |
47c122e3 | 751 | } |
522a7775 | 752 | |
22350ad7 CH |
753 | void blk_dump_rq_flags(struct request *rq, char *msg) |
754 | { | |
755 | printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, | |
f3fa33ac | 756 | rq->q->disk ? rq->q->disk->disk_name : "?", |
16458cf3 | 757 | (__force unsigned long long) rq->cmd_flags); |
22350ad7 CH |
758 | |
759 | printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", | |
760 | (unsigned long long)blk_rq_pos(rq), | |
761 | blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); | |
762 | printk(KERN_INFO " bio %p, biotail %p, len %u\n", | |
763 | rq->bio, rq->biotail, blk_rq_bytes(rq)); | |
764 | } | |
765 | EXPORT_SYMBOL(blk_dump_rq_flags); | |
766 | ||
9be3e06f JA |
767 | static void blk_account_io_completion(struct request *req, unsigned int bytes) |
768 | { | |
769 | if (req->part && blk_do_io_stat(req)) { | |
770 | const int sgrp = op_stat_group(req_op(req)); | |
771 | ||
772 | part_stat_lock(); | |
773 | part_stat_add(req->part, sectors[sgrp], bytes >> 9); | |
774 | part_stat_unlock(); | |
775 | } | |
776 | } | |
777 | ||
0d7a29a2 CH |
778 | static void blk_print_req_error(struct request *req, blk_status_t status) |
779 | { | |
780 | printk_ratelimited(KERN_ERR | |
781 | "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x " | |
782 | "phys_seg %u prio class %u\n", | |
783 | blk_status_to_str(status), | |
f3fa33ac | 784 | req->q->disk ? req->q->disk->disk_name : "?", |
16458cf3 BVA |
785 | blk_rq_pos(req), (__force u32)req_op(req), |
786 | blk_op_str(req_op(req)), | |
787 | (__force u32)(req->cmd_flags & ~REQ_OP_MASK), | |
0d7a29a2 CH |
788 | req->nr_phys_segments, |
789 | IOPRIO_PRIO_CLASS(req->ioprio)); | |
790 | } | |
791 | ||
5581a5dd JA |
792 | /* |
793 | * Fully end IO on a request. Does not support partial completions, or | |
794 | * errors. | |
795 | */ | |
796 | static void blk_complete_request(struct request *req) | |
797 | { | |
798 | const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0; | |
799 | int total_bytes = blk_rq_bytes(req); | |
800 | struct bio *bio = req->bio; | |
801 | ||
802 | trace_block_rq_complete(req, BLK_STS_OK, total_bytes); | |
803 | ||
804 | if (!bio) | |
805 | return; | |
806 | ||
5581a5dd | 807 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ) |
e9f5f44a | 808 | blk_integrity_complete(req, total_bytes); |
5581a5dd | 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 | ||
831 | /* | |
832 | * Reset counters so that the request stacking driver | |
833 | * can find how many bytes remain in the request | |
834 | * later. | |
835 | */ | |
ab3e1d3b JA |
836 | if (!req->end_io) { |
837 | req->bio = NULL; | |
838 | req->__data_len = 0; | |
839 | } | |
5581a5dd JA |
840 | } |
841 | ||
9be3e06f JA |
842 | /** |
843 | * blk_update_request - Complete multiple bytes without completing the request | |
844 | * @req: the request being processed | |
845 | * @error: block status code | |
846 | * @nr_bytes: number of bytes to complete for @req | |
847 | * | |
848 | * Description: | |
849 | * Ends I/O on a number of bytes attached to @req, but doesn't complete | |
850 | * the request structure even if @req doesn't have leftover. | |
851 | * If @req has leftover, sets it up for the next range of segments. | |
852 | * | |
853 | * Passing the result of blk_rq_bytes() as @nr_bytes guarantees | |
854 | * %false return from this function. | |
855 | * | |
856 | * Note: | |
857 | * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function | |
858 | * except in the consistency check at the end of this function. | |
859 | * | |
860 | * Return: | |
861 | * %false - this request doesn't have any more data | |
862 | * %true - this request has more data | |
863 | **/ | |
864 | bool blk_update_request(struct request *req, blk_status_t error, | |
865 | unsigned int nr_bytes) | |
866 | { | |
c0da26f9 DLM |
867 | bool is_flush = req->rq_flags & RQF_FLUSH_SEQ; |
868 | bool quiet = req->rq_flags & RQF_QUIET; | |
9be3e06f JA |
869 | int total_bytes; |
870 | ||
8a7d267b | 871 | trace_block_rq_complete(req, error, nr_bytes); |
9be3e06f JA |
872 | |
873 | if (!req->bio) | |
874 | return false; | |
875 | ||
9be3e06f JA |
876 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && |
877 | error == BLK_STS_OK) | |
e9f5f44a | 878 | blk_integrity_complete(req, nr_bytes); |
9be3e06f | 879 | |
9cd1e566 EB |
880 | /* |
881 | * Upper layers may call blk_crypto_evict_key() anytime after the last | |
882 | * bio_endio(). Therefore, the keyslot must be released before that. | |
883 | */ | |
884 | if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req)) | |
885 | __blk_crypto_rq_put_keyslot(req); | |
886 | ||
c0da26f9 DLM |
887 | if (unlikely(error && !blk_rq_is_passthrough(req) && !quiet) && |
888 | !test_bit(GD_DEAD, &req->q->disk->state)) { | |
9be3e06f | 889 | blk_print_req_error(req, error); |
d5869fdc YS |
890 | trace_block_rq_error(req, error, nr_bytes); |
891 | } | |
9be3e06f JA |
892 | |
893 | blk_account_io_completion(req, nr_bytes); | |
894 | ||
895 | total_bytes = 0; | |
896 | while (req->bio) { | |
897 | struct bio *bio = req->bio; | |
898 | unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); | |
899 | ||
c0da26f9 DLM |
900 | if (unlikely(error)) |
901 | bio->bi_status = error; | |
902 | ||
903 | if (bio_bytes == bio->bi_iter.bi_size) { | |
9be3e06f | 904 | req->bio = bio->bi_next; |
9b1ce7f0 | 905 | } else if (bio_is_zone_append(bio) && error == BLK_STS_OK) { |
c0da26f9 DLM |
906 | /* |
907 | * Partial zone append completions cannot be supported | |
908 | * as the BIO fragments may end up not being written | |
909 | * sequentially. | |
910 | */ | |
911 | bio->bi_status = BLK_STS_IOERR; | |
912 | } | |
9be3e06f JA |
913 | |
914 | /* Completion has already been traced */ | |
915 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
c0da26f9 DLM |
916 | if (unlikely(quiet)) |
917 | bio_set_flag(bio, BIO_QUIET); | |
918 | ||
919 | bio_advance(bio, bio_bytes); | |
920 | ||
921 | /* Don't actually finish bio if it's part of flush sequence */ | |
a0508c36 DLM |
922 | if (!bio->bi_iter.bi_size) { |
923 | blk_zone_update_request_bio(req, bio); | |
924 | if (!is_flush) | |
925 | bio_endio(bio); | |
c0da26f9 | 926 | } |
9be3e06f JA |
927 | |
928 | total_bytes += bio_bytes; | |
929 | nr_bytes -= bio_bytes; | |
930 | ||
931 | if (!nr_bytes) | |
932 | break; | |
933 | } | |
934 | ||
935 | /* | |
936 | * completely done | |
937 | */ | |
938 | if (!req->bio) { | |
939 | /* | |
940 | * Reset counters so that the request stacking driver | |
941 | * can find how many bytes remain in the request | |
942 | * later. | |
943 | */ | |
944 | req->__data_len = 0; | |
945 | return false; | |
946 | } | |
947 | ||
948 | req->__data_len -= total_bytes; | |
949 | ||
950 | /* update sector only for requests with clear definition of sector */ | |
951 | if (!blk_rq_is_passthrough(req)) | |
952 | req->__sector += total_bytes >> 9; | |
953 | ||
954 | /* mixed attributes always follow the first bio */ | |
955 | if (req->rq_flags & RQF_MIXED_MERGE) { | |
956 | req->cmd_flags &= ~REQ_FAILFAST_MASK; | |
957 | req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; | |
958 | } | |
959 | ||
960 | if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { | |
961 | /* | |
962 | * If total number of sectors is less than the first segment | |
963 | * size, something has gone terribly wrong. | |
964 | */ | |
965 | if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { | |
966 | blk_dump_rq_flags(req, "request botched"); | |
967 | req->__data_len = blk_rq_cur_bytes(req); | |
968 | } | |
969 | ||
970 | /* recalculate the number of segments */ | |
971 | req->nr_phys_segments = blk_recalc_rq_segments(req); | |
972 | } | |
973 | ||
974 | return true; | |
975 | } | |
976 | EXPORT_SYMBOL_GPL(blk_update_request); | |
977 | ||
450b7879 CH |
978 | static inline void blk_account_io_done(struct request *req, u64 now) |
979 | { | |
5a80bd07 HC |
980 | trace_block_io_done(req); |
981 | ||
450b7879 CH |
982 | /* |
983 | * Account IO completion. flush_rq isn't accounted as a | |
984 | * normal IO on queueing nor completion. Accounting the | |
985 | * containing request is enough. | |
986 | */ | |
987 | if (blk_do_io_stat(req) && req->part && | |
06965037 CK |
988 | !(req->rq_flags & RQF_FLUSH_SEQ)) { |
989 | const int sgrp = op_stat_group(req_op(req)); | |
450b7879 | 990 | |
06965037 CK |
991 | part_stat_lock(); |
992 | update_io_ticks(req->part, jiffies, true); | |
993 | part_stat_inc(req->part, ios[sgrp]); | |
994 | part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns); | |
99dc4223 YK |
995 | part_stat_local_dec(req->part, |
996 | in_flight[op_is_write(req_op(req))]); | |
06965037 CK |
997 | part_stat_unlock(); |
998 | } | |
450b7879 CH |
999 | } |
1000 | ||
1001 | static inline void blk_account_io_start(struct request *req) | |
1002 | { | |
5a80bd07 HC |
1003 | trace_block_io_start(req); |
1004 | ||
e165fb4d CK |
1005 | if (blk_do_io_stat(req)) { |
1006 | /* | |
1007 | * All non-passthrough requests are created from a bio with one | |
1008 | * exception: when a flush command that is part of a flush sequence | |
1009 | * generated by the state machine in blk-flush.c is cloned onto the | |
1010 | * lower device by dm-multipath we can get here without a bio. | |
1011 | */ | |
1012 | if (req->bio) | |
1013 | req->part = req->bio->bi_bdev; | |
1014 | else | |
1015 | req->part = req->q->disk->part0; | |
1016 | ||
1017 | part_stat_lock(); | |
1018 | update_io_ticks(req->part, jiffies, false); | |
99dc4223 YK |
1019 | part_stat_local_inc(req->part, |
1020 | in_flight[op_is_write(req_op(req))]); | |
e165fb4d CK |
1021 | part_stat_unlock(); |
1022 | } | |
450b7879 CH |
1023 | } |
1024 | ||
f794f335 | 1025 | static inline void __blk_mq_end_request_acct(struct request *rq, u64 now) |
320ae51f | 1026 | { |
54bdd67d | 1027 | if (rq->rq_flags & RQF_STATS) |
522a7775 | 1028 | blk_stat_add(rq, now); |
4bc6339a | 1029 | |
87890092 | 1030 | blk_mq_sched_completed_request(rq, now); |
522a7775 | 1031 | blk_account_io_done(rq, now); |
f794f335 | 1032 | } |
522a7775 | 1033 | |
f794f335 JA |
1034 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
1035 | { | |
1036 | if (blk_mq_need_time_stamp(rq)) | |
08420cf7 | 1037 | __blk_mq_end_request_acct(rq, blk_time_get_ns()); |
0d11e6ac | 1038 | |
e5c0ca13 CZ |
1039 | blk_mq_finish_request(rq); |
1040 | ||
91b63639 | 1041 | if (rq->end_io) { |
a7905043 | 1042 | rq_qos_done(rq->q, rq); |
de671d61 JA |
1043 | if (rq->end_io(rq, error) == RQ_END_IO_FREE) |
1044 | blk_mq_free_request(rq); | |
91b63639 | 1045 | } else { |
320ae51f | 1046 | blk_mq_free_request(rq); |
91b63639 | 1047 | } |
320ae51f | 1048 | } |
c8a446ad | 1049 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 1050 | |
2a842aca | 1051 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
1052 | { |
1053 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
1054 | BUG(); | |
c8a446ad | 1055 | __blk_mq_end_request(rq, error); |
63151a44 | 1056 | } |
c8a446ad | 1057 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 1058 | |
f794f335 JA |
1059 | #define TAG_COMP_BATCH 32 |
1060 | ||
1061 | static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx, | |
1062 | int *tag_array, int nr_tags) | |
1063 | { | |
1064 | struct request_queue *q = hctx->queue; | |
1065 | ||
b8643d68 | 1066 | blk_mq_sub_active_requests(hctx, nr_tags); |
3b87c6ea | 1067 | |
f794f335 JA |
1068 | blk_mq_put_tags(hctx->tags, tag_array, nr_tags); |
1069 | percpu_ref_put_many(&q->q_usage_counter, nr_tags); | |
1070 | } | |
1071 | ||
1072 | void blk_mq_end_request_batch(struct io_comp_batch *iob) | |
1073 | { | |
1074 | int tags[TAG_COMP_BATCH], nr_tags = 0; | |
02f7eab0 | 1075 | struct blk_mq_hw_ctx *cur_hctx = NULL; |
f794f335 JA |
1076 | struct request *rq; |
1077 | u64 now = 0; | |
1078 | ||
1079 | if (iob->need_ts) | |
08420cf7 | 1080 | now = blk_time_get_ns(); |
f794f335 JA |
1081 | |
1082 | while ((rq = rq_list_pop(&iob->req_list)) != NULL) { | |
1083 | prefetch(rq->bio); | |
1084 | prefetch(rq->rq_next); | |
1085 | ||
5581a5dd | 1086 | blk_complete_request(rq); |
f794f335 JA |
1087 | if (iob->need_ts) |
1088 | __blk_mq_end_request_acct(rq, now); | |
1089 | ||
e5c0ca13 CZ |
1090 | blk_mq_finish_request(rq); |
1091 | ||
98b26a0e JA |
1092 | rq_qos_done(rq->q, rq); |
1093 | ||
ab3e1d3b JA |
1094 | /* |
1095 | * If end_io handler returns NONE, then it still has | |
1096 | * ownership of the request. | |
1097 | */ | |
1098 | if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE) | |
1099 | continue; | |
1100 | ||
f794f335 | 1101 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
0a467d0f | 1102 | if (!req_ref_put_and_test(rq)) |
f794f335 JA |
1103 | continue; |
1104 | ||
1105 | blk_crypto_free_request(rq); | |
1106 | blk_pm_mark_last_busy(rq); | |
f794f335 | 1107 | |
02f7eab0 JA |
1108 | if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) { |
1109 | if (cur_hctx) | |
1110 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); | |
f794f335 | 1111 | nr_tags = 0; |
02f7eab0 | 1112 | cur_hctx = rq->mq_hctx; |
f794f335 JA |
1113 | } |
1114 | tags[nr_tags++] = rq->tag; | |
f794f335 JA |
1115 | } |
1116 | ||
1117 | if (nr_tags) | |
02f7eab0 | 1118 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); |
f794f335 JA |
1119 | } |
1120 | EXPORT_SYMBOL_GPL(blk_mq_end_request_batch); | |
1121 | ||
f9ab4918 | 1122 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 1123 | { |
f9ab4918 SAS |
1124 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
1125 | struct request *rq, *next; | |
c3077b5d | 1126 | |
f9ab4918 | 1127 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 1128 | rq->q->mq_ops->complete(rq); |
320ae51f | 1129 | } |
320ae51f | 1130 | |
e68ac2b4 | 1131 | static __latent_entropy void blk_done_softirq(void) |
320ae51f | 1132 | { |
f9ab4918 | 1133 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
1134 | } |
1135 | ||
c3077b5d CH |
1136 | static int blk_softirq_cpu_dead(unsigned int cpu) |
1137 | { | |
f9ab4918 | 1138 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
1139 | return 0; |
1140 | } | |
1141 | ||
40d09b53 | 1142 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 1143 | { |
f9ab4918 | 1144 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
1145 | } |
1146 | ||
96339526 CH |
1147 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
1148 | { | |
1149 | int cpu = raw_smp_processor_id(); | |
1150 | ||
1151 | if (!IS_ENABLED(CONFIG_SMP) || | |
1152 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
1153 | return false; | |
71425189 SAS |
1154 | /* |
1155 | * With force threaded interrupts enabled, raising softirq from an SMP | |
1156 | * function call will always result in waking the ksoftirqd thread. | |
1157 | * This is probably worse than completing the request on a different | |
1158 | * cache domain. | |
1159 | */ | |
91cc470e | 1160 | if (force_irqthreads()) |
71425189 | 1161 | return false; |
96339526 | 1162 | |
af550e4c | 1163 | /* same CPU or cache domain and capacity? Complete locally */ |
96339526 CH |
1164 | if (cpu == rq->mq_ctx->cpu || |
1165 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
af550e4c QY |
1166 | cpus_share_cache(cpu, rq->mq_ctx->cpu) && |
1167 | cpus_equal_capacity(cpu, rq->mq_ctx->cpu))) | |
96339526 CH |
1168 | return false; |
1169 | ||
1170 | /* don't try to IPI to an offline CPU */ | |
1171 | return cpu_online(rq->mq_ctx->cpu); | |
1172 | } | |
1173 | ||
f9ab4918 SAS |
1174 | static void blk_mq_complete_send_ipi(struct request *rq) |
1175 | { | |
f9ab4918 SAS |
1176 | unsigned int cpu; |
1177 | ||
1178 | cpu = rq->mq_ctx->cpu; | |
660e802c CZ |
1179 | if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu))) |
1180 | smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu)); | |
f9ab4918 SAS |
1181 | } |
1182 | ||
1183 | static void blk_mq_raise_softirq(struct request *rq) | |
1184 | { | |
1185 | struct llist_head *list; | |
1186 | ||
1187 | preempt_disable(); | |
1188 | list = this_cpu_ptr(&blk_cpu_done); | |
1189 | if (llist_add(&rq->ipi_list, list)) | |
1190 | raise_softirq(BLOCK_SOFTIRQ); | |
1191 | preempt_enable(); | |
1192 | } | |
1193 | ||
40d09b53 | 1194 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 1195 | { |
af78ff7c | 1196 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 1197 | |
4ab32bf3 | 1198 | /* |
f168420c LS |
1199 | * For request which hctx has only one ctx mapping, |
1200 | * or a polled request, always complete locally, | |
1201 | * it's pointless to redirect the completion. | |
4ab32bf3 | 1202 | */ |
30654614 ET |
1203 | if ((rq->mq_hctx->nr_ctx == 1 && |
1204 | rq->mq_ctx->cpu == raw_smp_processor_id()) || | |
1205 | rq->cmd_flags & REQ_POLLED) | |
40d09b53 | 1206 | return false; |
38535201 | 1207 | |
96339526 | 1208 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
1209 | blk_mq_complete_send_ipi(rq); |
1210 | return true; | |
3d6efbf6 | 1211 | } |
40d09b53 | 1212 | |
f9ab4918 SAS |
1213 | if (rq->q->nr_hw_queues == 1) { |
1214 | blk_mq_raise_softirq(rq); | |
1215 | return true; | |
1216 | } | |
1217 | return false; | |
40d09b53 CH |
1218 | } |
1219 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
1220 | ||
1221 | /** | |
1222 | * blk_mq_complete_request - end I/O on a request | |
1223 | * @rq: the request being processed | |
1224 | * | |
1225 | * Description: | |
1226 | * Complete a request by scheduling the ->complete_rq operation. | |
1227 | **/ | |
1228 | void blk_mq_complete_request(struct request *rq) | |
1229 | { | |
1230 | if (!blk_mq_complete_request_remote(rq)) | |
1231 | rq->q->mq_ops->complete(rq); | |
320ae51f | 1232 | } |
15f73f5b | 1233 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 1234 | |
105663f7 AA |
1235 | /** |
1236 | * blk_mq_start_request - Start processing a request | |
1237 | * @rq: Pointer to request to be started | |
1238 | * | |
1239 | * Function used by device drivers to notify the block layer that a request | |
1240 | * is going to be processed now, so blk layer can do proper initializations | |
1241 | * such as starting the timeout timer. | |
1242 | */ | |
e2490073 | 1243 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
1244 | { |
1245 | struct request_queue *q = rq->q; | |
1246 | ||
a54895fa | 1247 | trace_block_rq_issue(rq); |
320ae51f | 1248 | |
847c5bcd KK |
1249 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) && |
1250 | !blk_rq_is_passthrough(rq)) { | |
08420cf7 | 1251 | rq->io_start_time_ns = blk_time_get_ns(); |
3d244306 | 1252 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 1253 | rq->rq_flags |= RQF_STATS; |
a7905043 | 1254 | rq_qos_issue(q, rq); |
cf43e6be JA |
1255 | } |
1256 | ||
1d9bd516 | 1257 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 1258 | |
1d9bd516 | 1259 | blk_add_timer(rq); |
12f5b931 | 1260 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
217b613a | 1261 | rq->mq_hctx->tags->rqs[rq->tag] = rq; |
49f5baa5 | 1262 | |
54d4e6ab | 1263 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) |
e9f5f44a CH |
1264 | blk_integrity_prepare(rq); |
1265 | ||
3e08773c | 1266 | if (rq->bio && rq->bio->bi_opf & REQ_POLLED) |
f6c80cff | 1267 | WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num); |
320ae51f | 1268 | } |
e2490073 | 1269 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 1270 | |
a327c341 ML |
1271 | /* |
1272 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple | |
1273 | * queues. This is important for md arrays to benefit from merging | |
1274 | * requests. | |
1275 | */ | |
1276 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
1277 | { | |
1278 | if (plug->multiple_queues) | |
1279 | return BLK_MAX_REQUEST_COUNT * 2; | |
1280 | return BLK_MAX_REQUEST_COUNT; | |
1281 | } | |
1282 | ||
1283 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) | |
1284 | { | |
1285 | struct request *last = rq_list_peek(&plug->mq_list); | |
1286 | ||
1287 | if (!plug->rq_count) { | |
1288 | trace_block_plug(rq->q); | |
1289 | } else if (plug->rq_count >= blk_plug_max_rq_count(plug) || | |
1290 | (!blk_queue_nomerges(rq->q) && | |
1291 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1292 | blk_mq_flush_plug_list(plug, false); | |
878eb6e4 | 1293 | last = NULL; |
a327c341 ML |
1294 | trace_block_plug(rq->q); |
1295 | } | |
1296 | ||
1297 | if (!plug->multiple_queues && last && last->q != rq->q) | |
1298 | plug->multiple_queues = true; | |
c6b7a3a2 ML |
1299 | /* |
1300 | * Any request allocated from sched tags can't be issued to | |
1301 | * ->queue_rqs() directly | |
1302 | */ | |
1303 | if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS)) | |
a327c341 ML |
1304 | plug->has_elevator = true; |
1305 | rq->rq_next = NULL; | |
1306 | rq_list_add(&plug->mq_list, rq); | |
1307 | plug->rq_count++; | |
1308 | } | |
1309 | ||
4054cff9 CH |
1310 | /** |
1311 | * blk_execute_rq_nowait - insert a request to I/O scheduler for execution | |
4054cff9 CH |
1312 | * @rq: request to insert |
1313 | * @at_head: insert request at head or tail of queue | |
4054cff9 CH |
1314 | * |
1315 | * Description: | |
1316 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1317 | * for execution. Don't wait for completion. | |
1318 | * | |
1319 | * Note: | |
1320 | * This function will invoke @done directly if the queue is dead. | |
1321 | */ | |
e2e53086 | 1322 | void blk_execute_rq_nowait(struct request *rq, bool at_head) |
4054cff9 | 1323 | { |
f0dbe6e8 CH |
1324 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
1325 | ||
ae948fd6 CH |
1326 | WARN_ON(irqs_disabled()); |
1327 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 | 1328 | |
ae948fd6 | 1329 | blk_account_io_start(rq); |
110fdb44 | 1330 | |
f0dbe6e8 | 1331 | if (current->plug && !at_head) { |
ae948fd6 | 1332 | blk_add_rq_to_plug(current->plug, rq); |
f0dbe6e8 CH |
1333 | return; |
1334 | } | |
1335 | ||
710fa378 | 1336 | blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); |
65a558f6 | 1337 | blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING); |
4054cff9 CH |
1338 | } |
1339 | EXPORT_SYMBOL_GPL(blk_execute_rq_nowait); | |
1340 | ||
32ac5a9b CH |
1341 | struct blk_rq_wait { |
1342 | struct completion done; | |
1343 | blk_status_t ret; | |
1344 | }; | |
1345 | ||
de671d61 | 1346 | static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret) |
32ac5a9b CH |
1347 | { |
1348 | struct blk_rq_wait *wait = rq->end_io_data; | |
1349 | ||
1350 | wait->ret = ret; | |
1351 | complete(&wait->done); | |
de671d61 | 1352 | return RQ_END_IO_NONE; |
32ac5a9b CH |
1353 | } |
1354 | ||
c6e99ea4 | 1355 | bool blk_rq_is_poll(struct request *rq) |
4054cff9 CH |
1356 | { |
1357 | if (!rq->mq_hctx) | |
1358 | return false; | |
1359 | if (rq->mq_hctx->type != HCTX_TYPE_POLL) | |
1360 | return false; | |
4054cff9 CH |
1361 | return true; |
1362 | } | |
c6e99ea4 | 1363 | EXPORT_SYMBOL_GPL(blk_rq_is_poll); |
4054cff9 CH |
1364 | |
1365 | static void blk_rq_poll_completion(struct request *rq, struct completion *wait) | |
1366 | { | |
1367 | do { | |
f6c80cff | 1368 | blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0); |
4054cff9 CH |
1369 | cond_resched(); |
1370 | } while (!completion_done(wait)); | |
1371 | } | |
1372 | ||
1373 | /** | |
1374 | * blk_execute_rq - insert a request into queue for execution | |
4054cff9 CH |
1375 | * @rq: request to insert |
1376 | * @at_head: insert request at head or tail of queue | |
1377 | * | |
1378 | * Description: | |
1379 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1380 | * for execution and wait for completion. | |
1381 | * Return: The blk_status_t result provided to blk_mq_end_request(). | |
1382 | */ | |
b84ba30b | 1383 | blk_status_t blk_execute_rq(struct request *rq, bool at_head) |
4054cff9 | 1384 | { |
f0dbe6e8 | 1385 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
32ac5a9b CH |
1386 | struct blk_rq_wait wait = { |
1387 | .done = COMPLETION_INITIALIZER_ONSTACK(wait.done), | |
1388 | }; | |
4054cff9 | 1389 | |
ae948fd6 CH |
1390 | WARN_ON(irqs_disabled()); |
1391 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
4054cff9 CH |
1392 | |
1393 | rq->end_io_data = &wait; | |
ae948fd6 | 1394 | rq->end_io = blk_end_sync_rq; |
4054cff9 | 1395 | |
ae948fd6 | 1396 | blk_account_io_start(rq); |
710fa378 | 1397 | blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); |
f0dbe6e8 | 1398 | blk_mq_run_hw_queue(hctx, false); |
4054cff9 | 1399 | |
0eb4db47 | 1400 | if (blk_rq_is_poll(rq)) |
32ac5a9b | 1401 | blk_rq_poll_completion(rq, &wait.done); |
0eb4db47 KB |
1402 | else |
1403 | blk_wait_io(&wait.done); | |
4054cff9 | 1404 | |
32ac5a9b | 1405 | return wait.ret; |
4054cff9 CH |
1406 | } |
1407 | EXPORT_SYMBOL(blk_execute_rq); | |
1408 | ||
ed0791b2 | 1409 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
1410 | { |
1411 | struct request_queue *q = rq->q; | |
1412 | ||
923218f6 ML |
1413 | blk_mq_put_driver_tag(rq); |
1414 | ||
a54895fa | 1415 | trace_block_rq_requeue(rq); |
a7905043 | 1416 | rq_qos_requeue(q, rq); |
49f5baa5 | 1417 | |
12f5b931 KB |
1418 | if (blk_mq_request_started(rq)) { |
1419 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 1420 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 1421 | } |
320ae51f JA |
1422 | } |
1423 | ||
2b053aca | 1424 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 1425 | { |
214a4418 | 1426 | struct request_queue *q = rq->q; |
9a67aa52 | 1427 | unsigned long flags; |
214a4418 | 1428 | |
ed0791b2 | 1429 | __blk_mq_requeue_request(rq); |
ed0791b2 | 1430 | |
105976f5 ML |
1431 | /* this request will be re-inserted to io scheduler queue */ |
1432 | blk_mq_sched_requeue_request(rq); | |
1433 | ||
9a67aa52 CH |
1434 | spin_lock_irqsave(&q->requeue_lock, flags); |
1435 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
1436 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
214a4418 CH |
1437 | |
1438 | if (kick_requeue_list) | |
1439 | blk_mq_kick_requeue_list(q); | |
ed0791b2 CH |
1440 | } |
1441 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
1442 | ||
6fca6a61 CH |
1443 | static void blk_mq_requeue_work(struct work_struct *work) |
1444 | { | |
1445 | struct request_queue *q = | |
2849450a | 1446 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 | 1447 | LIST_HEAD(rq_list); |
9a67aa52 CH |
1448 | LIST_HEAD(flush_list); |
1449 | struct request *rq; | |
6fca6a61 | 1450 | |
18e9781d | 1451 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 1452 | list_splice_init(&q->requeue_list, &rq_list); |
9a67aa52 | 1453 | list_splice_init(&q->flush_list, &flush_list); |
18e9781d | 1454 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 | 1455 | |
9a67aa52 CH |
1456 | while (!list_empty(&rq_list)) { |
1457 | rq = list_entry(rq_list.next, struct request, queuelist); | |
aef1897c | 1458 | /* |
a1e948b8 CH |
1459 | * If RQF_DONTPREP ist set, the request has been started by the |
1460 | * driver already and might have driver-specific data allocated | |
1461 | * already. Insert it into the hctx dispatch list to avoid | |
1462 | * block layer merges for the request. | |
aef1897c | 1463 | */ |
a1e948b8 | 1464 | if (rq->rq_flags & RQF_DONTPREP) { |
a1e948b8 | 1465 | list_del_init(&rq->queuelist); |
2b597613 | 1466 | blk_mq_request_bypass_insert(rq, 0); |
9a67aa52 | 1467 | } else { |
a1e948b8 | 1468 | list_del_init(&rq->queuelist); |
710fa378 | 1469 | blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD); |
a1e948b8 | 1470 | } |
6fca6a61 CH |
1471 | } |
1472 | ||
9a67aa52 CH |
1473 | while (!list_empty(&flush_list)) { |
1474 | rq = list_entry(flush_list.next, struct request, queuelist); | |
6fca6a61 | 1475 | list_del_init(&rq->queuelist); |
710fa378 | 1476 | blk_mq_insert_request(rq, 0); |
6fca6a61 CH |
1477 | } |
1478 | ||
52d7f1b5 | 1479 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
1480 | } |
1481 | ||
6fca6a61 CH |
1482 | void blk_mq_kick_requeue_list(struct request_queue *q) |
1483 | { | |
ae943d20 | 1484 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
1485 | } |
1486 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
1487 | ||
2849450a MS |
1488 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
1489 | unsigned long msecs) | |
1490 | { | |
d4acf365 BVA |
1491 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
1492 | msecs_to_jiffies(msecs)); | |
2849450a MS |
1493 | } |
1494 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
1495 | ||
0e4237ae ML |
1496 | static bool blk_is_flush_data_rq(struct request *rq) |
1497 | { | |
1498 | return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq); | |
1499 | } | |
1500 | ||
2dd6532e | 1501 | static bool blk_mq_rq_inflight(struct request *rq, void *priv) |
ae879912 JA |
1502 | { |
1503 | /* | |
8ab30a33 JG |
1504 | * If we find a request that isn't idle we know the queue is busy |
1505 | * as it's checked in the iter. | |
1506 | * Return false to stop the iteration. | |
0e4237ae ML |
1507 | * |
1508 | * In case of queue quiesce, if one flush data request is completed, | |
1509 | * don't count it as inflight given the flush sequence is suspended, | |
1510 | * and the original flush data request is invisible to driver, just | |
1511 | * like other pending requests because of quiesce | |
ae879912 | 1512 | */ |
0e4237ae ML |
1513 | if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) && |
1514 | blk_is_flush_data_rq(rq) && | |
1515 | blk_mq_request_completed(rq))) { | |
ae879912 JA |
1516 | bool *busy = priv; |
1517 | ||
1518 | *busy = true; | |
1519 | return false; | |
1520 | } | |
1521 | ||
1522 | return true; | |
1523 | } | |
1524 | ||
3c94d83c | 1525 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
1526 | { |
1527 | bool busy = false; | |
1528 | ||
3c94d83c | 1529 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
1530 | return busy; |
1531 | } | |
3c94d83c | 1532 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 1533 | |
9bdb4833 | 1534 | static void blk_mq_rq_timed_out(struct request *req) |
320ae51f | 1535 | { |
da661267 | 1536 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
1537 | if (req->q->mq_ops->timeout) { |
1538 | enum blk_eh_timer_return ret; | |
1539 | ||
9bdb4833 | 1540 | ret = req->q->mq_ops->timeout(req); |
d1210d5a CH |
1541 | if (ret == BLK_EH_DONE) |
1542 | return; | |
1543 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 1544 | } |
d1210d5a CH |
1545 | |
1546 | blk_add_timer(req); | |
87ee7b11 | 1547 | } |
5b3f25fc | 1548 | |
82c22947 DJ |
1549 | struct blk_expired_data { |
1550 | bool has_timedout_rq; | |
1551 | unsigned long next; | |
1552 | unsigned long timeout_start; | |
1553 | }; | |
1554 | ||
1555 | static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired) | |
81481eb4 | 1556 | { |
12f5b931 | 1557 | unsigned long deadline; |
87ee7b11 | 1558 | |
12f5b931 KB |
1559 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
1560 | return false; | |
da661267 CH |
1561 | if (rq->rq_flags & RQF_TIMED_OUT) |
1562 | return false; | |
a7af0af3 | 1563 | |
079076b3 | 1564 | deadline = READ_ONCE(rq->deadline); |
82c22947 | 1565 | if (time_after_eq(expired->timeout_start, deadline)) |
12f5b931 | 1566 | return true; |
a7af0af3 | 1567 | |
82c22947 DJ |
1568 | if (expired->next == 0) |
1569 | expired->next = deadline; | |
1570 | else if (time_after(expired->next, deadline)) | |
1571 | expired->next = deadline; | |
12f5b931 | 1572 | return false; |
87ee7b11 JA |
1573 | } |
1574 | ||
2e315dc0 ML |
1575 | void blk_mq_put_rq_ref(struct request *rq) |
1576 | { | |
de671d61 JA |
1577 | if (is_flush_rq(rq)) { |
1578 | if (rq->end_io(rq, 0) == RQ_END_IO_FREE) | |
1579 | blk_mq_free_request(rq); | |
1580 | } else if (req_ref_put_and_test(rq)) { | |
2e315dc0 | 1581 | __blk_mq_free_request(rq); |
de671d61 | 1582 | } |
2e315dc0 ML |
1583 | } |
1584 | ||
2dd6532e | 1585 | static bool blk_mq_check_expired(struct request *rq, void *priv) |
1d9bd516 | 1586 | { |
82c22947 | 1587 | struct blk_expired_data *expired = priv; |
12f5b931 KB |
1588 | |
1589 | /* | |
c797b40c ML |
1590 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
1591 | * be reallocated underneath the timeout handler's processing, then | |
1592 | * the expire check is reliable. If the request is not expired, then | |
1593 | * it was completed and reallocated as a new request after returning | |
1594 | * from blk_mq_check_expired(). | |
1d9bd516 | 1595 | */ |
82c22947 DJ |
1596 | if (blk_mq_req_expired(rq, expired)) { |
1597 | expired->has_timedout_rq = true; | |
1598 | return false; | |
1599 | } | |
1600 | return true; | |
1601 | } | |
1602 | ||
1603 | static bool blk_mq_handle_expired(struct request *rq, void *priv) | |
1604 | { | |
1605 | struct blk_expired_data *expired = priv; | |
1606 | ||
1607 | if (blk_mq_req_expired(rq, expired)) | |
9bdb4833 | 1608 | blk_mq_rq_timed_out(rq); |
7baa8572 | 1609 | return true; |
1d9bd516 TH |
1610 | } |
1611 | ||
287922eb | 1612 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 1613 | { |
287922eb CH |
1614 | struct request_queue *q = |
1615 | container_of(work, struct request_queue, timeout_work); | |
82c22947 DJ |
1616 | struct blk_expired_data expired = { |
1617 | .timeout_start = jiffies, | |
1618 | }; | |
1d9bd516 | 1619 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 1620 | unsigned long i; |
320ae51f | 1621 | |
71f79fb3 GKB |
1622 | /* A deadlock might occur if a request is stuck requiring a |
1623 | * timeout at the same time a queue freeze is waiting | |
1624 | * completion, since the timeout code would not be able to | |
1625 | * acquire the queue reference here. | |
1626 | * | |
1627 | * That's why we don't use blk_queue_enter here; instead, we use | |
1628 | * percpu_ref_tryget directly, because we need to be able to | |
1629 | * obtain a reference even in the short window between the queue | |
1630 | * starting to freeze, by dropping the first reference in | |
1671d522 | 1631 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
1632 | * consumed, marked by the instant q_usage_counter reaches |
1633 | * zero. | |
1634 | */ | |
1635 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
1636 | return; |
1637 | ||
82c22947 DJ |
1638 | /* check if there is any timed-out request */ |
1639 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired); | |
1640 | if (expired.has_timedout_rq) { | |
1641 | /* | |
1642 | * Before walking tags, we must ensure any submit started | |
1643 | * before the current time has finished. Since the submit | |
1644 | * uses srcu or rcu, wait for a synchronization point to | |
1645 | * ensure all running submits have finished | |
1646 | */ | |
483239c7 | 1647 | blk_mq_wait_quiesce_done(q->tag_set); |
82c22947 DJ |
1648 | |
1649 | expired.next = 0; | |
1650 | blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired); | |
1651 | } | |
320ae51f | 1652 | |
82c22947 DJ |
1653 | if (expired.next != 0) { |
1654 | mod_timer(&q->timeout, expired.next); | |
0d2602ca | 1655 | } else { |
fcd36c36 BVA |
1656 | /* |
1657 | * Request timeouts are handled as a forward rolling timer. If | |
1658 | * we end up here it means that no requests are pending and | |
1659 | * also that no request has been pending for a while. Mark | |
1660 | * each hctx as idle. | |
1661 | */ | |
f054b56c ML |
1662 | queue_for_each_hw_ctx(q, hctx, i) { |
1663 | /* the hctx may be unmapped, so check it here */ | |
1664 | if (blk_mq_hw_queue_mapped(hctx)) | |
1665 | blk_mq_tag_idle(hctx); | |
1666 | } | |
0d2602ca | 1667 | } |
287922eb | 1668 | blk_queue_exit(q); |
320ae51f JA |
1669 | } |
1670 | ||
88459642 OS |
1671 | struct flush_busy_ctx_data { |
1672 | struct blk_mq_hw_ctx *hctx; | |
1673 | struct list_head *list; | |
1674 | }; | |
1675 | ||
1676 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1677 | { | |
1678 | struct flush_busy_ctx_data *flush_data = data; | |
1679 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1680 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1681 | enum hctx_type type = hctx->type; |
88459642 | 1682 | |
88459642 | 1683 | spin_lock(&ctx->lock); |
c16d6b5a | 1684 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1685 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1686 | spin_unlock(&ctx->lock); |
1687 | return true; | |
1688 | } | |
1689 | ||
1429d7c9 JA |
1690 | /* |
1691 | * Process software queues that have been marked busy, splicing them | |
1692 | * to the for-dispatch | |
1693 | */ | |
2c3ad667 | 1694 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1695 | { |
88459642 OS |
1696 | struct flush_busy_ctx_data data = { |
1697 | .hctx = hctx, | |
1698 | .list = list, | |
1699 | }; | |
1429d7c9 | 1700 | |
88459642 | 1701 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1702 | } |
2c3ad667 | 1703 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1704 | |
b347689f ML |
1705 | struct dispatch_rq_data { |
1706 | struct blk_mq_hw_ctx *hctx; | |
1707 | struct request *rq; | |
1708 | }; | |
1709 | ||
1710 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1711 | void *data) | |
1712 | { | |
1713 | struct dispatch_rq_data *dispatch_data = data; | |
1714 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1715 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1716 | enum hctx_type type = hctx->type; |
b347689f ML |
1717 | |
1718 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1719 | if (!list_empty(&ctx->rq_lists[type])) { |
1720 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1721 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1722 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1723 | sbitmap_clear_bit(sb, bitnr); |
1724 | } | |
1725 | spin_unlock(&ctx->lock); | |
1726 | ||
1727 | return !dispatch_data->rq; | |
1728 | } | |
1729 | ||
1730 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1731 | struct blk_mq_ctx *start) | |
1732 | { | |
f31967f0 | 1733 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1734 | struct dispatch_rq_data data = { |
1735 | .hctx = hctx, | |
1736 | .rq = NULL, | |
1737 | }; | |
1738 | ||
1739 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1740 | dispatch_rq_from_ctx, &data); | |
1741 | ||
1742 | return data.rq; | |
1743 | } | |
1744 | ||
b8643d68 | 1745 | bool __blk_mq_alloc_driver_tag(struct request *rq) |
570e9b73 | 1746 | { |
ae0f1a73 | 1747 | struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1748 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1749 | int tag; |
1750 | ||
568f2700 ML |
1751 | blk_mq_tag_busy(rq->mq_hctx); |
1752 | ||
570e9b73 | 1753 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
ae0f1a73 | 1754 | bt = &rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1755 | tag_offset = 0; |
28500850 ML |
1756 | } else { |
1757 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1758 | return false; | |
570e9b73 ML |
1759 | } |
1760 | ||
570e9b73 ML |
1761 | tag = __sbitmap_queue_get(bt); |
1762 | if (tag == BLK_MQ_NO_TAG) | |
1763 | return false; | |
1764 | ||
1765 | rq->tag = tag + tag_offset; | |
b8643d68 | 1766 | blk_mq_inc_active_requests(rq->mq_hctx); |
568f2700 | 1767 | return true; |
570e9b73 ML |
1768 | } |
1769 | ||
eb619fdb JA |
1770 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1771 | int flags, void *key) | |
da55f2cc OS |
1772 | { |
1773 | struct blk_mq_hw_ctx *hctx; | |
1774 | ||
1775 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1776 | ||
5815839b | 1777 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1778 | if (!list_empty(&wait->entry)) { |
1779 | struct sbitmap_queue *sbq; | |
1780 | ||
1781 | list_del_init(&wait->entry); | |
ae0f1a73 | 1782 | sbq = &hctx->tags->bitmap_tags; |
e8618575 JA |
1783 | atomic_dec(&sbq->ws_active); |
1784 | } | |
5815839b ML |
1785 | spin_unlock(&hctx->dispatch_wait_lock); |
1786 | ||
da55f2cc OS |
1787 | blk_mq_run_hw_queue(hctx, true); |
1788 | return 1; | |
1789 | } | |
1790 | ||
f906a6a0 JA |
1791 | /* |
1792 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1793 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1794 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1795 | * marking us as waiting. |
1796 | */ | |
2278d69f | 1797 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1798 | struct request *rq) |
da55f2cc | 1799 | { |
98b99e94 | 1800 | struct sbitmap_queue *sbq; |
5815839b | 1801 | struct wait_queue_head *wq; |
f906a6a0 JA |
1802 | wait_queue_entry_t *wait; |
1803 | bool ret; | |
da55f2cc | 1804 | |
47df9ce9 KS |
1805 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
1806 | !(blk_mq_is_shared_tags(hctx->flags))) { | |
684b7324 | 1807 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1808 | |
c27d53fb BVA |
1809 | /* |
1810 | * It's possible that a tag was freed in the window between the | |
1811 | * allocation failure and adding the hardware queue to the wait | |
1812 | * queue. | |
1813 | * | |
1814 | * Don't clear RESTART here, someone else could have set it. | |
1815 | * At most this will cost an extra queue run. | |
1816 | */ | |
8ab6bb9e | 1817 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1818 | } |
eb619fdb | 1819 | |
2278d69f | 1820 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1821 | if (!list_empty_careful(&wait->entry)) |
1822 | return false; | |
1823 | ||
98b99e94 KS |
1824 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) |
1825 | sbq = &hctx->tags->breserved_tags; | |
1826 | else | |
1827 | sbq = &hctx->tags->bitmap_tags; | |
e8618575 | 1828 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1829 | |
1830 | spin_lock_irq(&wq->lock); | |
1831 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1832 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1833 | spin_unlock(&hctx->dispatch_wait_lock); |
1834 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1835 | return false; |
eb619fdb JA |
1836 | } |
1837 | ||
e8618575 | 1838 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1839 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1840 | __add_wait_queue(wq, wait); | |
c27d53fb | 1841 | |
5266caaf ML |
1842 | /* |
1843 | * Add one explicit barrier since blk_mq_get_driver_tag() may | |
1844 | * not imply barrier in case of failure. | |
1845 | * | |
1846 | * Order adding us to wait queue and allocating driver tag. | |
1847 | * | |
1848 | * The pair is the one implied in sbitmap_queue_wake_up() which | |
1849 | * orders clearing sbitmap tag bits and waitqueue_active() in | |
1850 | * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless | |
1851 | * | |
1852 | * Otherwise, re-order of adding wait queue and getting driver tag | |
1853 | * may cause __sbitmap_queue_wake_up() to wake up nothing because | |
1854 | * the waitqueue_active() may not observe us in wait queue. | |
1855 | */ | |
1856 | smp_mb(); | |
1857 | ||
da55f2cc | 1858 | /* |
eb619fdb JA |
1859 | * It's possible that a tag was freed in the window between the |
1860 | * allocation failure and adding the hardware queue to the wait | |
1861 | * queue. | |
da55f2cc | 1862 | */ |
8ab6bb9e | 1863 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1864 | if (!ret) { |
5815839b ML |
1865 | spin_unlock(&hctx->dispatch_wait_lock); |
1866 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1867 | return false; |
eb619fdb | 1868 | } |
c27d53fb BVA |
1869 | |
1870 | /* | |
1871 | * We got a tag, remove ourselves from the wait queue to ensure | |
1872 | * someone else gets the wakeup. | |
1873 | */ | |
c27d53fb | 1874 | list_del_init(&wait->entry); |
e8618575 | 1875 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1876 | spin_unlock(&hctx->dispatch_wait_lock); |
1877 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1878 | |
1879 | return true; | |
da55f2cc OS |
1880 | } |
1881 | ||
6e768717 ML |
1882 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1883 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1884 | /* | |
1885 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1886 | * - EWMA is one simple way to compute running average value | |
1887 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1888 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1889 | * factor doesn't matter because EWMA decreases exponentially | |
1890 | */ | |
1891 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1892 | { | |
1893 | unsigned int ewma; | |
1894 | ||
6e768717 ML |
1895 | ewma = hctx->dispatch_busy; |
1896 | ||
1897 | if (!ewma && !busy) | |
1898 | return; | |
1899 | ||
1900 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1901 | if (busy) | |
1902 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1903 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1904 | ||
1905 | hctx->dispatch_busy = ewma; | |
1906 | } | |
1907 | ||
86ff7c2a ML |
1908 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1909 | ||
c92a4103 JT |
1910 | static void blk_mq_handle_dev_resource(struct request *rq, |
1911 | struct list_head *list) | |
1912 | { | |
c92a4103 JT |
1913 | list_add(&rq->queuelist, list); |
1914 | __blk_mq_requeue_request(rq); | |
1915 | } | |
1916 | ||
75383524 ML |
1917 | enum prep_dispatch { |
1918 | PREP_DISPATCH_OK, | |
1919 | PREP_DISPATCH_NO_TAG, | |
1920 | PREP_DISPATCH_NO_BUDGET, | |
1921 | }; | |
1922 | ||
1923 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1924 | bool need_budget) | |
1925 | { | |
1926 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1927 | int budget_token = -1; |
75383524 | 1928 | |
2a5a24aa ML |
1929 | if (need_budget) { |
1930 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1931 | if (budget_token < 0) { | |
1932 | blk_mq_put_driver_tag(rq); | |
1933 | return PREP_DISPATCH_NO_BUDGET; | |
1934 | } | |
1935 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1936 | } |
1937 | ||
1938 | if (!blk_mq_get_driver_tag(rq)) { | |
1939 | /* | |
1940 | * The initial allocation attempt failed, so we need to | |
1941 | * rerun the hardware queue when a tag is freed. The | |
1942 | * waitqueue takes care of that. If the queue is run | |
1943 | * before we add this entry back on the dispatch list, | |
1944 | * we'll re-run it below. | |
1945 | */ | |
1946 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1947 | /* |
1948 | * All budgets not got from this function will be put | |
1949 | * together during handling partial dispatch | |
1950 | */ | |
1951 | if (need_budget) | |
2a5a24aa | 1952 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1953 | return PREP_DISPATCH_NO_TAG; |
1954 | } | |
1955 | } | |
1956 | ||
1957 | return PREP_DISPATCH_OK; | |
1958 | } | |
1959 | ||
1fd40b5e ML |
1960 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1961 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1962 | struct list_head *list) |
1fd40b5e | 1963 | { |
2a5a24aa | 1964 | struct request *rq; |
1fd40b5e | 1965 | |
2a5a24aa ML |
1966 | list_for_each_entry(rq, list, queuelist) { |
1967 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1968 | |
2a5a24aa ML |
1969 | if (budget_token >= 0) |
1970 | blk_mq_put_dispatch_budget(q, budget_token); | |
1971 | } | |
1fd40b5e ML |
1972 | } |
1973 | ||
34c9f547 KS |
1974 | /* |
1975 | * blk_mq_commit_rqs will notify driver using bd->last that there is no | |
1976 | * more requests. (See comment in struct blk_mq_ops for commit_rqs for | |
1977 | * details) | |
1978 | * Attention, we should explicitly call this in unusual cases: | |
1979 | * 1) did not queue everything initially scheduled to queue | |
1980 | * 2) the last attempt to queue a request failed | |
1981 | */ | |
1982 | static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued, | |
1983 | bool from_schedule) | |
1984 | { | |
1985 | if (hctx->queue->mq_ops->commit_rqs && queued) { | |
1986 | trace_block_unplug(hctx->queue, queued, !from_schedule); | |
1987 | hctx->queue->mq_ops->commit_rqs(hctx); | |
1988 | } | |
1989 | } | |
1990 | ||
1f57f8d4 JA |
1991 | /* |
1992 | * Returns true if we did some work AND can potentially do more. | |
1993 | */ | |
445874e8 | 1994 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 1995 | unsigned int nr_budgets) |
320ae51f | 1996 | { |
75383524 | 1997 | enum prep_dispatch prep; |
445874e8 | 1998 | struct request_queue *q = hctx->queue; |
f1ce99f7 | 1999 | struct request *rq; |
4ea58fe4 | 2000 | int queued; |
86ff7c2a | 2001 | blk_status_t ret = BLK_STS_OK; |
9586e67b | 2002 | bool needs_resource = false; |
320ae51f | 2003 | |
81380ca1 OS |
2004 | if (list_empty(list)) |
2005 | return false; | |
2006 | ||
320ae51f JA |
2007 | /* |
2008 | * Now process all the entries, sending them to the driver. | |
2009 | */ | |
4ea58fe4 | 2010 | queued = 0; |
81380ca1 | 2011 | do { |
74c45052 | 2012 | struct blk_mq_queue_data bd; |
320ae51f | 2013 | |
f04c3df3 | 2014 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 2015 | |
445874e8 | 2016 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 2017 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 2018 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 2019 | break; |
de148297 | 2020 | |
320ae51f | 2021 | list_del_init(&rq->queuelist); |
320ae51f | 2022 | |
74c45052 | 2023 | bd.rq = rq; |
f1ce99f7 | 2024 | bd.last = list_empty(list); |
74c45052 | 2025 | |
1fd40b5e ML |
2026 | /* |
2027 | * once the request is queued to lld, no need to cover the | |
2028 | * budget any more | |
2029 | */ | |
2030 | if (nr_budgets) | |
2031 | nr_budgets--; | |
74c45052 | 2032 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
2033 | switch (ret) { |
2034 | case BLK_STS_OK: | |
2035 | queued++; | |
320ae51f | 2036 | break; |
7bf13729 | 2037 | case BLK_STS_RESOURCE: |
9586e67b NA |
2038 | needs_resource = true; |
2039 | fallthrough; | |
7bf13729 ML |
2040 | case BLK_STS_DEV_RESOURCE: |
2041 | blk_mq_handle_dev_resource(rq, list); | |
2042 | goto out; | |
7bf13729 | 2043 | default: |
e21ee5a6 | 2044 | blk_mq_end_request(rq, ret); |
320ae51f | 2045 | } |
81380ca1 | 2046 | } while (!list_empty(list)); |
7bf13729 | 2047 | out: |
632bfb63 | 2048 | /* If we didn't flush the entire list, we could have told the driver |
2049 | * there was more coming, but that turned out to be a lie. | |
2050 | */ | |
e4ef2e05 KS |
2051 | if (!list_empty(list) || ret != BLK_STS_OK) |
2052 | blk_mq_commit_rqs(hctx, queued, false); | |
2053 | ||
320ae51f JA |
2054 | /* |
2055 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
2056 | * that is where we will continue on next queue run. | |
2057 | */ | |
f04c3df3 | 2058 | if (!list_empty(list)) { |
86ff7c2a | 2059 | bool needs_restart; |
75383524 ML |
2060 | /* For non-shared tags, the RESTART check will suffice */ |
2061 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
47df9ce9 KS |
2062 | ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) || |
2063 | blk_mq_is_shared_tags(hctx->flags)); | |
86ff7c2a | 2064 | |
2a5a24aa ML |
2065 | if (nr_budgets) |
2066 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 2067 | |
320ae51f | 2068 | spin_lock(&hctx->lock); |
01e99aec | 2069 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 2070 | spin_unlock(&hctx->lock); |
f04c3df3 | 2071 | |
d7d8535f ML |
2072 | /* |
2073 | * Order adding requests to hctx->dispatch and checking | |
2074 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
2075 | * in blk_mq_sched_restart(). Avoid restart code path to | |
2076 | * miss the new added requests to hctx->dispatch, meantime | |
2077 | * SCHED_RESTART is observed here. | |
2078 | */ | |
2079 | smp_mb(); | |
2080 | ||
9ba52e58 | 2081 | /* |
710c785f BVA |
2082 | * If SCHED_RESTART was set by the caller of this function and |
2083 | * it is no longer set that means that it was cleared by another | |
2084 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 2085 | * |
eb619fdb JA |
2086 | * If 'no_tag' is set, that means that we failed getting |
2087 | * a driver tag with an I/O scheduler attached. If our dispatch | |
2088 | * waitqueue is no longer active, ensure that we run the queue | |
2089 | * AFTER adding our entries back to the list. | |
bd166ef1 | 2090 | * |
710c785f BVA |
2091 | * If no I/O scheduler has been configured it is possible that |
2092 | * the hardware queue got stopped and restarted before requests | |
2093 | * were pushed back onto the dispatch list. Rerun the queue to | |
2094 | * avoid starvation. Notes: | |
2095 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
2096 | * been stopped before rerunning a queue. | |
2097 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 2098 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 2099 | * and dm-rq. |
86ff7c2a ML |
2100 | * |
2101 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
2102 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 | 2103 | * that could otherwise occur if the queue is idle. We'll do |
9586e67b NA |
2104 | * similar if we couldn't get budget or couldn't lock a zone |
2105 | * and SCHED_RESTART is set. | |
bd166ef1 | 2106 | */ |
86ff7c2a | 2107 | needs_restart = blk_mq_sched_needs_restart(hctx); |
9586e67b NA |
2108 | if (prep == PREP_DISPATCH_NO_BUDGET) |
2109 | needs_resource = true; | |
86ff7c2a | 2110 | if (!needs_restart || |
eb619fdb | 2111 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 2112 | blk_mq_run_hw_queue(hctx, true); |
6d5e8d21 | 2113 | else if (needs_resource) |
86ff7c2a | 2114 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 2115 | |
6e768717 | 2116 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 2117 | return false; |
4ea58fe4 | 2118 | } |
f04c3df3 | 2119 | |
4ea58fe4 KS |
2120 | blk_mq_update_dispatch_busy(hctx, false); |
2121 | return true; | |
f04c3df3 JA |
2122 | } |
2123 | ||
f82ddf19 ML |
2124 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
2125 | { | |
2126 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
2127 | ||
2128 | if (cpu >= nr_cpu_ids) | |
2129 | cpu = cpumask_first(hctx->cpumask); | |
2130 | return cpu; | |
2131 | } | |
2132 | ||
a46c2702 ML |
2133 | /* |
2134 | * ->next_cpu is always calculated from hctx->cpumask, so simply use | |
2135 | * it for speeding up the check | |
2136 | */ | |
2137 | static bool blk_mq_hctx_empty_cpumask(struct blk_mq_hw_ctx *hctx) | |
2138 | { | |
2139 | return hctx->next_cpu >= nr_cpu_ids; | |
2140 | } | |
2141 | ||
506e931f JA |
2142 | /* |
2143 | * It'd be great if the workqueue API had a way to pass | |
2144 | * in a mask and had some smarts for more clever placement. | |
2145 | * For now we just round-robin here, switching for every | |
2146 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
2147 | */ | |
2148 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
2149 | { | |
7bed4595 | 2150 | bool tried = false; |
476f8c98 | 2151 | int next_cpu = hctx->next_cpu; |
7bed4595 | 2152 | |
a46c2702 ML |
2153 | /* Switch to unbound if no allowable CPUs in this hctx */ |
2154 | if (hctx->queue->nr_hw_queues == 1 || blk_mq_hctx_empty_cpumask(hctx)) | |
b657d7e6 | 2155 | return WORK_CPU_UNBOUND; |
506e931f JA |
2156 | |
2157 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 2158 | select_cpu: |
476f8c98 | 2159 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 2160 | cpu_online_mask); |
506e931f | 2161 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 2162 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2163 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2164 | } | |
2165 | ||
7bed4595 ML |
2166 | /* |
2167 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
2168 | * and it should only happen in the path of handling CPU DEAD. | |
2169 | */ | |
476f8c98 | 2170 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
2171 | if (!tried) { |
2172 | tried = true; | |
2173 | goto select_cpu; | |
2174 | } | |
2175 | ||
2176 | /* | |
2177 | * Make sure to re-select CPU next time once after CPUs | |
2178 | * in hctx->cpumask become online again. | |
2179 | */ | |
476f8c98 | 2180 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
2181 | hctx->next_cpu_batch = 1; |
2182 | return WORK_CPU_UNBOUND; | |
2183 | } | |
476f8c98 ML |
2184 | |
2185 | hctx->next_cpu = next_cpu; | |
2186 | return next_cpu; | |
506e931f JA |
2187 | } |
2188 | ||
105663f7 | 2189 | /** |
1aa8d875 | 2190 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. |
105663f7 | 2191 | * @hctx: Pointer to the hardware queue to run. |
fa94ba8a | 2192 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 | 2193 | * |
1aa8d875 | 2194 | * Run a hardware queue asynchronously with a delay of @msecs. |
105663f7 | 2195 | */ |
1aa8d875 | 2196 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
320ae51f | 2197 | { |
5435c023 | 2198 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f | 2199 | return; |
ae943d20 BVA |
2200 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
2201 | msecs_to_jiffies(msecs)); | |
7587a5ae | 2202 | } |
7587a5ae BVA |
2203 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); |
2204 | ||
105663f7 AA |
2205 | /** |
2206 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
2207 | * @hctx: Pointer to the hardware queue to run. | |
2208 | * @async: If we want to run the queue asynchronously. | |
2209 | * | |
2210 | * Check if the request queue is not in a quiesced state and if there are | |
2211 | * pending requests to be sent. If this is true, run the queue to send requests | |
2212 | * to hardware. | |
2213 | */ | |
626fb735 | 2214 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 2215 | { |
24f5a90f ML |
2216 | bool need_run; |
2217 | ||
4d5bba5b CH |
2218 | /* |
2219 | * We can't run the queue inline with interrupts disabled. | |
2220 | */ | |
2221 | WARN_ON_ONCE(!async && in_interrupt()); | |
2222 | ||
65a558f6 BVA |
2223 | might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING); |
2224 | ||
24f5a90f ML |
2225 | /* |
2226 | * When queue is quiesced, we may be switching io scheduler, or | |
2227 | * updating nr_hw_queues, or other things, and we can't run queue | |
2228 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
2229 | * | |
2230 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
2231 | * quiesced. | |
2232 | */ | |
41adf531 | 2233 | __blk_mq_run_dispatch_ops(hctx->queue, false, |
2a904d00 ML |
2234 | need_run = !blk_queue_quiesced(hctx->queue) && |
2235 | blk_mq_hctx_has_pending(hctx)); | |
24f5a90f | 2236 | |
1aa8d875 CH |
2237 | if (!need_run) |
2238 | return; | |
2239 | ||
65a558f6 | 2240 | if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) { |
1aa8d875 CH |
2241 | blk_mq_delay_run_hw_queue(hctx, 0); |
2242 | return; | |
2243 | } | |
2244 | ||
4d5bba5b CH |
2245 | blk_mq_run_dispatch_ops(hctx->queue, |
2246 | blk_mq_sched_dispatch_requests(hctx)); | |
320ae51f | 2247 | } |
5b727272 | 2248 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 2249 | |
b6e68ee8 JK |
2250 | /* |
2251 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
2252 | * scheduler. | |
2253 | */ | |
2254 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
2255 | { | |
5d05426e | 2256 | struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); |
b6e68ee8 JK |
2257 | /* |
2258 | * If the IO scheduler does not respect hardware queues when | |
2259 | * dispatching, we just don't bother with multiple HW queues and | |
2260 | * dispatch from hctx for the current CPU since running multiple queues | |
2261 | * just causes lock contention inside the scheduler and pointless cache | |
2262 | * bouncing. | |
2263 | */ | |
51ab80f0 | 2264 | struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT]; |
5d05426e | 2265 | |
b6e68ee8 JK |
2266 | if (!blk_mq_hctx_stopped(hctx)) |
2267 | return hctx; | |
2268 | return NULL; | |
2269 | } | |
2270 | ||
105663f7 | 2271 | /** |
24f7bb88 | 2272 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
2273 | * @q: Pointer to the request queue to run. |
2274 | * @async: If we want to run the queue asynchronously. | |
2275 | */ | |
b94ec296 | 2276 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 2277 | { |
b6e68ee8 | 2278 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2279 | unsigned long i; |
320ae51f | 2280 | |
b6e68ee8 | 2281 | sq_hctx = NULL; |
4d337ceb | 2282 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2283 | sq_hctx = blk_mq_get_sq_hctx(q); |
320ae51f | 2284 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 2285 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 2286 | continue; |
b6e68ee8 JK |
2287 | /* |
2288 | * Dispatch from this hctx either if there's no hctx preferred | |
2289 | * by IO scheduler or if it has requests that bypass the | |
2290 | * scheduler. | |
2291 | */ | |
2292 | if (!sq_hctx || sq_hctx == hctx || | |
2293 | !list_empty_careful(&hctx->dispatch)) | |
2294 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
2295 | } |
2296 | } | |
b94ec296 | 2297 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 2298 | |
b9151e7b DA |
2299 | /** |
2300 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
2301 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 2302 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
2303 | */ |
2304 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
2305 | { | |
b6e68ee8 | 2306 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
4f481208 | 2307 | unsigned long i; |
b9151e7b | 2308 | |
b6e68ee8 | 2309 | sq_hctx = NULL; |
4d337ceb | 2310 | if (blk_queue_sq_sched(q)) |
b6e68ee8 | 2311 | sq_hctx = blk_mq_get_sq_hctx(q); |
b9151e7b DA |
2312 | queue_for_each_hw_ctx(q, hctx, i) { |
2313 | if (blk_mq_hctx_stopped(hctx)) | |
2314 | continue; | |
8f5fea65 DJ |
2315 | /* |
2316 | * If there is already a run_work pending, leave the | |
2317 | * pending delay untouched. Otherwise, a hctx can stall | |
2318 | * if another hctx is re-delaying the other's work | |
2319 | * before the work executes. | |
2320 | */ | |
2321 | if (delayed_work_pending(&hctx->run_work)) | |
2322 | continue; | |
b6e68ee8 JK |
2323 | /* |
2324 | * Dispatch from this hctx either if there's no hctx preferred | |
2325 | * by IO scheduler or if it has requests that bypass the | |
2326 | * scheduler. | |
2327 | */ | |
2328 | if (!sq_hctx || sq_hctx == hctx || | |
2329 | !list_empty_careful(&hctx->dispatch)) | |
2330 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
2331 | } |
2332 | } | |
2333 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
2334 | ||
39a70c76 ML |
2335 | /* |
2336 | * This function is often used for pausing .queue_rq() by driver when | |
2337 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2338 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2339 | * |
2340 | * We do not guarantee that dispatch can be drained or blocked | |
2341 | * after blk_mq_stop_hw_queue() returns. Please use | |
2342 | * blk_mq_quiesce_queue() for that requirement. | |
2343 | */ | |
2719aa21 JA |
2344 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
2345 | { | |
641a9ed6 | 2346 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 2347 | |
641a9ed6 | 2348 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 2349 | } |
641a9ed6 | 2350 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 2351 | |
39a70c76 ML |
2352 | /* |
2353 | * This function is often used for pausing .queue_rq() by driver when | |
2354 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2355 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2356 | * |
2357 | * We do not guarantee that dispatch can be drained or blocked | |
2358 | * after blk_mq_stop_hw_queues() returns. Please use | |
2359 | * blk_mq_quiesce_queue() for that requirement. | |
2360 | */ | |
2719aa21 JA |
2361 | void blk_mq_stop_hw_queues(struct request_queue *q) |
2362 | { | |
641a9ed6 | 2363 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 2364 | unsigned long i; |
641a9ed6 ML |
2365 | |
2366 | queue_for_each_hw_ctx(q, hctx, i) | |
2367 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
2368 | } |
2369 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
2370 | ||
320ae51f JA |
2371 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
2372 | { | |
2373 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 2374 | |
65a558f6 | 2375 | blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING); |
320ae51f JA |
2376 | } |
2377 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
2378 | ||
2f268556 CH |
2379 | void blk_mq_start_hw_queues(struct request_queue *q) |
2380 | { | |
2381 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2382 | unsigned long i; |
2f268556 CH |
2383 | |
2384 | queue_for_each_hw_ctx(q, hctx, i) | |
2385 | blk_mq_start_hw_queue(hctx); | |
2386 | } | |
2387 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
2388 | ||
ae911c5e JA |
2389 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
2390 | { | |
2391 | if (!blk_mq_hctx_stopped(hctx)) | |
2392 | return; | |
2393 | ||
2394 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
2395 | blk_mq_run_hw_queue(hctx, async); | |
2396 | } | |
2397 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
2398 | ||
1b4a3258 | 2399 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
2400 | { |
2401 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 2402 | unsigned long i; |
320ae51f | 2403 | |
ae911c5e | 2404 | queue_for_each_hw_ctx(q, hctx, i) |
65a558f6 BVA |
2405 | blk_mq_start_stopped_hw_queue(hctx, async || |
2406 | (hctx->flags & BLK_MQ_F_BLOCKING)); | |
320ae51f JA |
2407 | } |
2408 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
2409 | ||
70f4db63 | 2410 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f | 2411 | { |
c20a1a2c CH |
2412 | struct blk_mq_hw_ctx *hctx = |
2413 | container_of(work, struct blk_mq_hw_ctx, run_work.work); | |
7b607814 | 2414 | |
4d5bba5b CH |
2415 | blk_mq_run_dispatch_ops(hctx->queue, |
2416 | blk_mq_sched_dispatch_requests(hctx)); | |
320ae51f JA |
2417 | } |
2418 | ||
105663f7 AA |
2419 | /** |
2420 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
2421 | * @rq: Pointer to request to be inserted. | |
2b597613 | 2422 | * @flags: BLK_MQ_INSERT_* |
105663f7 | 2423 | * |
157f377b JA |
2424 | * Should only be used carefully, when the caller knows we want to |
2425 | * bypass a potential IO scheduler on the target device. | |
2426 | */ | |
360f2648 | 2427 | static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags) |
157f377b | 2428 | { |
ea4f995e | 2429 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
2430 | |
2431 | spin_lock(&hctx->lock); | |
2b597613 | 2432 | if (flags & BLK_MQ_INSERT_AT_HEAD) |
01e99aec ML |
2433 | list_add(&rq->queuelist, &hctx->dispatch); |
2434 | else | |
2435 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b | 2436 | spin_unlock(&hctx->lock); |
157f377b JA |
2437 | } |
2438 | ||
05a93117 CH |
2439 | static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, |
2440 | struct blk_mq_ctx *ctx, struct list_head *list, | |
2441 | bool run_queue_async) | |
320ae51f | 2442 | { |
3f0cedc7 | 2443 | struct request *rq; |
c16d6b5a | 2444 | enum hctx_type type = hctx->type; |
3f0cedc7 | 2445 | |
94aa228c CH |
2446 | /* |
2447 | * Try to issue requests directly if the hw queue isn't busy to save an | |
2448 | * extra enqueue & dequeue to the sw queue. | |
2449 | */ | |
2450 | if (!hctx->dispatch_busy && !run_queue_async) { | |
2451 | blk_mq_run_dispatch_ops(hctx->queue, | |
2452 | blk_mq_try_issue_list_directly(hctx, list)); | |
2453 | if (list_empty(list)) | |
2454 | goto out; | |
2455 | } | |
2456 | ||
320ae51f JA |
2457 | /* |
2458 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
2459 | * offline now | |
2460 | */ | |
3f0cedc7 | 2461 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 2462 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 2463 | trace_block_rq_insert(rq); |
65a558f6 BVA |
2464 | if (rq->cmd_flags & REQ_NOWAIT) |
2465 | run_queue_async = true; | |
320ae51f | 2466 | } |
3f0cedc7 ML |
2467 | |
2468 | spin_lock(&ctx->lock); | |
c16d6b5a | 2469 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 2470 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 2471 | spin_unlock(&ctx->lock); |
94aa228c CH |
2472 | out: |
2473 | blk_mq_run_hw_queue(hctx, run_queue_async); | |
320ae51f JA |
2474 | } |
2475 | ||
710fa378 | 2476 | static void blk_mq_insert_request(struct request *rq, blk_insert_t flags) |
2bd215df CH |
2477 | { |
2478 | struct request_queue *q = rq->q; | |
2bd215df CH |
2479 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
2480 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2481 | ||
53548d2a CH |
2482 | if (blk_rq_is_passthrough(rq)) { |
2483 | /* | |
2484 | * Passthrough request have to be added to hctx->dispatch | |
2485 | * directly. The device may be in a situation where it can't | |
2486 | * handle FS request, and always returns BLK_STS_RESOURCE for | |
2487 | * them, which gets them added to hctx->dispatch. | |
2488 | * | |
2489 | * If a passthrough request is required to unblock the queues, | |
2490 | * and it is added to the scheduler queue, there is no chance to | |
2491 | * dispatch it given we prioritize requests in hctx->dispatch. | |
2492 | */ | |
2b597613 | 2493 | blk_mq_request_bypass_insert(rq, flags); |
be4c4278 | 2494 | } else if (req_op(rq) == REQ_OP_FLUSH) { |
2bd215df CH |
2495 | /* |
2496 | * Firstly normal IO request is inserted to scheduler queue or | |
2497 | * sw queue, meantime we add flush request to dispatch queue( | |
2498 | * hctx->dispatch) directly and there is at most one in-flight | |
2499 | * flush request for each hw queue, so it doesn't matter to add | |
2500 | * flush request to tail or front of the dispatch queue. | |
2501 | * | |
2502 | * Secondly in case of NCQ, flush request belongs to non-NCQ | |
2503 | * command, and queueing it will fail when there is any | |
2504 | * in-flight normal IO request(NCQ command). When adding flush | |
2505 | * rq to the front of hctx->dispatch, it is easier to introduce | |
2506 | * extra time to flush rq's latency because of S_SCHED_RESTART | |
2507 | * compared with adding to the tail of dispatch queue, then | |
2508 | * chance of flush merge is increased, and less flush requests | |
2509 | * will be issued to controller. It is observed that ~10% time | |
2510 | * is saved in blktests block/004 on disk attached to AHCI/NCQ | |
2511 | * drive when adding flush rq to the front of hctx->dispatch. | |
2512 | * | |
2513 | * Simply queue flush rq to the front of hctx->dispatch so that | |
2514 | * intensive flush workloads can benefit in case of NCQ HW. | |
2515 | */ | |
2b597613 | 2516 | blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD); |
53548d2a | 2517 | } else if (q->elevator) { |
2bd215df CH |
2518 | LIST_HEAD(list); |
2519 | ||
53548d2a CH |
2520 | WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG); |
2521 | ||
2bd215df | 2522 | list_add(&rq->queuelist, &list); |
93fffe16 | 2523 | q->elevator->type->ops.insert_requests(hctx, &list, flags); |
2bd215df | 2524 | } else { |
4ec5c055 CH |
2525 | trace_block_rq_insert(rq); |
2526 | ||
2bd215df | 2527 | spin_lock(&ctx->lock); |
710fa378 | 2528 | if (flags & BLK_MQ_INSERT_AT_HEAD) |
4ec5c055 CH |
2529 | list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]); |
2530 | else | |
2531 | list_add_tail(&rq->queuelist, | |
2532 | &ctx->rq_lists[hctx->type]); | |
a88db1e0 | 2533 | blk_mq_hctx_mark_pending(hctx, ctx); |
2bd215df CH |
2534 | spin_unlock(&ctx->lock); |
2535 | } | |
320ae51f JA |
2536 | } |
2537 | ||
14ccb66b CH |
2538 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
2539 | unsigned int nr_segs) | |
320ae51f | 2540 | { |
93f221ae EB |
2541 | int err; |
2542 | ||
f924cdde CH |
2543 | if (bio->bi_opf & REQ_RAHEAD) |
2544 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
2545 | ||
2546 | rq->__sector = bio->bi_iter.bi_sector; | |
44981351 | 2547 | rq->write_hint = bio->bi_write_hint; |
14ccb66b | 2548 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
2549 | |
2550 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
2551 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
2552 | WARN_ON_ONCE(err); | |
4b570521 | 2553 | |
b5af37ab | 2554 | blk_account_io_start(rq); |
320ae51f JA |
2555 | } |
2556 | ||
0f95549c | 2557 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2558 | struct request *rq, bool last) |
f984df1f | 2559 | { |
f984df1f | 2560 | struct request_queue *q = rq->q; |
f984df1f SL |
2561 | struct blk_mq_queue_data bd = { |
2562 | .rq = rq, | |
be94f058 | 2563 | .last = last, |
f984df1f | 2564 | }; |
f06345ad | 2565 | blk_status_t ret; |
0f95549c | 2566 | |
0f95549c MS |
2567 | /* |
2568 | * For OK queue, we are done. For error, caller may kill it. | |
2569 | * Any other error (busy), just add it to our list as we | |
2570 | * previously would have done. | |
2571 | */ | |
2572 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
2573 | switch (ret) { | |
2574 | case BLK_STS_OK: | |
6ce3dd6e | 2575 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2576 | break; |
2577 | case BLK_STS_RESOURCE: | |
86ff7c2a | 2578 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 2579 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
2580 | __blk_mq_requeue_request(rq); |
2581 | break; | |
2582 | default: | |
6ce3dd6e | 2583 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2584 | break; |
2585 | } | |
2586 | ||
2587 | return ret; | |
2588 | } | |
2589 | ||
2b71b877 | 2590 | static bool blk_mq_get_budget_and_tag(struct request *rq) |
0f95549c | 2591 | { |
2a5a24aa | 2592 | int budget_token; |
d964f04a | 2593 | |
2b71b877 | 2594 | budget_token = blk_mq_get_dispatch_budget(rq->q); |
2a5a24aa | 2595 | if (budget_token < 0) |
2b71b877 | 2596 | return false; |
2a5a24aa | 2597 | blk_mq_set_rq_budget_token(rq, budget_token); |
8ab6bb9e | 2598 | if (!blk_mq_get_driver_tag(rq)) { |
2b71b877 CH |
2599 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
2600 | return false; | |
88022d72 | 2601 | } |
2b71b877 | 2602 | return true; |
fd9c40f6 BVA |
2603 | } |
2604 | ||
105663f7 AA |
2605 | /** |
2606 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2607 | * @hctx: Pointer of the associated hardware queue. | |
2608 | * @rq: Pointer to request to be sent. | |
105663f7 AA |
2609 | * |
2610 | * If the device has enough resources to accept a new request now, send the | |
2611 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2612 | * we can try send it another time in the future. Requests inserted at this | |
2613 | * queue have higher priority. | |
2614 | */ | |
fd9c40f6 | 2615 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2616 | struct request *rq) |
fd9c40f6 | 2617 | { |
e1f44ac0 CH |
2618 | blk_status_t ret; |
2619 | ||
2620 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { | |
710fa378 | 2621 | blk_mq_insert_request(rq, 0); |
e1f44ac0 CH |
2622 | return; |
2623 | } | |
2624 | ||
dd6216bb | 2625 | if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) { |
710fa378 | 2626 | blk_mq_insert_request(rq, 0); |
65a558f6 | 2627 | blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT); |
e1f44ac0 CH |
2628 | return; |
2629 | } | |
fd9c40f6 | 2630 | |
e1f44ac0 CH |
2631 | ret = __blk_mq_issue_directly(hctx, rq, true); |
2632 | switch (ret) { | |
2633 | case BLK_STS_OK: | |
2634 | break; | |
2635 | case BLK_STS_RESOURCE: | |
2636 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2637 | blk_mq_request_bypass_insert(rq, 0); |
2394395c | 2638 | blk_mq_run_hw_queue(hctx, false); |
e1f44ac0 CH |
2639 | break; |
2640 | default: | |
fd9c40f6 | 2641 | blk_mq_end_request(rq, ret); |
e1f44ac0 CH |
2642 | break; |
2643 | } | |
fd9c40f6 BVA |
2644 | } |
2645 | ||
06c8c691 | 2646 | static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) |
fd9c40f6 | 2647 | { |
e1f44ac0 CH |
2648 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
2649 | ||
2650 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { | |
710fa378 | 2651 | blk_mq_insert_request(rq, 0); |
e1f44ac0 CH |
2652 | return BLK_STS_OK; |
2653 | } | |
2654 | ||
2655 | if (!blk_mq_get_budget_and_tag(rq)) | |
2656 | return BLK_STS_RESOURCE; | |
2657 | return __blk_mq_issue_directly(hctx, rq, last); | |
5eb6126e CH |
2658 | } |
2659 | ||
3e368fb0 | 2660 | static void blk_mq_plug_issue_direct(struct blk_plug *plug) |
b84c5b50 CH |
2661 | { |
2662 | struct blk_mq_hw_ctx *hctx = NULL; | |
2663 | struct request *rq; | |
2664 | int queued = 0; | |
0d617a83 | 2665 | blk_status_t ret = BLK_STS_OK; |
b84c5b50 CH |
2666 | |
2667 | while ((rq = rq_list_pop(&plug->mq_list))) { | |
2668 | bool last = rq_list_empty(plug->mq_list); | |
b84c5b50 CH |
2669 | |
2670 | if (hctx != rq->mq_hctx) { | |
34c9f547 KS |
2671 | if (hctx) { |
2672 | blk_mq_commit_rqs(hctx, queued, false); | |
2673 | queued = 0; | |
2674 | } | |
b84c5b50 CH |
2675 | hctx = rq->mq_hctx; |
2676 | } | |
2677 | ||
2678 | ret = blk_mq_request_issue_directly(rq, last); | |
2679 | switch (ret) { | |
2680 | case BLK_STS_OK: | |
2681 | queued++; | |
2682 | break; | |
2683 | case BLK_STS_RESOURCE: | |
2684 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2685 | blk_mq_request_bypass_insert(rq, 0); |
2394395c | 2686 | blk_mq_run_hw_queue(hctx, false); |
0d617a83 | 2687 | goto out; |
b84c5b50 CH |
2688 | default: |
2689 | blk_mq_end_request(rq, ret); | |
b84c5b50 CH |
2690 | break; |
2691 | } | |
2692 | } | |
2693 | ||
0d617a83 KS |
2694 | out: |
2695 | if (ret != BLK_STS_OK) | |
34c9f547 | 2696 | blk_mq_commit_rqs(hctx, queued, false); |
b84c5b50 CH |
2697 | } |
2698 | ||
518579a9 KB |
2699 | static void __blk_mq_flush_plug_list(struct request_queue *q, |
2700 | struct blk_plug *plug) | |
2701 | { | |
2702 | if (blk_queue_quiesced(q)) | |
2703 | return; | |
2704 | q->mq_ops->queue_rqs(&plug->mq_list); | |
2705 | } | |
2706 | ||
26fed4ac JA |
2707 | static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched) |
2708 | { | |
2709 | struct blk_mq_hw_ctx *this_hctx = NULL; | |
2710 | struct blk_mq_ctx *this_ctx = NULL; | |
2711 | struct request *requeue_list = NULL; | |
34e0a279 | 2712 | struct request **requeue_lastp = &requeue_list; |
26fed4ac | 2713 | unsigned int depth = 0; |
d97217e7 | 2714 | bool is_passthrough = false; |
26fed4ac JA |
2715 | LIST_HEAD(list); |
2716 | ||
2717 | do { | |
2718 | struct request *rq = rq_list_pop(&plug->mq_list); | |
2719 | ||
2720 | if (!this_hctx) { | |
2721 | this_hctx = rq->mq_hctx; | |
2722 | this_ctx = rq->mq_ctx; | |
d97217e7 ML |
2723 | is_passthrough = blk_rq_is_passthrough(rq); |
2724 | } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx || | |
2725 | is_passthrough != blk_rq_is_passthrough(rq)) { | |
34e0a279 | 2726 | rq_list_add_tail(&requeue_lastp, rq); |
26fed4ac JA |
2727 | continue; |
2728 | } | |
34e0a279 | 2729 | list_add(&rq->queuelist, &list); |
26fed4ac JA |
2730 | depth++; |
2731 | } while (!rq_list_empty(plug->mq_list)); | |
2732 | ||
2733 | plug->mq_list = requeue_list; | |
2734 | trace_block_unplug(this_hctx->queue, depth, !from_sched); | |
05a93117 CH |
2735 | |
2736 | percpu_ref_get(&this_hctx->queue->q_usage_counter); | |
d97217e7 | 2737 | /* passthrough requests should never be issued to the I/O scheduler */ |
2293cae7 ML |
2738 | if (is_passthrough) { |
2739 | spin_lock(&this_hctx->lock); | |
2740 | list_splice_tail_init(&list, &this_hctx->dispatch); | |
2741 | spin_unlock(&this_hctx->lock); | |
2742 | blk_mq_run_hw_queue(this_hctx, from_sched); | |
2743 | } else if (this_hctx->queue->elevator) { | |
05a93117 | 2744 | this_hctx->queue->elevator->type->ops.insert_requests(this_hctx, |
93fffe16 | 2745 | &list, 0); |
05a93117 CH |
2746 | blk_mq_run_hw_queue(this_hctx, from_sched); |
2747 | } else { | |
2748 | blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched); | |
2749 | } | |
2750 | percpu_ref_put(&this_hctx->queue->q_usage_counter); | |
26fed4ac JA |
2751 | } |
2752 | ||
b84c5b50 CH |
2753 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
2754 | { | |
3c67d44d | 2755 | struct request *rq; |
acc8c0a9 | 2756 | unsigned int depth; |
b84c5b50 | 2757 | |
70904263 RL |
2758 | /* |
2759 | * We may have been called recursively midway through handling | |
2760 | * plug->mq_list via a schedule() in the driver's queue_rq() callback. | |
2761 | * To avoid mq_list changing under our feet, clear rq_count early and | |
2762 | * bail out specifically if rq_count is 0 rather than checking | |
2763 | * whether the mq_list is empty. | |
2764 | */ | |
2765 | if (plug->rq_count == 0) | |
b84c5b50 | 2766 | return; |
acc8c0a9 | 2767 | depth = plug->rq_count; |
b84c5b50 CH |
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; | |
acc8c0a9 | 2775 | trace_block_unplug(q, depth, true); |
3c67d44d JA |
2776 | |
2777 | /* | |
2778 | * Peek first request and see if we have a ->queue_rqs() hook. | |
2779 | * If we do, we can dispatch the whole plug list in one go. We | |
2780 | * already know at this point that all requests belong to the | |
2781 | * same queue, caller must ensure that's the case. | |
3c67d44d | 2782 | */ |
434097ee | 2783 | if (q->mq_ops->queue_rqs) { |
3c67d44d | 2784 | blk_mq_run_dispatch_ops(q, |
518579a9 | 2785 | __blk_mq_flush_plug_list(q, plug)); |
3c67d44d JA |
2786 | if (rq_list_empty(plug->mq_list)) |
2787 | return; | |
2788 | } | |
73f3760e ML |
2789 | |
2790 | blk_mq_run_dispatch_ops(q, | |
3e368fb0 | 2791 | blk_mq_plug_issue_direct(plug)); |
b84c5b50 CH |
2792 | if (rq_list_empty(plug->mq_list)) |
2793 | return; | |
2794 | } | |
2795 | ||
b84c5b50 | 2796 | do { |
26fed4ac | 2797 | blk_mq_dispatch_plug_list(plug, from_schedule); |
b84c5b50 | 2798 | } while (!rq_list_empty(plug->mq_list)); |
b84c5b50 CH |
2799 | } |
2800 | ||
94aa228c | 2801 | static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
6ce3dd6e ML |
2802 | struct list_head *list) |
2803 | { | |
536167d4 | 2804 | int queued = 0; |
984ce0a7 | 2805 | blk_status_t ret = BLK_STS_OK; |
536167d4 | 2806 | |
6ce3dd6e | 2807 | while (!list_empty(list)) { |
6ce3dd6e ML |
2808 | struct request *rq = list_first_entry(list, struct request, |
2809 | queuelist); | |
2810 | ||
2811 | list_del_init(&rq->queuelist); | |
fd9c40f6 | 2812 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
27e8b2bb KS |
2813 | switch (ret) { |
2814 | case BLK_STS_OK: | |
536167d4 | 2815 | queued++; |
27e8b2bb KS |
2816 | break; |
2817 | case BLK_STS_RESOURCE: | |
2818 | case BLK_STS_DEV_RESOURCE: | |
2b597613 | 2819 | blk_mq_request_bypass_insert(rq, 0); |
2394395c CH |
2820 | if (list_empty(list)) |
2821 | blk_mq_run_hw_queue(hctx, false); | |
27e8b2bb KS |
2822 | goto out; |
2823 | default: | |
2824 | blk_mq_end_request(rq, ret); | |
2825 | break; | |
2826 | } | |
6ce3dd6e | 2827 | } |
d666ba98 | 2828 | |
27e8b2bb | 2829 | out: |
984ce0a7 KS |
2830 | if (ret != BLK_STS_OK) |
2831 | blk_mq_commit_rqs(hctx, queued, false); | |
6ce3dd6e ML |
2832 | } |
2833 | ||
b131f201 | 2834 | static bool blk_mq_attempt_bio_merge(struct request_queue *q, |
0c5bcc92 | 2835 | struct bio *bio, unsigned int nr_segs) |
900e0807 JA |
2836 | { |
2837 | if (!blk_queue_nomerges(q) && bio_mergeable(bio)) { | |
0c5bcc92 | 2838 | if (blk_attempt_plug_merge(q, bio, nr_segs)) |
900e0807 JA |
2839 | return true; |
2840 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) | |
2841 | return true; | |
2842 | } | |
2843 | return false; | |
2844 | } | |
2845 | ||
71539717 JA |
2846 | static struct request *blk_mq_get_new_requests(struct request_queue *q, |
2847 | struct blk_plug *plug, | |
0a5aa8d1 SK |
2848 | struct bio *bio, |
2849 | unsigned int nsegs) | |
71539717 JA |
2850 | { |
2851 | struct blk_mq_alloc_data data = { | |
2852 | .q = q, | |
2853 | .nr_tags = 1, | |
9d497e29 | 2854 | .cmd_flags = bio->bi_opf, |
71539717 JA |
2855 | }; |
2856 | struct request *rq; | |
2857 | ||
0a5aa8d1 SK |
2858 | rq_qos_throttle(q, bio); |
2859 | ||
71539717 JA |
2860 | if (plug) { |
2861 | data.nr_tags = plug->nr_ios; | |
2862 | plug->nr_ios = 1; | |
2863 | data.cached_rq = &plug->cached_rq; | |
2864 | } | |
2865 | ||
2866 | rq = __blk_mq_alloc_requests(&data); | |
373b5416 JA |
2867 | if (rq) |
2868 | return rq; | |
71539717 JA |
2869 | rq_qos_cleanup(q, bio); |
2870 | if (bio->bi_opf & REQ_NOWAIT) | |
2871 | bio_wouldblock_error(bio); | |
2872 | return NULL; | |
2873 | } | |
2874 | ||
309ce674 | 2875 | /* |
337e89fe | 2876 | * Check if there is a suitable cached request and return it. |
309ce674 | 2877 | */ |
337e89fe CH |
2878 | static struct request *blk_mq_peek_cached_request(struct blk_plug *plug, |
2879 | struct request_queue *q, blk_opf_t opf) | |
71539717 | 2880 | { |
337e89fe CH |
2881 | enum hctx_type type = blk_mq_get_hctx_type(opf); |
2882 | struct request *rq; | |
71539717 | 2883 | |
337e89fe CH |
2884 | if (!plug) |
2885 | return NULL; | |
2886 | rq = rq_list_peek(&plug->cached_rq); | |
2887 | if (!rq || rq->q != q) | |
2888 | return NULL; | |
2889 | if (type != rq->mq_hctx->type && | |
2890 | (type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT)) | |
2891 | return NULL; | |
2892 | if (op_is_flush(rq->cmd_flags) != op_is_flush(opf)) | |
2893 | return NULL; | |
2894 | return rq; | |
2895 | } | |
0a5aa8d1 | 2896 | |
337e89fe CH |
2897 | static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug, |
2898 | struct bio *bio) | |
2899 | { | |
2900 | WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq); | |
5b13bc8a | 2901 | |
2645672f JA |
2902 | /* |
2903 | * If any qos ->throttle() end up blocking, we will have flushed the | |
2904 | * plug and hence killed the cached_rq list as well. Pop this entry | |
2905 | * before we throttle. | |
2906 | */ | |
5b13bc8a | 2907 | plug->cached_rq = rq_list_next(rq); |
b0077e26 | 2908 | rq_qos_throttle(rq->q, bio); |
2645672f | 2909 | |
5c17f45e | 2910 | blk_mq_rq_time_init(rq, 0); |
b0077e26 | 2911 | rq->cmd_flags = bio->bi_opf; |
5b13bc8a | 2912 | INIT_LIST_HEAD(&rq->queuelist); |
71539717 JA |
2913 | } |
2914 | ||
0676c434 ML |
2915 | static bool bio_unaligned(const struct bio *bio, struct request_queue *q) |
2916 | { | |
2917 | unsigned int bs_mask = queue_logical_block_size(q) - 1; | |
2918 | ||
2919 | /* .bi_sector of any zero sized bio need to be initialized */ | |
2920 | if ((bio->bi_iter.bi_size & bs_mask) || | |
2921 | ((bio->bi_iter.bi_sector << SECTOR_SHIFT) & bs_mask)) | |
2922 | return true; | |
2923 | return false; | |
2924 | } | |
2925 | ||
105663f7 | 2926 | /** |
c62b37d9 | 2927 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2928 | * @bio: Bio pointer. |
2929 | * | |
2930 | * Builds up a request structure from @q and @bio and send to the device. The | |
2931 | * request may not be queued directly to hardware if: | |
2932 | * * This request can be merged with another one | |
2933 | * * We want to place request at plug queue for possible future merging | |
2934 | * * There is an IO scheduler active at this queue | |
2935 | * | |
2936 | * It will not queue the request if there is an error with the bio, or at the | |
2937 | * request creation. | |
105663f7 | 2938 | */ |
3e08773c | 2939 | void blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2940 | { |
ed6cddef | 2941 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
99a9476b | 2942 | struct blk_plug *plug = current->plug; |
ef295ecf | 2943 | const int is_sync = op_is_sync(bio->bi_opf); |
f0dbe6e8 | 2944 | struct blk_mq_hw_ctx *hctx; |
b35243a4 | 2945 | unsigned int nr_segs; |
72e84e90 | 2946 | struct request *rq; |
a892c8d5 | 2947 | blk_status_t ret; |
07068d5b | 2948 | |
dd291d77 DLM |
2949 | /* |
2950 | * If the plug has a cached request for this queue, try to use it. | |
2951 | */ | |
2952 | rq = blk_mq_peek_cached_request(plug, q, bio->bi_opf); | |
2953 | ||
2954 | /* | |
2955 | * A BIO that was released from a zone write plug has already been | |
2956 | * through the preparation in this function, already holds a reference | |
2957 | * on the queue usage counter, and is the only write BIO in-flight for | |
2958 | * the target zone. Go straight to preparing a request for it. | |
2959 | */ | |
2960 | if (bio_zone_write_plugging(bio)) { | |
2961 | nr_segs = bio->__bi_nr_segments; | |
2962 | if (rq) | |
2963 | blk_queue_exit(q); | |
2964 | goto new_request; | |
2965 | } | |
2966 | ||
51d798cd | 2967 | bio = blk_queue_bounce(bio, q); |
9c6227e0 | 2968 | |
72e84e90 | 2969 | /* |
72e84e90 CH |
2970 | * The cached request already holds a q_usage_counter reference and we |
2971 | * don't have to acquire a new one if we use it. | |
2972 | */ | |
72e84e90 | 2973 | if (!rq) { |
b0077e26 | 2974 | if (unlikely(bio_queue_enter(bio))) |
5b13bc8a | 2975 | return; |
b0077e26 CH |
2976 | } |
2977 | ||
0676c434 ML |
2978 | /* |
2979 | * Device reconfiguration may change logical block size, so alignment | |
2980 | * check has to be done with queue usage counter held | |
2981 | */ | |
2982 | if (unlikely(bio_unaligned(bio, q))) { | |
2983 | bio_io_error(bio); | |
2984 | goto queue_exit; | |
2985 | } | |
2986 | ||
b35243a4 CH |
2987 | bio = __bio_split_to_limits(bio, &q->limits, &nr_segs); |
2988 | if (!bio) | |
2989 | goto queue_exit; | |
2990 | ||
337e89fe CH |
2991 | if (!bio_integrity_prep(bio)) |
2992 | goto queue_exit; | |
b0077e26 | 2993 | |
0f299da5 CH |
2994 | if (blk_mq_attempt_bio_merge(q, bio, nr_segs)) |
2995 | goto queue_exit; | |
2996 | ||
dd291d77 DLM |
2997 | if (blk_queue_is_zoned(q) && blk_zone_plug_bio(bio, nr_segs)) |
2998 | goto queue_exit; | |
2999 | ||
3000 | new_request: | |
72e84e90 CH |
3001 | if (!rq) { |
3002 | rq = blk_mq_get_new_requests(q, plug, bio, nr_segs); | |
3003 | if (unlikely(!rq)) | |
3004 | goto queue_exit; | |
3005 | } else { | |
3006 | blk_mq_use_cached_rq(rq, plug, bio); | |
5b13bc8a | 3007 | } |
87760e5e | 3008 | |
e8a676d6 | 3009 | trace_block_getrq(bio); |
d6f1dda2 | 3010 | |
c1c80384 | 3011 | rq_qos_track(q, rq, bio); |
07068d5b | 3012 | |
970d168d BVA |
3013 | blk_mq_bio_to_request(rq, bio, nr_segs); |
3014 | ||
9cd1e566 | 3015 | ret = blk_crypto_rq_get_keyslot(rq); |
a892c8d5 ST |
3016 | if (ret != BLK_STS_OK) { |
3017 | bio->bi_status = ret; | |
3018 | bio_endio(bio); | |
3019 | blk_mq_free_request(rq); | |
3e08773c | 3020 | return; |
a892c8d5 ST |
3021 | } |
3022 | ||
096bc7ea DLM |
3023 | if (bio_zone_write_plugging(bio)) |
3024 | blk_zone_write_plug_init_request(rq); | |
3025 | ||
360f2648 | 3026 | if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq)) |
d92ca9d8 CH |
3027 | return; |
3028 | ||
f0dbe6e8 | 3029 | if (plug) { |
ce5b009c | 3030 | blk_add_rq_to_plug(plug, rq); |
f0dbe6e8 CH |
3031 | return; |
3032 | } | |
3033 | ||
3034 | hctx = rq->mq_hctx; | |
dd6216bb | 3035 | if ((rq->rq_flags & RQF_USE_SCHED) || |
f0dbe6e8 | 3036 | (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) { |
710fa378 | 3037 | blk_mq_insert_request(rq, 0); |
f0dbe6e8 CH |
3038 | blk_mq_run_hw_queue(hctx, true); |
3039 | } else { | |
3040 | blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq)); | |
3041 | } | |
0f299da5 CH |
3042 | return; |
3043 | ||
3044 | queue_exit: | |
72e84e90 CH |
3045 | /* |
3046 | * Don't drop the queue reference if we were trying to use a cached | |
3047 | * request and thus didn't acquire one. | |
3048 | */ | |
3049 | if (!rq) | |
3050 | blk_queue_exit(q); | |
320ae51f JA |
3051 | } |
3052 | ||
248c7933 | 3053 | #ifdef CONFIG_BLK_MQ_STACKING |
06c8c691 | 3054 | /** |
a5efda3c | 3055 | * blk_insert_cloned_request - Helper for stacking drivers to submit a request |
a5efda3c | 3056 | * @rq: the request being queued |
06c8c691 | 3057 | */ |
28db4711 | 3058 | blk_status_t blk_insert_cloned_request(struct request *rq) |
06c8c691 | 3059 | { |
28db4711 | 3060 | struct request_queue *q = rq->q; |
8d1dfd51 | 3061 | unsigned int max_sectors = blk_queue_get_max_sectors(rq); |
49d24398 | 3062 | unsigned int max_segments = blk_rq_get_max_segments(rq); |
a5efda3c | 3063 | blk_status_t ret; |
06c8c691 CH |
3064 | |
3065 | if (blk_rq_sectors(rq) > max_sectors) { | |
3066 | /* | |
3067 | * SCSI device does not have a good way to return if | |
3068 | * Write Same/Zero is actually supported. If a device rejects | |
3069 | * a non-read/write command (discard, write same,etc.) the | |
3070 | * low-level device driver will set the relevant queue limit to | |
3071 | * 0 to prevent blk-lib from issuing more of the offending | |
3072 | * operations. Commands queued prior to the queue limit being | |
3073 | * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O | |
3074 | * errors being propagated to upper layers. | |
3075 | */ | |
3076 | if (max_sectors == 0) | |
3077 | return BLK_STS_NOTSUPP; | |
3078 | ||
3079 | printk(KERN_ERR "%s: over max size limit. (%u > %u)\n", | |
3080 | __func__, blk_rq_sectors(rq), max_sectors); | |
3081 | return BLK_STS_IOERR; | |
3082 | } | |
3083 | ||
3084 | /* | |
3085 | * The queue settings related to segment counting may differ from the | |
3086 | * original queue. | |
3087 | */ | |
3088 | rq->nr_phys_segments = blk_recalc_rq_segments(rq); | |
49d24398 US |
3089 | if (rq->nr_phys_segments > max_segments) { |
3090 | printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n", | |
3091 | __func__, rq->nr_phys_segments, max_segments); | |
06c8c691 CH |
3092 | return BLK_STS_IOERR; |
3093 | } | |
3094 | ||
28db4711 | 3095 | if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq))) |
06c8c691 CH |
3096 | return BLK_STS_IOERR; |
3097 | ||
5b8562f0 EB |
3098 | ret = blk_crypto_rq_get_keyslot(rq); |
3099 | if (ret != BLK_STS_OK) | |
3100 | return ret; | |
06c8c691 CH |
3101 | |
3102 | blk_account_io_start(rq); | |
3103 | ||
3104 | /* | |
3105 | * Since we have a scheduler attached on the top device, | |
3106 | * bypass a potential scheduler on the bottom device for | |
3107 | * insert. | |
3108 | */ | |
28db4711 | 3109 | blk_mq_run_dispatch_ops(q, |
4cafe86c | 3110 | ret = blk_mq_request_issue_directly(rq, true)); |
592ee119 | 3111 | if (ret) |
08420cf7 | 3112 | blk_account_io_done(rq, blk_time_get_ns()); |
4cafe86c | 3113 | return ret; |
06c8c691 CH |
3114 | } |
3115 | EXPORT_SYMBOL_GPL(blk_insert_cloned_request); | |
3116 | ||
3117 | /** | |
3118 | * blk_rq_unprep_clone - Helper function to free all bios in a cloned request | |
3119 | * @rq: the clone request to be cleaned up | |
3120 | * | |
3121 | * Description: | |
3122 | * Free all bios in @rq for a cloned request. | |
3123 | */ | |
3124 | void blk_rq_unprep_clone(struct request *rq) | |
3125 | { | |
3126 | struct bio *bio; | |
3127 | ||
3128 | while ((bio = rq->bio) != NULL) { | |
3129 | rq->bio = bio->bi_next; | |
3130 | ||
3131 | bio_put(bio); | |
3132 | } | |
3133 | } | |
3134 | EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); | |
3135 | ||
3136 | /** | |
3137 | * blk_rq_prep_clone - Helper function to setup clone request | |
3138 | * @rq: the request to be setup | |
3139 | * @rq_src: original request to be cloned | |
3140 | * @bs: bio_set that bios for clone are allocated from | |
3141 | * @gfp_mask: memory allocation mask for bio | |
3142 | * @bio_ctr: setup function to be called for each clone bio. | |
3143 | * Returns %0 for success, non %0 for failure. | |
3144 | * @data: private data to be passed to @bio_ctr | |
3145 | * | |
3146 | * Description: | |
3147 | * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. | |
3148 | * Also, pages which the original bios are pointing to are not copied | |
3149 | * and the cloned bios just point same pages. | |
3150 | * So cloned bios must be completed before original bios, which means | |
3151 | * the caller must complete @rq before @rq_src. | |
3152 | */ | |
3153 | int blk_rq_prep_clone(struct request *rq, struct request *rq_src, | |
3154 | struct bio_set *bs, gfp_t gfp_mask, | |
3155 | int (*bio_ctr)(struct bio *, struct bio *, void *), | |
3156 | void *data) | |
3157 | { | |
3158 | struct bio *bio, *bio_src; | |
3159 | ||
3160 | if (!bs) | |
3161 | bs = &fs_bio_set; | |
3162 | ||
3163 | __rq_for_each_bio(bio_src, rq_src) { | |
abfc426d CH |
3164 | bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask, |
3165 | bs); | |
06c8c691 CH |
3166 | if (!bio) |
3167 | goto free_and_out; | |
3168 | ||
3169 | if (bio_ctr && bio_ctr(bio, bio_src, data)) | |
3170 | goto free_and_out; | |
3171 | ||
3172 | if (rq->bio) { | |
3173 | rq->biotail->bi_next = bio; | |
3174 | rq->biotail = bio; | |
3175 | } else { | |
3176 | rq->bio = rq->biotail = bio; | |
3177 | } | |
3178 | bio = NULL; | |
3179 | } | |
3180 | ||
3181 | /* Copy attributes of the original request to the clone request. */ | |
3182 | rq->__sector = blk_rq_pos(rq_src); | |
3183 | rq->__data_len = blk_rq_bytes(rq_src); | |
3184 | if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) { | |
3185 | rq->rq_flags |= RQF_SPECIAL_PAYLOAD; | |
3186 | rq->special_vec = rq_src->special_vec; | |
3187 | } | |
3188 | rq->nr_phys_segments = rq_src->nr_phys_segments; | |
3189 | rq->ioprio = rq_src->ioprio; | |
44981351 | 3190 | rq->write_hint = rq_src->write_hint; |
06c8c691 CH |
3191 | |
3192 | if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0) | |
3193 | goto free_and_out; | |
3194 | ||
3195 | return 0; | |
3196 | ||
3197 | free_and_out: | |
3198 | if (bio) | |
3199 | bio_put(bio); | |
3200 | blk_rq_unprep_clone(rq); | |
3201 | ||
3202 | return -ENOMEM; | |
3203 | } | |
3204 | EXPORT_SYMBOL_GPL(blk_rq_prep_clone); | |
248c7933 | 3205 | #endif /* CONFIG_BLK_MQ_STACKING */ |
06c8c691 | 3206 | |
f2b8f3ce CH |
3207 | /* |
3208 | * Steal bios from a request and add them to a bio list. | |
3209 | * The request must not have been partially completed before. | |
3210 | */ | |
3211 | void blk_steal_bios(struct bio_list *list, struct request *rq) | |
3212 | { | |
3213 | if (rq->bio) { | |
3214 | if (list->tail) | |
3215 | list->tail->bi_next = rq->bio; | |
3216 | else | |
3217 | list->head = rq->bio; | |
3218 | list->tail = rq->biotail; | |
3219 | ||
3220 | rq->bio = NULL; | |
3221 | rq->biotail = NULL; | |
3222 | } | |
3223 | ||
3224 | rq->__data_len = 0; | |
3225 | } | |
3226 | EXPORT_SYMBOL_GPL(blk_steal_bios); | |
3227 | ||
bd63141d ML |
3228 | static size_t order_to_size(unsigned int order) |
3229 | { | |
3230 | return (size_t)PAGE_SIZE << order; | |
3231 | } | |
3232 | ||
3233 | /* called before freeing request pool in @tags */ | |
f32e4eaf JG |
3234 | static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, |
3235 | struct blk_mq_tags *tags) | |
bd63141d | 3236 | { |
bd63141d ML |
3237 | struct page *page; |
3238 | unsigned long flags; | |
3239 | ||
76dd2980 YK |
3240 | /* |
3241 | * There is no need to clear mapping if driver tags is not initialized | |
3242 | * or the mapping belongs to the driver tags. | |
3243 | */ | |
3244 | if (!drv_tags || drv_tags == tags) | |
4f245d5b JG |
3245 | return; |
3246 | ||
bd63141d ML |
3247 | list_for_each_entry(page, &tags->page_list, lru) { |
3248 | unsigned long start = (unsigned long)page_address(page); | |
3249 | unsigned long end = start + order_to_size(page->private); | |
3250 | int i; | |
3251 | ||
f32e4eaf | 3252 | for (i = 0; i < drv_tags->nr_tags; i++) { |
bd63141d ML |
3253 | struct request *rq = drv_tags->rqs[i]; |
3254 | unsigned long rq_addr = (unsigned long)rq; | |
3255 | ||
3256 | if (rq_addr >= start && rq_addr < end) { | |
0a467d0f | 3257 | WARN_ON_ONCE(req_ref_read(rq) != 0); |
bd63141d ML |
3258 | cmpxchg(&drv_tags->rqs[i], rq, NULL); |
3259 | } | |
3260 | } | |
3261 | } | |
3262 | ||
3263 | /* | |
3264 | * Wait until all pending iteration is done. | |
3265 | * | |
3266 | * Request reference is cleared and it is guaranteed to be observed | |
3267 | * after the ->lock is released. | |
3268 | */ | |
3269 | spin_lock_irqsave(&drv_tags->lock, flags); | |
3270 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
3271 | } | |
3272 | ||
cc71a6f4 JA |
3273 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
3274 | unsigned int hctx_idx) | |
95363efd | 3275 | { |
f32e4eaf | 3276 | struct blk_mq_tags *drv_tags; |
e9b267d9 | 3277 | struct page *page; |
320ae51f | 3278 | |
e02657ea ML |
3279 | if (list_empty(&tags->page_list)) |
3280 | return; | |
3281 | ||
079a2e3e JG |
3282 | if (blk_mq_is_shared_tags(set->flags)) |
3283 | drv_tags = set->shared_tags; | |
e155b0c2 JG |
3284 | else |
3285 | drv_tags = set->tags[hctx_idx]; | |
f32e4eaf | 3286 | |
65de57bb | 3287 | if (tags->static_rqs && set->ops->exit_request) { |
e9b267d9 | 3288 | int i; |
320ae51f | 3289 | |
24d2f903 | 3290 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
3291 | struct request *rq = tags->static_rqs[i]; |
3292 | ||
3293 | if (!rq) | |
e9b267d9 | 3294 | continue; |
d6296d39 | 3295 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 3296 | tags->static_rqs[i] = NULL; |
e9b267d9 | 3297 | } |
320ae51f | 3298 | } |
320ae51f | 3299 | |
f32e4eaf | 3300 | blk_mq_clear_rq_mapping(drv_tags, tags); |
bd63141d | 3301 | |
24d2f903 CH |
3302 | while (!list_empty(&tags->page_list)) { |
3303 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 3304 | list_del_init(&page->lru); |
f75782e4 CM |
3305 | /* |
3306 | * Remove kmemleak object previously allocated in | |
273938bf | 3307 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
3308 | */ |
3309 | kmemleak_free(page_address(page)); | |
320ae51f JA |
3310 | __free_pages(page, page->private); |
3311 | } | |
cc71a6f4 | 3312 | } |
320ae51f | 3313 | |
e155b0c2 | 3314 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
cc71a6f4 | 3315 | { |
24d2f903 | 3316 | kfree(tags->rqs); |
cc71a6f4 | 3317 | tags->rqs = NULL; |
2af8cbe3 JA |
3318 | kfree(tags->static_rqs); |
3319 | tags->static_rqs = NULL; | |
320ae51f | 3320 | |
e155b0c2 | 3321 | blk_mq_free_tags(tags); |
320ae51f JA |
3322 | } |
3323 | ||
4d805131 ML |
3324 | static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set, |
3325 | unsigned int hctx_idx) | |
3326 | { | |
3327 | int i; | |
3328 | ||
3329 | for (i = 0; i < set->nr_maps; i++) { | |
3330 | unsigned int start = set->map[i].queue_offset; | |
3331 | unsigned int end = start + set->map[i].nr_queues; | |
3332 | ||
3333 | if (hctx_idx >= start && hctx_idx < end) | |
3334 | break; | |
3335 | } | |
3336 | ||
3337 | if (i >= set->nr_maps) | |
3338 | i = HCTX_TYPE_DEFAULT; | |
3339 | ||
3340 | return i; | |
3341 | } | |
3342 | ||
3343 | static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set, | |
3344 | unsigned int hctx_idx) | |
3345 | { | |
3346 | enum hctx_type type = hctx_idx_to_type(set, hctx_idx); | |
3347 | ||
3348 | return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx); | |
3349 | } | |
3350 | ||
63064be1 JG |
3351 | static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
3352 | unsigned int hctx_idx, | |
3353 | unsigned int nr_tags, | |
e155b0c2 | 3354 | unsigned int reserved_tags) |
320ae51f | 3355 | { |
4d805131 | 3356 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
24d2f903 | 3357 | struct blk_mq_tags *tags; |
320ae51f | 3358 | |
59f082e4 SL |
3359 | if (node == NUMA_NO_NODE) |
3360 | node = set->numa_node; | |
3361 | ||
e155b0c2 JG |
3362 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
3363 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
3364 | if (!tags) |
3365 | return NULL; | |
320ae51f | 3366 | |
590b5b7d | 3367 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 3368 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 3369 | node); |
7edfd681 JC |
3370 | if (!tags->rqs) |
3371 | goto err_free_tags; | |
320ae51f | 3372 | |
590b5b7d KC |
3373 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
3374 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
3375 | node); | |
7edfd681 JC |
3376 | if (!tags->static_rqs) |
3377 | goto err_free_rqs; | |
2af8cbe3 | 3378 | |
cc71a6f4 | 3379 | return tags; |
7edfd681 JC |
3380 | |
3381 | err_free_rqs: | |
3382 | kfree(tags->rqs); | |
3383 | err_free_tags: | |
3384 | blk_mq_free_tags(tags); | |
3385 | return NULL; | |
cc71a6f4 JA |
3386 | } |
3387 | ||
1d9bd516 TH |
3388 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
3389 | unsigned int hctx_idx, int node) | |
3390 | { | |
3391 | int ret; | |
3392 | ||
3393 | if (set->ops->init_request) { | |
3394 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
3395 | if (ret) | |
3396 | return ret; | |
3397 | } | |
3398 | ||
12f5b931 | 3399 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
3400 | return 0; |
3401 | } | |
3402 | ||
63064be1 JG |
3403 | static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, |
3404 | struct blk_mq_tags *tags, | |
3405 | unsigned int hctx_idx, unsigned int depth) | |
cc71a6f4 JA |
3406 | { |
3407 | unsigned int i, j, entries_per_page, max_order = 4; | |
4d805131 | 3408 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
cc71a6f4 | 3409 | size_t rq_size, left; |
59f082e4 | 3410 | |
59f082e4 SL |
3411 | if (node == NUMA_NO_NODE) |
3412 | node = set->numa_node; | |
cc71a6f4 JA |
3413 | |
3414 | INIT_LIST_HEAD(&tags->page_list); | |
3415 | ||
320ae51f JA |
3416 | /* |
3417 | * rq_size is the size of the request plus driver payload, rounded | |
3418 | * to the cacheline size | |
3419 | */ | |
24d2f903 | 3420 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 3421 | cache_line_size()); |
cc71a6f4 | 3422 | left = rq_size * depth; |
320ae51f | 3423 | |
cc71a6f4 | 3424 | for (i = 0; i < depth; ) { |
320ae51f JA |
3425 | int this_order = max_order; |
3426 | struct page *page; | |
3427 | int to_do; | |
3428 | void *p; | |
3429 | ||
b3a834b1 | 3430 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
3431 | this_order--; |
3432 | ||
3433 | do { | |
59f082e4 | 3434 | page = alloc_pages_node(node, |
36e1f3d1 | 3435 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 3436 | this_order); |
320ae51f JA |
3437 | if (page) |
3438 | break; | |
3439 | if (!this_order--) | |
3440 | break; | |
3441 | if (order_to_size(this_order) < rq_size) | |
3442 | break; | |
3443 | } while (1); | |
3444 | ||
3445 | if (!page) | |
24d2f903 | 3446 | goto fail; |
320ae51f JA |
3447 | |
3448 | page->private = this_order; | |
24d2f903 | 3449 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
3450 | |
3451 | p = page_address(page); | |
f75782e4 CM |
3452 | /* |
3453 | * Allow kmemleak to scan these pages as they contain pointers | |
3454 | * to additional allocations like via ops->init_request(). | |
3455 | */ | |
36e1f3d1 | 3456 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 3457 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 3458 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
3459 | left -= to_do * rq_size; |
3460 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
3461 | struct request *rq = p; |
3462 | ||
3463 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
3464 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
3465 | tags->static_rqs[i] = NULL; | |
3466 | goto fail; | |
e9b267d9 CH |
3467 | } |
3468 | ||
320ae51f JA |
3469 | p += rq_size; |
3470 | i++; | |
3471 | } | |
3472 | } | |
cc71a6f4 | 3473 | return 0; |
320ae51f | 3474 | |
24d2f903 | 3475 | fail: |
cc71a6f4 JA |
3476 | blk_mq_free_rqs(set, tags, hctx_idx); |
3477 | return -ENOMEM; | |
320ae51f JA |
3478 | } |
3479 | ||
bf0beec0 ML |
3480 | struct rq_iter_data { |
3481 | struct blk_mq_hw_ctx *hctx; | |
3482 | bool has_rq; | |
3483 | }; | |
3484 | ||
2dd6532e | 3485 | static bool blk_mq_has_request(struct request *rq, void *data) |
bf0beec0 ML |
3486 | { |
3487 | struct rq_iter_data *iter_data = data; | |
3488 | ||
3489 | if (rq->mq_hctx != iter_data->hctx) | |
3490 | return true; | |
3491 | iter_data->has_rq = true; | |
3492 | return false; | |
3493 | } | |
3494 | ||
3495 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
3496 | { | |
3497 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
3498 | hctx->sched_tags : hctx->tags; | |
3499 | struct rq_iter_data data = { | |
3500 | .hctx = hctx, | |
3501 | }; | |
3502 | ||
3503 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
3504 | return data.has_rq; | |
3505 | } | |
3506 | ||
a46c2702 ML |
3507 | static bool blk_mq_hctx_has_online_cpu(struct blk_mq_hw_ctx *hctx, |
3508 | unsigned int this_cpu) | |
bf0beec0 | 3509 | { |
a46c2702 ML |
3510 | enum hctx_type type = hctx->type; |
3511 | int cpu; | |
3512 | ||
3513 | /* | |
3514 | * hctx->cpumask has to rule out isolated CPUs, but userspace still | |
3515 | * might submit IOs on these isolated CPUs, so use the queue map to | |
3516 | * check if all CPUs mapped to this hctx are offline | |
3517 | */ | |
3518 | for_each_online_cpu(cpu) { | |
3519 | struct blk_mq_hw_ctx *h = blk_mq_map_queue_type(hctx->queue, | |
3520 | type, cpu); | |
3521 | ||
3522 | if (h != hctx) | |
3523 | continue; | |
3524 | ||
3525 | /* this hctx has at least one online CPU */ | |
3526 | if (this_cpu != cpu) | |
3527 | return true; | |
3528 | } | |
3529 | ||
3530 | return false; | |
bf0beec0 ML |
3531 | } |
3532 | ||
3533 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
3534 | { | |
3535 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3536 | struct blk_mq_hw_ctx, cpuhp_online); | |
3537 | ||
a46c2702 | 3538 | if (blk_mq_hctx_has_online_cpu(hctx, cpu)) |
bf0beec0 ML |
3539 | return 0; |
3540 | ||
3541 | /* | |
3542 | * Prevent new request from being allocated on the current hctx. | |
3543 | * | |
3544 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
3545 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
3546 | * seen once we return from the tag allocator. | |
3547 | */ | |
3548 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3549 | smp_mb__after_atomic(); | |
3550 | ||
3551 | /* | |
3552 | * Try to grab a reference to the queue and wait for any outstanding | |
3553 | * requests. If we could not grab a reference the queue has been | |
3554 | * frozen and there are no requests. | |
3555 | */ | |
3556 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
3557 | while (blk_mq_hctx_has_requests(hctx)) | |
3558 | msleep(5); | |
3559 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
3560 | } | |
3561 | ||
3562 | return 0; | |
3563 | } | |
3564 | ||
7b815817 ML |
3565 | /* |
3566 | * Check if one CPU is mapped to the specified hctx | |
3567 | * | |
3568 | * Isolated CPUs have been ruled out from hctx->cpumask, which is supposed | |
3569 | * to be used for scheduling kworker only. For other usage, please call this | |
3570 | * helper for checking if one CPU belongs to the specified hctx | |
3571 | */ | |
3572 | static bool blk_mq_cpu_mapped_to_hctx(unsigned int cpu, | |
3573 | const struct blk_mq_hw_ctx *hctx) | |
3574 | { | |
3575 | struct blk_mq_hw_ctx *mapped_hctx = blk_mq_map_queue_type(hctx->queue, | |
3576 | hctx->type, cpu); | |
3577 | ||
3578 | return mapped_hctx == hctx; | |
3579 | } | |
3580 | ||
bf0beec0 ML |
3581 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) |
3582 | { | |
3583 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3584 | struct blk_mq_hw_ctx, cpuhp_online); | |
3585 | ||
7b815817 | 3586 | if (blk_mq_cpu_mapped_to_hctx(cpu, hctx)) |
bf0beec0 ML |
3587 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); |
3588 | return 0; | |
3589 | } | |
3590 | ||
e57690fe JA |
3591 | /* |
3592 | * 'cpu' is going away. splice any existing rq_list entries from this | |
3593 | * software queue to the hw queue dispatch list, and ensure that it | |
3594 | * gets run. | |
3595 | */ | |
9467f859 | 3596 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 3597 | { |
9467f859 | 3598 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
3599 | struct blk_mq_ctx *ctx; |
3600 | LIST_HEAD(tmp); | |
c16d6b5a | 3601 | enum hctx_type type; |
484b4061 | 3602 | |
9467f859 | 3603 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
7b815817 | 3604 | if (!blk_mq_cpu_mapped_to_hctx(cpu, hctx)) |
bf0beec0 ML |
3605 | return 0; |
3606 | ||
e57690fe | 3607 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 3608 | type = hctx->type; |
484b4061 JA |
3609 | |
3610 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
3611 | if (!list_empty(&ctx->rq_lists[type])) { |
3612 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
3613 | blk_mq_hctx_clear_pending(hctx, ctx); |
3614 | } | |
3615 | spin_unlock(&ctx->lock); | |
3616 | ||
3617 | if (list_empty(&tmp)) | |
9467f859 | 3618 | return 0; |
484b4061 | 3619 | |
e57690fe JA |
3620 | spin_lock(&hctx->lock); |
3621 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
3622 | spin_unlock(&hctx->lock); | |
484b4061 JA |
3623 | |
3624 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 3625 | return 0; |
484b4061 JA |
3626 | } |
3627 | ||
9467f859 | 3628 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 3629 | { |
bf0beec0 ML |
3630 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3631 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3632 | &hctx->cpuhp_online); | |
9467f859 TG |
3633 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
3634 | &hctx->cpuhp_dead); | |
484b4061 JA |
3635 | } |
3636 | ||
364b6181 ML |
3637 | /* |
3638 | * Before freeing hw queue, clearing the flush request reference in | |
3639 | * tags->rqs[] for avoiding potential UAF. | |
3640 | */ | |
3641 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
3642 | unsigned int queue_depth, struct request *flush_rq) | |
3643 | { | |
3644 | int i; | |
3645 | unsigned long flags; | |
3646 | ||
3647 | /* The hw queue may not be mapped yet */ | |
3648 | if (!tags) | |
3649 | return; | |
3650 | ||
0a467d0f | 3651 | WARN_ON_ONCE(req_ref_read(flush_rq) != 0); |
364b6181 ML |
3652 | |
3653 | for (i = 0; i < queue_depth; i++) | |
3654 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
3655 | ||
3656 | /* | |
3657 | * Wait until all pending iteration is done. | |
3658 | * | |
3659 | * Request reference is cleared and it is guaranteed to be observed | |
3660 | * after the ->lock is released. | |
3661 | */ | |
3662 | spin_lock_irqsave(&tags->lock, flags); | |
3663 | spin_unlock_irqrestore(&tags->lock, flags); | |
3664 | } | |
3665 | ||
c3b4afca | 3666 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
3667 | static void blk_mq_exit_hctx(struct request_queue *q, |
3668 | struct blk_mq_tag_set *set, | |
3669 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
3670 | { | |
364b6181 ML |
3671 | struct request *flush_rq = hctx->fq->flush_rq; |
3672 | ||
8ab0b7dc ML |
3673 | if (blk_mq_hw_queue_mapped(hctx)) |
3674 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 3675 | |
6cfeadbf ML |
3676 | if (blk_queue_init_done(q)) |
3677 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], | |
3678 | set->queue_depth, flush_rq); | |
f70ced09 | 3679 | if (set->ops->exit_request) |
364b6181 | 3680 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 3681 | |
08e98fc6 ML |
3682 | if (set->ops->exit_hctx) |
3683 | set->ops->exit_hctx(hctx, hctx_idx); | |
3684 | ||
9467f859 | 3685 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 | 3686 | |
4e5cc99e ML |
3687 | xa_erase(&q->hctx_table, hctx_idx); |
3688 | ||
2f8f1336 ML |
3689 | spin_lock(&q->unused_hctx_lock); |
3690 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
3691 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
3692 | } |
3693 | ||
624dbe47 ML |
3694 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
3695 | struct blk_mq_tag_set *set, int nr_queue) | |
3696 | { | |
3697 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 3698 | unsigned long i; |
624dbe47 ML |
3699 | |
3700 | queue_for_each_hw_ctx(q, hctx, i) { | |
3701 | if (i == nr_queue) | |
3702 | break; | |
08e98fc6 | 3703 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 3704 | } |
624dbe47 ML |
3705 | } |
3706 | ||
08e98fc6 ML |
3707 | static int blk_mq_init_hctx(struct request_queue *q, |
3708 | struct blk_mq_tag_set *set, | |
3709 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 3710 | { |
7c6c5b7c ML |
3711 | hctx->queue_num = hctx_idx; |
3712 | ||
bf0beec0 ML |
3713 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3714 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3715 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
3716 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
3717 | ||
3718 | hctx->tags = set->tags[hctx_idx]; | |
3719 | ||
3720 | if (set->ops->init_hctx && | |
3721 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
3722 | goto unregister_cpu_notifier; | |
08e98fc6 | 3723 | |
7c6c5b7c ML |
3724 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
3725 | hctx->numa_node)) | |
3726 | goto exit_hctx; | |
4e5cc99e ML |
3727 | |
3728 | if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL)) | |
3729 | goto exit_flush_rq; | |
3730 | ||
7c6c5b7c ML |
3731 | return 0; |
3732 | ||
4e5cc99e ML |
3733 | exit_flush_rq: |
3734 | if (set->ops->exit_request) | |
3735 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); | |
7c6c5b7c ML |
3736 | exit_hctx: |
3737 | if (set->ops->exit_hctx) | |
3738 | set->ops->exit_hctx(hctx, hctx_idx); | |
3739 | unregister_cpu_notifier: | |
3740 | blk_mq_remove_cpuhp(hctx); | |
3741 | return -1; | |
3742 | } | |
3743 | ||
3744 | static struct blk_mq_hw_ctx * | |
3745 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
3746 | int node) | |
3747 | { | |
3748 | struct blk_mq_hw_ctx *hctx; | |
3749 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
3750 | ||
704b914f | 3751 | hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node); |
7c6c5b7c ML |
3752 | if (!hctx) |
3753 | goto fail_alloc_hctx; | |
3754 | ||
3755 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
3756 | goto free_hctx; | |
3757 | ||
3758 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 3759 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
3760 | node = set->numa_node; |
3761 | hctx->numa_node = node; | |
08e98fc6 | 3762 | |
9f993737 | 3763 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
3764 | spin_lock_init(&hctx->lock); |
3765 | INIT_LIST_HEAD(&hctx->dispatch); | |
3766 | hctx->queue = q; | |
51db1c37 | 3767 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 3768 | |
2f8f1336 ML |
3769 | INIT_LIST_HEAD(&hctx->hctx_list); |
3770 | ||
320ae51f | 3771 | /* |
08e98fc6 ML |
3772 | * Allocate space for all possible cpus to avoid allocation at |
3773 | * runtime | |
320ae51f | 3774 | */ |
d904bfa7 | 3775 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 3776 | gfp, node); |
08e98fc6 | 3777 | if (!hctx->ctxs) |
7c6c5b7c | 3778 | goto free_cpumask; |
320ae51f | 3779 | |
5b202853 | 3780 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 3781 | gfp, node, false, false)) |
08e98fc6 | 3782 | goto free_ctxs; |
08e98fc6 | 3783 | hctx->nr_ctx = 0; |
320ae51f | 3784 | |
5815839b | 3785 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
3786 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
3787 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
3788 | ||
754a1572 | 3789 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 3790 | if (!hctx->fq) |
7c6c5b7c | 3791 | goto free_bitmap; |
320ae51f | 3792 | |
7c6c5b7c | 3793 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 3794 | |
7c6c5b7c | 3795 | return hctx; |
320ae51f | 3796 | |
08e98fc6 | 3797 | free_bitmap: |
88459642 | 3798 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
3799 | free_ctxs: |
3800 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
3801 | free_cpumask: |
3802 | free_cpumask_var(hctx->cpumask); | |
3803 | free_hctx: | |
3804 | kfree(hctx); | |
3805 | fail_alloc_hctx: | |
3806 | return NULL; | |
08e98fc6 | 3807 | } |
320ae51f | 3808 | |
320ae51f JA |
3809 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
3810 | unsigned int nr_hw_queues) | |
3811 | { | |
b3c661b1 JA |
3812 | struct blk_mq_tag_set *set = q->tag_set; |
3813 | unsigned int i, j; | |
320ae51f JA |
3814 | |
3815 | for_each_possible_cpu(i) { | |
3816 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
3817 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 3818 | int k; |
320ae51f | 3819 | |
320ae51f JA |
3820 | __ctx->cpu = i; |
3821 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
3822 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
3823 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
3824 | ||
320ae51f JA |
3825 | __ctx->queue = q; |
3826 | ||
320ae51f JA |
3827 | /* |
3828 | * Set local node, IFF we have more than one hw queue. If | |
3829 | * not, we remain on the home node of the device | |
3830 | */ | |
b3c661b1 JA |
3831 | for (j = 0; j < set->nr_maps; j++) { |
3832 | hctx = blk_mq_map_queue_type(q, j, i); | |
3833 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 3834 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 3835 | } |
320ae51f JA |
3836 | } |
3837 | } | |
3838 | ||
63064be1 JG |
3839 | struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3840 | unsigned int hctx_idx, | |
3841 | unsigned int depth) | |
cc71a6f4 | 3842 | { |
63064be1 JG |
3843 | struct blk_mq_tags *tags; |
3844 | int ret; | |
cc71a6f4 | 3845 | |
e155b0c2 | 3846 | tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); |
63064be1 JG |
3847 | if (!tags) |
3848 | return NULL; | |
cc71a6f4 | 3849 | |
63064be1 JG |
3850 | ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); |
3851 | if (ret) { | |
e155b0c2 | 3852 | blk_mq_free_rq_map(tags); |
63064be1 JG |
3853 | return NULL; |
3854 | } | |
cc71a6f4 | 3855 | |
63064be1 | 3856 | return tags; |
cc71a6f4 JA |
3857 | } |
3858 | ||
63064be1 JG |
3859 | static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3860 | int hctx_idx) | |
cc71a6f4 | 3861 | { |
079a2e3e JG |
3862 | if (blk_mq_is_shared_tags(set->flags)) { |
3863 | set->tags[hctx_idx] = set->shared_tags; | |
1c0706a7 | 3864 | |
e155b0c2 | 3865 | return true; |
bd166ef1 | 3866 | } |
e155b0c2 | 3867 | |
63064be1 JG |
3868 | set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, |
3869 | set->queue_depth); | |
3870 | ||
3871 | return set->tags[hctx_idx]; | |
cc71a6f4 JA |
3872 | } |
3873 | ||
645db34e JG |
3874 | void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3875 | struct blk_mq_tags *tags, | |
3876 | unsigned int hctx_idx) | |
cc71a6f4 | 3877 | { |
645db34e JG |
3878 | if (tags) { |
3879 | blk_mq_free_rqs(set, tags, hctx_idx); | |
e155b0c2 | 3880 | blk_mq_free_rq_map(tags); |
bd166ef1 | 3881 | } |
cc71a6f4 JA |
3882 | } |
3883 | ||
e155b0c2 JG |
3884 | static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3885 | unsigned int hctx_idx) | |
3886 | { | |
079a2e3e | 3887 | if (!blk_mq_is_shared_tags(set->flags)) |
e155b0c2 JG |
3888 | blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); |
3889 | ||
3890 | set->tags[hctx_idx] = NULL; | |
cc71a6f4 JA |
3891 | } |
3892 | ||
4b855ad3 | 3893 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 3894 | { |
4f481208 ML |
3895 | unsigned int j, hctx_idx; |
3896 | unsigned long i; | |
320ae51f JA |
3897 | struct blk_mq_hw_ctx *hctx; |
3898 | struct blk_mq_ctx *ctx; | |
2a34c087 | 3899 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
3900 | |
3901 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 3902 | cpumask_clear(hctx->cpumask); |
320ae51f | 3903 | hctx->nr_ctx = 0; |
d416c92c | 3904 | hctx->dispatch_from = NULL; |
320ae51f JA |
3905 | } |
3906 | ||
3907 | /* | |
4b855ad3 | 3908 | * Map software to hardware queues. |
4412efec ML |
3909 | * |
3910 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 3911 | */ |
20e4d813 | 3912 | for_each_possible_cpu(i) { |
4412efec | 3913 | |
897bb0c7 | 3914 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 3915 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
3916 | if (!set->map[j].nr_queues) { |
3917 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3918 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 3919 | continue; |
bb94aea1 | 3920 | } |
fd689871 ML |
3921 | hctx_idx = set->map[j].mq_map[i]; |
3922 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
3923 | if (!set->tags[hctx_idx] && | |
63064be1 | 3924 | !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { |
fd689871 ML |
3925 | /* |
3926 | * If tags initialization fail for some hctx, | |
3927 | * that hctx won't be brought online. In this | |
3928 | * case, remap the current ctx to hctx[0] which | |
3929 | * is guaranteed to always have tags allocated | |
3930 | */ | |
3931 | set->map[j].mq_map[i] = 0; | |
3932 | } | |
e5edd5f2 | 3933 | |
b3c661b1 | 3934 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 3935 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
3936 | /* |
3937 | * If the CPU is already set in the mask, then we've | |
3938 | * mapped this one already. This can happen if | |
3939 | * devices share queues across queue maps. | |
3940 | */ | |
3941 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
3942 | continue; | |
3943 | ||
3944 | cpumask_set_cpu(i, hctx->cpumask); | |
3945 | hctx->type = j; | |
3946 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
3947 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
3948 | ||
3949 | /* | |
3950 | * If the nr_ctx type overflows, we have exceeded the | |
3951 | * amount of sw queues we can support. | |
3952 | */ | |
3953 | BUG_ON(!hctx->nr_ctx); | |
3954 | } | |
bb94aea1 JW |
3955 | |
3956 | for (; j < HCTX_MAX_TYPES; j++) | |
3957 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3958 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 3959 | } |
506e931f JA |
3960 | |
3961 | queue_for_each_hw_ctx(q, hctx, i) { | |
a46c2702 ML |
3962 | int cpu; |
3963 | ||
4412efec ML |
3964 | /* |
3965 | * If no software queues are mapped to this hardware queue, | |
3966 | * disable it and free the request entries. | |
3967 | */ | |
3968 | if (!hctx->nr_ctx) { | |
3969 | /* Never unmap queue 0. We need it as a | |
3970 | * fallback in case of a new remap fails | |
3971 | * allocation | |
3972 | */ | |
e155b0c2 JG |
3973 | if (i) |
3974 | __blk_mq_free_map_and_rqs(set, i); | |
4412efec ML |
3975 | |
3976 | hctx->tags = NULL; | |
3977 | continue; | |
3978 | } | |
484b4061 | 3979 | |
2a34c087 ML |
3980 | hctx->tags = set->tags[i]; |
3981 | WARN_ON(!hctx->tags); | |
3982 | ||
889fa31f CY |
3983 | /* |
3984 | * Set the map size to the number of mapped software queues. | |
3985 | * This is more accurate and more efficient than looping | |
3986 | * over all possibly mapped software queues. | |
3987 | */ | |
88459642 | 3988 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 3989 | |
a46c2702 ML |
3990 | /* |
3991 | * Rule out isolated CPUs from hctx->cpumask to avoid | |
3992 | * running block kworker on isolated CPUs | |
3993 | */ | |
3994 | for_each_cpu(cpu, hctx->cpumask) { | |
3995 | if (cpu_is_isolated(cpu)) | |
3996 | cpumask_clear_cpu(cpu, hctx->cpumask); | |
3997 | } | |
3998 | ||
484b4061 JA |
3999 | /* |
4000 | * Initialize batch roundrobin counts | |
4001 | */ | |
f82ddf19 | 4002 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
4003 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
4004 | } | |
320ae51f JA |
4005 | } |
4006 | ||
8e8320c9 JA |
4007 | /* |
4008 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
4009 | * the queue isn't live yet. | |
4010 | */ | |
2404e607 | 4011 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
4012 | { |
4013 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 4014 | unsigned long i; |
0d2602ca | 4015 | |
2404e607 | 4016 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 4017 | if (shared) { |
51db1c37 | 4018 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
4019 | } else { |
4020 | blk_mq_tag_idle(hctx); | |
51db1c37 | 4021 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 4022 | } |
2404e607 JM |
4023 | } |
4024 | } | |
4025 | ||
655ac300 HR |
4026 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
4027 | bool shared) | |
2404e607 JM |
4028 | { |
4029 | struct request_queue *q; | |
0d2602ca | 4030 | |
705cda97 BVA |
4031 | lockdep_assert_held(&set->tag_list_lock); |
4032 | ||
0d2602ca JA |
4033 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4034 | blk_mq_freeze_queue(q); | |
2404e607 | 4035 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
4036 | blk_mq_unfreeze_queue(q); |
4037 | } | |
4038 | } | |
4039 | ||
4040 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
4041 | { | |
4042 | struct blk_mq_tag_set *set = q->tag_set; | |
4043 | ||
0d2602ca | 4044 | mutex_lock(&set->tag_list_lock); |
08c875cb | 4045 | list_del(&q->tag_set_list); |
2404e607 JM |
4046 | if (list_is_singular(&set->tag_list)) { |
4047 | /* just transitioned to unshared */ | |
51db1c37 | 4048 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 4049 | /* update existing queue */ |
655ac300 | 4050 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 4051 | } |
0d2602ca | 4052 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 4053 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
4054 | } |
4055 | ||
4056 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
4057 | struct request_queue *q) | |
4058 | { | |
0d2602ca | 4059 | mutex_lock(&set->tag_list_lock); |
2404e607 | 4060 | |
ff821d27 JA |
4061 | /* |
4062 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
4063 | */ | |
4064 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
4065 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
4066 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 4067 | /* update existing queue */ |
655ac300 | 4068 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 4069 | } |
51db1c37 | 4070 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 4071 | queue_set_hctx_shared(q, true); |
08c875cb | 4072 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 4073 | |
0d2602ca JA |
4074 | mutex_unlock(&set->tag_list_lock); |
4075 | } | |
4076 | ||
1db4909e ML |
4077 | /* All allocations will be freed in release handler of q->mq_kobj */ |
4078 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
4079 | { | |
4080 | struct blk_mq_ctxs *ctxs; | |
4081 | int cpu; | |
4082 | ||
4083 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
4084 | if (!ctxs) | |
4085 | return -ENOMEM; | |
4086 | ||
4087 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
4088 | if (!ctxs->queue_ctx) | |
4089 | goto fail; | |
4090 | ||
4091 | for_each_possible_cpu(cpu) { | |
4092 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
4093 | ctx->ctxs = ctxs; | |
4094 | } | |
4095 | ||
4096 | q->mq_kobj = &ctxs->kobj; | |
4097 | q->queue_ctx = ctxs->queue_ctx; | |
4098 | ||
4099 | return 0; | |
4100 | fail: | |
4101 | kfree(ctxs); | |
4102 | return -ENOMEM; | |
4103 | } | |
4104 | ||
e09aae7e ML |
4105 | /* |
4106 | * It is the actual release handler for mq, but we do it from | |
4107 | * request queue's release handler for avoiding use-after-free | |
4108 | * and headache because q->mq_kobj shouldn't have been introduced, | |
4109 | * but we can't group ctx/kctx kobj without it. | |
4110 | */ | |
4111 | void blk_mq_release(struct request_queue *q) | |
4112 | { | |
2f8f1336 | 4113 | struct blk_mq_hw_ctx *hctx, *next; |
4f481208 | 4114 | unsigned long i; |
e09aae7e | 4115 | |
2f8f1336 ML |
4116 | queue_for_each_hw_ctx(q, hctx, i) |
4117 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
4118 | ||
4119 | /* all hctx are in .unused_hctx_list now */ | |
4120 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
4121 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 4122 | kobject_put(&hctx->kobj); |
c3b4afca | 4123 | } |
e09aae7e | 4124 | |
4e5cc99e | 4125 | xa_destroy(&q->hctx_table); |
e09aae7e | 4126 | |
7ea5fe31 ML |
4127 | /* |
4128 | * release .mq_kobj and sw queue's kobject now because | |
4129 | * both share lifetime with request queue. | |
4130 | */ | |
4131 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
4132 | } |
4133 | ||
8023e144 CH |
4134 | static bool blk_mq_can_poll(struct blk_mq_tag_set *set) |
4135 | { | |
4136 | return set->nr_maps > HCTX_TYPE_POLL && | |
4137 | set->map[HCTX_TYPE_POLL].nr_queues; | |
4138 | } | |
4139 | ||
9ac4dd8c CH |
4140 | struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set, |
4141 | struct queue_limits *lim, void *queuedata) | |
b62c21b7 | 4142 | { |
9ac4dd8c | 4143 | struct queue_limits default_lim = { }; |
26a9750a CH |
4144 | struct request_queue *q; |
4145 | int ret; | |
b62c21b7 | 4146 | |
cdb24979 CH |
4147 | if (!lim) |
4148 | lim = &default_lim; | |
f76af42f | 4149 | lim->features |= BLK_FEAT_IO_STAT | BLK_FEAT_NOWAIT; |
8023e144 CH |
4150 | if (blk_mq_can_poll(set)) |
4151 | lim->features |= BLK_FEAT_POLL; | |
cdb24979 CH |
4152 | |
4153 | q = blk_alloc_queue(lim, set->numa_node); | |
ad751ba1 CH |
4154 | if (IS_ERR(q)) |
4155 | return q; | |
26a9750a CH |
4156 | q->queuedata = queuedata; |
4157 | ret = blk_mq_init_allocated_queue(set, q); | |
4158 | if (ret) { | |
6f8191fd | 4159 | blk_put_queue(q); |
26a9750a CH |
4160 | return ERR_PTR(ret); |
4161 | } | |
b62c21b7 MS |
4162 | return q; |
4163 | } | |
9ac4dd8c | 4164 | EXPORT_SYMBOL(blk_mq_alloc_queue); |
b62c21b7 | 4165 | |
6f8191fd CH |
4166 | /** |
4167 | * blk_mq_destroy_queue - shutdown a request queue | |
4168 | * @q: request queue to shutdown | |
4169 | * | |
9ac4dd8c | 4170 | * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future |
81ea42b9 | 4171 | * requests will be failed with -ENODEV. The caller is responsible for dropping |
9ac4dd8c | 4172 | * the reference from blk_mq_alloc_queue() by calling blk_put_queue(). |
6f8191fd CH |
4173 | * |
4174 | * Context: can sleep | |
4175 | */ | |
4176 | void blk_mq_destroy_queue(struct request_queue *q) | |
4177 | { | |
4178 | WARN_ON_ONCE(!queue_is_mq(q)); | |
4179 | WARN_ON_ONCE(blk_queue_registered(q)); | |
4180 | ||
4181 | might_sleep(); | |
4182 | ||
4183 | blk_queue_flag_set(QUEUE_FLAG_DYING, q); | |
4184 | blk_queue_start_drain(q); | |
56c1ee92 | 4185 | blk_mq_freeze_queue_wait(q); |
6f8191fd CH |
4186 | |
4187 | blk_sync_queue(q); | |
4188 | blk_mq_cancel_work_sync(q); | |
4189 | blk_mq_exit_queue(q); | |
6f8191fd CH |
4190 | } |
4191 | EXPORT_SYMBOL(blk_mq_destroy_queue); | |
4192 | ||
27e32cd2 CH |
4193 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, |
4194 | struct queue_limits *lim, void *queuedata, | |
4dcc4874 | 4195 | struct lock_class_key *lkclass) |
9316a9ed JA |
4196 | { |
4197 | struct request_queue *q; | |
b461dfc4 | 4198 | struct gendisk *disk; |
9316a9ed | 4199 | |
27e32cd2 | 4200 | q = blk_mq_alloc_queue(set, lim, queuedata); |
b461dfc4 CH |
4201 | if (IS_ERR(q)) |
4202 | return ERR_CAST(q); | |
9316a9ed | 4203 | |
4a1fa41d | 4204 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 | 4205 | if (!disk) { |
0a3e5cc7 | 4206 | blk_mq_destroy_queue(q); |
2b3f056f | 4207 | blk_put_queue(q); |
b461dfc4 | 4208 | return ERR_PTR(-ENOMEM); |
9316a9ed | 4209 | } |
6f8191fd | 4210 | set_bit(GD_OWNS_QUEUE, &disk->state); |
b461dfc4 | 4211 | return disk; |
9316a9ed | 4212 | } |
b461dfc4 | 4213 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 4214 | |
6f8191fd CH |
4215 | struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q, |
4216 | struct lock_class_key *lkclass) | |
4217 | { | |
22c17e27 CH |
4218 | struct gendisk *disk; |
4219 | ||
6f8191fd CH |
4220 | if (!blk_get_queue(q)) |
4221 | return NULL; | |
22c17e27 CH |
4222 | disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass); |
4223 | if (!disk) | |
4224 | blk_put_queue(q); | |
4225 | return disk; | |
6f8191fd CH |
4226 | } |
4227 | EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue); | |
4228 | ||
34d11ffa JW |
4229 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
4230 | struct blk_mq_tag_set *set, struct request_queue *q, | |
4231 | int hctx_idx, int node) | |
4232 | { | |
2f8f1336 | 4233 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 4234 | |
2f8f1336 ML |
4235 | /* reuse dead hctx first */ |
4236 | spin_lock(&q->unused_hctx_lock); | |
4237 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
4238 | if (tmp->numa_node == node) { | |
4239 | hctx = tmp; | |
4240 | break; | |
4241 | } | |
4242 | } | |
4243 | if (hctx) | |
4244 | list_del_init(&hctx->hctx_list); | |
4245 | spin_unlock(&q->unused_hctx_lock); | |
4246 | ||
4247 | if (!hctx) | |
4248 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 4249 | if (!hctx) |
7c6c5b7c | 4250 | goto fail; |
34d11ffa | 4251 | |
7c6c5b7c ML |
4252 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
4253 | goto free_hctx; | |
34d11ffa JW |
4254 | |
4255 | return hctx; | |
7c6c5b7c ML |
4256 | |
4257 | free_hctx: | |
4258 | kobject_put(&hctx->kobj); | |
4259 | fail: | |
4260 | return NULL; | |
34d11ffa JW |
4261 | } |
4262 | ||
868f2f0b KB |
4263 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
4264 | struct request_queue *q) | |
320ae51f | 4265 | { |
4e5cc99e ML |
4266 | struct blk_mq_hw_ctx *hctx; |
4267 | unsigned long i, j; | |
ac0d6b92 | 4268 | |
fb350e0a ML |
4269 | /* protect against switching io scheduler */ |
4270 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 4271 | for (i = 0; i < set->nr_hw_queues; i++) { |
306f13ee | 4272 | int old_node; |
4d805131 | 4273 | int node = blk_mq_get_hctx_node(set, i); |
4e5cc99e | 4274 | struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i); |
868f2f0b | 4275 | |
306f13ee ML |
4276 | if (old_hctx) { |
4277 | old_node = old_hctx->numa_node; | |
4278 | blk_mq_exit_hctx(q, set, old_hctx, i); | |
4279 | } | |
868f2f0b | 4280 | |
4e5cc99e | 4281 | if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) { |
306f13ee | 4282 | if (!old_hctx) |
34d11ffa | 4283 | break; |
306f13ee ML |
4284 | pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n", |
4285 | node, old_node); | |
4e5cc99e ML |
4286 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node); |
4287 | WARN_ON_ONCE(!hctx); | |
868f2f0b | 4288 | } |
320ae51f | 4289 | } |
e01ad46d JW |
4290 | /* |
4291 | * Increasing nr_hw_queues fails. Free the newly allocated | |
4292 | * hctxs and keep the previous q->nr_hw_queues. | |
4293 | */ | |
4294 | if (i != set->nr_hw_queues) { | |
4295 | j = q->nr_hw_queues; | |
e01ad46d JW |
4296 | } else { |
4297 | j = i; | |
e01ad46d JW |
4298 | q->nr_hw_queues = set->nr_hw_queues; |
4299 | } | |
34d11ffa | 4300 | |
4e5cc99e ML |
4301 | xa_for_each_start(&q->hctx_table, j, hctx, j) |
4302 | blk_mq_exit_hctx(q, set, hctx, j); | |
fb350e0a | 4303 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
4304 | } |
4305 | ||
26a9750a CH |
4306 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
4307 | struct request_queue *q) | |
868f2f0b | 4308 | { |
66841672 ML |
4309 | /* mark the queue as mq asap */ |
4310 | q->mq_ops = set->ops; | |
4311 | ||
1db4909e | 4312 | if (blk_mq_alloc_ctxs(q)) |
54bdd67d | 4313 | goto err_exit; |
868f2f0b | 4314 | |
737f98cf ML |
4315 | /* init q->mq_kobj and sw queues' kobjects */ |
4316 | blk_mq_sysfs_init(q); | |
4317 | ||
2f8f1336 ML |
4318 | INIT_LIST_HEAD(&q->unused_hctx_list); |
4319 | spin_lock_init(&q->unused_hctx_lock); | |
4320 | ||
4e5cc99e ML |
4321 | xa_init(&q->hctx_table); |
4322 | ||
868f2f0b KB |
4323 | blk_mq_realloc_hw_ctxs(set, q); |
4324 | if (!q->nr_hw_queues) | |
4325 | goto err_hctxs; | |
320ae51f | 4326 | |
287922eb | 4327 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 4328 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 4329 | |
a8908939 | 4330 | q->tag_set = set; |
320ae51f | 4331 | |
94eddfbe | 4332 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 4333 | |
2849450a | 4334 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
9a67aa52 | 4335 | INIT_LIST_HEAD(&q->flush_list); |
6fca6a61 CH |
4336 | INIT_LIST_HEAD(&q->requeue_list); |
4337 | spin_lock_init(&q->requeue_lock); | |
4338 | ||
eba71768 JA |
4339 | q->nr_requests = set->queue_depth; |
4340 | ||
24d2f903 | 4341 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 4342 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 4343 | blk_mq_map_swqueue(q); |
26a9750a | 4344 | return 0; |
18741986 | 4345 | |
320ae51f | 4346 | err_hctxs: |
943f45b9 | 4347 | blk_mq_release(q); |
c7de5726 ML |
4348 | err_exit: |
4349 | q->mq_ops = NULL; | |
26a9750a | 4350 | return -ENOMEM; |
320ae51f | 4351 | } |
b62c21b7 | 4352 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 4353 | |
c7e2d94b ML |
4354 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
4355 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 4356 | { |
630ef623 | 4357 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 4358 | |
630ef623 | 4359 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 4360 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
4361 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
4362 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 4363 | } |
320ae51f | 4364 | |
a5164405 JA |
4365 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
4366 | { | |
4367 | int i; | |
4368 | ||
079a2e3e JG |
4369 | if (blk_mq_is_shared_tags(set->flags)) { |
4370 | set->shared_tags = blk_mq_alloc_map_and_rqs(set, | |
e155b0c2 JG |
4371 | BLK_MQ_NO_HCTX_IDX, |
4372 | set->queue_depth); | |
079a2e3e | 4373 | if (!set->shared_tags) |
e155b0c2 JG |
4374 | return -ENOMEM; |
4375 | } | |
4376 | ||
8229cca8 | 4377 | for (i = 0; i < set->nr_hw_queues; i++) { |
63064be1 | 4378 | if (!__blk_mq_alloc_map_and_rqs(set, i)) |
a5164405 | 4379 | goto out_unwind; |
8229cca8 XT |
4380 | cond_resched(); |
4381 | } | |
a5164405 JA |
4382 | |
4383 | return 0; | |
4384 | ||
4385 | out_unwind: | |
4386 | while (--i >= 0) | |
e155b0c2 JG |
4387 | __blk_mq_free_map_and_rqs(set, i); |
4388 | ||
079a2e3e JG |
4389 | if (blk_mq_is_shared_tags(set->flags)) { |
4390 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 | 4391 | BLK_MQ_NO_HCTX_IDX); |
645db34e | 4392 | } |
a5164405 | 4393 | |
a5164405 JA |
4394 | return -ENOMEM; |
4395 | } | |
4396 | ||
4397 | /* | |
4398 | * Allocate the request maps associated with this tag_set. Note that this | |
4399 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
4400 | * will be updated to reflect the allocated depth. | |
4401 | */ | |
63064be1 | 4402 | static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) |
a5164405 JA |
4403 | { |
4404 | unsigned int depth; | |
4405 | int err; | |
4406 | ||
4407 | depth = set->queue_depth; | |
4408 | do { | |
4409 | err = __blk_mq_alloc_rq_maps(set); | |
4410 | if (!err) | |
4411 | break; | |
4412 | ||
4413 | set->queue_depth >>= 1; | |
4414 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
4415 | err = -ENOMEM; | |
4416 | break; | |
4417 | } | |
4418 | } while (set->queue_depth); | |
4419 | ||
4420 | if (!set->queue_depth || err) { | |
4421 | pr_err("blk-mq: failed to allocate request map\n"); | |
4422 | return -ENOMEM; | |
4423 | } | |
4424 | ||
4425 | if (depth != set->queue_depth) | |
4426 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
4427 | depth, set->queue_depth); | |
4428 | ||
4429 | return 0; | |
4430 | } | |
4431 | ||
a4e1d0b7 | 4432 | static void blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
ebe8bddb | 4433 | { |
6e66b493 BVA |
4434 | /* |
4435 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
4436 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
4437 | * number of hardware queues. | |
4438 | */ | |
4439 | if (set->nr_maps == 1) | |
4440 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
4441 | ||
ec30b461 | 4442 | if (set->ops->map_queues) { |
b3c661b1 JA |
4443 | int i; |
4444 | ||
7d4901a9 ML |
4445 | /* |
4446 | * transport .map_queues is usually done in the following | |
4447 | * way: | |
4448 | * | |
4449 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
4450 | * mask = get_cpu_mask(queue) | |
4451 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 4452 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
4453 | * } |
4454 | * | |
4455 | * When we need to remap, the table has to be cleared for | |
4456 | * killing stale mapping since one CPU may not be mapped | |
4457 | * to any hw queue. | |
4458 | */ | |
b3c661b1 JA |
4459 | for (i = 0; i < set->nr_maps; i++) |
4460 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 4461 | |
a4e1d0b7 | 4462 | set->ops->map_queues(set); |
b3c661b1 JA |
4463 | } else { |
4464 | BUG_ON(set->nr_maps > 1); | |
a4e1d0b7 | 4465 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 4466 | } |
ebe8bddb OS |
4467 | } |
4468 | ||
f7e76dbc | 4469 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
ee9d5521 | 4470 | int new_nr_hw_queues) |
f7e76dbc BVA |
4471 | { |
4472 | struct blk_mq_tags **new_tags; | |
e1dd7bc9 | 4473 | int i; |
f7e76dbc | 4474 | |
6be6d112 | 4475 | if (set->nr_hw_queues >= new_nr_hw_queues) |
d4b2e0d4 | 4476 | goto done; |
f7e76dbc BVA |
4477 | |
4478 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
4479 | GFP_KERNEL, set->numa_node); | |
4480 | if (!new_tags) | |
4481 | return -ENOMEM; | |
4482 | ||
4483 | if (set->tags) | |
ee9d5521 | 4484 | memcpy(new_tags, set->tags, set->nr_hw_queues * |
f7e76dbc BVA |
4485 | sizeof(*set->tags)); |
4486 | kfree(set->tags); | |
4487 | set->tags = new_tags; | |
7222657e CZ |
4488 | |
4489 | for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) { | |
4490 | if (!__blk_mq_alloc_map_and_rqs(set, i)) { | |
4491 | while (--i >= set->nr_hw_queues) | |
4492 | __blk_mq_free_map_and_rqs(set, i); | |
4493 | return -ENOMEM; | |
4494 | } | |
4495 | cond_resched(); | |
4496 | } | |
4497 | ||
d4b2e0d4 | 4498 | done: |
f7e76dbc | 4499 | set->nr_hw_queues = new_nr_hw_queues; |
f7e76dbc BVA |
4500 | return 0; |
4501 | } | |
4502 | ||
a4391c64 JA |
4503 | /* |
4504 | * Alloc a tag set to be associated with one or more request queues. | |
4505 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 4506 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
4507 | * value will be stored in set->queue_depth. |
4508 | */ | |
24d2f903 CH |
4509 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
4510 | { | |
b3c661b1 | 4511 | int i, ret; |
da695ba2 | 4512 | |
205fb5f5 BVA |
4513 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
4514 | ||
24d2f903 CH |
4515 | if (!set->nr_hw_queues) |
4516 | return -EINVAL; | |
a4391c64 | 4517 | if (!set->queue_depth) |
24d2f903 CH |
4518 | return -EINVAL; |
4519 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
4520 | return -EINVAL; | |
4521 | ||
7d7e0f90 | 4522 | if (!set->ops->queue_rq) |
24d2f903 CH |
4523 | return -EINVAL; |
4524 | ||
de148297 ML |
4525 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
4526 | return -EINVAL; | |
4527 | ||
a4391c64 JA |
4528 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
4529 | pr_info("blk-mq: reduced tag depth to %u\n", | |
4530 | BLK_MQ_MAX_DEPTH); | |
4531 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
4532 | } | |
24d2f903 | 4533 | |
b3c661b1 JA |
4534 | if (!set->nr_maps) |
4535 | set->nr_maps = 1; | |
4536 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
4537 | return -EINVAL; | |
4538 | ||
6637fadf SL |
4539 | /* |
4540 | * If a crashdump is active, then we are potentially in a very | |
ec30b461 ML |
4541 | * memory constrained environment. Limit us to 64 tags to prevent |
4542 | * using too much memory. | |
6637fadf | 4543 | */ |
ec30b461 | 4544 | if (is_kdump_kernel()) |
6637fadf | 4545 | set->queue_depth = min(64U, set->queue_depth); |
ec30b461 | 4546 | |
868f2f0b | 4547 | /* |
392546ae JA |
4548 | * There is no use for more h/w queues than cpus if we just have |
4549 | * a single map | |
868f2f0b | 4550 | */ |
392546ae | 4551 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4552 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 4553 | |
80bd4a7a CH |
4554 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4555 | set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL); | |
4556 | if (!set->srcu) | |
4557 | return -ENOMEM; | |
4558 | ret = init_srcu_struct(set->srcu); | |
4559 | if (ret) | |
4560 | goto out_free_srcu; | |
4561 | } | |
24d2f903 | 4562 | |
da695ba2 | 4563 | ret = -ENOMEM; |
5ee20298 CH |
4564 | set->tags = kcalloc_node(set->nr_hw_queues, |
4565 | sizeof(struct blk_mq_tags *), GFP_KERNEL, | |
4566 | set->numa_node); | |
4567 | if (!set->tags) | |
80bd4a7a | 4568 | goto out_cleanup_srcu; |
24d2f903 | 4569 | |
b3c661b1 JA |
4570 | for (i = 0; i < set->nr_maps; i++) { |
4571 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 4572 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
4573 | GFP_KERNEL, set->numa_node); |
4574 | if (!set->map[i].mq_map) | |
4575 | goto out_free_mq_map; | |
ec30b461 | 4576 | set->map[i].nr_queues = set->nr_hw_queues; |
b3c661b1 | 4577 | } |
bdd17e75 | 4578 | |
a4e1d0b7 | 4579 | blk_mq_update_queue_map(set); |
da695ba2 | 4580 | |
63064be1 | 4581 | ret = blk_mq_alloc_set_map_and_rqs(set); |
da695ba2 | 4582 | if (ret) |
bdd17e75 | 4583 | goto out_free_mq_map; |
24d2f903 | 4584 | |
0d2602ca JA |
4585 | mutex_init(&set->tag_list_lock); |
4586 | INIT_LIST_HEAD(&set->tag_list); | |
4587 | ||
24d2f903 | 4588 | return 0; |
bdd17e75 CH |
4589 | |
4590 | out_free_mq_map: | |
b3c661b1 JA |
4591 | for (i = 0; i < set->nr_maps; i++) { |
4592 | kfree(set->map[i].mq_map); | |
4593 | set->map[i].mq_map = NULL; | |
4594 | } | |
5676e7b6 RE |
4595 | kfree(set->tags); |
4596 | set->tags = NULL; | |
80bd4a7a CH |
4597 | out_cleanup_srcu: |
4598 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4599 | cleanup_srcu_struct(set->srcu); | |
4600 | out_free_srcu: | |
4601 | if (set->flags & BLK_MQ_F_BLOCKING) | |
4602 | kfree(set->srcu); | |
da695ba2 | 4603 | return ret; |
24d2f903 CH |
4604 | } |
4605 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
4606 | ||
cdb14e0f CH |
4607 | /* allocate and initialize a tagset for a simple single-queue device */ |
4608 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
4609 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
4610 | unsigned int set_flags) | |
4611 | { | |
4612 | memset(set, 0, sizeof(*set)); | |
4613 | set->ops = ops; | |
4614 | set->nr_hw_queues = 1; | |
4615 | set->nr_maps = 1; | |
4616 | set->queue_depth = queue_depth; | |
4617 | set->numa_node = NUMA_NO_NODE; | |
4618 | set->flags = set_flags; | |
4619 | return blk_mq_alloc_tag_set(set); | |
4620 | } | |
4621 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
4622 | ||
24d2f903 CH |
4623 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
4624 | { | |
b3c661b1 | 4625 | int i, j; |
24d2f903 | 4626 | |
f7e76dbc | 4627 | for (i = 0; i < set->nr_hw_queues; i++) |
e155b0c2 | 4628 | __blk_mq_free_map_and_rqs(set, i); |
484b4061 | 4629 | |
079a2e3e JG |
4630 | if (blk_mq_is_shared_tags(set->flags)) { |
4631 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 JG |
4632 | BLK_MQ_NO_HCTX_IDX); |
4633 | } | |
32bc15af | 4634 | |
b3c661b1 JA |
4635 | for (j = 0; j < set->nr_maps; j++) { |
4636 | kfree(set->map[j].mq_map); | |
4637 | set->map[j].mq_map = NULL; | |
4638 | } | |
bdd17e75 | 4639 | |
981bd189 | 4640 | kfree(set->tags); |
5676e7b6 | 4641 | set->tags = NULL; |
80bd4a7a CH |
4642 | if (set->flags & BLK_MQ_F_BLOCKING) { |
4643 | cleanup_srcu_struct(set->srcu); | |
4644 | kfree(set->srcu); | |
4645 | } | |
24d2f903 CH |
4646 | } |
4647 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
4648 | ||
e3a2b3f9 JA |
4649 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
4650 | { | |
4651 | struct blk_mq_tag_set *set = q->tag_set; | |
4652 | struct blk_mq_hw_ctx *hctx; | |
4f481208 ML |
4653 | int ret; |
4654 | unsigned long i; | |
e3a2b3f9 | 4655 | |
af281414 CH |
4656 | if (WARN_ON_ONCE(!q->mq_freeze_depth)) |
4657 | return -EINVAL; | |
4658 | ||
bd166ef1 | 4659 | if (!set) |
e3a2b3f9 JA |
4660 | return -EINVAL; |
4661 | ||
e5fa8140 AZ |
4662 | if (q->nr_requests == nr) |
4663 | return 0; | |
4664 | ||
24f5a90f | 4665 | blk_mq_quiesce_queue(q); |
70f36b60 | 4666 | |
e3a2b3f9 JA |
4667 | ret = 0; |
4668 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
4669 | if (!hctx->tags) |
4670 | continue; | |
bd166ef1 JA |
4671 | /* |
4672 | * If we're using an MQ scheduler, just update the scheduler | |
4673 | * queue depth. This is similar to what the old code would do. | |
4674 | */ | |
f6adcef5 | 4675 | if (hctx->sched_tags) { |
70f36b60 | 4676 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
f6adcef5 | 4677 | nr, true); |
f6adcef5 JG |
4678 | } else { |
4679 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
4680 | false); | |
70f36b60 | 4681 | } |
e3a2b3f9 JA |
4682 | if (ret) |
4683 | break; | |
77f1e0a5 JA |
4684 | if (q->elevator && q->elevator->type->ops.depth_updated) |
4685 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 4686 | } |
d97e594c | 4687 | if (!ret) { |
e3a2b3f9 | 4688 | q->nr_requests = nr; |
079a2e3e | 4689 | if (blk_mq_is_shared_tags(set->flags)) { |
8fa04464 | 4690 | if (q->elevator) |
079a2e3e | 4691 | blk_mq_tag_update_sched_shared_tags(q); |
8fa04464 | 4692 | else |
079a2e3e | 4693 | blk_mq_tag_resize_shared_tags(set, nr); |
8fa04464 | 4694 | } |
d97e594c | 4695 | } |
e3a2b3f9 | 4696 | |
24f5a90f | 4697 | blk_mq_unquiesce_queue(q); |
70f36b60 | 4698 | |
e3a2b3f9 JA |
4699 | return ret; |
4700 | } | |
4701 | ||
d48ece20 JW |
4702 | /* |
4703 | * request_queue and elevator_type pair. | |
4704 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
4705 | * the elevator_type associated with a request_queue. | |
4706 | */ | |
4707 | struct blk_mq_qe_pair { | |
4708 | struct list_head node; | |
4709 | struct request_queue *q; | |
4710 | struct elevator_type *type; | |
4711 | }; | |
4712 | ||
4713 | /* | |
4714 | * Cache the elevator_type in qe pair list and switch the | |
4715 | * io scheduler to 'none' | |
4716 | */ | |
4717 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
4718 | struct request_queue *q) | |
4719 | { | |
4720 | struct blk_mq_qe_pair *qe; | |
4721 | ||
d48ece20 JW |
4722 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); |
4723 | if (!qe) | |
4724 | return false; | |
4725 | ||
5fd7a84a ML |
4726 | /* q->elevator needs protection from ->sysfs_lock */ |
4727 | mutex_lock(&q->sysfs_lock); | |
4728 | ||
24516565 ML |
4729 | /* the check has to be done with holding sysfs_lock */ |
4730 | if (!q->elevator) { | |
4731 | kfree(qe); | |
4732 | goto unlock; | |
4733 | } | |
4734 | ||
d48ece20 JW |
4735 | INIT_LIST_HEAD(&qe->node); |
4736 | qe->q = q; | |
4737 | qe->type = q->elevator->type; | |
dd6f7f17 CH |
4738 | /* keep a reference to the elevator module as we'll switch back */ |
4739 | __elevator_get(qe->type); | |
d48ece20 | 4740 | list_add(&qe->node, head); |
64b36075 | 4741 | elevator_disable(q); |
24516565 | 4742 | unlock: |
d48ece20 JW |
4743 | mutex_unlock(&q->sysfs_lock); |
4744 | ||
4745 | return true; | |
4746 | } | |
4747 | ||
4a3b666e JK |
4748 | static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head, |
4749 | struct request_queue *q) | |
d48ece20 JW |
4750 | { |
4751 | struct blk_mq_qe_pair *qe; | |
d48ece20 JW |
4752 | |
4753 | list_for_each_entry(qe, head, node) | |
4a3b666e JK |
4754 | if (qe->q == q) |
4755 | return qe; | |
d48ece20 | 4756 | |
4a3b666e JK |
4757 | return NULL; |
4758 | } | |
d48ece20 | 4759 | |
4a3b666e JK |
4760 | static void blk_mq_elv_switch_back(struct list_head *head, |
4761 | struct request_queue *q) | |
4762 | { | |
4763 | struct blk_mq_qe_pair *qe; | |
4764 | struct elevator_type *t; | |
4765 | ||
4766 | qe = blk_lookup_qe_pair(head, q); | |
4767 | if (!qe) | |
4768 | return; | |
4769 | t = qe->type; | |
d48ece20 JW |
4770 | list_del(&qe->node); |
4771 | kfree(qe); | |
4772 | ||
4773 | mutex_lock(&q->sysfs_lock); | |
8237c01f | 4774 | elevator_switch(q, t); |
8ed40ee3 JC |
4775 | /* drop the reference acquired in blk_mq_elv_switch_none */ |
4776 | elevator_put(t); | |
d48ece20 JW |
4777 | mutex_unlock(&q->sysfs_lock); |
4778 | } | |
4779 | ||
e4dc2b32 KB |
4780 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
4781 | int nr_hw_queues) | |
868f2f0b KB |
4782 | { |
4783 | struct request_queue *q; | |
d48ece20 | 4784 | LIST_HEAD(head); |
6be6d112 CZ |
4785 | int prev_nr_hw_queues = set->nr_hw_queues; |
4786 | int i; | |
868f2f0b | 4787 | |
705cda97 BVA |
4788 | lockdep_assert_held(&set->tag_list_lock); |
4789 | ||
392546ae | 4790 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4791 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
4792 | if (nr_hw_queues < 1) |
4793 | return; | |
4794 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
4795 | return; |
4796 | ||
4797 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4798 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
4799 | /* |
4800 | * Switch IO scheduler to 'none', cleaning up the data associated | |
4801 | * with the previous scheduler. We will switch back once we are done | |
4802 | * updating the new sw to hw queue mappings. | |
4803 | */ | |
4804 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4805 | if (!blk_mq_elv_switch_none(&head, q)) | |
4806 | goto switch_back; | |
868f2f0b | 4807 | |
477e19de JW |
4808 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4809 | blk_mq_debugfs_unregister_hctxs(q); | |
eaa870f9 | 4810 | blk_mq_sysfs_unregister_hctxs(q); |
477e19de JW |
4811 | } |
4812 | ||
ee9d5521 | 4813 | if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0) |
f7e76dbc BVA |
4814 | goto reregister; |
4815 | ||
e01ad46d | 4816 | fallback: |
aa880ad6 | 4817 | blk_mq_update_queue_map(set); |
868f2f0b | 4818 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
8023e144 CH |
4819 | struct queue_limits lim; |
4820 | ||
868f2f0b | 4821 | blk_mq_realloc_hw_ctxs(set, q); |
8023e144 | 4822 | |
e01ad46d | 4823 | if (q->nr_hw_queues != set->nr_hw_queues) { |
a846a8e6 YB |
4824 | int i = prev_nr_hw_queues; |
4825 | ||
e01ad46d JW |
4826 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", |
4827 | nr_hw_queues, prev_nr_hw_queues); | |
a846a8e6 YB |
4828 | for (; i < set->nr_hw_queues; i++) |
4829 | __blk_mq_free_map_and_rqs(set, i); | |
4830 | ||
e01ad46d | 4831 | set->nr_hw_queues = prev_nr_hw_queues; |
e01ad46d JW |
4832 | goto fallback; |
4833 | } | |
8023e144 CH |
4834 | lim = queue_limits_start_update(q); |
4835 | if (blk_mq_can_poll(set)) | |
4836 | lim.features |= BLK_FEAT_POLL; | |
4837 | else | |
4838 | lim.features &= ~BLK_FEAT_POLL; | |
4839 | if (queue_limits_commit_update(q, &lim) < 0) | |
4840 | pr_warn("updating the poll flag failed\n"); | |
477e19de JW |
4841 | blk_mq_map_swqueue(q); |
4842 | } | |
4843 | ||
f7e76dbc | 4844 | reregister: |
477e19de | 4845 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
eaa870f9 | 4846 | blk_mq_sysfs_register_hctxs(q); |
477e19de | 4847 | blk_mq_debugfs_register_hctxs(q); |
868f2f0b KB |
4848 | } |
4849 | ||
d48ece20 JW |
4850 | switch_back: |
4851 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4852 | blk_mq_elv_switch_back(&head, q); | |
4853 | ||
868f2f0b KB |
4854 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
4855 | blk_mq_unfreeze_queue(q); | |
6be6d112 CZ |
4856 | |
4857 | /* Free the excess tags when nr_hw_queues shrink. */ | |
4858 | for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++) | |
4859 | __blk_mq_free_map_and_rqs(set, i); | |
868f2f0b | 4860 | } |
e4dc2b32 KB |
4861 | |
4862 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
4863 | { | |
4864 | mutex_lock(&set->tag_list_lock); | |
4865 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
4866 | mutex_unlock(&set->tag_list_lock); | |
4867 | } | |
868f2f0b KB |
4868 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
4869 | ||
f6c80cff KB |
4870 | static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx, |
4871 | struct io_comp_batch *iob, unsigned int flags) | |
bbd7bb70 | 4872 | { |
c6699d6f CH |
4873 | long state = get_current_state(); |
4874 | int ret; | |
bbd7bb70 | 4875 | |
aa61bec3 | 4876 | do { |
5a72e899 | 4877 | ret = q->mq_ops->poll(hctx, iob); |
bbd7bb70 | 4878 | if (ret > 0) { |
849a3700 | 4879 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4880 | return ret; |
bbd7bb70 JA |
4881 | } |
4882 | ||
4883 | if (signal_pending_state(state, current)) | |
849a3700 | 4884 | __set_current_state(TASK_RUNNING); |
b03fbd4f | 4885 | if (task_is_running(current)) |
85f4d4b6 | 4886 | return 1; |
c6699d6f | 4887 | |
ef99b2d3 | 4888 | if (ret < 0 || (flags & BLK_POLL_ONESHOT)) |
bbd7bb70 JA |
4889 | break; |
4890 | cpu_relax(); | |
aa61bec3 | 4891 | } while (!need_resched()); |
bbd7bb70 | 4892 | |
67b4110f | 4893 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4894 | return 0; |
bbd7bb70 | 4895 | } |
1052b8ac | 4896 | |
f6c80cff KB |
4897 | int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, |
4898 | struct io_comp_batch *iob, unsigned int flags) | |
4899 | { | |
4900 | struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie); | |
4901 | ||
4902 | return blk_hctx_poll(q, hctx, iob, flags); | |
4903 | } | |
4904 | ||
4905 | int blk_rq_poll(struct request *rq, struct io_comp_batch *iob, | |
4906 | unsigned int poll_flags) | |
4907 | { | |
4908 | struct request_queue *q = rq->q; | |
4909 | int ret; | |
4910 | ||
4911 | if (!blk_rq_is_poll(rq)) | |
4912 | return 0; | |
4913 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
4914 | return 0; | |
4915 | ||
4916 | ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags); | |
4917 | blk_queue_exit(q); | |
4918 | ||
4919 | return ret; | |
4920 | } | |
4921 | EXPORT_SYMBOL_GPL(blk_rq_poll); | |
4922 | ||
9cf2bab6 JA |
4923 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4924 | { | |
4925 | return rq->mq_ctx->cpu; | |
4926 | } | |
4927 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4928 | ||
2a19b28f ML |
4929 | void blk_mq_cancel_work_sync(struct request_queue *q) |
4930 | { | |
219cf43c JC |
4931 | struct blk_mq_hw_ctx *hctx; |
4932 | unsigned long i; | |
2a19b28f | 4933 | |
219cf43c | 4934 | cancel_delayed_work_sync(&q->requeue_work); |
2a19b28f | 4935 | |
219cf43c JC |
4936 | queue_for_each_hw_ctx(q, hctx, i) |
4937 | cancel_delayed_work_sync(&hctx->run_work); | |
2a19b28f ML |
4938 | } |
4939 | ||
320ae51f JA |
4940 | static int __init blk_mq_init(void) |
4941 | { | |
c3077b5d CH |
4942 | int i; |
4943 | ||
4944 | for_each_possible_cpu(i) | |
f9ab4918 | 4945 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
660e802c CZ |
4946 | for_each_possible_cpu(i) |
4947 | INIT_CSD(&per_cpu(blk_cpu_csd, i), | |
4948 | __blk_mq_complete_request_remote, NULL); | |
c3077b5d CH |
4949 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4950 | ||
4951 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4952 | "block/softirq:dead", NULL, | |
4953 | blk_softirq_cpu_dead); | |
9467f859 TG |
4954 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4955 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4956 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4957 | blk_mq_hctx_notify_online, | |
4958 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4959 | return 0; |
4960 | } | |
4961 | subsys_initcall(blk_mq_init); |